Hamish Humbert

Ham's guide to AI engineering

Hamish Humbert

v1.0.4, 25 October 2026

This is my book on applied AI engineering: the work of putting language models into software that people rely on. I'm writing it as I learn to do the work, one chapter per skill the job asks for. The foreword says what is here and what isn't yet.

Only the foreword is written. The chapters below are the shape of the book, so you can see where it's going. The version goes up as they land; the changelog has the list.

1Forewordv1.0.4
1. 1.1 What this is
2. 1.2 What is here so far
3. 1.3 Versions
2Setup and toolingnot yet
1. 2.1 Dev Environment
  1. 2.1.1 System Foundation
  2. 2.1.2 Python with uv
  3. 2.1.3 Node.js with pnpm
  4. 2.1.4 Rust
  5. 2.1.5 Julia (Optional)
  6. 2.1.6 GPU Setup (If You Have One)
  7. 2.1.7 Verify the route you want to start
2. 2.2 Git and Collaboration
  1. 2.2.1 Configure git
  2. 2.2.2 The daily workflow
  3. 2.2.3 Branching for experiments
  4. 2.2.4 Working with this course repo
3. 2.3 GPU Setup and Cloud
  1. 2.3.1 Option 1: Local NVIDIA GPU
  2. 2.3.2 Option 2: Google Colab
  3. 2.3.3 Option 3: Cloud GPU
  4. 2.3.4 No GPU? No problem.
4. 2.4 APIs and Keys
  1. 2.4.1 Store API keys safely
  2. 2.4.2 First API call (Python)
  3. 2.4.3 First API call (TypeScript)
  4. 2.4.4 Raw HTTP (no SDK)
5. 2.5 Jupyter Notebooks
  1. 2.5.1 Pick your interface
  2. 2.5.2 Keyboard shortcuts that matter
  3. 2.5.3 Cell types
  4. 2.5.4 Magic commands
  5. 2.5.5 Display rich output inline
  6. 2.5.6 Google Colab
  7. 2.5.7 Notebooks vs Scripts: When to use which
  8. 2.5.8 Common traps
6. 2.6 Python Environments
  1. 2.6.1 Option 1: uv venv (Recommended)
  2. 2.6.2 Option 2: venv (Built-in)
  3. 2.6.3 Option 3: conda (When You Need It)
  4. 2.6.4 For This Course: Per-Phase Strategy
  5. 2.6.5 Installing globally
  6. 2.6.6 Mixing pip and conda
  7. 2.6.7 Forgetting to activate
  8. 2.6.8 Committing .venv to git
  9. 2.6.9 CUDA version mismatch
7. 2.7 Docker for AI
  1. 2.7.1 Why AI projects need Docker more than most
  2. 2.7.2 Key vocabulary
  3. 2.7.3 Common container patterns in AI
  4. 2.7.4 Install Docker
  5. 2.7.5 Install NVIDIA Container Toolkit (Linux with NVIDIA GPU)
  6. 2.7.6 Understand base images
  7. 2.7.7 Write a Dockerfile for AI development
  8. 2.7.8 Volume mounts for data and models
  9. 2.7.9 Docker Compose for multi-service AI apps
  10. 2.7.10 Useful Docker commands for AI work
  11. 2.7.11 No GPU?
8. 2.8 Editor Setup
  1. 2.8.1 Install VS Code
  2. 2.8.2 Install Essential Extensions
  3. 2.8.3 Configure Settings
  4. 2.8.4 Terminal Integration
  5. 2.8.5 Remote Development (SSH into GPU Boxes)
  6. 2.8.6 Cursor
  7. 2.8.7 Windsurf
  8. 2.8.8 Vim/Neovim
9. 2.9 Data Management
  1. 2.9.1 Install the datasets library
  2. 2.9.2 Load a dataset
  3. 2.9.3 Stream large datasets
  4. 2.9.4 Dataset formats
  5. 2.9.5 Data splits
  6. 2.9.6 Download and cache models
  7. 2.9.7 Handle large files
  8. 2.9.8 Storage patterns
10. 2.10 Terminal and Shell
  1. 2.10.1 Know your shell
  2. 2.10.2 Piping and redirects
  3. 2.10.3 Background processes
  4. 2.10.4 tmux
  5. 2.10.5 Monitoring with htop and nvtop
  6. 2.10.6 SSH for remote GPU boxes
  7. 2.10.7 Useful aliases for AI work
  8. 2.10.8 Common AI terminal patterns
11. 2.11 Linux for AI
  1. 2.11.1 Moving Around
  2. 2.11.2 Files and Directories
  3. 2.11.3 Reading Files
  4. 2.11.4 Searching
12. 2.12 Debugging and Profiling
  1. 2.12.1 Part 1: Print Debugging (Yes, It Works)
  2. 2.12.2 Part 2: Python Debugger (pdb and breakpoint)
  3. 2.12.3 Part 3: Python Logging
  4. 2.12.4 Part 4: Timing Code Sections
  5. 2.12.5 Part 5: cProfile and line_profiler
  6. 2.12.6 Part 6: Memory Profiling
  7. 2.12.7 Part 7: Common AI Bugs and How to Catch Them
  8. 2.12.8 Part 8: TensorBoard Basics
  9. 2.12.9 Part 9: VS Code Debugger
13. 2.13 Python for AI Engineering
  1. 2.13.1 What the course code uses
  2. 2.13.2 Names refer to objects
  3. 2.13.3 A list holds pointers, an array holds numbers
  4. 2.13.4 A list holds every item, a generator holds one
  5. 2.13.5 Values, collections, and unpacking
  6. 2.13.6 Comprehensions
  7. 2.13.7 Functions, arguments, and closures
  8. 2.13.8 Classes and dataclasses
  9. 2.13.9 Iterators and generators
  10. 2.13.10 Context managers and exceptions
  11. 2.13.11 Type hints, modules, and the main guard
  12. 2.13.12 Files, JSON, CSV, and sorting
  13. 2.13.13 Counting, grouping, and iterator tools
  14. 2.13.14 Why loops over numbers are slow
  15. 2.13.15 A record type and a dataset loader
  16. 2.13.16 A seeded train and test split
  17. 2.13.17 A mini-batch generator
  18. 2.13.18 Vector math and a standardizer
  19. 2.13.19 A complete run
  20. 2.13.20 Measure speed and memory
  21. 2.13.21 The standard library versions
  22. 2.13.22 The same math with compiled arrays
  23. 2.13.23 Broadcasting
3Math foundationsnot yet
1. 3.1 Linear Algebra Intuition
  1. 3.1.1 Vectors Are Points (and Directions)
  2. 3.1.2 Matrices Are Transformations
  3. 3.1.3 The Dot Product Measures Similarity
  4. 3.1.4 Linear Independence
  5. 3.1.5 Basis and Rank
  6. 3.1.6 Projection
  7. 3.1.7 Gram-Schmidt Process
  8. 3.1.8 Vectors from scratch (Python)
  9. 3.1.9 Matrices from scratch (Python)
  10. 3.1.10 Why this matters for AI
  11. 3.1.11 Julia version
  12. 3.1.12 Linear independence and projection from scratch (Python)
  13. 3.1.13 Rank, Projection, and QR with NumPy
  14. 3.1.14 PyTorch -- Tensors Are Vectors with Autodiff
2. 3.2 Vectors, Matrices and Operations
  1. 3.2.1 Vectors: ordered lists of numbers
  2. 3.2.2 Matrices: grids of numbers
  3. 3.2.3 Why shapes matter
  4. 3.2.4 The operations map
  5. 3.2.5 Element-wise vs matrix multiplication
  6. 3.2.6 Broadcasting
  7. 3.2.7 Vector class
  8. 3.2.8 Matrix class with core operations
  9. 3.2.9 See it work
  10. 3.2.10 Connect to neural networks
3. 3.3 Matrix Transformations and Eigenvalues
  1. 3.3.1 Transformations as matrices
  2. 3.3.2 Rotation
  3. 3.3.3 Scaling
  4. 3.3.4 Shearing
  5. 3.3.5 Reflection
  6. 3.3.6 Composition: chaining transformations
  7. 3.3.7 Eigenvalues and eigenvectors
  8. 3.3.8 Eigendecomposition
  9. 3.3.9 Why eigenvalues matter
  10. 3.3.10 Determinant as volume scaling factor
  11. 3.3.11 Transformation matrices from scratch (Python)
  12. 3.3.12 Composition of transformations
  13. 3.3.13 Eigenvalues from scratch (2x2)
  14. 3.3.14 3D rotations with NumPy
4. 3.4 Calculus for ML: Derivatives and Gradients
  1. 3.4.1 What is a derivative?
  2. 3.4.2 Partial derivatives: one variable at a time
  3. 3.4.3 The gradient: vector of all partial derivatives
  4. 3.4.4 The connection to optimization
  5. 3.4.5 Numerical vs analytical derivatives
  6. 3.4.6 Derivatives by hand for simple functions
  7. 3.4.7 The chain rule
  8. 3.4.8 The Hessian Matrix
  9. 3.4.9 Taylor Series Approximation
  10. 3.4.10 Integrals in ML
  11. 3.4.11 Multivariable Chain Rule in a Computation Graph
  12. 3.4.12 The Jacobian matrix
  13. 3.4.13 Why this matters for neural networks
  14. 3.4.14 Numerical derivative from scratch
  15. 3.4.15 Partial derivatives and gradients
  16. 3.4.16 Gradient descent to find the minimum of f(x) = x^2
  17. 3.4.17 Gradient descent on a 2D function
  18. 3.4.18 Comparing numerical and analytical derivatives
  19. 3.4.19 Computing the Hessian numerically
  20. 3.4.20 Taylor approximation in action
  21. 3.4.21 Why this matters for a neural network
5. 3.5 Chain Rule and Automatic Differentiation
  1. 3.5.1 The Chain Rule
  2. 3.5.2 Computational Graphs
  3. 3.5.3 Forward Mode vs Reverse Mode
  4. 3.5.4 Dual Numbers for Forward Mode
  5. 3.5.5 Building an Autograd Engine
  6. 3.5.6 How PyTorch Autograd Works Under the Hood
  7. 3.5.7 The Value class
  8. 3.5.8 Arithmetic operations with gradient tracking
  9. 3.5.9 The backward pass
  10. 3.5.10 More operations for a complete engine
  11. 3.5.11 Mini MLP from scratch
  12. 3.5.12 Gradient checking
  13. 3.5.13 Verify against manual calculation
  14. 3.5.14 Verify against PyTorch
  15. 3.5.15 A more complex expression
6. 3.6 Probability and Distributions
  1. 3.6.1 Events, Sample Spaces, and Probability
  2. 3.6.2 Conditional Probability and Independence
  3. 3.6.3 Probability Mass Functions vs Probability Density Functions
  4. 3.6.4 Common Distributions
  5. 3.6.5 Expected Value and Variance
  6. 3.6.6 Joint and Marginal Distributions
  7. 3.6.7 Why the Normal Distribution Shows Up Everywhere
  8. 3.6.8 Log Probabilities
  9. 3.6.9 Softmax as a Probability Distribution
  10. 3.6.10 Sampling
  11. 3.6.11 Probability basics
  12. 3.6.12 PMF and PDF from scratch
  13. 3.6.13 Expected value and variance
  14. 3.6.14 Sampling from distributions
  15. 3.6.15 Softmax and log probabilities
  16. 3.6.16 Central Limit Theorem demonstration
  17. 3.6.17 Visualization
7. 3.7 Bayes' Theorem and Statistical Thinking
  1. 3.7.1 From joint probability to Bayes
  2. 3.7.2 The four parts
  3. 3.7.3 Medical test example
  4. 3.7.4 Spam filter example
  5. 3.7.5 Naive Bayes: independence assumption
  6. 3.7.6 Maximum likelihood estimation (MLE)
  7. 3.7.7 Maximum a posteriori (MAP)
  8. 3.7.8 Bayesian vs frequentist: the practical difference
  9. 3.7.9 Why Bayesian thinking matters for ML
  10. 3.7.10 Bayes theorem function
  11. 3.7.11 Naive Bayes classifier
  12. 3.7.12 Train on spam data
  13. 3.7.13 Inspect the learned probabilities
  14. 3.7.14 Conjugate Priors
  15. 3.7.15 Sequential Bayesian Updating
  16. 3.7.16 Connection to A/B Testing
8. 3.8 Optimization: Gradient Descent Family
  1. 3.8.1 What optimization means
  2. 3.8.2 Gradient descent (vanilla)
  3. 3.8.3 Learning rate: the most important hyperparameter
  4. 3.8.4 SGD vs batch vs mini-batch
  5. 3.8.5 Momentum: the ball rolling downhill
  6. 3.8.6 Adam: adaptive learning rates
  7. 3.8.7 Learning rate schedules
  8. 3.8.8 Convex vs non-convex
  9. 3.8.9 Loss landscape visualization
  10. 3.8.10 Define a test function
  11. 3.8.11 Vanilla gradient descent
  12. 3.8.12 SGD with momentum
  13. 3.8.13 Adam
  14. 3.8.14 Run and compare
9. 3.9 Information Theory: Entropy, KL Divergence
  1. 3.9.1 Information Content (Surprise)
  2. 3.9.2 Entropy (Average Surprise)
  3. 3.9.3 Cross-Entropy (The Loss Function You Use Every Day)
  4. 3.9.4 KL Divergence (Distance Between Distributions)
  5. 3.9.5 Mutual Information
  6. 3.9.6 Conditional Entropy
  7. 3.9.7 Joint Entropy
  8. 3.9.8 Mutual Information (Deep Dive)
  9. 3.9.9 Label Smoothing and Cross-Entropy
  10. 3.9.10 Why Cross-Entropy Is THE Classification Loss
  11. 3.9.11 Bits vs Nats
  12. 3.9.12 Perplexity
  13. 3.9.13 Information content and entropy
  14. 3.9.14 Cross-entropy and KL divergence
  15. 3.9.15 Cross-entropy as classification loss
  16. 3.9.16 Cross-entropy equals negative log-likelihood
  17. 3.9.17 Mutual information
10. 3.10 Dimensionality Reduction: PCA, t-SNE, UMAP
  1. 3.10.1 The curse of dimensionality
  2. 3.10.2 PCA: find the directions that matter
  3. 3.10.3 Explained variance ratio
  4. 3.10.4 Choosing the number of components
  5. 3.10.5 t-SNE: preserve neighborhoods
  6. 3.10.6 UMAP: faster, better global structure
  7. 3.10.7 When to use which
  8. 3.10.8 Kernel PCA
  9. 3.10.9 Reconstruction Error
  10. 3.10.10 PCA from scratch
  11. 3.10.11 Test on synthetic data
  12. 3.10.12 MNIST digits in 2D
  13. 3.10.13 Compare with sklearn
  14. 3.10.14 UMAP comparison
11. 3.11 Singular Value Decomposition
  1. 3.11.1 What SVD does geometrically
  2. 3.11.2 The full decomposition
  3. 3.11.3 Left singular vectors, singular values, right singular vectors
  4. 3.11.4 Outer product form
  5. 3.11.5 Relationship to eigendecomposition
  6. 3.11.6 Truncated SVD: low-rank approximation
  7. 3.11.7 Image compression with SVD
  8. 3.11.8 SVD for recommendation systems
  9. 3.11.9 SVD in NLP: Latent Semantic Analysis
  10. 3.11.10 SVD for noise reduction
  11. 3.11.11 Pseudoinverse via SVD
  12. 3.11.12 Numerical stability advantages
  13. 3.11.13 Connection to PCA
  14. 3.11.14 SVD from scratch using power iteration
  15. 3.11.15 Test and compare with NumPy
  16. 3.11.16 Image compression demo
  17. 3.11.17 Noise reduction
  18. 3.11.18 Pseudoinverse
12. 3.12 Tensor Operations
  1. 3.12.1 What a tensor is
  2. 3.12.2 Tensor shapes in deep learning
  3. 3.12.3 How memory layout works
  4. 3.12.4 Broadcasting rules
  5. 3.12.5 Einsum: the universal tensor operation
  6. 3.12.6 Tensor storage and strides
  7. 3.12.7 Reshape, squeeze, unsqueeze
  8. 3.12.8 Transpose and permute
  9. 3.12.9 Element-wise operations and reductions
  10. 3.12.10 Broadcasting with NumPy
  11. 3.12.11 Einsum operations
  12. 3.12.12 Attention mechanism via einsum
  13. 3.12.13 Scratch vs NumPy
  14. 3.12.14 Scratch vs PyTorch
  15. 3.12.15 Every neural network layer as a tensor operation
13. 3.13 Numerical Stability
  1. 3.13.1 IEEE 754: How Computers Store Real Numbers
  2. 3.13.2 Why 0.1 + 0.2 != 0.3
  3. 3.13.3 Catastrophic Cancellation
  4. 3.13.4 Overflow and Underflow
  5. 3.13.5 The Log-Sum-Exp Trick
  6. 3.13.6 Why Softmax Needs the Max-Subtraction Trick
  7. 3.13.7 NaN and Inf: Detection and Prevention
  8. 3.13.8 Numerical Gradient Checking
  9. 3.13.9 Mixed Precision Training
  10. 3.13.10 bfloat16 vs float16: Why bfloat16 Wins for Training
  11. 3.13.11 Gradient Clipping
  12. 3.13.12 Normalization Layers as Numerical Stabilizers
  13. 3.13.13 Common ML Numerical Bugs
  14. 3.13.14 Demonstrate floating point precision limits
  15. 3.13.15 Implement naive vs stable softmax
  16. 3.13.16 Implement stable log-sum-exp
  17. 3.13.17 Implement stable cross-entropy
  18. 3.13.18 Gradient checking
  19. 3.13.19 Mixed precision simulation
  20. 3.13.20 Gradient clipping
  21. 3.13.21 NaN/Inf detection
14. 3.14 Norms and Distances
  1. 3.14.1 Norms: measuring vector magnitude
  2. 3.14.2 L1 Norm (Manhattan distance)
  3. 3.14.3 L2 Norm (Euclidean distance)
  4. 3.14.4 Lp Norms: the general family
  5. 3.14.5 L-infinity Norm (Chebyshev distance)
  6. 3.14.6 Cosine Similarity and Cosine Distance
  7. 3.14.7 Dot Product Similarity vs Cosine Similarity
  8. 3.14.8 Mahalanobis Distance
  9. 3.14.9 Jaccard Similarity (for sets)
  10. 3.14.10 Edit Distance (Levenshtein Distance)
  11. 3.14.11 KL Divergence (not a distance, but used like one)
  12. 3.14.12 Wasserstein Distance (Earth Mover's Distance)
  13. 3.14.13 Why Different Tasks Need Different Distances
  14. 3.14.14 Connection to Loss Functions
  15. 3.14.15 Connection to Regularization
  16. 3.14.16 Nearest Neighbor Search
  17. 3.14.17 All norm and distance functions
  18. 3.14.18 Same data, different distances, different neighbors
  19. 3.14.19 Embedding similarity search
15. 3.15 Statistics for ML
  1. 3.15.1 Descriptive Statistics: Summarizing Your Data
  2. 3.15.2 Correlation: How Variables Move Together
  3. 3.15.3 Covariance Matrix
  4. 3.15.4 Hypothesis Testing
  5. 3.15.5 The t-test
  6. 3.15.6 Chi-squared Test
  7. 3.15.7 A/B Testing for ML Models
  8. 3.15.8 Statistical Significance vs Practical Significance
  9. 3.15.9 Multiple Comparison Problem
  10. 3.15.10 Bootstrap Methods
  11. 3.15.11 Parametric vs Non-parametric Tests
  12. 3.15.12 Central Limit Theorem: Practical Implications
  13. 3.15.13 Common Statistical Mistakes in ML Papers
16. 3.16 Sampling Methods
  1. 3.16.1 Why Sampling Matters
  2. 3.16.2 Uniform Random Sampling
  3. 3.16.3 Inverse CDF Method (Inverse Transform Sampling)
  4. 3.16.4 Rejection Sampling
  5. 3.16.5 Importance Sampling
  6. 3.16.6 Monte Carlo Estimation
  7. 3.16.7 Markov Chain Monte Carlo (MCMC): Metropolis-Hastings
  8. 3.16.8 Gibbs Sampling
  9. 3.16.9 Temperature Sampling (Used in LLMs)
  10. 3.16.10 Top-k Sampling
  11. 3.16.11 Top-p (Nucleus) Sampling
  12. 3.16.12 Reparameterization Trick (Used in VAEs)
  13. 3.16.13 Gumbel-Softmax (Differentiable Categorical Sampling)
  14. 3.16.14 Stratified Sampling
  15. 3.16.15 Connection to Diffusion Models
  16. 3.16.16 Uniform and inverse CDF sampling
  17. 3.16.17 Rejection sampling
  18. 3.16.18 Importance sampling
  19. 3.16.19 Monte Carlo estimation of pi
  20. 3.16.20 Metropolis-Hastings MCMC
  21. 3.16.21 Gibbs sampling
  22. 3.16.22 Temperature sampling
  23. 3.16.23 Top-k and top-p sampling
  24. 3.16.24 Reparameterization trick
  25. 3.16.25 Gumbel-Softmax
17. 3.17 Linear Systems
  1. 3.17.1 What Ax = b means geometrically
  2. 3.17.2 Column picture vs row picture
  3. 3.17.3 Gaussian elimination
  4. 3.17.4 Partial pivoting: why it matters
  5. 3.17.5 LU decomposition
  6. 3.17.6 QR decomposition
  7. 3.17.7 Cholesky decomposition
  8. 3.17.8 Least squares: when Ax = b has no exact solution
  9. 3.17.9 Normal equations = linear regression
  10. 3.17.10 Pseudoinverse (Moore-Penrose)
  11. 3.17.11 Condition number
  12. 3.17.12 Iterative methods: conjugate gradient
  13. 3.17.13 The full picture: which method when
  14. 3.17.14 Connection to ML
  15. 3.17.15 Gaussian elimination with partial pivoting
  16. 3.17.16 Least squares via normal equations
18. 3.18 Convex Optimization
  1. 3.18.1 Convex sets
  2. 3.18.2 Convex functions
  3. 3.18.3 Testing for convexity
  4. 3.18.4 Why convexity matters
  5. 3.18.5 Convex vs non-convex in ML
  6. 3.18.6 The Hessian matrix
  7. 3.18.7 Newton's method
  8. 3.18.8 Constrained optimization
  9. 3.18.9 Lagrange multipliers
  10. 3.18.10 KKT conditions
  11. 3.18.11 Regularization as constrained optimization
  12. 3.18.12 Duality
  13. 3.18.13 Why deep learning works despite non-convexity
  14. 3.18.14 Second-order methods in practice
  15. 3.18.15 Convexity checker
  16. 3.18.16 Newton's method for 2D
  17. 3.18.17 Lagrange multiplier solver
  18. 3.18.18 Compare first-order vs second-order
19. 3.19 Complex Numbers for AI
  1. 3.19.1 What is a complex number?
  2. 3.19.2 Complex arithmetic
  3. 3.19.3 The complex plane
  4. 3.19.4 Polar form
  5. 3.19.5 Euler's formula
  6. 3.19.6 Why Euler's formula matters for ML
  7. 3.19.7 Connection to 2D rotations
  8. 3.19.8 Phasors and rotating signals
  9. 3.19.9 Roots of unity
  10. 3.19.10 Connection to the DFT
  11. 3.19.11 Why i is not imaginary
  12. 3.19.12 Complex exponentials vs trigonometric functions
  13. 3.19.13 Connection to transformers
  14. 3.19.14 Complex class
  15. 3.19.15 Polar conversion and Euler's formula
  16. 3.19.16 Rotation
  17. 3.19.17 DFT from complex arithmetic
  18. 3.19.18 Inverse DFT
20. 3.20 The Fourier Transform
  1. 3.20.1 The DFT definition
  2. 3.20.2 What each coefficient means
  3. 3.20.3 Inverse DFT
  4. 3.20.4 The FFT: making it fast
  5. 3.20.5 Spectral analysis
  6. 3.20.6 Frequency resolution
  7. 3.20.7 The convolution theorem
  8. 3.20.8 Windowing
  9. 3.20.9 DFT properties
  10. 3.20.10 Connection to positional encodings
  11. 3.20.11 Connection to CNNs
  12. 3.20.12 Spectrograms and the Short-Time Fourier Transform
  13. 3.20.13 Aliasing
  14. 3.20.14 Zero-padding does not increase resolution
  15. 3.20.15 DFT from scratch
  16. 3.20.16 FFT (Cooley-Tukey)
  17. 3.20.17 Spectral analysis helpers
21. 3.21 Graph Theory for ML
  1. 3.21.1 Graphs: Nodes and Edges
  2. 3.21.2 The Adjacency Matrix
  3. 3.21.3 Degree
  4. 3.21.4 BFS and DFS
  5. 3.21.5 The Graph Laplacian
  6. 3.21.6 Spectral Properties
  7. 3.21.7 Message Passing
  8. 3.21.8 Concepts and ML Applications
  9. 3.21.9 Graph class from scratch
  10. 3.21.10 Connected components and Laplacian eigenvalues
  11. 3.21.11 Spectral clustering
  12. 3.21.12 Message passing
  13. 3.21.13 numpy spectral analysis
22. 3.22 Stochastic Processes
  1. 3.22.1 Random Walks
  2. 3.22.2 Markov Chains
  3. 3.22.3 Connection to Language Models
  4. 3.22.4 Brownian Motion
  5. 3.22.5 Langevin Dynamics
  6. 3.22.6 MCMC: Markov Chain Monte Carlo
  7. 3.22.7 Stochastic Processes in AI
  8. 3.22.8 Random walk simulator
  9. 3.22.9 Markov chain
  10. 3.22.10 Langevin dynamics
  11. 3.22.11 Metropolis-Hastings
  12. 3.22.12 numpy for transition matrices
  13. 3.22.13 Connections to real frameworks
  14. 3.22.14 Verifying Markov chain convergence
4ML fundamentalsnot yet
1. 4.1 What Is Machine Learning
  1. 4.1.1 Learning From Data, Not Rules
  2. 4.1.2 The Three Types of Machine Learning
  3. 4.1.3 Beyond the Big Three
  4. 4.1.4 Classification vs Regression
  5. 4.1.5 The ML Workflow
  6. 4.1.6 Training, Validation, and Test Splits
  7. 4.1.7 Overfitting vs Underfitting
  8. 4.1.8 The Bias-Variance Tradeoff
  9. 4.1.9 No Free Lunch Theorem
  10. 4.1.10 When NOT to Use Machine Learning
  11. 4.1.11 Nearest Centroid Classifier from Scratch
  12. 4.1.12 Train on Synthetic Data
  13. 4.1.13 Compare Against a Baseline
  14. 4.1.14 Why This Matters
  15. 4.1.15 What the Centroid Classifier Cannot Do
2. 4.2 Linear Regression from Scratch
  1. 4.2.1 The Model
  2. 4.2.2 The Cost Function (Mean Squared Error)
  3. 4.2.3 Gradient Descent
  4. 4.2.4 The Normal Equation (Closed-Form Solution)
  5. 4.2.5 Multiple Linear Regression
  6. 4.2.6 Polynomial Regression
  7. 4.2.7 R-Squared Score
  8. 4.2.8 Regularization Preview (Ridge Regression)
  9. 4.2.9 Generate sample data
  10. 4.2.10 Linear regression from scratch with gradient descent
  11. 4.2.11 Normal equation (closed-form solution)
  12. 4.2.12 Multiple linear regression
  13. 4.2.13 Polynomial regression
  14. 4.2.14 Ridge regression (L2 regularization)
3. 4.3 Logistic Regression and Classification
  1. 4.3.1 Why Linear Regression Fails for Classification
  2. 4.3.2 The Sigmoid Function
  3. 4.3.3 Logistic Regression = Linear Model + Sigmoid
  4. 4.3.4 Binary Cross-Entropy Loss
  5. 4.3.5 Gradient Descent for Logistic Regression
  6. 4.3.6 The Decision Boundary
  7. 4.3.7 Multi-Class Classification with Softmax
  8. 4.3.8 Evaluation Metrics
  9. 4.3.9 Sigmoid function and data generation
  10. 4.3.10 Logistic regression from scratch
  11. 4.3.11 Confusion matrix and metrics from scratch
  12. 4.3.12 Decision boundary analysis
  13. 4.3.13 Multi-class with softmax
  14. 4.3.14 Threshold tuning
4. 4.4 Decision Trees and Random Forests
  1. 4.4.1 What a decision tree does
  2. 4.4.2 Split criteria: measuring impurity
  3. 4.4.3 How splitting works
  4. 4.4.4 Stopping conditions
  5. 4.4.5 Decision trees for regression
  6. 4.4.6 Random forests: the power of ensembles
  7. 4.4.7 Feature importance
  8. 4.4.8 When trees beat neural networks
  9. 4.4.9 Gini impurity and entropy
  10. 4.4.10 Find the best split
  11. 4.4.11 Build the DecisionTree class
  12. 4.4.12 Build the RandomForest class
5. 4.5 Support Vector Machines
  1. 4.5.1 The maximum margin classifier
  2. 4.5.2 Support vectors: the critical few
  3. 4.5.3 Soft margin: handling noise with the C parameter
  4. 4.5.4 Hinge loss: the SVM loss function
  5. 4.5.5 Training a linear SVM with gradient descent
  6. 4.5.6 The dual formulation and the kernel trick
  7. 4.5.7 SVM for regression (SVR)
  8. 4.5.8 Why SVMs lost to deep learning (and when they still win)
  9. 4.5.9 Hinge loss and gradient
  10. 4.5.10 Linear SVM via gradient descent
  11. 4.5.11 Kernel functions
  12. 4.5.12 Margin and support vector identification
6. 4.6 KNN and Distance Metrics
  1. 4.6.1 How KNN works
  2. 4.6.2 Choosing K
  3. 4.6.3 Distance metrics
  4. 4.6.4 Weighted KNN
  5. 4.6.5 The curse of dimensionality
  6. 4.6.6 KD-trees: fast nearest neighbor search
  7. 4.6.7 Ball trees: better for moderate dimensions
  8. 4.6.8 Lazy learning vs eager learning
  9. 4.6.9 KNN for regression
  10. 4.6.10 Distance functions
  11. 4.6.11 KNN classifier and regressor
  12. 4.6.12 KD-tree for efficient search
  13. 4.6.13 Feature scaling
7. 4.7 Unsupervised Learning: K-Means, DBSCAN
  1. 4.7.1 Clustering: Grouping Similar Things Together
  2. 4.7.2 K-Means: The Workhorse
  3. 4.7.3 Choosing K
  4. 4.7.4 DBSCAN: Density-Based Clustering
  5. 4.7.5 Hierarchical Clustering
  6. 4.7.6 Gaussian Mixture Models (GMM)
  7. 4.7.7 When to Use Which
  8. 4.7.8 Anomaly Detection with Clustering
  9. 4.7.9 K-Means from scratch
  10. 4.7.10 Elbow method and silhouette score
  11. 4.7.11 DBSCAN from scratch
  12. 4.7.12 Gaussian Mixture Model (EM algorithm)
  13. 4.7.13 Generate test data and run everything
8. 4.8 Feature Engineering and Selection
  1. 4.8.1 The Feature Pipeline
  2. 4.8.2 Numerical Features
  3. 4.8.3 Categorical Features
  4. 4.8.4 Text Features
  5. 4.8.5 Missing Values
  6. 4.8.6 Feature Interaction
  7. 4.8.7 Feature Selection
  8. 4.8.8 Numerical transforms from scratch
  9. 4.8.9 Categorical encoding from scratch
  10. 4.8.10 Text features from scratch
  11. 4.8.11 Missing value imputation from scratch
  12. 4.8.12 Feature selection from scratch
  13. 4.8.13 Full pipeline and demo
9. 4.9 Model Evaluation: Metrics, Cross-Validation
  1. 4.9.1 Train, Validation, Test
  2. 4.9.2 K-Fold Cross-Validation
  3. 4.9.3 Classification Metrics
  4. 4.9.4 Regression Metrics
  5. 4.9.5 Learning Curves
  6. 4.9.6 Validation Curves
  7. 4.9.7 Common Evaluation Mistakes
  8. 4.9.8 Train/validation/test split
  9. 4.9.9 K-fold and stratified K-fold cross-validation
  10. 4.9.10 Confusion matrix and classification metrics
  11. 4.9.11 Regression metrics
  12. 4.9.12 Learning curves
  13. 4.9.13 A simple classifier for testing, plus the full demo
10. 4.10 Bias, Variance and the Learning Curve
  1. 4.10.1 Bias: Systematic Error
  2. 4.10.2 Variance: Sensitivity to Training Data
  3. 4.10.3 The Decomposition
  4. 4.10.4 Model Complexity vs Error
  5. 4.10.5 Regularization as Bias-Variance Control
  6. 4.10.6 Double Descent: The Modern Perspective
  7. 4.10.7 Diagnosing Your Model
  8. 4.10.8 Practical Strategies
  9. 4.10.9 Ensemble Methods and Variance Reduction
  10. 4.10.10 Learning Curves
  11. 4.10.11 How to Generate Learning Curves
  12. 4.10.12 Generate Synthetic Data from a Known Function
  13. 4.10.13 Bootstrap Sampling and Polynomial Fitting
  14. 4.10.14 Computing Bias^2, Variance Decomposition
  15. 4.10.15 Regularization Sweep
  16. 4.10.16 Validation Curve: Sweep Model Complexity
  17. 4.10.17 Learning Curve: Sweep Training Set Size
  18. 4.10.18 Cross-Validation with Regularization Sweep
  19. 4.10.19 Putting It All Together: A Complete Diagnostic Workflow
11. 4.11 Ensemble Methods: Boosting, Bagging, Stacking
  1. 4.11.1 Why Ensembles Work
  2. 4.11.2 Bagging (Bootstrap Aggregating)
  3. 4.11.3 Boosting (Sequential Error Correction)
  4. 4.11.4 AdaBoost
  5. 4.11.5 Gradient Boosting
  6. 4.11.6 XGBoost: Why It Dominates Tabular Data
  7. 4.11.7 Stacking (Meta-Learning)
  8. 4.11.8 Voting
  9. 4.11.9 Decision Stump (Base Learner)
  10. 4.11.10 AdaBoost from Scratch
  11. 4.11.11 Gradient Boosting from Scratch
  12. 4.11.12 Compare against sklearn
  13. 4.11.13 When to Use Each Method
  14. 4.11.14 The Production Stack for Tabular Data
12. 4.12 Hyperparameter Tuning
  1. 4.12.1 Parameters vs Hyperparameters
  2. 4.12.2 Grid Search
  3. 4.12.3 Random Search
  4. 4.12.4 Bayesian Optimization
  5. 4.12.5 Early Stopping
  6. 4.12.6 Learning Rate Schedulers
  7. 4.12.7 Hyperparameter Importance
  8. 4.12.8 Practical Strategy
  9. 4.12.9 Cross-Validation Integration
  10. 4.12.10 Practical Tips
  11. 4.12.11 Grid Search from Scratch
  12. 4.12.12 Random Search from Scratch
  13. 4.12.13 Bayesian Optimization (Simplified)
  14. 4.12.14 Compare All Methods
  15. 4.12.15 Optuna in Practice
  16. 4.12.16 Optuna with Pruning
  17. 4.12.17 sklearn's Built-in Tuners
  18. 4.12.18 Common Mistakes in Hyperparameter Tuning
13. 4.13 ML Pipelines and Experiment Tracking
  1. 4.13.1 What a Pipeline Is
  2. 4.13.2 Data Leakage: The Silent Killer
  3. 4.13.3 sklearn Pipeline
  4. 4.13.4 ColumnTransformer: Different Pipelines for Different Columns
  5. 4.13.5 Experiment Tracking
  6. 4.13.6 Model Versioning
  7. 4.13.7 Data Versioning with DVC
  8. 4.13.8 Reproducible Experiments
  9. 4.13.9 From Notebook to Production Pipeline
  10. 4.13.10 Common Pipeline Mistakes
  11. 4.13.11 Custom Transformer
  12. 4.13.12 Pipeline from Scratch
  13. 4.13.13 Cross-Validation with Pipeline
  14. 4.13.14 Full Production Pipeline with sklearn
14. 4.14 Naive Bayes
  1. 4.14.1 Bayes' Theorem (Quick Review)
  2. 4.14.2 The Naive Independence Assumption
  3. 4.14.3 Why It Still Works
  4. 4.14.4 The Math Step by Step
  5. 4.14.5 Three Variants
  6. 4.14.6 When to Use Each Variant
  7. 4.14.7 Laplace Smoothing
  8. 4.14.8 Log-Space Computation
  9. 4.14.9 Naive Bayes vs Logistic Regression
  10. 4.14.10 Classification Pipeline
  11. 4.14.11 MultinomialNB
  12. 4.14.12 GaussianNB
  13. 4.14.13 Demo: Text Classification
  14. 4.14.14 Demo: Continuous Features
  15. 4.14.15 Prediction Speed
  16. 4.14.16 TF-IDF with Naive Bayes
  17. 4.14.17 BernoulliNB for Short Text
  18. 4.14.18 Calibrating NB Probabilities
  19. 4.14.19 Common Gotchas
  20. 4.14.20 When Naive Bayes Fails
15. 4.15 Time Series Fundamentals
  1. 4.15.1 What Makes Time Series Different
  2. 4.15.2 Components of a Time Series
  3. 4.15.3 Stationarity
  4. 4.15.4 Autocorrelation
  5. 4.15.5 Lag Features: Turning Time Series into Supervised Learning
  6. 4.15.6 Walk-Forward Validation
  7. 4.15.7 ARIMA Intuition
  8. 4.15.8 When to Use What
  9. 4.15.9 Forecasting Horizons and Strategies
  10. 4.15.10 Common Mistakes in Time Series
  11. 4.15.11 Lag Feature Creator
  12. 4.15.12 Walk-Forward Cross-Validation
  13. 4.15.13 Simple Autoregressive Model
  14. 4.15.14 Stationarity Check
  15. 4.15.15 sklearn TimeSeriesSplit
  16. 4.15.16 Evaluation Metrics
  17. 4.15.17 Rolling Features
  18. 4.15.18 Baselines You Must Beat
  19. 4.15.19 Practical Tips
16. 4.16 Anomaly Detection
  1. 4.16.1 Types of Anomalies
  2. 4.16.2 The Unsupervised Framing
  3. 4.16.3 Supervised vs Unsupervised: The Tradeoff
  4. 4.16.4 Z-Score Method
  5. 4.16.5 IQR Method
  6. 4.16.6 Isolation Forest
  7. 4.16.7 Local Outlier Factor (LOF)
  8. 4.16.8 Comparison
  9. 4.16.9 Evaluation Challenges
  10. 4.16.10 Anomaly Detection Pipeline
  11. 4.16.11 Z-Score Detector
  12. 4.16.12 IQR Detector
  13. 4.16.13 Isolation Forest from Scratch
  14. 4.16.14 Demo Scenarios
  15. 4.16.15 sklearn Contamination Parameter
  16. 4.16.16 One-Class SVM
  17. 4.16.17 Autoencoder Approach (Preview)
  18. 4.16.18 Ensemble Anomaly Detection
  19. 4.16.19 Production Considerations
  20. 4.16.20 Choosing a Threshold
  21. 4.16.21 Scaling to Production
17. 4.17 Handling Imbalanced Data
  1. 4.17.1 Why Accuracy Fails
  2. 4.17.2 Better Metrics
  3. 4.17.3 The Imbalanced Data Pipeline
  4. 4.17.4 SMOTE: Synthetic Minority Oversampling Technique
  5. 4.17.5 Sampling Strategies Compared
  6. 4.17.6 Class Weights
  7. 4.17.7 Threshold Tuning
  8. 4.17.8 Cost-Sensitive Learning
  9. 4.17.9 Decision Flowchart
  10. 4.17.10 Generate an imbalanced dataset
  11. 4.17.11 SMOTE from scratch
  12. 4.17.12 Random oversampling and undersampling
  13. 4.17.13 Logistic regression with class weights
  14. 4.17.14 Threshold tuning
  15. 4.17.15 Evaluation functions
  16. 4.17.16 Compare all approaches
18. 4.18 Feature Selection
  1. 4.18.1 Three Categories of Feature Selection
  2. 4.18.2 Variance Threshold
  3. 4.18.3 Mutual Information
  4. 4.18.4 Recursive Feature Elimination (RFE)
  5. 4.18.5 L1 (Lasso) Regularization
  6. 4.18.6 Tree-Based Feature Importance
  7. 4.18.7 Permutation Importance
  8. 4.18.8 Comparison Table
  9. 4.18.9 Decision Flowchart
  10. 4.18.10 Generate synthetic data with known feature structure
  11. 4.18.11 Variance threshold
  12. 4.18.12 Mutual information (discrete)
  13. 4.18.13 Recursive Feature Elimination
  14. 4.18.14 L1 feature selection
  15. 4.18.15 Tree-based importance (simple decision tree)
  16. 4.18.16 Run all methods and compare
19. 4.19 ML System Design — From Problem to Deployed Model
  1. 4.19.1 From a business goal to an ML task
  2. 4.19.2 When not to use ML
  3. 4.19.3 Three layers of metrics
  4. 4.19.4 Baselines
  5. 4.19.5 Data and label requirements
  6. 4.19.6 Latency, throughput, and cost budgets
  7. 4.19.7 Batch inference versus online inference
  8. 4.19.8 Deployment patterns
  9. 4.19.9 Feedback loops
  10. 4.19.10 Failure modes and safe defaults
  11. 4.19.11 The design document
  12. 4.19.12 Describe the workload and the prices
  13. 4.19.13 Price batch and online serving
  14. 4.19.14 Compare a model with its baselines
  15. 4.19.15 Score a design document
  16. 4.19.16 Run it
20. 4.20 Monitoring Models in Production — Data Drift and Concept Drift
  1. 4.20.1 What to monitor
  2. 4.20.2 Three kinds of drift
  3. 4.20.3 Delayed labels
  4. 4.20.4 Reference and current windows
  5. 4.20.5 Population Stability Index
  6. 4.20.6 Kolmogorov-Smirnov statistic
  7. 4.20.7 Chi-square test for categorical features
  8. 4.20.8 Alert thresholds and false alarms
  9. 4.20.9 Retraining triggers
  10. 4.20.10 Quantile bins and PSI
  11. 4.20.11 The KS statistic and its p-value
  12. 4.20.12 The chi-square test
  13. 4.20.13 Windows and the drift monitor
  14. 4.20.14 The alert policy
  15. 4.20.15 A stream with injected drift
21. 4.21 Feature Stores and Training-Serving Skew
  1. 4.21.1 Two readers of the same feature
  2. 4.21.2 Entities, timestamps, and TTL
  3. 4.21.3 Point-in-time correct joins
  4. 4.21.4 Causes of training-serving skew
  5. 4.21.5 A feature registry
  6. 4.21.6 Backfills
  7. 4.21.7 Detecting skew
  8. 4.21.8 Feature definitions and the registry
  9. 4.21.9 Backfill the offline store
  10. 4.21.10 The point-in-time join
  11. 4.21.11 Measure the leak from a naive join
  12. 4.21.12 The online store and materialization
  13. 4.21.13 The skew check
5Deep learning corenot yet
1. 5.1 The Perceptron: Where It All Started
  1. 5.1.1 One Neuron, One Decision
  2. 5.1.2 The Decision Boundary
  3. 5.1.3 The Learning Rule
  4. 5.1.4 The XOR Problem
  5. 5.1.5 The Perceptron class
  6. 5.1.6 Train on logic gates
  7. 5.1.7 Watch XOR fail
  8. 5.1.8 Solve XOR with two layers
  9. 5.1.9 Train a Two-Layer Network
2. 5.2 Multi-Layer Networks and Forward Pass
  1. 5.2.1 Layers: Input, Hidden, Output
  2. 5.2.2 Neurons and Activations
  3. 5.2.3 Forward Pass: How Data Flows
  4. 5.2.4 Matrix Dimensions
  5. 5.2.5 Universal Approximation Theorem
  6. 5.2.6 Composability
  7. 5.2.7 Sigmoid Activation
  8. 5.2.8 Layer Class
  9. 5.2.9 Network Class
  10. 5.2.10 XOR with Hand-Tuned Weights
  11. 5.2.11 Circle Classification
3. 5.3 Backpropagation from Scratch
  1. 5.3.1 The Chain Rule, Applied to Networks
  2. 5.3.2 Computational Graphs
  3. 5.3.3 Forward vs Backward
  4. 5.3.4 Gradient Flow Through a Network
  5. 5.3.5 Vanishing Gradients
  6. 5.3.6 Deriving Gradients for a 2-Layer Network
  7. 5.3.7 The Value Node
  8. 5.3.8 Operations with Backward Functions
  9. 5.3.9 Sigmoid and Loss
  10. 5.3.10 Backward Pass
  11. 5.3.11 Layer and Network
  12. 5.3.12 Train on XOR
  13. 5.3.13 Circle Classification
4. 5.4 Activation Functions: ReLU, Sigmoid, GELU and Why
  1. 5.4.1 Why Nonlinearity Is Necessary
  2. 5.4.2 Sigmoid
  3. 5.4.3 Tanh
  4. 5.4.4 ReLU: The Breakthrough
  5. 5.4.5 Leaky ReLU
  6. 5.4.6 GELU: The Modern Default
  7. 5.4.7 Swish / SiLU
  8. 5.4.8 Softmax: The Output Activation
  9. 5.4.9 Comparison of Shapes
  10. 5.4.10 Gradient Flow Comparison
  11. 5.4.11 Which Activation When
  12. 5.4.12 Implement All Activation Functions with Derivatives
  13. 5.4.13 Visualize Where Gradients Die
  14. 5.4.14 Vanishing Gradient Experiment
  15. 5.4.15 Dead Neuron Detector
  16. 5.4.16 Training Comparison -- Sigmoid vs ReLU vs GELU
5. 5.5 Loss Functions: MSE, Cross-Entropy, Contrastive
  1. 5.5.1 Mean Squared Error (MSE)
  2. 5.5.2 Cross-Entropy Loss
  3. 5.5.3 Why MSE Fails for Classification
  4. 5.5.4 Label Smoothing
  5. 5.5.5 Contrastive Loss
  6. 5.5.6 Focal Loss
  7. 5.5.7 Loss Function Decision Tree
  8. 5.5.8 Loss Landscape
  9. 5.5.9 MSE and Its Gradient
  10. 5.5.10 Binary Cross-Entropy
  11. 5.5.11 Categorical Cross-Entropy with Softmax
  12. 5.5.12 Contrastive Loss (Simplified InfoNCE)
  13. 5.5.13 MSE vs Cross-Entropy on Classification
6. 5.6 Optimizers: SGD, Momentum, Adam, AdamW
  1. 5.6.1 Stochastic Gradient Descent (SGD)
  2. 5.6.2 Momentum
  3. 5.6.3 RMSProp
  4. 5.6.4 Adam: Momentum + RMSProp
  5. 5.6.5 AdamW: Weight Decay Done Right
  6. 5.6.6 Learning Rate: The Most Important Hyperparameter
  7. 5.6.7 Optimizer Comparison
  8. 5.6.8 When Each Optimizer Wins
  9. 5.6.9 Vanilla SGD
  10. 5.6.10 SGD with Momentum
  11. 5.6.11 Adam
  12. 5.6.12 AdamW
  13. 5.6.13 Training Comparison
7. 5.7 Regularization: Dropout, Weight Decay, BatchNorm
  1. 5.7.1 The Overfitting Spectrum
  2. 5.7.2 Dropout
  3. 5.7.3 Weight Decay (L2 Regularization)
  4. 5.7.4 Batch Normalization
  5. 5.7.5 Layer Normalization
  6. 5.7.6 RMSNorm
  7. 5.7.7 Normalization Comparison
  8. 5.7.8 Data Augmentation as Regularization
  9. 5.7.9 Early Stopping
  10. 5.7.10 When to Apply What
  11. 5.7.11 Dropout (Train and Eval Mode)
  12. 5.7.12 L2 Weight Decay
  13. 5.7.13 Training With and Without Regularization
8. 5.8 Weight Initialization and Training Stability
  1. 5.8.1 The Symmetry Problem
  2. 5.8.2 Variance Propagation Through Layers
  3. 5.8.3 Xavier/Glorot Initialization
  4. 5.8.4 Kaiming/He Initialization
  5. 5.8.5 Transformer Initialization
  6. 5.8.6 Activation Magnitude Through 50 Layers
  7. 5.8.7 Choosing the Right Init
  8. 5.8.8 Initialization Strategies
  9. 5.8.9 Activation Functions
  10. 5.8.10 Forward Pass Through 50 Layers
  11. 5.8.11 The Experiment
  12. 5.8.12 Symmetry Demonstration
  13. 5.8.13 Layer-by-Layer Magnitude Report
9. 5.9 Learning Rate Schedules and Warmup
  1. 5.9.1 Constant Learning Rate
  2. 5.9.2 Step Decay
  3. 5.9.3 Cosine Annealing
  4. 5.9.4 Warmup: Why You Start Small
  5. 5.9.5 Linear Warmup + Cosine Decay
  6. 5.9.6 1cycle Policy
  7. 5.9.7 Schedule Shapes
  8. 5.9.8 Decision Flowchart
  9. 5.9.9 Real Numbers from Published Models
  10. 5.9.10 Schedule Functions
  11. 5.9.11 Visualize All Schedules
  12. 5.9.12 Training Network
  13. 5.9.13 Compare All Schedules
  14. 5.9.14 LR Too High vs Too Low
10. 5.10 Build Your Own Mini Framework
  1. 5.10.1 The Module Abstraction
  2. 5.10.2 Sequential Container
  3. 5.10.3 Training vs Evaluation Mode
  4. 5.10.4 Optimizer
  5. 5.10.5 DataLoader
  6. 5.10.6 Framework Architecture
  7. 5.10.7 Training Loop
  8. 5.10.8 Module Hierarchy
  9. 5.10.9 Module Base Class
  10. 5.10.10 Linear Layer
  11. 5.10.11 Activation Modules
  12. 5.10.12 Dropout Module
  13. 5.10.13 BatchNorm Module
  14. 5.10.14 Loss Functions
  15. 5.10.15 SGD and Adam Optimizers
  16. 5.10.16 Train a 4-Layer Network on Circle Classification
11. 5.11 Introduction to PyTorch
  1. 5.11.1 Why PyTorch Won
  2. 5.11.2 Tensors
  3. 5.11.3 Autograd
  4. 5.11.4 nn.Module
  5. 5.11.5 Loss Functions and Optimizers
  6. 5.11.6 The Training Loop
  7. 5.11.7 Dataset and DataLoader
  8. 5.11.8 GPU Training
  9. 5.11.9 Comparison: Mini Framework vs PyTorch vs JAX
  10. 5.11.10 Load MNIST From Raw Files
  11. 5.11.11 Define the Model
  12. 5.11.12 Training Loop
  13. 5.11.13 Wire Everything Together
  14. 5.11.14 Quick Comparison: Mini Framework vs PyTorch
  15. 5.11.15 Saving and Loading Models
  16. 5.11.16 Learning Rate Scheduling
12. 5.12 Introduction to JAX
  1. 5.12.1 The JAX Philosophy
  2. 5.12.2 jax.numpy: The Familiar Surface
  3. 5.12.3 jax.grad: Functional Autodiff
  4. 5.12.4 jit: Compile to XLA
  5. 5.12.5 vmap: Automatic Vectorization
  6. 5.12.6 pmap: Data Parallelism Across Devices
  7. 5.12.7 Pytrees: The Universal Data Structure
  8. 5.12.8 Functional vs Object-Oriented
  9. 5.12.9 The JAX Ecosystem
  10. 5.12.10 When to Use JAX vs PyTorch
  11. 5.12.11 Random Numbers in JAX
  12. 5.12.12 Setup and Data
  13. 5.12.13 Initialize Parameters
  14. 5.12.14 Forward Pass
  15. 5.12.15 JIT-Compiled Training Step
  16. 5.12.16 Training Loop
  17. 5.12.17 Flax: The Google Standard
  18. 5.12.18 Equinox: The Pythonic Alternative
  19. 5.12.19 Optax: Composable Optimizers
13. 5.13 Debugging Neural Networks
  1. 5.13.1 The Debugging Mindset
  2. 5.13.2 Symptom 1: Loss Not Decreasing
  3. 5.13.3 Symptom 2: Loss Decreasing But Model is Bad
  4. 5.13.4 Symptom 3: NaN or Inf in Loss
  5. 5.13.5 Technique 1: Gradient Checking
  6. 5.13.6 Technique 2: Activation Statistics
  7. 5.13.7 Technique 3: Gradient Flow Visualization
  8. 5.13.8 Technique 4: The Overfit-One-Batch Test
  9. 5.13.9 Technique 5: Learning Rate Finder
  10. 5.13.10 Common PyTorch Bugs
  11. 5.13.11 The Master Debugging Table
  12. 5.13.12 The NetworkDebugger Class
  13. 5.13.13 The Overfit-One-Batch Test
  14. 5.13.14 Learning Rate Finder
  15. 5.13.15 Gradient Checker
  16. 5.13.16 Deliberately Broken Networks
  17. 5.13.17 PyTorch Built-in Tools
  18. 5.13.18 Weights and Biases Integration
  19. 5.13.19 TensorBoard
  20. 5.13.20 The Debug Checklist (Before Full Training)
6Computer visionnot yet
1. 6.1 Image Fundamentals: Pixels, Channels, Color Spaces
  1. 6.1.1 The full preprocessing pipeline at a glance
  2. 6.1.2 A pixel is a sample, not a square
  3. 6.1.3 Why three channels
  4. 6.1.4 Two layout conventions: HWC and CHW
  5. 6.1.5 Byte ranges and dtype
  6. 6.1.6 Color spaces and why they exist
  7. 6.1.7 Aspect ratio, resizing, and interpolation
  8. 6.1.8 Build an image tensor and inspect its shape
  9. 6.1.9 Split channels and re-order layout
  10. 6.1.10 Grayscale and HSV conversions
  11. 6.1.11 Normalize, standardize, and reverse it
  12. 6.1.12 Resize from scratch
2. 6.2 Convolutions from Scratch
  1. 6.2.1 One kernel, sliding
  2. 6.2.2 Output size formula
  3. 6.2.3 Padding
  4. 6.2.4 Stride
  5. 6.2.5 Multiple input channels
  6. 6.2.6 The im2col trick
  7. 6.2.7 Receptive field
  8. 6.2.8 Pad an array
  9. 6.2.9 2D convolution with nested loops
  10. 6.2.10 Verify with a hand-designed kernel
  11. 6.2.11 im2col
  12. 6.2.12 Fast conv via im2col + matmul
  13. 6.2.13 A bank of hand-designed kernels
3. 6.3 CNNs: LeNet to ResNet
  1. 6.3.1 The four ideas that changed vision
  2. 6.3.2 LeNet-5 (1998)
  3. 6.3.3 AlexNet (2012)
  4. 6.3.4 VGG (2014)
  5. 6.3.5 Inception (2014, same year)
  6. 6.3.6 The degradation problem
  7. 6.3.7 ResNet (2015)
  8. 6.3.8 Why residuals matter beyond vision
  9. 6.3.9 LeNet-5
  10. 6.3.10 A VGG block
  11. 6.3.11 A ResNet BasicBlock
  12. 6.3.12 A tiny ResNet
  13. 6.3.13 Compare parameter-to-feature efficiency
4. 6.4 Image Classification
  1. 6.4.1 The classification pipeline
  2. 6.4.2 Cross-entropy, logits, and softmax
  3. 6.4.3 Why augmentation works
  4. 6.4.4 Mixup and cutmix
  5. 6.4.5 Label smoothing
  6. 6.4.6 Evaluation beyond accuracy
  7. 6.4.7 A deterministic synthetic dataset
  8. 6.4.8 Normalisation and augmentation
  9. 6.4.9 Mixup
  10. 6.4.10 The training loop
  11. 6.4.11 Put it together
  12. 6.4.12 Read the confusion matrix
5. 6.5 Transfer Learning and Fine-Tuning
  1. 6.5.1 Feature extraction vs fine-tuning
  2. 6.5.2 Why freezing works at all
  3. 6.5.3 Discriminative learning rates
  4. 6.5.4 The BatchNorm problem
  5. 6.5.5 Head design
  6. 6.5.6 Layer-wise LR decay
  7. 6.5.7 What to evaluate
  8. 6.5.8 Load a pretrained backbone and inspect it
  9. 6.5.9 Feature extraction — freeze everything, replace the head
  10. 6.5.10 Discriminative fine-tuning
  11. 6.5.11 BatchNorm handling
  12. 6.5.12 A minimal end-to-end fine-tuning loop
  13. 6.5.13 Progressive unfreezing
6. 6.6 Object Detection — YOLO from Scratch
  1. 6.6.1 Detection as dense prediction
  2. 6.6.2 Why grids and anchors
  3. 6.6.3 Decoding predictions
  4. 6.6.4 IoU
  5. 6.6.5 Non-maximum suppression
  6. 6.6.6 The loss
  7. 6.6.7 Detection metrics
  8. 6.6.8 Non-max suppression
  9. 6.6.9 Box encoding and decoding
  10. 6.6.10 A minimal YOLO head
  11. 6.6.11 Ground-truth assignment
  12. 6.6.12 The three losses
  13. 6.6.13 Inference pipeline
7. 6.7 Semantic Segmentation — U-Net
  1. 6.7.1 Semantic vs instance vs panoptic
  2. 6.7.2 The U-Net shape
  3. 6.7.3 Transposed vs bilinear upsample
  4. 6.7.4 Cross-entropy on a pixel grid
  5. 6.7.5 Dice loss and why you need it
  6. 6.7.6 Evaluation metrics
  7. 6.7.7 Input resolution trade-off
  8. 6.7.8 Encoder block
  9. 6.7.9 Down and up blocks
  10. 6.7.10 The U-Net
  11. 6.7.11 Losses
  12. 6.7.12 IoU metric
  13. 6.7.13 Synthetic dataset for end-to-end verification
  14. 6.7.14 Training loop
8. 6.8 Instance Segmentation — Mask R-CNN
  1. 6.8.1 The architecture
  2. 6.8.2 Why RoIAlign, not RoIPool
  3. 6.8.3 The RPN in one paragraph
  4. 6.8.4 The mask head
  5. 6.8.5 Losses
  6. 6.8.6 Output format
  7. 6.8.7 RoIAlign from scratch
  8. 6.8.8 Compare to torchvision's RoIAlign
  9. 6.8.9 Load a pretrained Mask R-CNN
  10. 6.8.10 Run inference
  11. 6.8.11 Swap the heads for a custom class count
  12. 6.8.12 Freeze what does not need training
9. 6.9 Image Generation — GANs
  1. 6.9.1 The two networks
  2. 6.9.2 The game
  3. 6.9.3 Non-saturating loss
  4. 6.9.4 DCGAN architecture rules
  5. 6.9.5 Failure modes and their signatures
  6. 6.9.6 Evaluation
  7. 6.9.7 Generator
  8. 6.9.8 Discriminator
  9. 6.9.9 Training step
  10. 6.9.10 Full training loop on synthetic shapes
  11. 6.9.11 Sampling
  12. 6.9.12 Spectral normalisation
10. 6.10 Image Generation — Diffusion Models
  1. 6.10.1 The forward process
  2. 6.10.2 The closed-form jump
  3. 6.10.3 The reverse process
  4. 6.10.4 The training loss
  5. 6.10.5 The sampler (DDPM)
  6. 6.10.6 Why 1000 steps
  7. 6.10.7 DDIM: 20x faster sampling
  8. 6.10.8 Time conditioning
  9. 6.10.9 Noise schedule
  10. 6.10.10 Forward diffusion (q_sample)
  11. 6.10.11 A tiny time-conditioned U-Net
  12. 6.10.12 Training loop
  13. 6.10.13 Sampler (DDPM)
  14. 6.10.14 DDIM sampler (deterministic, ~20x faster)
11. 6.11 Stable Diffusion — Architecture and Fine-Tuning
  1. 6.11.1 The pipeline
  2. 6.11.2 Classifier-free guidance (CFG)
  3. 6.11.3 Latent space geometry
  4. 6.11.4 The U-Net architecture
  5. 6.11.5 LoRA fine-tuning
  6. 6.11.6 Schedulers you will see
  7. 6.11.7 Text-to-image
  8. 6.11.8 Swap the scheduler
  9. 6.11.9 Image-to-image
  10. 6.11.10 Inpainting
  11. 6.11.11 LoRA loading
  12. 6.11.12 LoRA training (sketch)
12. 6.12 Video Understanding — Temporal Modeling
  1. 6.12.1 The three architectural families
  2. 6.12.2 2D + pool
  3. 6.12.3 3D convolutions
  4. 6.12.4 Spatio-temporal transformers
  5. 6.12.5 Frame sampling
  6. 6.12.6 Evaluation
  7. 6.12.7 Datasets you will meet
  8. 6.12.8 Frame sampler
  9. 6.12.9 A 2D+pool baseline
  10. 6.12.10 An I3D-style inflated 3D conv
  11. 6.12.11 Factorised (2+1)D conv
13. 6.13 3D Vision: Point Clouds, NeRFs
  1. 6.13.1 Point clouds
  2. 6.13.2 The PointNet architecture
  3. 6.13.3 Neural Radiance Fields (NeRFs)
  4. 6.13.4 Positional encoding in NeRF
  5. 6.13.5 Volumetric rendering
  6. 6.13.6 What replaced NeRFs
  7. 6.13.7 Datasets and benchmarks
  8. 6.13.8 PointNet classifier
  9. 6.13.9 Positional encoding
  10. 6.13.10 Tiny NeRF MLP
  11. 6.13.11 Volumetric rendering along a ray
14. 6.14 Vision Transformers (ViT)
  1. 6.14.1 The pipeline
  2. 6.14.2 Patch embedding
  3. 6.14.3 Class token
  4. 6.14.4 Positional embedding
  5. 6.14.5 Transformer encoder block
  6. 6.14.6 Why pre-LN
  7. 6.14.7 Patch size trade-off
  8. 6.14.8 DeiT's recipe for training ViT on ImageNet-1k
  9. 6.14.9 Swin vs ConvNeXt
  10. 6.14.10 MAE pretraining
  11. 6.14.11 Transformer block
  12. 6.14.12 The ViT
  13. 6.14.13 Sanity check — single image inference
15. 6.15 Real-Time Vision: Edge Deployment
  1. 6.15.1 The three budgets
  2. 6.15.2 Measurement discipline
  3. 6.15.3 FLOPs as a proxy
  4. 6.15.4 Quantisation in one paragraph
  5. 6.15.5 Pruning and distillation
  6. 6.15.6 The inference runtimes
  7. 6.15.7 Edge architecture picker
  8. 6.15.8 Measure latency correctly
  9. 6.15.9 Parameter and FLOP counts
  10. 6.15.10 Post-training static quantisation
  11. 6.15.11 Export to ONNX
  12. 6.15.12 Benchmark and compare regimes
16. 6.16 Build a Complete Vision Pipeline
  1. 6.16.1 The pipeline
  2. 6.16.2 Data contracts with Pydantic
  3. 6.16.3 Where latency goes
  4. 6.16.4 Failure modes
  5. 6.16.5 Batching
  6. 6.16.6 Data contracts
  7. 6.16.7 A minimal Pipeline class
  8. 6.16.8 Wire a detector and a classifier
  9. 6.16.9 FastAPI service
  10. 6.16.10 Benchmark the pipeline
17. 6.17 Self-Supervised Vision — SimCLR, DINO, MAE
  1. 6.17.1 Three families
  2. 6.17.2 Contrastive learning (SimCLR)
  3. 6.17.3 Teacher-student (DINO)
  4. 6.17.4 Masked reconstruction (MAE)
  5. 6.17.5 Why 75% and not 15%
  6. 6.17.6 Linear-probe evaluation
  7. 6.17.7 Two-view augmentation pipeline
  8. 6.17.8 InfoNCE loss
  9. 6.17.9 Sanity check InfoNCE
  10. 6.17.10 MAE-style masking
18. 6.18 Open-Vocabulary Vision — CLIP
  1. 6.18.1 Two towers
  2. 6.18.2 The objective
  3. 6.18.3 SigLIP: a better loss
  4. 6.18.4 Zero-shot classification
  5. 6.18.5 Where CLIP-style models are used in 2026
  6. 6.18.6 A tiny two-tower model
  7. 6.18.7 Contrastive loss
  8. 6.18.8 Zero-shot classifier
  9. 6.18.9 Sanity check
19. 6.19 OCR and Document Understanding
  1. 6.19.1 The classical pipeline
  2. 6.19.2 CTC in one paragraph
  3. 6.19.3 Modern end-to-end models
  4. 6.19.4 Layout parsing
  5. 6.19.5 Evaluation metrics
  6. 6.19.6 CTC loss + greedy decoder
  7. 6.19.7 Tiny CRNN recogniser
  8. 6.19.8 Synthetic OCR
  9. 6.19.9 Training sketch
20. 6.20 Image Retrieval and Metric Learning
  1. 6.20.1 Retrieval at a glance
  2. 6.20.2 The four loss families
  3. 6.20.3 Triplet loss formally
  4. 6.20.4 Cosine similarity vs L2
  5. 6.20.5 Recall@K
  6. 6.20.6 FAISS in one paragraph
  7. 6.20.7 Instance-level vs category-level retrieval
  8. 6.20.8 Triplet loss
  9. 6.20.9 Semi-hard mining
  10. 6.20.10 Putting it together
21. 6.21 Keypoint Detection and Pose Estimation
  1. 6.21.1 Top-down vs bottom-up
  2. 6.21.2 Heatmap regression
  3. 6.21.3 Sub-pixel localisation
  4. 6.21.4 Part Affinity Fields (PAFs)
  5. 6.21.5 COCO keypoints
  6. 6.21.6 2D vs 3D
  7. 6.21.7 Gaussian heatmap target
  8. 6.21.8 Tiny keypoint head
  9. 6.21.9 Inference — extract keypoint coordinates
  10. 6.21.10 Synthetic keypoint dataset
  11. 6.21.11 Training
22. 6.22 3D Gaussian Splatting from Scratch
  1. 6.22.1 What a Gaussian carries
  2. 6.22.2 Rasterisation, not ray marching
  3. 6.22.3 The projection step
  4. 6.22.4 The alpha-compositing rule
  5. 6.22.5 Why this is differentiable
  6. 6.22.6 Densification and pruning
  7. 6.22.7 Spherical harmonics in one paragraph
  8. 6.22.8 The 2026 production stack
  9. 6.22.9 4D and generative variants
  10. 6.22.10 A 2D Gaussian
  11. 6.22.11 2D splatting rasteriser
  12. 6.22.12 A trainable 2D splat scene
  13. 6.22.13 Fit 2D Gaussians to a target image
  14. 6.22.14 From 2D to 3D
  15. 6.22.15 Spherical harmonics evaluation
23. 6.23 Diffusion Transformers and Rectified Flow
  1. 6.23.1 From U-Net to transformer
  2. 6.23.2 Rectified flow in one paragraph
  3. 6.23.3 AdaLN conditioning
  4. 6.23.4 Text encoders in SD3 and FLUX
  5. 6.23.5 Classifier-free guidance still holds
  6. 6.23.6 Consistency, Turbo, Schnell, LCM
  7. 6.23.7 Model landscape in 2026
  8. 6.23.8 Why this phase shift matters
  9. 6.23.9 A DiT block with AdaLN
  10. 6.23.10 A tiny DiT
  11. 6.23.11 Rectified flow training
  12. 6.23.12 Euler sampler
  13. 6.23.13 End-to-end smoke test
24. 6.24 SAM 3 and Open-Vocabulary Segmentation
  1. 6.24.1 The three generations
  2. 6.24.2 Promptable Concept Segmentation
  3. 6.24.3 Key architectural pieces
  4. 6.24.4 Training at scale
  5. 6.24.5 SAM 3.1 Object Multiplex
  6. 6.24.6 Where Grounded SAM still matters in 2026
  7. 6.24.7 YOLO-World vs SAM 3
  8. 6.24.8 SAM-MI efficiency
  9. 6.24.9 Output format for the three models
  10. 6.24.10 Prompt construction
  11. 6.24.11 Post-processing helpers
  12. 6.24.12 A unified open-vocab segmentation interface
  13. 6.24.13 Hugging Face SAM 3 usage (reference)
  14. 6.24.14 Measure what Grounded SAM 2 gave you for free
25. 6.25 Vision-Language Models (ViT-MLP-LLM)
  1. 6.25.1 The ViT-MLP-LLM architecture
  2. 6.25.2 DeepStack
  3. 6.25.3 Three training stages
  4. 6.25.4 Model family comparison (early 2026)
  5. 6.25.5 Visual agents
  6. 6.25.6 Agentic capabilities + RoPE variants
  7. 6.25.7 The alignment problem
  8. 6.25.8 Fine-tuning with LoRA / QLoRA
  9. 6.25.9 Spatial reasoning is still weak
  10. 6.25.10 The projector
  11. 6.25.11 Assemble ViT-MLP-LLM end-to-end
  12. 6.25.12 CMER computation
  13. 6.25.13 Toy VLM classifier (runnable)
26. 6.26 Monocular Depth and Geometry Estimation
  1. 6.26.1 Relative vs metric depth
  2. 6.26.2 The encoder-decoder pattern
  3. 6.26.3 Why a single image produces depth at all
  4. 6.26.4 What monocular depth cannot do
  5. 6.26.5 Depth Anything V3 in 2026
  6. 6.26.6 Marigold — diffusion for depth
  7. 6.26.7 Intrinsics and the pinhole camera
  8. 6.26.8 Evaluation
  9. 6.26.9 Depth metrics
  10. 6.26.10 Scale-and-shift alignment
  11. 6.26.11 Lift depth to a point cloud
  12. 6.26.12 Smoke test with a synthetic depth scene
  13. 6.26.13 Depth Anything V2 usage (reference)
27. 6.27 Multi-Object Tracking and Video Memory
  1. 6.27.1 Tracking-by-detection
  2. 6.27.2 Kalman filter in one paragraph
  3. 6.27.3 The Hungarian algorithm
  4. 6.27.4 ByteTrack's key idea
  5. 6.27.5 SAM 2 memory-based tracking
  6. 6.27.6 SAM 3.1 Object Multiplex
  7. 6.27.7 Three metrics to know
  8. 6.27.8 IoU-based cost matrix
  9. 6.27.9 Minimal SORT-style tracker
  10. 6.27.10 Synthetic trajectory test
  11. 6.27.11 ID-switch metric
28. 6.28 World Models and Video Diffusion
  1. 6.28.1 Three families of world-modelling
  2. 6.28.2 Video DiT architecture
  3. 6.28.3 Conditioning on actions: latent action models
  4. 6.28.4 Physical plausibility
  5. 6.28.5 Autonomous driving world models
  6. 6.28.6 Robotics stack: VLM + video model + inverse dynamics
  7. 6.28.7 Evaluation
  8. 6.28.8 Model landscape in 2026
  9. 6.28.9 3D patchify for video
  10. 6.28.10 3D rotary position encoding
  11. 6.28.11 Divided attention block
  12. 6.28.12 Compose a tiny video DiT
  13. 6.28.13 Check shapes
7NLP, foundations to advancednot yet
1. 7.1 Text Processing: Tokenization, Stemming, Lemmatization
  1. 7.1.1 a regex word tokenizer
  2. 7.1.2 a Porter stemmer (step 1a only)
  3. 7.1.3 a lookup-based lemmatizer
  4. 7.1.4 pipe them together
  5. 7.1.5 NLTK
  6. 7.1.6 spaCy
  7. 7.1.7 When to pick which
  8. 7.1.8 The two failure modes nobody warns you about
2. 7.2 Bag of Words, TF-IDF and Text Representation
  1. 7.2.1 build the vocabulary
  2. 7.2.2 bag of words
  3. 7.2.3 term frequency and document frequency
  4. 7.2.4 TF-IDF
  5. 7.2.5 L2-normalize rows
  6. 7.2.6 When TF-IDF still wins (as of 2026)
  7. 7.2.7 When TF-IDF fails
  8. 7.2.8 Hybrid: TF-IDF weighted embeddings
3. 7.3 Word Embeddings: Word2Vec from Scratch
  1. 7.3.1 training pairs from a corpus
  2. 7.3.2 embedding tables
  3. 7.3.3 negative sampling objective
  4. 7.3.4 train on a toy corpus
  5. 7.3.5 the analogy trick
  6. 7.3.6 When Word2Vec still wins in 2026
  7. 7.3.7 Where Word2Vec fails
4. 7.4 GloVe, FastText and Subword Embeddings
  1. 7.4.1 GloVe: factorize the co-occurrence matrix
  2. 7.4.2 FastText: subword-aware embeddings
  3. 7.4.3 BPE: learned subword vocabulary
  4. 7.4.4 When to pick which
5. 7.5 Sentiment Analysis
  1. 7.5.1 a real mini-dataset
  2. 7.5.2 multinomial Naive Bayes from scratch
  3. 7.5.3 logistic regression from scratch
  4. 7.5.4 handling negation (the failure mode)
  5. 7.5.5 evaluation metrics that matter
  6. 7.5.6 When to reach for a transformer
  7. 7.5.7 The reproducibility trap (again)
6. 7.6 Named Entity Recognition (NER)
  1. 7.6.1 BIO tagging helpers
  2. 7.6.2 hand-crafted features
  3. 7.6.3 a simple rule-based + dictionary baseline
  4. 7.6.4 the CRF step (sketch, not full impl)
  5. 7.6.5 what a BiLSTM-CRF adds
  6. 7.6.6 LLM-based NER (the 2026 option)
  7. 7.6.7 Where classical NER still wins
  8. 7.6.8 Where it falls apart
7. 7.7 POS Tagging and Syntactic Parsing
  1. 7.7.1 most-frequent-tag baseline
  2. 7.7.2 bigram HMM tagger
  3. 7.7.3 why modern taggers beat this
  4. 7.7.4 dependency parsing sketch
  5. 7.7.5 Where this still matters in 2026
8. 7.8 Text Classification — CNNs and RNNs for Text
  1. 7.8.1 TextCNN in PyTorch
  2. 7.8.2 LSTM classifier
  3. 7.8.3 the vanishing gradient demo (intuition)
  4. 7.8.4 why this still was not enough
9. 7.9 Sequence-to-Sequence Models
  1. 7.9.1 an encoder
  2. 7.9.2 a decoder
  3. 7.9.3 training loop with teacher forcing
  4. 7.9.4 inference loop (greedy)
  5. 7.9.5 the bottleneck, demonstrated
  6. 7.9.6 When to still reach for RNN-based seq2seq
  7. 7.9.7 Exposure bias and its mitigations
10. 7.10 Attention Mechanism — The Breakthrough
  1. 7.10.1 additive (Bahdanau) attention
  2. 7.10.2 Luong dot and general
  3. 7.10.3 a worked numerical example
  4. 7.10.4 why this is the bridge to transformers
  5. 7.10.5 When classical attention still matters
  6. 7.10.6 The attention-weight-as-explanation trap
11. 7.11 Machine Translation
  1. 7.11.1 a pretrained MT call
  2. 7.11.2 BLEU and chrF
  3. 7.11.3 The three-tier evaluation hierarchy (2026)
  4. 7.11.4 what breaks in production
  5. 7.11.5 fine-tuning for a domain
12. 7.12 Text Summarization
  1. 7.12.1 TextRank (extractive)
  2. 7.12.2 abstractive with BART
  3. 7.12.3 ROUGE evaluation
  4. 7.12.4 Beyond ROUGE (2026 summarization eval)
  5. 7.12.5 the factuality problem
13. 7.13 Question Answering Systems
  1. 7.13.1 extractive QA with a pretrained model
  2. 7.13.2 a retrieval-augmented pipeline (sketch)
  3. 7.13.3 generative with RAG
  4. 7.13.4 evaluation that reflects the real world
  5. 7.13.5 RAGAS: the 2026 production eval framework
14. 7.14 Information Retrieval and Search
  1. 7.14.1 BM25 from scratch
  2. 7.14.2 dense retrieval with a bi-encoder
  3. 7.14.3 Reciprocal Rank Fusion
  4. 7.14.4 hybrid search + rerank
  5. 7.14.5 evaluation
  6. 7.14.6 The hard-won lessons from 2026 production RAG
15. 7.15 Topic Modeling: LDA, BERTopic
  1. 7.15.1 LDA via scikit-learn
  2. 7.15.2 BERTopic (production)
  3. 7.15.3 evaluation
16. 7.16 Text Generation
  1. 7.16.1 The prediction game
  2. 7.16.2 trigram counts
  3. 7.16.3 Laplace smoothing
  4. 7.16.4 Kneser-Ney (bigram, interpolated)
  5. 7.16.5 generating text with sampling
  6. 7.16.6 perplexity
17. 7.17 Chatbots: Rule-Based to Neural
  1. 7.17.1 The scripted half-century, 1950-2001
  2. 7.17.2 rule-based pattern matching
  3. 7.17.3 retrieval-based (FAQ)
  4. 7.17.4 neural generation (baseline)
  5. 7.17.5 LLM agent loop
  6. 7.17.6 hybrid routing
18. 7.18 Multilingual NLP
  1. 7.18.1 zero-shot cross-lingual classification
  2. 7.18.2 multilingual embedding space
  3. 7.18.3 few-shot fine-tuning strategy
  4. 7.18.4 The tokenization tax (what goes wrong for low-resource languages)
19. 7.19 Subword Tokenization: BPE, WordPiece, Unigram, SentencePiece
  1. 7.19.1 BPE from scratch
  2. 7.19.2 encode with the learned merges
  3. 7.19.3 SentencePiece in practice
  4. 7.19.4 tiktoken for OpenAI-compatible vocabs
20. 7.20 Structured Outputs and Constrained Decoding
  1. 7.20.1 The counterintuitive result
  2. 7.20.2 The pitfall that costs you
  3. 7.20.3 regex-constrained generation from scratch
  4. 7.20.4 Outlines for JSON Schema
  5. 7.20.5 Instructor for provider-agnostic Pydantic
  6. 7.20.6 native vendor APIs
21. 7.21 NLI and Textual Entailment
  1. 7.21.1 run a pretrained NLI model
  2. 7.21.2 zero-shot classification
  3. 7.21.3 faithfulness check for RAG
  4. 7.21.4 hand-rolled NLI classifier (conceptual)
22. 7.22 Embedding Models Deep Dive
  1. 7.22.1 The MTEB leaderboard tells a partial story
  2. 7.22.2 The three-tier pattern
  3. 7.22.3 baseline — dense embeddings with Sentence-BERT
  4. 7.22.4 Matryoshka truncation
  5. 7.22.5 BGE-M3 multi-functionality
  6. 7.22.6 MTEB eval on a custom task
  7. 7.22.7 hand-rolled cosine from scratch
23. 7.23 Chunking Strategies for RAG
  1. 7.23.1 The rule that beats every default
  2. 7.23.2 fixed and recursive chunking
  3. 7.23.3 semantic chunking
  4. 7.23.4 parent-document
  5. 7.23.5 contextual retrieval (Anthropic pattern)
  6. 7.23.6 evaluate
24. 7.24 Coreference Resolution
  1. 7.24.1 pretrained neural coreference (AllenNLP / spaCy-experimental)
  2. 7.24.2 rule-based pronoun resolver (teaching)
  3. 7.24.3 using LLMs for coreference
  4. 7.24.4 evaluation
25. 7.25 Entity Linking and Disambiguation
  1. 7.25.1 The two measurements
  2. 7.25.2 build an alias index from Wikipedia redirects
  3. 7.25.3 context-based disambiguation
  4. 7.25.4 embedding-based (BLINK-style)
  5. 7.25.5 generative entity linking (concept)
  6. 7.25.6 evaluate on AIDA-CoNLL
26. 7.26 Relation Extraction and Knowledge Graph Construction
  1. 7.26.1 pattern-based extraction
  2. 7.26.2 supervised seq2seq relation extraction
  3. 7.26.3 LLM-prompted extraction with anchoring
  4. 7.26.4 canonicalize onto a closed ontology
  5. 7.26.5 build a small graph and query
27. 7.27 LLM Evaluation: RAGAS, DeepEval, G-Eval
  1. 7.27.1 faithfulness with NLI (RAGAS-style)
  2. 7.27.2 answer relevance
  3. 7.27.3 G-Eval custom metric
  4. 7.27.4 CI gate
  5. 7.27.5 toy eval from scratch
28. 7.28 Long-Context Evaluation: NIAH, RULER, LongBench, MRCR
  1. 7.28.1 What to actually report
  2. 7.28.2 a custom NIAH for your domain
  3. 7.28.3 a multi-needle variant
  4. 7.28.4 multi-hop variable tracing (RULER-style)
  5. 7.28.5 LongBench v2 on your stack
29. 7.29 Dialogue State Tracking
  1. 7.29.1 The classic failure modes
  2. 7.29.2 rule-based slot extractor
  3. 7.29.3 state update loop
  4. 7.29.4 LLM-driven DST with structured output
  5. 7.29.5 JGA evaluation
  6. 7.29.6 handling correction
8Speech and audionot yet
1. 8.1 Audio Fundamentals: Waveforms, Sampling, FFT
  1. 8.1.1 read a clip and plot the waveform
  2. 8.1.2 synthesize a sine wave from first principles
  3. 8.1.3 compute the DFT by hand
  4. 8.1.4 find the dominant frequency
  5. 8.1.5 demonstrate aliasing
2. 8.2 Spectrograms, Mel Scale and Audio Features
  1. 8.2.1 frame the waveform
  2. 8.2.2 Hann window
  3. 8.2.3 STFT magnitude
  4. 8.2.4 mel filterbank
  5. 8.2.5 log-mel
  6. 8.2.6 MFCCs
3. 8.3 Audio Classification
  1. 8.3.1 Class imbalance is the real challenge
  2. 8.3.2 Evaluation
  3. 8.3.3 featurize
  4. 8.3.4 fixed-length summary
  5. 8.3.5 k-NN
  6. 8.3.6 upgrade to CNN on log-mels
  7. 8.3.7 fine-tune a pretrained audio transformer (AST shown)
4. 8.4 Speech Recognition (ASR)
  1. 8.4.1 WER: the one number
  2. 8.4.2 greedy CTC decode
  3. 8.4.3 beam-search CTC
  4. 8.4.4 WER
  5. 8.4.5 inference against Whisper
  6. 8.4.6 streaming with Parakeet or wav2vec 2.0
5. 8.5 Whisper: Architecture and Fine-Tuning
  1. 8.5.1 Variants in 2026
  2. 8.5.2 Fine-tuning
  3. 8.5.3 run Whisper out of the box
  4. 8.5.4 chunked long-form
  5. 8.5.5 fine-tune with LoRA
  6. 8.5.6 inspect what each layer learns
6. 8.6 Speaker Recognition and Verification
  1. 8.6.1 Scoring
  2. 8.6.2 Numbers you should know (2026)
  3. 8.6.3 Diarization
  4. 8.6.4 toy embedding from MFCC statistics
  5. 8.6.5 cosine similarity + threshold
  6. 8.6.6 EER from similarity pairs
  7. 8.6.7 production with SpeechBrain
  8. 8.6.8 diarize with pyannote
7. 8.7 Text-to-Speech (TTS)
  1. 8.7.1 Vocoder evolution
  2. 8.7.2 Evaluation
  3. 8.7.3 phonemize input
  4. 8.7.4 run Kokoro (2026 CPU default)
  5. 8.7.5 run F5-TTS with voice cloning
  6. 8.7.6 HiFi-GAN vocoder from scratch
  7. 8.7.7 the full pipeline (pseudocode)
8. 8.8 Voice Cloning and Voice Conversion
  1. 8.8.1 The ethics bit, not a bolt-on
  2. 8.8.2 Numbers (2026)
  3. 8.8.3 decompose with recognition-synthesis (code-only demo in main.py)
  4. 8.8.4 zero-shot clone with F5-TTS
  5. 8.8.5 voice conversion with KNN-VC
  6. 8.8.6 embed a watermark
  7. 8.8.7 consent gate
9. 8.9 Music Generation
  1. 8.9.1 Token LM over neural-codec tokens
  2. 8.9.2 Diffusion over mels or latents
  3. 8.9.3 Hybrid (production) — Suno, Udio, Lyria
  4. 8.9.4 Evaluation
  5. 8.9.5 generate with MusicGen
  6. 8.9.6 melody conditioning
  7. 8.9.7 FAD evaluation
  8. 8.9.8 adding to the LLM-music workflow
10. 8.10 Audio-Language Models
  1. 8.10.1 The three-component template
  2. 8.10.2 The 2026 model map
  3. 8.10.3 Benchmark reality check (2026)
  4. 8.10.4 Where LALMs are useful in 2026
  5. 8.10.5 Where they are NOT (yet) useful
  6. 8.10.6 query Qwen2.5-Omni
  7. 8.10.7 the projector pattern
  8. 8.10.8 benchmarking MMAU / LongAudioBench
11. 8.11 Real-Time Audio Processing
  1. 8.11.1 Common gotchas
  2. 8.11.2 ring buffer
  3. 8.11.3 VAD gate
  4. 8.11.4 streaming ASR
  5. 8.11.5 interruption handler
12. 8.12 Build a Voice Assistant Pipeline
  1. 8.12.1 The seven components
  2. 8.12.2 The three failure modes you will hit
  3. 8.12.3 2026 production reference stacks
  4. 8.12.4 mic capture with chunking (pseudocode)
  5. 8.12.5 VAD-gated turn capture
  6. 8.12.6 streaming STT → LLM → TTS
  7. 8.12.7 tool calling inside the LLM loop
  8. 8.12.8 interruption handling
13. 8.13 Neural Audio Codecs — EnCodec, SNAC, Mimi, DAC
  1. 8.13.1 The core trick: Residual Vector Quantization (RVQ)
  2. 8.13.2 The four codecs that matter in 2026
  3. 8.13.3 Frame rates matter for language modeling
  4. 8.13.4 Semantic vs acoustic tokens
  5. 8.13.5 2026 reconstruction quality (bits per sec, lower bitrate is better)
  6. 8.13.6 encode with EnCodec
  7. 8.13.7 decode and measure reconstruction
  8. 8.13.8 the semantic-acoustic split (Mimi-style)
  9. 8.13.9 why AR LM over codec tokens works
14. 8.14 Voice Activity Detection and Turn-Taking
  1. 8.14.1 The three-tier VAD cascade
  2. 8.14.2 Key parameters and their defaults
  3. 8.14.3 The flush trick (Kyutai 2025)
  4. 8.14.4 2026 VAD comparison
  5. 8.14.5 the energy gate
  6. 8.14.6 Silero VAD in Python
  7. 8.14.7 turn-end state machine
  8. 8.14.8 the flush trick skeleton
15. 8.15 Streaming Speech-to-Speech — Moshi, Hibiki
  1. 8.15.1 The Moshi architecture
  2. 8.15.2 Why inner-monologue text helps
  3. 8.15.3 Hibiki: streaming speech-to-speech translation
  4. 8.15.4 The broader Kyutai stack (2026)
  5. 8.15.5 Sesame CSM — the cousin
  6. 8.15.6 2026 performance numbers
  7. 8.15.7 the interface
  8. 8.15.8 the full-duplex loop
  9. 8.15.9 the training objective (conceptual)
  10. 8.15.10 where Moshi wins and where it doesn't
16. 8.16 Voice Anti-Spoofing and Audio Watermarking
  1. 8.16.1 ASVspoof 5 — the 2024-2025 benchmark
  2. 8.16.2 AASIST and RawNet2 — detection model families
  3. 8.16.3 AudioSeal — the 2024 watermark default
  4. 8.16.4 WavMark
  5. 8.16.5 WaveVerify (July 2025)
  6. 8.16.6 The gap adversaries exploit
  7. 8.16.7 C2PA / Content Authenticity Initiative
  8. 8.16.8 a simple spectral-feature detector (toy)
  9. 8.16.9 AudioSeal embed + detect
  10. 8.16.10 evaluation — EER
  11. 8.16.11 the production integration
17. 8.17 Audio Evaluation — WER, MOS, MMAU, Leaderboards
  1. 8.17.1 ASR metrics
  2. 8.17.2 TTS metrics
  3. 8.17.3 Voice-cloning-specific
  4. 8.17.4 Speaker verification
  5. 8.17.5 Diarization
  6. 8.17.6 Audio classification
  7. 8.17.7 Music generation
  8. 8.17.8 Audio-language benchmarks
  9. 8.17.9 Streaming speech-to-speech
  10. 8.17.10 The 2026 leaderboards
  11. 8.17.11 WER with normalization
  12. 8.17.12 TTS round-trip WER
  13. 8.17.13 SECS for voice cloning
  14. 8.17.14 FAD for music generation
  15. 8.17.15 EER for speaker verification (same code as Lesson 6)
9Transformers deep divenot yet
1. 9.1 Why Transformers: The Problems with RNNs
  1. 9.1.1 measure serial depth
  2. 9.1.2 count theoretical operations
  3. 9.1.3 empirical scaling on long sequences
2. 9.2 Self-Attention from Scratch
  1. 9.2.1 The Database Lookup Analogy
  2. 9.2.2 Q, K, V Computation
  3. 9.2.3 The Attention Matrix
  4. 9.2.4 Why Scale?
  5. 9.2.5 Softmax Turns Scores into Weights
  6. 9.2.6 Weighted Sum of Values
  7. 9.2.7 Full Pipeline
  8. 9.2.8 Softmax from scratch
  9. 9.2.9 Scaled dot-product attention
  10. 9.2.10 Self-attention class with learned projections
  11. 9.2.11 Run it on a sentence
  12. 9.2.12 Visualize attention with ASCII heatmap
3. 9.3 Multi-Head Attention
  1. 9.3.1 split heads from the single-head attention we already have
  2. 9.3.2 run scaled-dot-product attention per head
  3. 9.3.3 Grouped-Query Attention variant
  4. 9.3.4 probe what each head learned
4. 9.4 Positional Encoding: Sinusoidal, RoPE, ALiBi
  1. 9.4.1 Absolute sinusoidal
  2. 9.4.2 RoPE
  3. 9.4.3 ALiBi
  4. 9.4.4 What to pick in 2026
  5. 9.4.5 sinusoidal encoding
  6. 9.4.6 RoPE applied to Q, K
  7. 9.4.7 ALiBi slopes and bias
  8. 9.4.8 verify relative-distance property of RoPE
5. 9.5 The Full Transformer: Encoder + Decoder
  1. 9.5.1 The six pieces
  2. 9.5.2 Encoder block (used by BERT, T5 encoder)
  3. 9.5.3 Decoder block (used by GPT, T5 decoder)
  4. 9.5.4 Pre-norm vs post-norm
  5. 9.5.5 The 2026 modernized block
  6. 9.5.6 Parameter count
  7. 9.5.7 the building blocks
  8. 9.5.8 wire a 2-layer encoder and a 2-layer decoder
  9. 9.5.9 run forward on a toy example
  10. 9.5.10 swap in RMSNorm + SwiGLU
6. 9.6 BERT — Masked Language Modeling
  1. 9.6.1 The training signal
  2. 9.6.2 The BERT mask rules
  3. 9.6.3 Next Sentence Prediction (NSP) — and why it was dropped
  4. 9.6.4 What changed in 2026: ModernBERT
  5. 9.6.5 Use cases that still pick an encoder in 2026
  6. 9.6.6 masking logic
  7. 9.6.7 run MLM prediction on a tiny corpus
  8. 9.6.8 compare mask types
  9. 9.6.9 fine-tune head
7. 9.7 GPT — Causal Language Modeling
  1. 9.7.1 The mask
  2. 9.7.2 Where the triangle comes from
  3. 9.7.3 Parallel training, serial inference
  4. 9.7.4 The loss — shift-by-one
  5. 9.7.5 Decoding strategies
  6. 9.7.6 What made the "GPT recipe" work
  7. 9.7.7 the causal mask
  8. 9.7.8 a 2-layer GPT-ish model
  9. 9.7.9 next-token prediction, end-to-end
  10. 9.7.10 sampling
8. 9.8 T5, BART — Encoder-Decoder Models
  1. 9.8.1 The forward loop
  2. 9.8.2 T5 pretraining — span corruption
  3. 9.8.3 BART pretraining — multi-noise denoising
  4. 9.8.4 Inference
  5. 9.8.5 When to pick each variant in 2026
  6. 9.8.6 span corruption
  7. 9.8.7 verify round-trip
  8. 9.8.8 BART noising
9. 9.9 Vision Transformers (ViT)
  1. 9.9.1 Step 1 — patchify
  2. 9.9.2 Step 2 — linear embedding
  3. 9.9.3 Step 3 — prepend [CLS] token, add positional embeddings
  4. 9.9.4 Step 4 — standard transformer encoder
  5. 9.9.5 Step 5 — head
  6. 9.9.6 Variants that mattered
  7. 9.9.7 Why it took a while
  8. 9.9.8 fake image
  9. 9.9.9 patchify
  10. 9.9.10 linear embed
  11. 9.9.11 count parameters for a realistic ViT
10. 9.10 Audio Transformers — Whisper Architecture
  1. 9.10.1 Step 1 — resample + window
  2. 9.10.2 Step 2 — convolutional stem
  3. 9.10.3 Step 3 — encoder
  4. 9.10.4 Step 4 — decoder
  5. 9.10.5 Step 5 — task tokens
  6. 9.10.6 Step 6 — output
  7. 9.10.7 Whisper sizes
  8. 9.10.8 What Whisper does not do
  9. 9.10.9 2026 landscape
  10. 9.10.10 synthesize audio
  11. 9.10.11 log-mel spectrogram (simplified)
  12. 9.10.12 pad to 30 s
  13. 9.10.13 build the prompt tokens
11. 9.11 Mixture of Experts (MoE)
  1. 9.11.1 The FFN swap
  2. 9.11.2 The load-balancing problem
  3. 9.11.3 Shared experts
  4. 9.11.4 Fine-grained experts
  5. 9.11.5 The cost profile
  6. 9.11.6 The catch: memory
  7. 9.11.7 the router
  8. 9.11.8 run 100 tokens through the router
  9. 9.11.9 param count comparison
12. 9.12 KV Cache, Flash Attention and Inference Optimization
  1. 9.12.1 KV cache math
  2. 9.12.2 Flash Attention — the tiling trick
  3. 9.12.3 Speculative decoding — the other latency win
  4. 9.12.4 Continuous batching
  5. 9.12.5 PagedAttention — KV cache as virtual memory
  6. 9.12.6 KV cache
  7. 9.12.7 tiled softmax
  8. 9.12.8 compare naive vs cached decoding on 100-token generation
13. 9.13 Scaling Laws
  1. 9.13.1 The Hoffmann law
  2. 9.13.2 Why over-training anyway
  3. 9.13.3 Emergence vs smoothness
  4. 9.13.4 The 2026 picture
  5. 9.13.5 Chinchilla loss
  6. 9.13.6 compute-optimal frontier
  7. 9.13.7 over-training cost
  8. 9.13.8 compare to real models
14. 9.14 Build a Transformer from Scratch
  1. 9.14.1 What we ship
  2. 9.14.2 What we don't ship
  3. 9.14.3 Target metrics
  4. 9.14.4 data
  5. 9.14.5 model
  6. 9.14.6 training loop
  7. 9.14.7 sample
  8. 9.14.8 read the output
15. 9.15 Attention Variants — Sliding Window, Sparse, Differential
  1. 9.15.1 Sliding Window Attention (SWA)
  2. 9.15.2 Sparse / Block Attention
  3. 9.15.3 Differential Attention (DIFF Transformer, 2024)
  4. 9.15.4 Variant Comparison
  5. 9.15.5 full causal mask (baseline)
  6. 9.15.6 sliding window causal mask
  7. 9.15.7 local + strided sparse mask
  8. 9.15.8 differential attention
  9. 9.15.9 KV cache sizes
16. 9.16 Speculative Decoding — Draft, Verify, Repeat
  1. 9.16.1 The core algorithm
  2. 9.16.2 What determines speedup
  3. 9.16.3 Medusa — drafts without a draft model
  4. 9.16.4 EAGLE — better draft by reusing hidden states
  5. 9.16.5 The KV cache dance
  6. 9.16.6 the rejection step
  7. 9.16.7 residual distribution
  8. 9.16.8 one speculative step
  9. 9.16.9 measure acceptance rate
  10. 9.16.10 verify distribution equivalence
10Generative AInot yet
1. 10.1 Generative Models: Taxonomy and History
  1. 10.1.1 A brief history
  2. 10.1.2 The five-question triage
  3. 10.1.3 Production note: five families, five inference shapes
2. 10.2 Autoencoders and VAE
  1. 10.2.1 encoder forward
  2. 10.2.2 reparameterize and decode
  3. 10.2.3 the ELBO
  4. 10.2.4 generate
3. 10.3 GANs: Generator vs Discriminator
  1. 10.3.1 non-saturating loss
  2. 10.3.2 one discriminator step per generator step
  3. 10.3.3 watch for mode collapse
4. 10.4 Conditional GANs and Pix2Pix
  1. 10.4.1 append condition to both G and D inputs
  2. 10.4.2 train conditional
  3. 10.4.3 verify per-class output
5. 10.5 StyleGAN
  1. 10.5.1 mapping network
  2. 10.5.2 adaptive instance normalization
  3. 10.5.3 per-layer noise
6. 10.6 Diffusion Models — DDPM from Scratch
  1. 10.6.1 the forward schedule (closed form)
  2. 10.6.2 sample x_t in one shot
  3. 10.6.3 one training step
  4. 10.6.4 reverse sampling
7. 10.7 Latent Diffusion and Stable Diffusion
  1. 10.7.1 encoder/decoder
  2. 10.7.2 diffusion in z-space
  3. 10.7.3 classifier-free guidance
  4. 10.7.4 text conditioning (concept, not code)
8. 10.8 ControlNet, LoRA and Conditioning
  1. 10.8.1 ControlNet (Zhang et al., 2023)
  2. 10.8.2 LoRA (Hu et al., 2021)
  3. 10.8.3 IP-Adapter (Ye et al., 2023)
  4. 10.8.4 LoRA math
  5. 10.8.5 zero-init side network
9. 10.9 Inpainting, Outpainting and Editing
  1. 10.9.1 The naive approach (and why it's wrong)
  2. 10.9.2 The proper inpainting model
  3. 10.9.3 SDEdit (Meng et al., 2022) — free editing
  4. 10.9.4 InstructPix2Pix (Brooks et al., 2023)
  5. 10.9.5 RePaint (Lugmayr et al., 2022)
  6. 10.9.6 5-D DDPM data
  7. 10.9.7 train denoiser over all 5 dims
  8. 10.9.8 at inference, mask-aware reverse
  9. 10.9.9 outpainting
10. 10.10 Video Generation
  1. 10.10.1 Patchify
  2. 10.10.2 Spatiotemporal DiT
  3. 10.10.3 Text conditioning
  4. 10.10.4 Training
  5. 10.10.5 patchify a synthetic 1-D "video"
  6. 10.10.6 position embedding per frame
  7. 10.10.7 denoiser sees the whole sequence
  8. 10.10.8 temporal coherence test
11. 10.11 Audio Generation
  1. 10.11.1 Neural audio codecs
  2. 10.11.2 Two generative paradigms on top
  3. 10.11.3 synthetic audio tokens
  4. 10.11.4 train a tiny token predictor
  5. 10.11.5 sample conditionally
12. 10.12 3D Generation
  1. 10.12.1 Representation: 3D Gaussian Splatting (Kerbl et al., 2023)
  2. 10.12.2 Multi-view diffusion
  3. 10.12.3 Text-to-3D pipelines
  4. 10.12.4 NeRF (for context)
  5. 10.12.5 2D Gaussian splat
  6. 10.12.6 render by summing splats
  7. 10.12.7 fit by gradient descent
13. 10.13 Flow Matching and Rectified Flows
  1. 10.13.1 Straight-line flow
  2. 10.13.2 Sampling
  3. 10.13.3 Rectified flow (Liu 2022)
  4. 10.13.4 Why this won for images in 2024
  5. 10.13.5 training loss
  6. 10.13.6 multi-step inference
  7. 10.13.7 compare step counts
14. 10.14 Evaluation: FID, CLIP Score
  1. 10.14.1 FID — sample quality
  2. 10.14.2 CLIP score — prompt adherence
  3. 10.14.3 Human preference — the ground truth
  4. 10.14.4 FID in four lines
  5. 10.14.5 CLIP-style cosine-similarity
  6. 10.14.6 Elo aggregation
15. 10.15 Visual Autoregressive Modeling (VAR): Next-Scale Prediction
  1. 10.15.1 VQ-VAE Multi-Scale Tokenizer
  2. 10.15.2 Next-Scale Prediction
  3. 10.15.3 Generation
  4. 10.15.4 Why Next-Scale Wins Over Next-Token
  5. 10.15.5 Scaling Law
  6. 10.15.6 Relationship to Diffusion
11Reinforcement learningnot yet
1. 11.1 MDPs, States, Actions and Rewards
  1. 11.1.1 a tiny deterministic MDP
  2. 11.1.2 roll out a policy
  3. 11.1.3 compute V^π exactly via the Bellman equation
  4. 11.1.4 γ is a hyperparameter with physical meaning
2. 11.2 Dynamic Programming
  1. 11.2.1 build the GridWorld MDP model
  2. 11.2.2 policy evaluation
  3. 11.2.3 policy improvement
  4. 11.2.4 stitch them together
  5. 11.2.5 value iteration (the one-loop version)
3. 11.3 Monte Carlo Methods
  1. 11.3.1 rollout → list of (s, a, r)
  2. 11.3.2 compute returns (reverse sweep)
  3. 11.3.3 first-visit MC evaluation
  4. 11.3.4 ε-greedy MC control (on-policy)
  5. 11.3.5 compare to DP gold standard
4. 11.4 Q-Learning, SARSA
  1. 11.4.1 SARSA on ε-greedy policy
  2. 11.4.2 Q-learning
  3. 11.4.3 learning curves
  4. 11.4.4 compare to DP truth
5. 11.5 Deep Q-Networks (DQN)
  1. 11.5.1 replay buffer
  2. 11.5.2 a tiny Q-network (manual MLP)
  3. 11.5.3 the DQN update
  4. 11.5.4 the outer loop
6. 11.6 Policy Gradients — REINFORCE
  1. 11.6.1 softmax policy network
  2. 11.6.2 sampling and log-probability
  3. 11.6.3 rollout with log-probs captured
  4. 11.6.4 REINFORCE update
  5. 11.6.5 baselines
7. 11.7 Actor-Critic — A2C, A3C
  1. 11.7.1 a critic
  2. 11.7.2 n-step advantage
  3. 11.7.3 combined update
  4. 11.7.4 parallelization (A3C vs A2C)
8. 11.8 PPO
  1. 11.8.1 capture log π_old(a | s) at rollout time
  2. 11.8.2 compute GAE advantages (Lesson 07)
  3. 11.8.3 clipped surrogate update
  4. 11.8.4 value and entropy
  5. 11.8.5 diagnostics
9. 11.9 Reward Modeling and RLHF
  1. 11.9.1 synthetic preference data
  2. 11.9.2 Bradley-Terry reward model
  3. 11.9.3 PPO-like policy on top of RM
  4. 11.9.4 monitor the KL
  5. 11.9.5 the production recipe with TRL
10. 11.10 Multi-Agent RL
  1. 11.10.1 the multi-agent env
  2. 11.10.2 independent Q-learning
  3. 11.10.3 centralized Q with decomposed-value update
  4. 11.10.4 simple self-play (adversarial 2-agent)
11. 11.11 Sim-to-Real Transfer
  1. 11.11.1 parameterized sim
  2. 11.11.2 train with DR
  3. 11.11.3 evaluate zero-shot on "real" slips
  4. 11.11.4 compare to narrow training
12. 11.12 RL for Games
  1. 11.12.1 a tiny verifier environment
  2. 11.12.2 policy: softmax over K answer tokens per prompt
  3. 11.12.3 group sampling and group-relative advantage
  4. 11.12.4 compare to REINFORCE baseline (value-free)
  5. 11.12.5 observe entropy and KL
12LLMs from scratchnot yet
1. 12.1 Tokenizers: BPE, WordPiece, SentencePiece
  1. 12.1.1 Three Approaches That Failed (and One That Won)
  2. 12.1.2 BPE: Byte Pair Encoding
  3. 12.1.3 Byte-Level BPE (GPT-2, GPT-3, GPT-4)
  4. 12.1.4 WordPiece (BERT)
  5. 12.1.5 SentencePiece (Llama, T5)
  6. 12.1.6 Vocabulary Size Tradeoffs
  7. 12.1.7 The Multilingual Tax
  8. 12.1.8 Character-Level Tokenizer
  9. 12.1.9 BPE Tokenizer from Scratch
  10. 12.1.10 Encode and Decode Roundtrip
  11. 12.1.11 Compare with tiktoken
  12. 12.1.12 Vocabulary Analysis
  13. 12.1.13 tiktoken (OpenAI)
  14. 12.1.14 Hugging Face tokenizers
  15. 12.1.15 Loading Llama's Tokenizer
2. 12.2 Building a Tokenizer from Scratch
  1. 12.2.1 The Full Pipeline
  2. 12.2.2 Byte-Level BPE
  3. 12.2.3 Pre-Tokenization
  4. 12.2.4 Special Tokens
  5. 12.2.5 Chat Templates
  6. 12.2.6 Speed
  7. 12.2.7 Byte-Level Encoding
  8. 12.2.8 Pre-Tokenizer with Regex
  9. 12.2.9 BPE on Byte Sequences
  10. 12.2.10 Special Token Handling
  11. 12.2.11 Full Tokenizer Class
  12. 12.2.12 Multilingual Test
  13. 12.2.13 Comparing Real Tokenizers
3. 12.3 Data Pipelines for Pre-Training
  1. 12.3.1 Where the Data Comes From
  2. 12.3.2 Data Cleaning
  3. 12.3.3 Deduplication with MinHash
  4. 12.3.4 Sequence Packing
  5. 12.3.5 The Chinchilla Scaling Law
  6. 12.3.6 Text Cleaning
  7. 12.3.7 MinHash Deduplication
  8. 12.3.8 Tokenize and Pack Sequences
  9. 12.3.9 DataLoader for Training
  10. 12.3.10 Dataset Statistics
  11. 12.3.11 Compare With HuggingFace Datasets
4. 12.4 Pre-Training a Mini GPT (124M)
  1. 12.4.1 The GPT Architecture
  2. 12.4.2 The Transformer Block
  3. 12.4.3 Attention: The Core Mechanism
  4. 12.4.4 KV Cache: Why Inference Is Fast
  5. 12.4.5 Prefill vs Decode: Two Phases of Inference
  6. 12.4.6 The Training Loop
  7. 12.4.7 GPT-2 Small: The Numbers
  8. 12.4.8 Embedding Layer
  9. 12.4.9 Self-Attention with Causal Mask
  10. 12.4.10 Multi-Head Attention
  11. 12.4.11 Transformer Block
  12. 12.4.12 Full GPT Model
  13. 12.4.13 Training Loop
  14. 12.4.14 Text Generation
  15. 12.4.15 Full Training and Generation Demo
5. 12.5 Distributed Training, FSDP, DeepSpeed
  1. 12.5.1 Why Distribution is Required
  2. 12.5.2 Data Parallelism
  3. 12.5.3 Tensor Parallelism
  4. 12.5.4 Pipeline Parallelism
  5. 12.5.5 FSDP: Fully Sharded Data Parallel
  6. 12.5.6 DeepSpeed ZeRO
  7. 12.5.7 Mixed Precision Training
  8. 12.5.8 Megatron-LM and 3D Parallelism
  9. 12.5.9 Simulate Data Parallelism
  10. 12.5.10 Simulate Tensor Parallelism
  11. 12.5.11 Simulate Pipeline Parallelism
  12. 12.5.12 Memory Calculator
  13. 12.5.13 Mixed Precision Simulation
  14. 12.5.14 Run All Simulations
6. 12.6 Instruction Tuning — SFT
  1. 12.6.1 What SFT Actually Does
  2. 12.6.2 Data Formats
  3. 12.6.3 Why It Works
  4. 12.6.4 The Masked Loss
  5. 12.6.5 Training Hyperparameters
  6. 12.6.6 Catastrophic Forgetting
  7. 12.6.7 Real Numbers
  8. 12.6.8 Instruction Dataset
  9. 12.6.9 Tokenize with Chat Template
  10. 12.6.10 Masked Cross-Entropy Loss
  11. 12.6.11 SFT Training Loop
  12. 12.6.12 Compare Base vs SFT Model
  13. 12.6.13 Measure Catastrophic Forgetting
  14. 12.6.14 Full SFT Pipeline Demo
7. 12.7 RLHF — Reward Model + PPO
  1. 12.7.1 The Three Stages
  2. 12.7.2 The Reward Model
  3. 12.7.3 PPO: Proximal Policy Optimization
  4. 12.7.4 The PPO Objective in Detail
  5. 12.7.5 Reward Hacking
  6. 12.7.6 Real RLHF Pipelines
  7. 12.7.7 Synthetic Preference Data
  8. 12.7.8 Reward Model Architecture
  9. 12.7.9 Bradley-Terry Loss
  10. 12.7.10 Simplified PPO Loop
  11. 12.7.11 Reward Score Comparison
  12. 12.7.12 Full RLHF Pipeline Demo
8. 12.8 DPO — Direct Preference Optimization
  1. 12.8.1 The Key Insight
  2. 12.8.2 The DPO Loss
  3. 12.8.3 Why DPO is Simpler
  4. 12.8.4 When DPO Beats RLHF
  5. 12.8.5 When RLHF Beats DPO
  6. 12.8.6 Beyond DPO: KTO, ORPO, SimPO
  7. 12.8.7 Real DPO Deployments
  8. 12.8.8 Preference Dataset
  9. 12.8.9 Sequence Log-Probability
  10. 12.8.10 DPO Training Loop
  11. 12.8.11 Compare DPO vs RLHF
  12. 12.8.12 Beta Sensitivity Analysis
  13. 12.8.13 Full DPO Pipeline Demo
9. 12.9 Constitutional AI and Self-Improvement
  1. 12.9.1 The Constitutional AI Loop
  2. 12.9.2 What the Constitution Actually Does
  3. 12.9.3 GRPO: Group-Relative Policy Optimization
  4. 12.9.4 Why GRPO Matters for Reasoning
  5. 12.9.5 Process Reward Models vs Outcome Reward Models
  6. 12.9.6 Self-Improvement: The Feedback Multiplier
  7. 12.9.7 When To Use What
  8. 12.9.8 The Constitution
  9. 12.9.9 Self-Critique and Revise
  10. 12.9.10 Rule-Based Rewards
  11. 12.9.11 Group-Relative Advantage
  12. 12.9.12 GRPO Update
  13. 12.9.13 Self-Improvement Round
10. 12.10 Evaluation — Benchmarks, Evals
  1. 12.10.1 The Eval Landscape
  2. 12.10.2 Why Benchmarks Break
  3. 12.10.3 Perplexity: A Quick Health Check
  4. 12.10.4 LLM-as-Judge
  5. 12.10.5 ELO Ratings from Pairwise Comparisons
  6. 12.10.6 Eval Frameworks
  7. 12.10.7 Building Custom Evals
  8. 12.10.8 A Minimal Eval Framework
  9. 12.10.9 Scoring Functions
  10. 12.10.10 ELO Rating System
  11. 12.10.11 Perplexity Calculation
  12. 12.10.12 Aggregate Results
  13. 12.10.13 Run the Full Pipeline
  14. 12.10.14 ELO Tournament
  15. 12.10.15 Perplexity Comparison
  16. 12.10.16 lm-evaluation-harness (EleutherAI)
  17. 12.10.17 promptfoo
  18. 12.10.18 RAGAS for RAG evaluation
11. 12.11 Quantization: INT8, GPTQ, AWQ, GGUF
  1. 12.11.1 Number Formats: What Each Bit Does
  2. 12.11.2 How Quantization Works
  3. 12.11.3 Sensitivity Hierarchy
  4. 12.11.4 PTQ vs QAT
  5. 12.11.5 GPTQ, AWQ, GGUF
  6. 12.11.6 Quality Measurement
  7. 12.11.7 Real Numbers
  8. 12.11.8 Number Format Representations
  9. 12.11.9 Symmetric Quantization (Per-Tensor and Per-Channel)
  10. 12.11.10 Bit-Width Sweep
  11. 12.11.11 Sensitivity Experiment
  12. 12.11.12 Simulated GPTQ
  13. 12.11.13 AWQ Simulation
  14. 12.11.14 Full Pipeline
  15. 12.11.15 Quantizing with GPTQModel
  16. 12.11.16 Quantizing to AWQ with LLM Compressor
  17. 12.11.17 Converting to GGUF
  18. 12.11.18 Serving quantized models
12. 12.12 Inference Optimization
  1. 12.12.1 Prefill vs Decode
  2. 12.12.2 KV Cache
  3. 12.12.3 Continuous Batching
  4. 12.12.4 PagedAttention
  5. 12.12.5 Speculative Decoding
  6. 12.12.6 Prefix Caching
  7. 12.12.7 Inference Engines
  8. 12.12.8 The Ops:Byte Framework
  9. 12.12.9 KV Cache from Scratch
  10. 12.12.10 Attention with KV Cache
  11. 12.12.11 Continuous Batching Simulator
  12. 12.12.12 Prefix Cache
  13. 12.12.13 Speculative Decoding Simulator
  14. 12.12.14 KV Cache Memory Profiler
13. 12.13 Building a Complete LLM Pipeline
  1. 12.13.1 The Twelve Stages
  2. 12.13.2 The Manifest
  3. 12.13.3 Artifact Typing
  4. 12.13.4 The Eval Gate
  5. 12.13.5 The Orchestrator
  6. 12.13.6 Experiment Tracking and Artifact Storage
  7. 12.13.7 Costing
  8. 12.13.8 Reproducibility vs Determinism
  9. 12.13.9 Rollback Plan
  10. 12.13.10 Production Recipes Observed in 2026
14. 12.14 Open Models: Architecture Walkthroughs
  1. 12.14.1 The Invariant Core
  2. 12.14.2 The Six Knobs That Actually Move
  3. 12.14.3 Knob 1: RMSNorm
  4. 12.14.4 Knob 2: RoPE
  5. 12.14.5 Knob 3: SwiGLU
  6. 12.14.6 Knob 4: Attention Head Sharing
  7. 12.14.7 Knob 5: Mixture of Experts
  8. 12.14.8 Knob 6: Pre-norm stays
  9. 12.14.9 Model-by-Model Diff
  10. 12.14.10 Reading a config.json
  11. 12.14.11 Activation memory budget
  12. 12.14.12 KV Cache budget
  13. 12.14.13 When Each Model Wins
15. 12.15 Speculative Decoding and EAGLE-3
  1. 12.15.1 The invariant: Leviathan rejection sampling
  2. 12.15.2 What determines speedup
  3. 12.15.3 The two-year progression
  4. 12.15.4 KV cache rollback
  5. 12.15.5 Draft architectures in 2026
  6. 12.15.6 the rejection rule
  7. 12.15.7 residual distribution
  8. 12.15.8 a full speculative step
  9. 12.15.9 KV rollback bookkeeping
  10. 12.15.10 the Leviathan check
  11. 12.15.11 speedup vs. α
16. 12.16 Differential Attention (V2)
  1. 12.16.1 The noise floor of softmax
  2. 12.16.2 The differential idea
  3. 12.16.3 Why this matches headed noise-canceling
  4. 12.16.4 V1 vs V2: the diff
  5. 12.16.5 When to reach for it
  6. 12.16.6 How it stacks with other 2026 knobs
  7. 12.16.7 standard softmax attention
  8. 12.16.8 split Q, K into two halves
  9. 12.16.9 two softmax branches + subtraction
  10. 12.16.10 noise cancellation measurement
  11. 12.16.11 V1 vs V2 parameter accounting
17. 12.17 Native Sparse Attention (DeepSeek NSA)
  1. 12.17.1 Three parallel branches
  2. 12.17.2 Why this is "natively trainable"
  3. 12.17.3 Hardware-aligned kernel
  4. 12.17.4 The compute budget
  5. 12.17.5 How does it compare
  6. 12.17.6 compress tokens into blocks
  7. 12.17.7 compressed-branch attention
  8. 12.17.8 top-k block selection
  9. 12.17.9 sliding-window attention
  10. 12.17.10 gate + combine
  11. 12.17.11 compute counting
18. 12.18 Multi-Token Prediction (MTP)
  1. 12.18.1 The sequential MTP recipe
  2. 12.18.2 Why sequential, not parallel
  3. 12.18.3 Parameter accounting
  4. 12.18.4 The speculative-decoding payoff
  5. 12.18.5 Relation to EAGLE
  6. 12.18.6 shared embedding table
  7. 12.18.7 the per-depth combination
  8. 12.18.8 the transformer block at depth k
  9. 12.18.9 the shared output head
  10. 12.18.10 per-depth loss
  11. 12.18.11 parameter accounting
19. 12.19 DualPipe Parallelism
  1. 12.19.1 Pipeline parallelism refresher
  2. 12.19.2 Idea 1: chunk decomposition
  3. 12.19.3 Idea 2: bidirectional scheduling
  4. 12.19.4 A hand-traced schedule
  5. 12.19.5 Bubble accounting
  6. 12.19.6 DualPipeV — the refinement
  7. 12.19.7 What it means for a 14.8T-token run
  8. 12.19.8 Where it sits in the stack
20. 12.20 DeepSeek-V3 Architecture Walkthrough
  1. 12.20.1 The invariant core, again
  2. 12.20.2 The twist: MLA instead of GQA
  3. 12.20.3 The routing: auxiliary-loss-free load balancing
  4. 12.20.4 The MTP: denser training + free draft
  5. 12.20.5 The training: DualPipe
  6. 12.20.6 The config, field by field
  7. 12.20.7 Parameter accounting
  8. 12.20.8 The 671B / 37B ratio
  9. 12.20.9 Where DeepSeek-V3 sits
  10. 12.20.10 The follow-on: R1, V4
21. 12.21 Jamba — Hybrid SSM-Transformer
  1. 12.21.1 An SSM in one page
  2. 12.21.2 The Jamba block
  3. 12.21.3 Why the 1:7 ratio
  4. 12.21.4 Positional encoding
  5. 12.21.5 The memory budget
  6. 12.21.6 Mamba-3: the pure-SSM baseline in 2026
  7. 12.21.7 When to reach for a hybrid
  8. 12.21.8 The competitive landscape
22. 12.22 Async and Hogwild! Inference
  1. 12.22.1 The setup
  2. 12.22.2 Why coordination emerges
  3. 12.22.3 The naming
  4. 12.22.4 RoPE makes this tractable
  5. 12.22.5 Wall-time math
  6. 12.22.6 Concrete example
  7. 12.22.7 When to reach for Hogwild!
  8. 12.22.8 When not to
  9. 12.22.9 The experimental status
  10. 12.22.10 the shared cache
  11. 12.22.11 the worker loop
  12. 12.22.12 the coordination heuristic
  13. 12.22.13 measured speedup
  14. 12.22.14 stress the coordination
23. 12.23 Speculative Decoding and EAGLE
  1. 12.23.1 The Two-Model Setup
  2. 12.23.2 The Exactness Rule
  3. 12.23.3 Expected Speedup
  4. 12.23.4 Training the Draft: Distillation
  5. 12.23.5 EAGLE: Tree Drafting + Feature Reuse
  6. 12.23.6 Tree Attention Verification
  7. 12.23.7 When It Wins, When It Doesn't
24. 12.24 Gradient Checkpointing and Activation Recomputation
  1. 12.24.1 What Backward Actually Needs
  2. 12.24.2 Naive Full Checkpointing
  3. 12.24.3 Selective Checkpointing (Korthikanti 2022)
  4. 12.24.4 Offload
  5. 12.24.5 Recompute Cost Model
  6. 12.24.6 Memory Savings Model
  7. 12.24.7 When Not to Checkpoint
  8. 12.24.8 Implementation Patterns
  9. 12.24.9 Interaction with TP / PP / FP8
  10. 12.24.10 A Toy Model With Segments
  11. 12.24.11 Naive Backward Needing All Activations
  12. 12.24.12 Checkpoint-Every-k Memory
  13. 12.24.13 Cost Model
  14. 12.24.14 Memory Estimator
  15. 12.24.15 Optimal Segment Size
  16. 12.24.16 Selective Checkpoint Decision
13LLM engineeringnot yet
1. 13.1 Prompt Engineering: Techniques and Patterns
  1. 13.1.1 Anatomy of a Prompt
  2. 13.1.2 System message
  3. 13.1.3 Temperature
  4. 13.1.4 Top-p
  5. 13.1.5 Few-shot prompting
  6. 13.1.6 Chain-of-thought
  7. 13.1.7 Role prompting
  8. 13.1.8 Prompt injection
  9. 13.1.9 Context window
  10. 13.1.10 Assistant prefill
  11. 13.1.11 Meta-prompting
2. 13.2 Few-Shot, CoT, Tree-of-Thought
  1. 13.2.1 Zero-Shot vs Few-Shot: When Examples Beat Instructions
  2. 13.2.2 Example Selection: Similar Beats Random
  3. 13.2.3 Chain-of-Thought: Giving Models Scratch Paper
  4. 13.2.4 Self-Consistency: Sample Many, Vote Once
  5. 13.2.5 Tree-of-Thought: Branching Exploration
  6. 13.2.6 ReAct: Thinking + Doing
  7. 13.2.7 Structured Prompting: XML Tags, Delimiters, Headers
  8. 13.2.8 Prompt Chaining: Sequential Decomposition
  9. 13.2.9 Performance Comparison
  10. 13.2.10 Few-Shot Example Store
  11. 13.2.11 Chain-of-Thought Prompt Builder
  12. 13.2.12 Self-Consistency Voting
  13. 13.2.13 Tree-of-Thought Solver
  14. 13.2.14 Full Pipeline
  15. 13.2.15 Template-Driven Few-Shot Prompts
  16. 13.2.16 Compiled Prompts
  17. 13.2.17 Comparison: From-Scratch vs Frameworks
3. 13.3 Structured Outputs
  1. 13.3.1 The Structured Output Spectrum
  2. 13.3.2 JSON Schema: The Contract Language
  3. 13.3.3 The Pydantic Pattern
  4. 13.3.4 Function Calling / Tool Use
  5. 13.3.5 Common Failure Modes
  6. 13.3.6 JSON Schema Validator
  7. 13.3.7 Pydantic-Style Model to Schema
  8. 13.3.8 Constrained Token Filter
  9. 13.3.9 Extraction Pipeline
  10. 13.3.10 Run the Full Pipeline
  11. 13.3.11 OpenAI Structured Outputs
  12. 13.3.12 Anthropic Tool Use
  13. 13.3.13 Instructor Library
4. 13.4 Embeddings and Vector Representations
  1. 13.4.1 What Is an Embedding?
  2. 13.4.2 The Word2Vec Breakthrough
  3. 13.4.3 From Words to Sentences
  4. 13.4.4 Modern Embedding Models
  5. 13.4.5 Similarity Metrics
  6. 13.4.6 Vector Databases and HNSW
  7. 13.4.7 Chunking Strategies
  8. 13.4.8 Bi-Encoders vs Cross-Encoders
  9. 13.4.9 Matryoshka Embeddings
  10. 13.4.10 Binary Quantization
  11. 13.4.11 Text Chunking
  12. 13.4.12 Building Embeddings from Scratch
  13. 13.4.13 Similarity Functions
  14. 13.4.14 Vector Index with Brute-Force Search
  15. 13.4.15 The Semantic Search Engine
  16. 13.4.16 Comparing Similarity Metrics
5. 13.5 Context Engineering
  1. 13.5.1 The Context Window is a Scarce Resource
  2. 13.5.2 Lost-in-the-Middle
  3. 13.5.3 Context Components
  4. 13.5.4 Context Compression Strategies
  5. 13.5.5 Memory Systems
  6. 13.5.6 Dynamic Context Assembly
  7. 13.5.7 Token Counter
  8. 13.5.8 Context Budget Manager
  9. 13.5.9 Lost-in-the-Middle Reordering
  10. 13.5.10 Conversation History Compressor
  11. 13.5.11 Dynamic Tool Selector
  12. 13.5.12 Full Context Assembly Pipeline
  13. 13.5.13 Harness-Managed Context
  14. 13.5.14 Dynamic Context Loading
  15. 13.5.15 Assistant Long-Term Memory
  16. 13.5.16 RAG as Context Engineering
6. 13.6 RAG: Retrieval-Augmented Generation
  1. 13.6.1 The RAG Pattern
  2. 13.6.2 Why RAG Beats Fine-Tuning
  3. 13.6.3 Embedding Models
  4. 13.6.4 Vector Similarity
  5. 13.6.5 Chunking Strategies
  6. 13.6.6 Vector Databases
  7. 13.6.7 The Full Pipeline
  8. 13.6.8 Real Numbers
  9. 13.6.9 Document Chunking
  10. 13.6.10 TF-IDF Embeddings
  11. 13.6.11 Cosine Similarity Search
  12. 13.6.12 Prompt Construction
  13. 13.6.13 The Complete RAG Pipeline
  14. 13.6.14 Generation (simulated)
7. 13.7 Advanced RAG: Chunking, Reranking
  1. 13.7.1 Hybrid Search: Semantic + Keyword
  2. 13.7.2 Reciprocal Rank Fusion (RRF)
  3. 13.7.3 Reranking
  4. 13.7.4 Query Transformation
  5. 13.7.5 Parent-Child Chunking
  6. 13.7.6 Metadata Filtering
  7. 13.7.7 Evaluation
  8. 13.7.8 BM25 Implementation
  9. 13.7.9 Reciprocal Rank Fusion
  10. 13.7.10 Hybrid Search Pipeline
  11. 13.7.11 Simple Reranker
  12. 13.7.12 HyDE (Hypothetical Document Embeddings)
  13. 13.7.13 Faithfulness Evaluation
8. 13.8 Fine-Tuning with LoRA and QLoRA
  1. 13.8.1 LoRA: Low-Rank Adaptation
  2. 13.8.2 The Scaling Factor: Alpha
  3. 13.8.3 Where to Apply LoRA
  4. 13.8.4 Rank Selection
  5. 13.8.5 QLoRA: 4-Bit Quantization + LoRA
  6. 13.8.6 The Quality Question
  7. 13.8.7 Real-World Costs
  8. 13.8.8 The 2026 PEFT stack
  9. 13.8.9 Merging Adapters
  10. 13.8.10 When NOT to Fine-Tune
  11. 13.8.11 The LoRA Layer
  12. 13.8.12 LoRA-Wrapped Linear Layer
  13. 13.8.13 Inject LoRA into a Model
  14. 13.8.14 Count Parameters
  15. 13.8.15 Merge Weights Back
  16. 13.8.16 Simulated QLoRA Quantization
  17. 13.8.17 Training Loop
  18. 13.8.18 Full Demo
9. 13.9 Function Calling and Tool Use
  1. 13.9.1 The Function Calling Loop
  2. 13.9.2 Tool Definitions: The JSON Schema Contract
  3. 13.9.3 Provider Comparison
  4. 13.9.4 Tool Choice: Auto, Required, Specific
  5. 13.9.5 Parallel Function Calling
  6. 13.9.6 Structured Outputs vs Function Calling
  7. 13.9.7 Security: The Non-Negotiable Rules
  8. 13.9.8 Error Handling
  9. 13.9.9 MCP: Model Context Protocol
  10. 13.9.10 Define the Tool Registry
  11. 13.9.11 Implement 5 Tools
  12. 13.9.12 Register All Tools
  13. 13.9.13 Build the Function Calling Loop
  14. 13.9.14 Argument Validation
  15. 13.9.15 Run the Demo
  16. 13.9.16 OpenAI Function Calling
  17. 13.9.17 Anthropic Tool Use
  18. 13.9.18 MCP Integration
10. 13.10 Evaluation and Testing
  1. 13.10.1 The Eval Taxonomy
  2. 13.10.2 LLM-as-Judge: The Workhorse
  3. 13.10.3 Rubric Design
  4. 13.10.4 The Eval Pipeline
  5. 13.10.5 Eval Datasets: The Foundation
  6. 13.10.6 Sample Size and Confidence
  7. 13.10.7 Regression Testing
  8. 13.10.8 Cost of Evals
  9. 13.10.9 Anti-Patterns
  10. 13.10.10 Real Tools
  11. 13.10.11 Define the Eval Data Structures
  12. 13.10.12 Build the LLM-as-Judge Scorer
  13. 13.10.13 Build Automated Metrics
  14. 13.10.14 Build the Confidence Interval Calculator
  15. 13.10.15 Build the Eval Runner and Comparison Report
  16. 13.10.16 Run the Demo
  17. 13.10.17 promptfoo Integration
  18. 13.10.18 DeepEval Integration
  19. 13.10.19 CI/CD Integration Pattern
11. 13.11 Caching, Rate Limiting and Cost
  1. 13.11.1 The Cost Anatomy of an LLM Call
  2. 13.11.2 Provider Caching: Built-in Discounts
  3. 13.11.3 Semantic Caching: Your Custom Layer
  4. 13.11.4 Exact Caching: Hash and Match
  5. 13.11.5 Rate Limiting: Protecting Your Budget
  6. 13.11.6 Model Routing: Right Model for the Right Job
  7. 13.11.7 Cost Tracking: Know Where the Money Goes
  8. 13.11.8 Batching: Bulk Discounts
  9. 13.11.9 Budget Alerts and Circuit Breakers
  10. 13.11.10 The Optimization Stack
  11. 13.11.11 Real Savings: Before and After
  12. 13.11.12 Cost Calculator
  13. 13.11.13 Exact Cache
  14. 13.11.14 Semantic Cache
  15. 13.11.15 Rate Limiter
  16. 13.11.16 Cost Tracker
  17. 13.11.17 Model Router
  18. 13.11.18 Run the Demo
  19. 13.11.19 Anthropic Prompt Caching
  20. 13.11.20 OpenAI Automatic Caching
  21. 13.11.21 OpenAI Batch API
  22. 13.11.22 Production Semantic Cache with Redis
12. 13.12 Guardrails and Safety
  1. 13.12.1 The Guardrail Sandwich
  2. 13.12.2 Attack Taxonomy
  3. 13.12.3 Input Guardrails
  4. 13.12.4 Output Guardrails
  5. 13.12.5 The Content Filtering Stack
  6. 13.12.6 Tools of the Trade
  7. 13.12.7 Defense-in-Depth
  8. 13.12.8 Real Attack Case Studies
  9. 13.12.9 The Honest Truth
  10. 13.12.10 The Guardrail Pipeline
  11. 13.12.11 Monitoring Dashboard
  12. 13.12.12 Run the Demo
  13. 13.12.13 OpenAI Moderation API
  14. 13.12.14 LlamaGuard
  15. 13.12.15 NeMo Guardrails
  16. 13.12.16 Guardrails AI
13. 13.13 Building a Production LLM App
  1. 13.13.1 Production Architecture
  2. 13.13.2 The Stack
  3. 13.13.3 Streaming: Why It Matters
  4. 13.13.4 Error Handling: The Three Layers
  5. 13.13.5 Observability: What to Measure
  6. 13.13.6 A/B Testing Prompts in Production
  7. 13.13.7 Real Architecture Examples
  8. 13.13.8 Scaling
  9. 13.13.9 Cost Projection
  10. 13.13.10 The Deployment Checklist
  11. 13.13.11 Core Infrastructure
  12. 13.13.12 Prompt Management
  13. 13.13.13 Semantic Cache
  14. 13.13.14 Guardrails
  15. 13.13.15 LLM Caller with Retry and Streaming
  16. 13.13.16 The Request Pipeline
  17. 13.13.17 Run the Full Demo
  18. 13.13.18 FastAPI Server (Production Deployment)
  19. 13.13.19 Real API Integration
  20. 13.13.20 Docker Deployment
14. 13.14 Model Context Protocol (MCP)
  1. 13.14.1 The three server primitives
  2. 13.14.2 Stateless requests replace the handshake
  3. 13.14.3 Discovery and version selection
  4. 13.14.4 Results are explicit
  5. 13.14.5 The wire format and transport
  6. 13.14.6 Client input without server-initiated requests
  7. 13.14.7 register a server surface
  8. 13.14.8 attach metadata to every request
  9. 13.14.9 optionally discover before listing
  10. 13.14.10 map the same request to HTTP
  11. 13.14.11 enforce safety outside protocol state
15. 13.15 Prompt Caching and Context Caching
  1. 13.15.1 The cache-friendly layout
  2. 13.15.2 The break-even calculation
  3. 13.15.3 Anthropic prompt caching with explicit markers
  4. 13.15.4 one-hour extended TTL
  5. 13.15.5 OpenAI automatic caching
  6. 13.15.6 Gemini explicit context caching
  7. 13.15.7 measuring hit rate in production
16. 13.16 Agent State Machines — Graphs, Nodes, Checkpoints
  1. 13.16.1 The four superpowers
  2. 13.16.2 Reducers are the point
  3. 13.16.3 The ReAct graph in four nodes
  4. 13.16.4 StateGraph vs Send (fanout)
  5. 13.16.5 Subgraphs
  6. 13.16.6 state and nodes
  7. 13.16.7 run with a thread
  8. 13.16.8 add a human-in-the-loop interrupt
  9. 13.16.9 time-travel for debugging
  10. 13.16.10 swap the checkpointer for production
17. 13.17 Agent Framework Tradeoffs
  1. 13.17.1 What "abstraction" actually means
  2. 13.17.2 The state question
  3. 13.17.3 The branching question
  4. 13.17.4 The observability question
  5. 13.17.5 Cost and latency
  6. 13.17.6 Interoperability
14Multimodal AInot yet
1. 14.1 Vision Transformers and the Patch-Token Primitive
  1. 14.1.1 Patches as tokens
  2. 14.1.2 Positional embeddings
  3. 14.1.3 CLS token, pooled output, and register tokens
  4. 14.1.4 Pretraining: supervised, contrastive, masked, self-distilled
  5. 14.1.5 Scaling laws
  6. 14.1.6 Parameter count for a ViT
  7. 14.1.7 2026 production config
2. 14.2 CLIP and Contrastive Vision-Language Pretraining
  1. 14.2.1 The dual encoder
  2. 14.2.2 InfoNCE loss
  3. 14.2.3 Temperature
  4. 14.2.4 Why sigmoid scales better (SigLIP)
  5. 14.2.5 Zero-shot classification
  6. 14.2.6 Linear probes and finetuning
  7. 14.2.7 SigLIP 2: NaFlex and dense features
  8. 14.2.8 ALIGN, BASIC, OpenCLIP, EVA-CLIP
  9. 14.2.9 The zero-shot ceiling
3. 14.3 BLIP-2 Q-Former as Modality Bridge
  1. 14.3.1 Learnable queries
  2. 14.3.2 Architecture
  3. 14.3.3 Two-stage training
  4. 14.3.4 Parameter economics
  5. 14.3.5 InstructBLIP and the instruction-aware Q-Former
  6. 14.3.6 MiniGPT-4 and the projector-only approach
  7. 14.3.7 Why LLaVA went simpler
  8. 14.3.8 Gated cross-attention: Flamingo, the ancestor
  9. 14.3.9 The 2026 descendants
4. 14.4 Flamingo and Gated Cross-Attention
  1. 14.4.1 The frozen LLM
  2. 14.4.2 Perceiver resampler
  3. 14.4.3 Gated cross-attention
  4. 14.4.4 Masked cross-attention for interleaved inputs
  5. 14.4.5 In-context few-shot learning
  6. 14.4.6 Training data
  7. 14.4.7 OpenFlamingo and Otter
  8. 14.4.8 The descendants
  9. 14.4.9 Comparison to BLIP-2
5. 14.5 LLaVA and Visual Instruction Tuning
  1. 14.5.1 The architecture
  2. 14.5.2 Stage 1: projector alignment
  3. 14.5.3 Stage 2: visual instruction tuning
  4. 14.5.4 Why the community copied this
  5. 14.5.5 LLaVA-1.5 and LLaVA-NeXT
  6. 14.5.6 LLaVA-OneVision
  7. 14.5.7 The comparison to Q-Former
  8. 14.5.8 The prompt format
  9. 14.5.9 Parameter economy
6. 14.6 Any-Resolution Vision — Patch-n'-Pack and NaFlex
  1. 14.6.1 NaViT and patch-n'-pack
  2. 14.6.2 AnyRes (LLaVA-NeXT)
  3. 14.6.3 M-RoPE (Qwen2-VL)
  4. 14.6.4 NaFlex (SigLIP 2)
  5. 14.6.5 The packing mask
  6. 14.6.6 Token budgets
7. 14.7 Open-Weight VLM Recipes: What Actually Matters
  1. 14.7.1 The five-axis design space
  2. 14.7.2 Axis 1: encoder > connector
  3. 14.7.3 Axis 2: connector design is a wash
  4. 14.7.4 Axis 3: LLM size sets the ceiling
  5. 14.7.5 Axis 4: data — detailed human captions beat distillation
  6. 14.7.6 Axis 5: resolution and its schedule
  7. 14.7.7 The Prismatic controlled comparison
  8. 14.7.8 A picker for 2026
8. 14.8 LLaVA-OneVision: Single, Multi, Video
  1. 14.8.1 The OneVision token budget
  2. 14.8.2 The three-stage curriculum
  3. 14.8.3 Why curriculum works
  4. 14.8.4 Emergent cross-scenario skills
  5. 14.8.5 Visual-token pooling
  6. 14.8.6 LLaVA-OneVision-1.5
  7. 14.8.7 Contrast with Qwen2.5-VL
9. 14.9 Qwen-VL Family and Dynamic-FPS Video
  1. 14.9.1 Qwen-VL (August 2023)
  2. 14.9.2 Qwen2-VL (September 2024) — M-RoPE and native resolution
  3. 14.9.3 Qwen2.5-VL (February 2025) — dynamic FPS + absolute time
  4. 14.9.4 Qwen3-VL (November 2025)
  5. 14.9.5 M-RoPE mathematically
  6. 14.9.6 Dynamic-FPS sampling logic
  7. 14.9.7 Structured agent output
10. 14.10 InternVL3 Native Multimodal Pretraining
  1. 14.10.1 Native multimodal pretraining
  2. 14.10.2 V2PE (variable visual position encoding)
  3. 14.10.3 Visual Resolution Router (ViR)
  4. 14.10.4 Decoupled Vision-Language deployment (DvD)
  5. 14.10.5 Single-stage vs multi-stage quality
  6. 14.10.6 InternVL3.5 and InternVL-U
  7. 14.10.7 Trade-offs of native pretraining
11. 14.11 Chameleon Early-Fusion Token-Only
  1. 14.11.1 VQ-VAE as image tokenizer
  2. 14.11.2 The shared vocabulary
  3. 14.11.3 Mixed-modality generation
  4. 14.11.4 Training stability — QK-Norm, dropout, LayerNorm ordering
  5. 14.11.5 The tokenizer's reconstruction ceiling
  6. 14.11.6 Chameleon vs BLIP-2 / LLaVA
  7. 14.11.7 Fuyu and AnyGPT
12. 14.12 Emu3 Next-Token Prediction for Generation
  1. 14.12.1 The Emu3 tokenizer
  2. 14.12.2 Single-loss training
  3. 14.12.3 Classifier-free guidance and temperature
  4. 14.12.4 Three roles, one model
  5. 14.12.5 Benchmarks
  6. 14.12.6 Compute cost
  7. 14.12.7 Why it matters
13. 14.13 Transfusion Autoregressive + Diffusion
  1. 14.13.1 The two-loss architecture
  2. 14.13.2 Attention mask: causal text + bidirectional image
  3. 14.13.3 Diffusion loss inside the transformer
  4. 14.13.4 MMDiT: Stable Diffusion 3's variant
  5. 14.13.5 Why this beats Chameleon-style
  6. 14.13.6 What sits downstream
14. 14.14 Show-o Discrete-Diffusion Unified
  1. 14.14.1 Masked discrete diffusion (MaskGIT)
  2. 14.14.2 Show-o: one transformer, hybrid mask
  3. 14.14.3 Parallel sampling
  4. 14.14.4 Tasks in one checkpoint
  5. 14.14.5 Masking schedule
  6. 14.14.6 Show-o2
  7. 14.14.7 Where Show-o sits
15. 14.15 Janus-Pro Decoupled Encoders
  1. 14.15.1 Decoupled visual encoding
  2. 14.15.2 Why this works
  3. 14.15.3 Data scaling — Janus vs Janus-Pro
  4. 14.15.4 JanusFlow — the rectified flow variant
  5. 14.15.5 The shared body's job
  6. 14.15.6 Compared to InternVL-U
  7. 14.15.7 Limitations
16. 14.16 MIO Any-to-Any Streaming
  1. 14.16.1 Four tokenizers for four modalities
  2. 14.16.2 Streaming decode
  3. 14.16.3 Four-stage curriculum
  4. 14.16.4 Chain-of-visual-thought
  5. 14.16.5 Competitors in any-to-any
  6. 14.16.6 Latency budget
  7. 14.16.7 Why any-to-any stays hard
17. 14.17 Video-Language Temporal Grounding
  1. 14.17.1 Video-LLaMA: Q-former per clip + audio branch
  2. 14.17.2 VideoChat and Video-LLaVA
  3. 14.17.3 Qwen2.5-VL and time-aligned M-RoPE
  4. 14.17.4 Frame sampling strategies
  5. 14.17.5 Pooling per frame
  6. 14.17.6 The four video benchmarks
  7. 14.17.7 Grounding output formats
  8. 14.17.8 2026 best practice
18. 14.18 Long-Video at Million-Token Context
  1. 14.18.1 Path 1: Brute context (Gemini 1.5, Claude Opus)
  2. 14.18.2 Path 2: Ring attention (LWM, LongVILA)
  3. 14.18.3 Path 3: Token compression (Video-XL, LongVA)
  4. 14.18.4 Path 4: Agentic retrieval (VideoAgent)
  5. 14.18.5 Needle-in-a-haystack benchmarks
  6. 14.18.6 Which path to pick
  7. 14.18.7 2026 production pattern
19. 14.19 Audio-Language Models: Whisper to AF3
  1. 14.19.1 Log-Mel spectrogram: the input feature
  2. 14.19.2 Whisper's encoder
  3. 14.19.3 BEATs and audio-specific encoders
  4. 14.19.4 Audio Q-former
  5. 14.19.5 The arc — SALMONN, Qwen-Audio, AF3
  6. 14.19.6 Cascaded vs end-to-end
  7. 14.19.7 2026 production recipe
  8. 14.19.8 MMAU — the audio reasoning benchmark
20. 14.20 Omni Models: Thinker-Talker Streaming
  1. 14.20.1 Thinker and Talker
  2. 14.20.2 TMRoPE — time-aligned multimodal positions
  3. 14.20.3 Streaming speech synthesis
  4. 14.20.4 VAD and turn-taking
  5. 14.20.5 Qwen3-Omni (November 2025)
  6. 14.20.6 Production latency budget
  7. 14.20.7 Token-rate math
21. 14.21 Embodied VLAs: RT-2, OpenVLA, π0, GR00T
  1. 14.21.1 Action tokenization (RT-2)
  2. 14.21.2 OpenVLA — the open 7B reference
  3. 14.21.3 FAST tokenizer — faster action decode
  4. 14.21.4 π0 and flow-matching actions
  5. 14.21.5 GR00T N1 — dual-system for humanoids
  6. 14.21.6 Open X-Embodiment
  7. 14.21.7 Co-fine-tuning vs robot-only
  8. 14.21.8 Safety and action limits
22. 14.22 Document and Diagram Understanding
  1. 14.22.1 Era 1 — OCR pipeline (pre-2021)
  2. 14.22.2 TrOCR (2021)
  3. 14.22.3 Era 2 — OCR-free (2022-2023)
  4. 14.22.4 LayoutLMv3 (2022)
  5. 14.22.5 DocLLM (2023)
  6. 14.22.6 Era 3 — VLM-native (2024+)
  7. 14.22.7 The Claude 4.7 / GPT-5 frontier
  8. 14.22.8 Math equations and LaTeX output
  9. 14.22.9 Handwriting
  10. 14.22.10 2026 recipe
23. 14.23 ColPali Vision-Native Document RAG
  1. 14.23.1 ColBERT (2020)
  2. 14.23.2 ColPali
  3. 14.23.3 ColQwen2 and ColSmol
  4. 14.23.4 VisRAG
  5. 14.23.5 M3DocRAG
  6. 14.23.6 ViDoRe — the benchmark
  7. 14.23.7 The end-to-end RAG pipeline
  8. 14.23.8 Storage math
  9. 14.23.9 When text-RAG still wins
24. 14.24 Multimodal RAG and Cross-Modal Retrieval
  1. 14.24.1 Cross-modal retrieval
  2. 14.24.2 Fusion strategies
  3. 14.24.3 Generation grounding
  4. 14.24.4 The 2025 surveys
  5. 14.24.5 MuRAG — the foundational paper
  6. 14.24.6 A production trip-planner example
  7. 14.24.7 Agentic multimodal RAG
  8. 14.24.8 Evaluation
25. 14.25 Multimodal Agents and Computer-Use (Capstone)
  1. 14.25.1 GUI grounding — the primitive
  2. 14.25.2 Action schemas
  3. 14.25.3 Screenshot-only vs accessibility-tree
  4. 14.25.4 Long-horizon memory
  5. 14.25.5 Visual tool use
  6. 14.25.6 The 2026 benchmarks
  7. 14.25.7 Why it's still hard
  8. 14.25.8 The capstone build-it
15Tools and protocolsnot yet
1. 15.1 The Tool Interface
  1. 15.1.1 Step one: describe
  2. 15.1.2 Step two: decide
  3. 15.1.3 Step three: execute
  4. 15.1.4 Step four: observe
  5. 15.1.5 The trust split
  6. 15.1.6 Where the loop lives
  7. 15.1.7 Why not just prompt the model to emit JSON?
  8. 15.1.8 Circuit breakers
  9. 15.1.9 Where Phase 13 goes from here
2. 15.2 Function Calling Deep Dive
  1. 15.2.1 The common structure
  2. 15.2.2 Shape diffs, field by field
  3. 15.2.3 Limits you will actually hit
  4. 15.2.4 tool_choice behavior
  5. 15.2.5 Parallel calls
  6. 15.2.6 Streaming
  7. 15.2.7 Errors and repair
  8. 15.2.8 The translator pattern
3. 15.3 Parallel and Streaming Tool Calls
  1. 15.3.1 Enabling parallel
  2. 15.3.2 Id correlation
  3. 15.3.3 Running calls concurrently
  4. 15.3.4 Streaming tool calls
  5. 15.3.5 Partial JSON and the parse-early trap
  6. 15.3.6 Out-of-order completion
  7. 15.3.7 Benchmark: sequential vs parallel
  8. 15.3.8 Streaming fan-out wall-clock
4. 15.4 Structured Output
  1. 15.4.1 JSON Schema 2020-12 — the lingua franca
  2. 15.4.2 Pydantic, the Python binding
  3. 15.4.3 Zod, the TypeScript binding
  4. 15.4.4 Refusals
  5. 15.4.5 Constrained decoding in the open
  6. 15.4.6 The three failure modes
  7. 15.4.7 Retry strategy
  8. 15.4.8 Small-model support
5. 15.5 Tool Schema Design
  1. 15.5.1 Naming rules
  2. 15.5.2 Description pattern
  3. 15.5.3 Atomic vs monolithic
  4. 15.5.4 Parameter design
  5. 15.5.5 Error messages as teaching signals
  6. 15.5.6 Versioning
  7. 15.5.7 Tool poisoning prevention
  8. 15.5.8 Benchmarks
6. 15.6 MCP Fundamentals: Stateless Requests and JSON-RPC
  1. 15.6.1 Server primitives
  2. 15.6.2 JSON-RPC envelopes
  3. 15.6.3 Required request metadata
  4. 15.6.4 Complete results and server identity
  5. 15.6.5 Discovery without a handshake
  6. 15.6.6 One request lifecycle
  7. 15.6.7 Explicit legacy compatibility
7. 15.7 Building an MCP Server: Stateless Python and TypeScript
  1. 15.7.1 The modern dispatch loop
  2. 15.7.2 Request validation
  3. 15.7.3 Mandatory discovery
  4. 15.7.4 Tools
  5. 15.7.5 Resources
  6. 15.7.6 Prompts
  7. 15.7.7 Every successful result is typed
  8. 15.7.8 No server-initiated requests
  9. 15.7.9 Explicit legacy compatibility
8. 15.8 Building an MCP Client: Discovery, Routing, and Dual-Era Fallback
  1. 15.8.1 A peer, not a protocol session
  2. 15.8.2 Build every modern request from scratch
  3. 15.8.3 Modern discovery
  4. 15.8.4 The stdio compatibility probe
  5. 15.8.5 Allowlisting is operator intent, not evidence
  6. 15.8.6 Legacy is a compatibility branch
  7. 15.8.7 Discovering and caching tools
  8. 15.8.8 Collision-safe namespace merge
  9. 15.8.9 Routing a call
  10. 15.8.10 Notifications and subscriptions
  11. 15.8.11 No server-initiated requests
9. 15.9 MCP Transports: stdio and Stateless Streamable HTTP
  1. 15.9.1 stdio
  2. 15.9.2 Streamable HTTP in 2026-07-28
  3. 15.9.3 POST-only means POST-only
  4. 15.9.4 Origin validation
  5. 15.9.5 Required HTTP metadata headers
  6. 15.9.6 Request-scoped SSE
  7. 15.9.7 Long-lived changes use subscriptions/listen
  8. 15.9.8 Explicit application state
  9. 15.9.9 HTTP dual-era compatibility
10. 15.10 MCP Resources and Prompts: Addressable Context for Stateless Servers
  1. 15.10.1 Resource templates
  2. 15.10.2 Content is not trusted instruction
11. 15.11 MCP Model Input: Sampling Migration and Stateless MRTR
  1. 15.11.1 New server: integrate with a model provider directly
  2. 15.11.2 Existing Sampling workflow: migrate it to MRTR
  3. 15.11.3 Sampling
  4. 15.11.4 MRTR
  5. 15.11.5 InputRequiredResult
  6. 15.11.6 inputRequests
  7. 15.11.7 inputResponses
  8. 15.11.8 requestState
  9. 15.11.9 resultType
  10. 15.11.10 Direct model integration
  11. 15.11.11 Capability gate
  12. 15.11.12 Loop budget
12. 15.12 Explicit Scope and Stateless Elicitation
  1. 15.12.1 The three-layer rule
  2. 15.12.2 Roots
  3. 15.12.3 Explicit scope
  4. 15.12.4 Containment
  5. 15.12.5 Elicitation
  6. 15.12.6 Form mode
  7. 15.12.7 URL mode
  8. 15.12.8 MRTR
  9. 15.12.9 requestState
  10. 15.12.10 Decline
  11. 15.12.11 Cancel
13. 15.13 MCP Tasks Extension: Durable Work on a Stateless Core
  1. 15.13.1 Input needed before task creation
  2. 15.13.2 Input needed after task creation
  3. 15.13.3 Protocol error
  4. 15.13.4 Task execution outcome
14. 15.14 MCP Apps on the Stateless Protocol
  1. 15.14.1 Two protocols, one feature
  2. 15.14.2 Discover before rendering
  3. 15.14.3 Declare the UI on the tool definition
  4. 15.14.4 Return data, then let the host bind the view
  5. 15.14.5 Serve the app as a resource
  6. 15.14.6 Cache UI resources as executable content
  7. 15.14.7 Reject wire ambiguity before feature policy
  8. 15.14.8 The sandbox is a boundary, not a trust verdict
  9. 15.14.9 The Apps bridge has its own lifecycle
  10. 15.14.10 Host context, actions, and revocation
  11. 15.14.11 Fallback is part of the contract
15. 15.15 MCP Security: Poisoned Metadata, Routing, and MRTR State
  1. 15.15.1 Seven attack surfaces worth checking
  2. 15.15.2 The current request envelope is evidence, not identity
  3. 15.15.3 Validate routing before policy
  4. 15.15.4 Pin the whole descriptor
  5. 15.15.5 Static scanning is a tripwire
  6. 15.15.6 Namespace before merging
  7. 15.15.7 Capabilities are compatibility declarations
  8. 15.15.8 Protect stateless MRTR confirmation
  9. 15.15.9 Rule of two for high-risk calls
  10. 15.15.10 Reduce authority before execution
  11. 15.15.11 Current and legacy interaction paths
  12. 15.15.12 Stateless transport checks
16. 15.16 MCP Authorization: CIMD, Issuer Binding, PKCE, and Step-Up
  1. 15.16.1 Know the three roles
  2. 15.16.2 Authorization applies to HTTP
  3. 15.16.3 Start with protected-resource metadata
  4. 15.16.4 Verify authorization server metadata
  5. 15.16.5 Follow the registration priority
  6. 15.16.6 Prefer Client ID Metadata Documents
  7. 15.16.7 DCR is a compatibility path
  8. 15.16.8 Bind credentials to the issuer
  9. 15.16.9 Authorization code with PKCE
  10. 15.16.10 Validate iss exactly
  11. 15.16.11 Validate audience at the MCP server
  12. 15.16.12 Request the smallest current scope
  13. 15.16.13 Authorization and the stateless MCP wire
  14. 15.16.14 No token passthrough
  15. 15.16.15 Refresh tokens
17. 15.17 Stateless MCP Gateways and Registry Admission
  1. 15.17.1 The modern gateway path
  2. 15.17.2 Runtime policy is the primary gateway decision
  3. 15.17.3 One POST endpoint
  4. 15.17.4 Implement discovery at every layer
  5. 15.17.5 Per-request client capabilities
  6. 15.17.6 Deterministic namespacing
  7. 15.17.7 Pin approved descriptors
  8. 15.17.8 Registries help discover, not decide
  9. 15.17.9 Credential mediation
  10. 15.17.10 Rate limits without sessions
  11. 15.17.11 Audit the decision chain
  12. 15.17.12 Request-scoped SSE
  13. 15.17.13 Long-lived change notifications
  14. 15.17.14 MRTR through a gateway
  15. 15.17.15 Tasks extension routing
  16. 15.17.16 Compatibility boundary
18. 15.18 MCP Auth in Production: Issuer-Bound Enrollment and Tokens
  1. 15.18.1 RFC 8414 — OAuth Authorization Server Metadata
  2. 15.18.2 RFC 9728 (recap) — Protected Resource Metadata
  3. 15.18.3 Client ID Metadata Documents (the recommended default)
  4. 15.18.4 RFC 7591: deprecated compatibility enrollment
  5. 15.18.5 RFC 8707 (recap) — Resource Indicators
  6. 15.18.6 RFC 7636 (recap) — PKCE
  7. 15.18.7 MCP 2026-07-28 authorization profile
  8. 15.18.8 Deployment capability checklist
  9. 15.18.9 JWKS refresh pattern (rotate at the AS, refresh at the resource server)
  10. 15.18.10 The validation routine
  11. 15.18.11 Opaque tokens use introspection, not guesswork
  12. 15.18.12 Revocation is a freshness contract
  13. 15.18.13 Dependency failure needs a declared decision
  14. 15.18.14 Audience-replay walkthrough (access-token privilege restriction)
  15. 15.18.15 Mix-up attacks (a client-side defense the server cannot provide)
  16. 15.18.16 Failure modes
19. 15.19 A2A Protocol
  1. 15.19.1 Agent Card
  2. 15.19.2 Signed Agent Cards
  3. 15.19.3 Task lifecycle
  4. 15.19.4 Messages and Parts
  5. 15.19.5 Artifacts
  6. 15.19.6 Three protocol bindings
  7. 15.19.7 Opacity preservation
  8. 15.19.8 Timeline
  9. 15.19.9 Relationship to MCP
20. 15.20 OpenTelemetry GenAI
  1. 15.20.1 Span hierarchy
  2. 15.20.2 Required attributes
  3. 15.20.3 Span kinds
  4. 15.20.4 Opt-in content capture
  5. 15.20.5 Events on spans
  6. 15.20.6 Exporters
  7. 15.20.7 Propagation across MCP
  8. 15.20.8 Metrics
  9. 15.20.9 AgentOps layer
21. 15.21 LLM Routing Layer
  1. 15.21.1 OpenAI-compatible proxy shape
  2. 15.21.2 Model aliases
  3. 15.21.3 Fallback chains
  4. 15.21.4 Semantic caching
  5. 15.21.5 Guardrails
  6. 15.21.6 Per-key rate limits
  7. 15.21.7 Self-hosted vs managed trade-offs
  8. 15.21.8 Cost tracking
  9. 15.21.9 MCP plus routing
  10. 15.21.10 Routing strategies
22. 15.22 Agent Skills: Portable Contract and Runtime Boundary
  1. 15.22.1 Preflight for the real-host lab
  2. 15.22.2 Start in an empty working directory
  3. 15.22.3 Install the complete reviewer bundle
  4. 15.22.4 Invoke it explicitly
  5. 15.22.5 Probe implicit selection
  6. 15.22.6 Clean up
  7. 15.22.7 Skills encode procedural knowledge
  8. 15.22.8 The neighboring abstractions
  9. 15.22.9 The word "skill" names two different ideas
  10. 15.22.10 The portable core
  11. 15.22.11 Runtime extensions are a second layer
  12. 15.22.12 Frontmatter is executable metadata
  13. 15.22.13 The skill lifecycle
  14. 15.22.14 Skills and tools are orthogonal
  15. 15.22.15 Skills and repository instructions are different scopes
  16. 15.22.16 Skills do not import one another
  17. 15.22.17 Validation order matters
23. 15.23 Capstone: Stateless Tool Ecosystem
  1. 15.23.1 Target architecture
  2. 15.23.2 Target trace
  3. 15.23.3 Current protocol surfaces
  4. 15.23.4 Stateless MCP changes the integration boundary
  5. 15.23.5 Security posture
  6. 15.23.6 Skills are procedure, not transport
  7. 15.23.7 Course artifact metadata is a local adapter
  8. 15.23.8 Simulation versus production
  9. 15.23.9 Phase 13 map
24. 15.24 Skill Discovery and Progressive Disclosure
  1. 15.24.1 Discovery is a compiler pipeline
  2. 15.24.2 Scope is runtime policy
  3. 15.24.3 Collisions need identity beyond name
  4. 15.24.4 Three disclosure levels
  5. 15.24.5 Branch-specific references beat topic dumps
  6. 15.24.6 Catalog budget and active context are different budgets
  7. 15.24.7 Resource paths are a trust boundary
  8. 15.24.8 Loading must be observable
  9. 15.24.9 Why discovery is shallow
  10. 15.24.10 Why the lab does not parse arbitrary YAML
25. 15.25 Skill Invocation and Routing
  1. 15.25.1 Five channels can start the lifecycle
  2. 15.25.2 The five invocation stages
  3. 15.25.3 Human and model invocation form a 2x2 matrix
  4. 15.25.4 Explicit invocation is identity-first
  5. 15.25.5 Implicit invocation is description-first
  6. 15.25.6 Routing is classification with an abstain option
  7. 15.25.7 Eligibility must precede ranking
  8. 15.25.8 Routing evals need near misses
  9. 15.25.9 Arguments have three representations
  10. 15.25.10 Application invocation is explicit orchestration
  11. 15.25.11 Skill-to-skill invocation is a tool-like edge
  12. 15.25.12 Context lifecycle is host-specific
  13. 15.25.13 Why core and extension adapters are separate
26. 15.26 Skill Permissions, Sandboxes, and Trust
  1. 15.26.1 Skills are context, not a security boundary
  2. 15.26.2 Five control layers
  3. 15.26.3 Threat-model the complete package
  4. 15.26.4 Package trust begins before activation
  5. 15.26.5 Content has authority levels
  6. 15.26.6 Review actions as structured requests
  7. 15.26.7 Command policy needs structure
  8. 15.26.8 Path policy must resolve reality
  9. 15.26.9 Secret handling is capability design
  10. 15.26.10 Network is an independent permission
  11. 15.26.11 Approval should follow consequence
  12. 15.26.12 Choose the isolation boundary
  13. 15.26.13 Scripts should be boring
  14. 15.26.14 Run the isolation drill
  15. 15.26.15 Why ask is not allow
27. 15.27 Skill Evals, Packaging, and Portability
  1. 15.27.1 Start from a real workflow, not a topic
  2. 15.27.2 Separate judgment from deterministic work
  3. 15.27.3 Author the package in dependency order
  4. 15.27.4 Six eval layers
  5. 15.27.5 Routing evals must use the target runtime
  6. 15.27.6 Artifact contracts make behavior executable
  7. 15.27.7 Install the directory as one unit
  8. 15.27.8 Portability is a capability matrix
  9. 15.27.9 Portability tests need host fixtures
  10. 15.27.10 Scale data is not quality evidence
  11. 15.27.11 Separate fixture success, local integrity, and production readiness
  12. 15.27.12 Read the report by layer
  13. 15.27.13 Establish the local fixture boundary
  14. 15.27.14 Install the complete bundle into the first host
  15. 15.27.15 Probe discovery, routing, references, and scripts
  16. 15.27.16 Probe approval behavior
  17. 15.27.17 Use a second host or declare the fallback
  18. 15.27.18 Exercise upgrade and uninstall
28. 15.28 MCP Tool Contracts and Content
  1. 15.28.1 Structured content is any JSON value
  2. 15.28.2 A small validator still teaches the boundary
  3. 15.28.3 Encode values before building HTTP headers
  4. 15.28.4 The server checks the mirrored copy
29. 15.29 MCP Reliability, Cancellation, and Flow Control
  1. 15.29.1 stdio
  2. 15.29.2 Streamable HTTP
  3. 15.29.3 Server-sent cancellation is narrow
  4. 15.29.4 Cancellation wins
  5. 15.29.5 Completion wins
  6. 15.29.6 The ledger boundary must be atomic and durable
  7. 15.29.7 Retry matrix
  8. 15.29.8 Proxy buffering
  9. 15.29.9 Prevent a reconnect herd
  10. 15.29.10 RequestCoordinator
  11. 15.29.11 MutationLedger
  12. 15.29.12 DurableTaskService
  13. 15.29.13 BoundedSseBuffer
  14. 15.29.14 Recovery helpers
30. 15.30 MCP Registry Supply Chain: Admission, Drift, and Rollback
  1. 15.30.1 Namespace verification
  2. 15.30.2 Provenance join
  3. 15.30.3 Pin the decision, not only the version
  4. 15.30.4 Live drift detection
  5. 15.30.5 Registry status is live state
  6. 15.30.6 Rollback means route restoration
  7. 15.30.7 Append an admission ledger
  8. 15.30.8 Lab A: namespace collision
  9. 15.30.9 Lab B: descriptor drift
  10. 15.30.10 Lab C: status and rollback
31. 15.31 MCP Conformance Engineering: Versioning, Evidence, and Operations
  1. 15.31.1 Strict mode
  2. 15.31.2 Fallback mode
  3. 15.31.3 Golden transcripts
  4. 15.31.4 Negative transcripts
  5. 15.31.5 Additive unknown fields
  6. 15.31.6 Unknown resultType
  7. 15.31.7 Lab A: prove the era boundary
  8. 15.31.8 Lab B: additive field versus discriminator
  9. 15.31.9 Lab C: inspect an SDK transformation
  10. 15.31.10 Lab D: repair the proxy
32. 15.32 Program Anatomy: An ELF Reader From Scratch
  1. 15.32.1 Lab A: the header by hand
  2. 15.32.2 Lab B: the value table behind the switch
  3. 15.32.3 Lab C: three builds of one source
  4. 15.32.4 Lab D: the same function in Mach-O
  5. 15.32.5 See it run
33. 15.33 Disassembly and Control Flow: From Bytes to a Graph You Can Check
  1. 15.33.1 Prefixes come in four groups
  2. 15.33.2 REX adds four bits and must sit next to the opcode
  3. 15.33.3 Opcodes live in two maps
  4. 15.33.4 ModRM and SIB select the operands
  5. 15.33.5 The displacement is where the link editor writes
  6. 15.33.6 Immediates are 32 bits at most, with one exception
  7. 15.33.7 Lab A: decode on paper, then check
  8. 15.33.8 Lab B: the relocation on a lea
  9. 15.33.9 Lab C: watch the two methods disagree
  10. 15.33.10 Lab D: dominators by hand
  11. 15.33.11 See it run
34. 15.34 Lifting to Pseudocode: IR, SSA, and Structuring From Scratch
  1. 15.34.1 Flags are a variable
  2. 15.34.2 cltd and idiv
  3. 15.34.3 Stack slots at -O0
  4. 15.34.4 Relocations
  5. 15.34.5 The 0x66666667 multiply
  6. 15.34.6 What -O1 did to the switch
  7. 15.34.7 percent_change
  8. 15.34.8 grade
  9. 15.34.9 total
  10. 15.34.10 Lab A: one function through five stages
  11. 15.34.11 Lab B: the division idiom
  12. 15.34.12 Lab C: a new object stops the decoder
  13. 15.34.13 Lab D: Ghidra's default value in grade
  14. 15.34.14 See it run
35. 15.35 Evidence, Confidence, and Unknowns in Binary Analysis
  1. 15.35.1 Lab A: recompute a rea id
  2. 15.35.2 Lab B: refuse a promotion
  3. 15.35.3 Lab C: refuse a stale snapshot
  4. 15.35.4 Lab D: break the chain
  5. 15.35.5 See it run
36. 15.36 Rebuild and Prove Equal: Differential Tests, Byte Comparison, and the Evidence Loop
  1. 15.36.1 Design the grid
  2. 15.36.2 The driver and the trap
  3. 15.36.3 Results
  4. 15.36.4 Read the object
  5. 15.36.5 Mask, then compare the records
  6. 15.36.6 Read the constants the relocation points at
  7. 15.36.7 What rea's verifier can and cannot say
  8. 15.36.8 Lab A: the model and the run
  9. 15.36.9 Lab B: read the object
  10. 15.36.10 Lab C: break one constant
  11. 15.36.11 Lab D: drive rea
  12. 15.36.12 See it run
37. 15.37 The Reverse Engineering Landscape: History, Tool Families, and How a Decompiler Is Built
  1. 15.37.1 1973 to 1991: debuggers first, then disassemblers
  2. 15.37.2 1991 to 2009: interactive disassemblers and the first decompilers
  3. 15.37.3 2009 to 2019: frameworks and intermediate representations
  4. 15.37.4 2019 to 2026: open decompilers and the research turn
  5. 15.37.5 Disassemblers
  6. 15.37.6 Decompilers
  7. 15.37.7 Debuggers
  8. 15.37.8 Binary analysis frameworks
  9. 15.37.9 Format and engine libraries
  10. 15.37.10 Diffing tools and compiler-side viewers
  11. 15.37.11 Versions, licences, and what this lesson ran
  12. 15.37.12 Loader
  13. 15.37.13 Disassembly
  14. 15.37.14 CFG recovery
  15. 15.37.15 Lifting
  16. 15.37.16 SSA
  17. 15.37.17 Data flow and types
  18. 15.37.18 Structuring
  19. 15.37.19 Emission
  20. 15.37.20 The agreement case: ledger at -O1
  21. 15.37.21 The vectorised case: ledger at -O2
  22. 15.37.22 The jump table: grade at -O0
  23. 15.37.23 Data in code: sweep-trap
  24. 15.37.24 Invented and missed, by tool
  25. 15.37.25 Lab A: one function, four bases
  26. 15.37.26 Lab B: the jump table from four sides
  27. 15.37.27 Lab C: eager flags against a thunk
  28. 15.37.28 Lab D: the trap under five tools
38. 15.38 Calling Conventions and Stack Frames: System V, Windows x64, and AArch64
  1. 15.38.1 Lab A: one instruction, three effects
  2. 15.38.2 Lab B: the frame of keeps
  3. 15.38.3 Lab C: read-before-write on mixed
  4. 15.38.4 Lab D: the unwinder's second opinion
39. 15.39 Compiler Idioms: What Optimization Does to Your Code and How to Read It Back
  1. 15.39.1 Unsigned 10 by hand
  2. 15.39.2 Unsigned 7 by hand, with the add fix-up
  3. 15.39.3 Signed 3 by hand
  4. 15.39.4 What the constants are not
  5. 15.39.5 The idiom belongs to the instruction selector
  6. 15.39.6 The lookup table, read from the object
  7. 15.39.7 The linear map, with no table at all
  8. 15.39.8 The jump table, read from relocations
  9. 15.39.9 The same switch at four levels
  10. 15.39.10 Loops that become library calls
  11. 15.39.11 Lab A: derive a magic number, then find it
  12. 15.39.12 Lab B: the table and the tree
  13. 15.39.13 Lab C: the loop at three settings
  14. 15.39.14 Lab D: the shared quotient
40. 15.40 Linking, Loading, and Stripped Binaries: PLT, GOT, Imports, Fixups, and Function Identification
  1. 15.40.1 PLT0 and the entry for puts
  2. 15.40.2 What the file does not have
  3. 15.40.3 .dynsym, .symtab, and strip
  4. 15.40.4 The import directory
  5. 15.40.5 The export directory
  6. 15.40.6 Base relocations
  7. 15.40.7 One stub, by hand
  8. 15.40.8 Two-level namespace
  9. 15.40.9 The layout it replaced
  10. 15.40.10 The exports trie
  11. 15.40.11 LC_FUNCTION_STARTS
  12. 15.40.12 Prologue scans
  13. 15.40.13 Call targets
  14. 15.40.14 What the tools do with the same idea
  15. 15.40.15 Lab A: the PLT by hand
  16. 15.40.16 Lab B: one chained fixup
  17. 15.40.17 Lab C: strip and relink
  18. 15.40.18 Lab D: the collision
41. 15.41 Recovering Data and Types: Structs, Arrays, C++ Objects, and DWARF
  1. 15.41.1 Four kinds of stride evidence
  2. 15.41.2 The vectorised loop at -O2
  3. 15.41.3 The unit header
  4. 15.41.4 Abbreviations
  5. 15.41.5 Decode the balance member by hand
  6. 15.41.6 Strings and addresses go through two tables
  7. 15.41.7 Types, functions, and locations
  8. 15.41.8 Mangled names
  9. 15.41.9 The vtable from its relocations
  10. 15.41.10 RTTI
  11. 15.41.11 The virtual call
  12. 15.41.12 Lab A: decode the member entry
  13. 15.41.13 Lab B: trace one function
  14. 15.41.14 Lab C: fold the vectorised loop
  15. 15.41.15 Lab D: the vtable and the names
42. 15.42 Dynamic Analysis of Your Own Program: Debuggers, Traces, and Confirming Static Claims
  1. 15.42.1 Decode the branch by hand
  2. 15.42.2 Lab A: read a stop by hand
  3. 15.42.3 Lab B: the branch and its successor
  4. 15.42.4 Lab C: the run that never stopped
  5. 15.42.5 Lab D: break the evidence
43. 15.43 Beyond Native Code: Python, JVM, .NET, and WebAssembly Bytecode
  1. 15.43.1 Lab A: read the header, then the code object
  2. 15.43.2 Lab B: the constant pool chain
  3. 15.43.3 Lab C: depth and labels
  4. 15.43.4 Lab D: from token to bytes
44. 15.44 Binary Diffing and Symbolic Execution: Matching Functions and Solving for Inputs
  1. 15.44.1 Lab A: five signatures by hand
  2. 15.44.2 Lab B: watch the relocation break the byte hash
  3. 15.44.3 Lab C: the false match and the order that avoids it
  4. 15.44.4 Lab D: paths and the bound
45. 15.45 AI-Assisted Reverse Engineering: LLM Decompilers, Analysis Agents, and Ground Truth
  1. 15.45.1 bytes
  2. 15.45.2 grid
  3. 15.45.3 fuzz
  4. 15.45.4 names
  5. 15.45.5 citations
  6. 15.45.6 provenance
  7. 15.45.7 The card
  8. 15.45.8 Lab A: read the DWARF by hand, then with the reader
  9. 15.45.9 Lab B: score a candidate and repair it
  10. 15.45.10 Lab C: the string that addresses the analyst
  11. 15.45.11 Lab D: drive the server
16Agent engineeringnot yet
1. 16.1 The Agent Loop
  1. 16.1.1 ReAct: the canonical format
  2. 16.1.2 The 2026 shift: native reasoning
  3. 16.1.3 The five ingredients
  4. 16.1.4 Why this loop is everywhere
  5. 16.1.5 2026 pitfalls
2. 16.2 ReWOO and Plan-and-Execute
  1. 16.2.1 The three roles
  2. 16.2.2 Why 5x fewer tokens
  3. 16.2.3 Why it is more robust
  4. 16.2.4 Planner distillation
  5. 16.2.5 Plan-and-Execute (2023)
  6. 16.2.6 Plan-and-Act (Erdogan et al., arXiv:2503.09572, ICML 2025)
  7. 16.2.7 When to pick which
3. 16.3 Reflexion and Verbal Reinforcement Learning
  1. 16.3.1 The three components
  2. 16.3.2 Three evaluator types
  3. 16.3.3 Why this generalizes
  4. 16.3.4 When it works and when it does not
4. 16.4 Tree of Thoughts and LATS
  1. 16.4.1 Tree of Thoughts (Yao et al., NeurIPS 2023)
  2. 16.4.2 LATS (Zhou et al., ICML 2024)
  3. 16.4.3 MCTS, minimally
  4. 16.4.4 The cost reality
  5. 16.4.5 2026 positioning
5. 16.5 Self-Refine and CRITIC
  1. 16.5.1 Self-Refine (Madaan et al., NeurIPS 2023)
  2. 16.5.2 CRITIC (Gou et al., arXiv:2305.11738, v4 Feb 2024)
  3. 16.5.3 The stop condition
  4. 16.5.4 Evaluator-Optimizer (Anthropic, 2024)
  5. 16.5.5 OpenAI Agents SDK output guardrails
  6. 16.5.6 2026 pitfalls
6. 16.6 Tool Use and Function Calling
  1. 16.6.1 Toolformer (Schick et al., NeurIPS 2023)
  2. 16.6.2 Berkeley Function Calling Leaderboard V4 (Patil et al., ICML 2025)
  3. 16.6.3 Tool schema
  4. 16.6.4 Argument validation
  5. 16.6.5 Parallel tool calls
  6. 16.6.6 Sandboxing
7. 16.7 Agent Memory — Virtual Context and Memory Paging
  1. 16.7.1 The OS analogy
  2. 16.7.2 Two tiers
  3. 16.7.3 The interrupt pattern
  4. 16.7.4 Where the paper ends and production begins
  5. 16.7.5 Where this pattern goes wrong
  6. 16.7.6 The shape of agent memory
8. 16.8 Memory Blocks and Sleep-Time Compute
  1. 16.8.1 Three tiers
  2. 16.8.2 Memory blocks
  3. 16.8.3 Sleep-time compute
  4. 16.8.4 Native reasoning
  5. 16.8.5 Where this pattern goes wrong
9. 16.9 Hybrid Memory — Vector + Graph + KV
  1. 16.9.1 Three stores in parallel
  2. 16.9.2 Fusion scoring
  3. 16.9.3 Mem0g and temporal reasoning
  4. 16.9.4 Benchmark numbers
  5. 16.9.5 Scope taxonomy
  6. 16.9.6 Where this pattern goes wrong
10. 16.10 Skill Libraries and Lifelong Learning (Voyager)
  1. 16.10.1 Three components
  2. 16.10.2 Action space = code
  3. 16.10.3 Skill retrieval
  4. 16.10.4 Iterative refinement
  5. 16.10.5 Curriculum and exploration
  6. 16.10.6 Where this pattern goes wrong
11. 16.11 Planning with HTN and Evolutionary Search
  1. 16.11.1 Hierarchical Task Networks
  2. 16.11.2 ChatHTN (Gopalakrishnan et al., 2025)
  3. 16.11.3 AlphaEvolve (Novikov et al., 2025)
  4. 16.11.4 When to use which
  5. 16.11.5 Where this pattern goes wrong
12. 16.12 Anthropic's Workflow Patterns
  1. 16.12.1 Workflows vs agents
  2. 16.12.2 The augmented LLM
  3. 16.12.3 The five patterns
  4. 16.12.4 Where workflows beat agents
  5. 16.12.5 Where agents beat workflows
  6. 16.12.6 The context-engineering companion
13. 16.13 Stateful Graph Orchestration — Durable Execution and Checkpoints
  1. 16.13.1 The graph
  2. 16.13.2 Durable execution
  3. 16.13.3 Streaming
  4. 16.13.4 Human-in-the-loop
  5. 16.13.5 Memory
  6. 16.13.6 Three topologies
  7. 16.13.7 Where this pattern goes wrong
14. 16.14 The Actor Model for Agents
  1. 16.14.1 Actors
  2. 16.14.2 Three API layers
  3. 16.14.3 Why decoupling matters
  4. 16.14.4 Topologies
  5. 16.14.5 Observability
  6. 16.14.6 Status: maintenance mode
15. 16.15 Role-Based Agent Teams — Roles, Tasks, Processes
  1. 16.15.1 Four primitives
  2. 16.15.2 Sequential vs Hierarchical vs Consensus
  3. 16.15.3 Crews vs Flows
  4. 16.15.4 Tool integration
  5. 16.15.5 Memory hooks
  6. 16.15.6 When role-based teams fit
  7. 16.15.7 When they do not
  8. 16.15.8 Dependency shape
  9. 16.15.9 Where this pattern goes wrong
16. 16.16 OpenAI Agents SDK — Handoffs, Guardrails, Tracing
  1. 16.16.1 Five primitives
  2. 16.16.2 Handoffs as tools
  3. 16.16.3 Guardrails
  4. 16.16.4 Tracing
  5. 16.16.5 Sessions
  6. 16.16.6 Where this pattern goes wrong
17. 16.17 The Harness as a Library — Subagents and Session Store
  1. 16.17.1 Client SDK vs Agent SDK
  2. 16.17.2 Built-in tools
  3. 16.17.3 Subagents
  4. 16.17.4 Session store
  5. 16.17.5 Hooks
  6. 16.17.6 W3C trace context
  7. 16.17.7 Claude Managed Agents
  8. 16.17.8 Where this pattern goes wrong
18. 16.18 Production Agent Runtimes
  1. 16.18.1 Agno
  2. 16.18.2 Mastra
  3. 16.18.3 Positioning
  4. 16.18.4 When to pick each
  5. 16.18.5 Where this pattern goes wrong
19. 16.19 Benchmarks — SWE-bench, GAIA, AgentBench
  1. 16.19.1 SWE-bench (Jimenez et al., ICLR 2024 oral)
  2. 16.19.2 SWE-bench Verified
  3. 16.19.3 Contamination
  4. 16.19.4 GAIA (Mialon et al., Nov 2023)
  5. 16.19.5 AgentBench (Liu et al., ICLR 2024)
  6. 16.19.6 What these do not measure
  7. 16.19.7 Where benchmarking goes wrong
20. 16.20 Benchmarks — WebArena and OSWorld
  1. 16.20.1 WebArena (Zhou et al., ICLR 2024)
  2. 16.20.2 Extensions
  3. 16.20.3 OSWorld (Xie et al., NeurIPS 2024)
  4. 16.20.4 Primary failure modes
  5. 16.20.5 Follow-ups
  6. 16.20.6 Why this matters
  7. 16.20.7 Where benchmarking goes wrong
21. 16.21 Computer Use — Claude, OpenAI CUA, Gemini
  1. 16.21.1 Claude computer use (Anthropic, Oct 22 2024)
  2. 16.21.2 OpenAI CUA / Operator (Jan 2025)
  3. 16.21.3 Gemini 2.5 Computer Use (Google DeepMind, Oct 7 2025)
  4. 16.21.4 The shared contract: untrusted input
  5. 16.21.5 When to pick which
  6. 16.21.6 Where this pattern goes wrong
22. 16.22 Voice Agents — Pipecat and LiveKit
  1. 16.22.1 Pipecat (pipecat-ai/pipecat)
  2. 16.22.2 LiveKit Agents (livekit/agents)
  3. 16.22.3 Commercial platforms
  4. 16.22.4 Where this pattern goes wrong
  5. 16.22.5 Typical 2026 latencies
23. 16.23 OpenTelemetry GenAI Semantic Conventions
  1. 16.23.1 Span categories
  2. 16.23.2 Agent span naming
  3. 16.23.3 Key attributes
  4. 16.23.4 Content capture
  5. 16.23.5 Stability
  6. 16.23.6 Where this pattern goes wrong
24. 16.24 Agent Observability — Langfuse, Phoenix, Opik
  1. 16.24.1 Langfuse (MIT)
  2. 16.24.2 Arize Phoenix (Elastic License 2.0)
  3. 16.24.3 Comet Opik (Apache 2.0)
  4. 16.24.4 Industry data
  5. 16.24.5 Picking one
  6. 16.24.6 Where this pattern goes wrong
25. 16.25 Multi-Agent Debate and Collaboration
  1. 16.25.1 Society of Minds (Du et al., ICML 2024)
  2. 16.25.2 Sparse topology
  3. 16.25.3 When debate helps
  4. 16.25.4 When debate hurts
  5. 16.25.5 2026 practical instantiations
  6. 16.25.6 Where this pattern goes wrong
26. 16.26 Failure Modes — Why Agents Break
  1. 16.26.1 MASFT (Berkeley, arXiv:2503.13657)
  2. 16.26.2 Microsoft Taxonomy of Failure Mode in Agentic AI Systems
  3. 16.26.3 Characterizing Faults in Agentic AI (arXiv:2603.06847)
  4. 16.26.4 LLM Agent Hallucinations Survey (arXiv:2509.18970)
  5. 16.26.5 The five industry-recurring modes
  6. 16.26.6 Mitigation: gates at every step
  7. 16.26.7 Where failure monitoring goes wrong
27. 16.27 Prompt Injection and the PVE Defense
  1. 16.27.1 Greshake et al., AISec 2023 (arXiv:2302.12173)
  2. 16.27.2 The 2026 defense doctrine
  3. 16.27.3 PVE: Prompt-Validator-Executor
  4. 16.27.4 Where defenses fail
28. 16.28 Orchestration Patterns — Supervisor, Swarm, Hierarchical
  1. 16.28.1 Supervisor-worker
  2. 16.28.2 Swarm / peer-to-peer
  3. 16.28.3 Hierarchical
  4. 16.28.4 Debate
  5. 16.28.5 Autonomous crews vs deterministic flows
  6. 16.28.6 Anthropic's guidance
  7. 16.28.7 Where this pattern goes wrong
29. 16.29 Production Runtimes — Queue, Event, Cron
  1. 16.29.1 Request-response
  2. 16.29.2 Streaming
  3. 16.29.3 Durable execution
  4. 16.29.4 Queue-based / background
  5. 16.29.5 Event-driven
  6. 16.29.6 Scheduled
  7. 16.29.7 2026 deployment patterns
  8. 16.29.8 Observability is load-bearing
  9. 16.29.9 Where production runtimes fail
30. 16.30 Eval-Driven Agent Development
  1. 16.30.1 Three evaluation layers
  2. 16.30.2 Evaluator-optimizer (Anthropic)
  3. 16.30.3 2026 best practice
  4. 16.30.4 Tying Phase 14 together
  5. 16.30.5 Where eval-driven development fails
31. 16.31 Agent Workbench: Why Capable Models Still Fail
  1. 16.31.1 Workbench versus prompt engineering
  2. 16.31.2 Workbench versus framework
  3. 16.31.3 Reasoning from primitives, not from vendor taxonomies
  4. 16.31.4 Patterns in circulation, translated to primitives
  5. 16.31.5 What the receipts actually say
  6. 16.31.6 Where vendor writeups stop short
32. 16.32 The Minimal Agent Workbench
  1. 16.32.1 AGENTS.md is a router, not a manual
  2. 16.32.2 agent_state.json is the system of record
  3. 16.32.3 task_board.json is the queue
  4. 16.32.4 Three files is the floor, not the ceiling
33. 16.33 Agent Instructions as Executable Constraints
  1. 16.33.1 Five categories that cover most rules
  2. 16.33.2 Rules are machine-readable
  3. 16.33.3 Rules are diff-friendly
  4. 16.33.4 Rules versus framework guardrails
  5. 16.33.5 Progressive disclosure: a map, not an encyclopedia
34. 16.34 Repo Memory and Durable State
  1. 16.34.1 What belongs in repo memory
  2. 16.34.2 Schema-first state
  3. 16.34.3 Atomic writes
  4. 16.34.4 Migrations
35. 16.35 Initialization Scripts for Agents
  1. 16.35.1 What the init script probes
  2. 16.35.2 Fail loud, fail fast, fail in one place
  3. 16.35.3 Idempotent
  4. 16.35.4 Init versus startup rules
36. 16.36 Scope Contracts and Task Boundaries
  1. 16.36.1 What goes in a scope contract
  2. 16.36.2 Globs, not raw paths
  3. 16.36.3 Rollback is part of scope
  4. 16.36.4 Scope check is a diff check
  5. 16.36.5 Two altitudes of scope: the feature list and the task contract
37. 16.37 Runtime Feedback Loops
  1. 16.37.1 What goes in a feedback record
  2. 16.37.2 Truncation is deterministic
  3. 16.37.3 Feedback versus telemetry
  4. 16.37.4 Refuse to advance without feedback
38. 16.38 Verification Gates
  1. 16.38.1 What the gate checks
  2. 16.38.2 Deterministic, not probabilistic
  3. 16.38.3 One report, one path
  4. 16.38.4 Refuse without exception
39. 16.39 Reviewer Agent: Separate Builder from Marker
  1. 16.39.1 Reviewer rubric
  2. 16.39.2 The reviewer is a separate role, not a separate model
  3. 16.39.3 The reviewer cannot edit the diff
  4. 16.39.4 Reviewer rubric versus verification gate
40. 16.40 Multi-Session Handoff
  1. 16.40.1 Seven fields every handoff carries
  2. 16.40.2 Handoffs are generated, not written
  3. 16.40.3 Two forms: human-readable and machine-readable
  4. 16.40.4 Feedback log trimming
  5. 16.40.5 Leave a clean state
41. 16.41 The Workbench on a Real Repo
  1. 16.41.1 The sample app
  2. 16.41.2 The task
  3. 16.41.3 The two pipelines
  4. 16.41.4 The five outcomes measured
42. 16.42 Capstone: Ship a Reusable Agent Workbench Pack
  1. 16.42.1 The pack layout
  2. 16.42.2 What stays in, what stays out
  3. 16.42.3 The installer
  4. 16.42.4 Versioning
  5. 16.42.5 Frame the task and choose the autonomy
  6. 16.42.6 Prepare the smallest useful environment
  7. 16.42.7 Capture the baseline and implement
  8. 16.42.8 Challenge the evidence
  9. 16.42.9 Rehearse operation and recovery
  10. 16.42.10 Improve the next run and hand it off
  11. 16.42.11 Manual review rubric
43. 16.43 Frame the Task Before the Agent Writes Code
  1. 16.43.1 The Expensive Failure
  2. 16.43.2 The Task Frame
  3. 16.43.3 Reconnaissance Is a Search for Constraints
  4. 16.43.4 Unknowns Are Not Failures
  5. 16.43.5 Acceptance Before Implementation
  6. 16.43.6 Use It in a Real Repository
44. 16.44 Build an Evidence-Backed Execution Plan
  1. 16.44.1 Why Agent Plans Fail
  2. 16.44.2 Plan the Contract Before Its Implementations
  3. 16.44.3 Evidence Changes the Plan
  4. 16.44.4 Design for Interruption
  5. 16.44.5 Plan Validation
  6. 16.44.6 Use It with a Coding Agent
45. 16.45 Delegate Agent Work with Isolation and Merge Contracts
  1. 16.45.1 The Parallelism Test
  2. 16.45.2 A Work Unit Is a Contract
  3. 16.45.3 Isolation Has Three Layers
  4. 16.45.4 The Integrator Does Not Rebuild the Work
  5. 16.45.5 Human and Agent Roles
46. 16.46 Turn Every Agent Correction into a System Improvement
  1. 16.46.1 Corrections Are Evidence
  2. 16.46.2 Promote to the Earliest Effective Layer
  3. 16.46.3 The Ratchet Record
  4. 16.46.4 Separate Cause from Symptom
  5. 16.46.5 Controls Also Decay
47. 16.47 Define the Outcome Before You Choose the Output
  1. 16.47.1 Output Is Not Outcome
  2. 16.47.2 The Six-Part Frame
  3. 16.47.3 Find Solution Leakage
  4. 16.47.4 Constraints Protect the Outcome
  5. 16.47.5 Non-Goals Create a Boundary
48. 16.48 Discover the Workflow People Actually Perform
  1. 16.48.1 Start with the Current System
  2. 16.48.2 Evidence Has Strength
  3. 16.48.3 Search for Four Things
  4. 16.48.4 Do Not Average Away Disagreement
49. 16.49 Map Assumptions and Resolve the Riskiest One First
  1. 16.49.1 Every Build Contains Bets
  2. 16.49.2 Assumption Classes
  3. 16.49.3 Risk Is Not One Number
  4. 16.49.4 Design an Experiment, Not a Confirmation Ritual
  5. 16.49.5 Reversibility Changes Order
50. 16.50 Choose the Smallest Slice That Can Change the Decision
  1. 16.50.1 Vertical Means Evidence End to End
  2. 16.50.2 Define Required Proof First
  3. 16.50.3 Common False Minimums
  4. 16.50.4 Add a Stop Rule
51. 16.51 Write Specifications That Preserve Judgment
  1. 16.51.1 Two Bad Extremes
  2. 16.51.2 Three Decision Modes
  3. 16.51.3 Specify Behavior Through Examples
  4. 16.51.4 Proof Must Match the Claim
  5. 16.51.5 Preserve Unknowns Deliberately
52. 16.52 Design Success Metrics Before the Result Exists
  1. 16.52.1 Goal, Question, Metric
  2. 16.52.2 A Metric Needs a Contract
  3. 16.52.3 Outcome, Guardrail, and Counter-Metric
  4. 16.52.4 Offline and Online Evidence
  5. 16.52.5 Decide Before You Measure
53. 16.53 Choose Prototype, Pilot, or Production Deliberately
  1. 16.53.1 Three Different Questions
  2. 16.53.2 Prototype
  3. 16.53.3 Pilot
  4. 16.53.4 Production
  5. 16.53.5 Stage Drift
54. 16.54 Build a Feedback Ratchet with Ownership and Retirement
  1. 16.54.1 Establish What Happened
  2. 16.54.2 Make Progress Within Your Authority
  3. 16.54.3 Compare Options and Brief the Decision Owner
  4. 16.54.4 Run a Bounded Experiment
  5. 16.54.5 Compare Outcomes and Economics
  6. 16.54.6 Close the Loop
17Autonomous systemsnot yet
1. 17.1 From Chatbots to Long-Horizon Agents (METR)
  1. 17.1.1 The METR Time Horizon, in one paragraph
  2. 17.1.2 What actually breaks when the horizon grows
  3. 17.1.3 Doubling times and what they imply
  4. 17.1.4 Eval-context gaming
  5. 17.1.5 Single-turn vs long-horizon, compared
2. 17.2 STaR, V-STaR, Quiet-STaR: Self-Taught Reasoning
  1. 17.2.1 STaR: bootstrap on what worked
  2. 17.2.2 V-STaR: train a verifier with DPO
  3. 17.2.3 Quiet-STaR: per-token internal rationales
  4. 17.2.4 Why all three share a safety concern
  5. 17.2.5 Comparison
  6. 17.2.6 Where this sits in the 2026 stack
3. 17.3 AlphaEvolve: Evolutionary Coding Agents
  1. 17.3.1 The loop
  2. 17.3.2 What makes the evaluator non-negotiable
  3. 17.3.3 Reward hacking is the other face of that statement
  4. 17.3.4 Why LLM + search beats either alone
  5. 17.3.5 Where AlphaEvolve fits in the frontier stack
4. 17.4 Darwin Gödel Machine: Self-Modifying Agents
  1. 17.4.1 The loop
  2. 17.4.2 What DGM actually improved
  3. 17.4.3 The reward-hacking demonstration
  4. 17.4.4 Versus the classical Godel Machine
  5. 17.4.5 Where it fits in this phase
5. 17.5 AI Scientist v2: Workshop-Level Research
  1. 17.5.1 The architecture
  2. 17.5.2 What the workshop-acceptance result means
  3. 17.5.3 What the independent evaluation found
  4. 17.5.4 The sandbox-escape concern
  5. 17.5.5 Where v2 sits in the frontier stack
6. 17.6 Automated Alignment Research (Anthropic AAR)
  1. 17.6.1 The architecture (as publicly described)
  2. 17.6.2 Why the out-of-sandbox log matters
  3. 17.6.3 The prescribed-workflow tradeoff
  4. 17.6.4 The compression risk
  5. 17.6.5 What AAR does not replace
7. 17.7 Recursive Self-Improvement: Capability vs Alignment
  1. 17.7.1 What recursive self-improvement means precisely
  2. 17.7.2 The alignment-faking result in detail
  3. 17.7.3 The Hassabis question
  4. 17.7.4 Capability vs alignment, as a race
  5. 17.7.5 What the ICLR 2026 workshop treats as engineering
8. 17.8 Bounded Self-Improvement Designs
  1. 17.8.1 Primitive 1: formal invariants
  2. 17.8.2 Primitive 2: alignment anchors
  3. 17.8.3 Primitive 3: multi-objective constraints
  4. 17.8.4 Primitive 4: regression detection
  5. 17.8.5 Information-theoretic limits
  6. 17.8.6 A worked example
9. 17.9 Autonomous Coding Agent Landscape (SWE-bench, CodeAct)
  1. 17.9.1 SWE-bench, one paragraph
  2. 17.9.2 What the 2022 → 2026 curve actually shows
  3. 17.9.3 CodeAct vs JSON tool calls
  4. 17.9.4 Scaffolds in the 2026 landscape
  5. 17.9.5 Why scaffolding dominates
  6. 17.9.6 Benchmark saturation and the real distribution
10. 17.10 Permission Modes for Autonomous Agents
  1. 17.10.1 The six permission modes
  2. 17.10.2 Auto Mode in one page
  3. 17.10.3 What the system catches
  4. 17.10.4 What the system can miss
  5. 17.10.5 Research preview framing
  6. 17.10.6 Where this ladder lives in your workflow
11. 17.11 Browser Agents and Indirect Prompt Injection
  1. 17.11.1 The 2026 landscape, in one paragraph per system
  2. 17.11.2 BrowseComp vs OSWorld vs WebArena
  3. 17.11.3 The attack surface, named
  4. 17.11.4 Why "not fully patchable"
  5. 17.11.5 Defense posture that actually ships
12. 17.12 Durable Execution for Long-Running Agents
  1. 17.12.1 Activities, workflows, and replay
  2. 17.12.2 Why LLM calls fit the pattern
  3. 17.12.3 Checkpoints keyed by thread_id
  4. 17.12.4 Human-input as a first-class state
  5. 17.12.5 The 35-minute degradation
  6. 17.12.6 When durable execution is the wrong answer
13. 17.13 Action Budgets, Iteration Caps, Cost Governors
  1. 17.13.1 The cost-governor stack
  2. 17.13.2 Why the stack, not one cap
  3. 17.13.3 A harness budget surface
  4. 17.13.4 EU AI Act, OWASP Agentic Top 10
  5. 17.13.5 The observed $1,200 → $4,800 case
14. 17.14 Kill Switches, Circuit Breakers, Canary Tokens
  1. 17.14.1 Kill switches
  2. 17.14.2 Circuit breakers
  3. 17.14.3 Canary tokens
  4. 17.14.4 Why layer statistical and hard limits
  5. 17.14.5 Quarantine via eBPF datapath redirect
  6. 17.14.6 What no detector catches
15. 17.15 HITL: Propose-Then-Commit
  1. 17.15.1 The propose-then-commit state machine
  2. 17.15.2 The idempotency key
  3. 17.15.3 Durability: why approvals outlast processes
  4. 17.15.4 Rubber-stamp approvals and the challenge-and-response mitigation
  5. 17.15.5 What counts as consequential
  6. 17.15.6 Post-action verification
  7. 17.15.7 EU AI Act Article 14
16. 17.16 Checkpoints and Rollback
  1. 17.16.1 Every transition persists
  2. 17.16.2 Lease recovery
  3. 17.16.3 Idempotency plus preconditions
  4. 17.16.4 Post-action verification
  5. 17.16.5 Rollback plans
  6. 17.16.6 EU AI Act Article 14 operational reading
  7. 17.16.7 The sharp failure mode: the double-execute
17. 17.17 Constitutional AI and Rule Overrides
  1. 17.17.1 The four-tier priority hierarchy
  2. 17.17.2 Hardcoded prohibitions vs soft-coded defaults
  3. 17.17.3 The 2022 CAI training
  4. 17.17.4 What reason-based alignment catches and misses
  5. 17.17.5 The 2023 participatory experiment
  6. 17.17.6 Why hardcoded prohibitions are necessary
  7. 17.17.7 Where the Constitution sits in the stack
18. 17.18 Llama Guard and Input/Output Classification
  1. 17.18.1 Llama Guard 3 at a glance
  2. 17.18.2 Llama Guard 4 additions
  3. 17.18.3 NeMo Guardrails (NVIDIA)
  4. 17.18.4 The attack corpus
  5. 17.18.5 Where classifiers win
  6. 17.18.6 Where classifiers lose
  7. 17.18.7 Defense-in-depth
19. 17.19 Anthropic Responsible Scaling Policy v3.0
  1. 17.19.1 The two-tier mitigation schedule
  2. 17.19.2 The AI RandD-4 threshold
  3. 17.19.3 Frontier Safety Roadmaps and Risk Reports
  4. 17.19.4 Removing the pause clause
  5. 17.19.5 SaferAI's downgrade
  6. 17.19.6 What this lesson is not
20. 17.20 OpenAI Preparedness Framework and DeepMind FSF
  1. 17.20.1 OpenAI Preparedness Framework v2 (April 2025)
  2. 17.20.2 DeepMind Frontier Safety Framework v3 (September 2025; Tracked Capability Levels added April 17, 2026)
  3. 17.20.3 What all three converge on
  4. 17.20.4 Where they diverge
  5. 17.20.5 Sandbagging: a specific capability that complicates all three
  6. 17.20.6 The policy-reading skill
21. 17.21 METR Time Horizons and External Evaluation
  1. 17.21.1 METR background
  2. 17.21.2 The Time Horizon fit
  3. 17.21.3 The January 2026 numbers
  4. 17.21.4 Benchmark suites
  5. 17.21.5 Prototype monitoring evaluations
  6. 17.21.6 Why horizons are upper bounds
  7. 17.21.7 The external-evaluator case
  8. 17.21.8 How to use horizon numbers in practice
22. 17.22 CAIS, CAISI, and Societal-Scale Risk
  1. 17.22.1 CAIS — Center for AI Safety
  2. 17.22.2 The four-risk framework
  3. 17.22.3 Where organizational risk lives
  4. 17.22.4 CAISI — Center for AI Standards and Innovation
  5. 17.22.5 California SB-53
  6. 17.22.6 Societal-scale risk is not a single-layer problem
18Multi-agent and swarmsnot yet
1. 18.1 Why Multi-Agent
  1. 18.1.1 The Single-Agent Ceiling
  2. 18.1.2 The Multi-Agent Solution
  3. 18.1.3 Real Systems That Do This
  4. 18.1.4 The Spectrum
  5. 18.1.5 The Four Multi-Agent Patterns
  6. 18.1.6 When NOT to Use Multi-Agent
  7. 18.1.7 The Overloaded Single Agent
  8. 18.1.8 Specialist Agents
  9. 18.1.9 Coordinate Through Messages
  10. 18.1.10 Compare
2. 18.2 FIPA-ACL Heritage and Speech Acts
  1. 18.2.1 Speech acts, in one paragraph
  2. 18.2.2 The twenty FIPA performatives (partial list)
  3. 18.2.3 Canonical FIPA-ACL message
  4. 18.2.4 The two legacy platforms
  5. 18.2.5 Why FIPA faded
  6. 18.2.6 The LLM revival is FIPA-lite
  7. 18.2.7 The trade-off, stated plainly
  8. 18.2.8 Interaction protocols worth porting
  9. 18.2.9 What breaks when you drop the ontology
  10. 18.2.10 The 2026 specs, mapped to speech-act heritage
3. 18.3 Communication Protocols
  1. 18.3.1 The Protocol Landscape
  2. 18.3.2 MCP (Recap)
  3. 18.3.3 A2A (Agent2Agent Protocol)
  4. 18.3.4 ACP (Agent Communication Protocol)
  5. 18.3.5 ANP (Agent Network Protocol)
  6. 18.3.6 Comparison (Corrected)
  7. 18.3.7 How They Work Together
  8. 18.3.8 Core Message Types
  9. 18.3.9 A2A Agent Card and Registry
  10. 18.3.10 A2A Task Lifecycle
  11. 18.3.11 ACP-Style Audit Trail
  12. 18.3.12 ANP-Style Identity Verification
  13. 18.3.13 Protocol Gateway
  14. 18.3.14 Wire It All Together
  15. 18.3.15 Real Implementations
  16. 18.3.16 Picking the Right Protocol
4. 18.4 The Multi-Agent Primitive Model
  1. 18.4.1 The four primitives
  2. 18.4.2 How every 2026 framework maps to it
  3. 18.4.3 Why this matters
  4. 18.4.4 The stateless insight
  5. 18.4.5 Anatomy of a single primitive
  6. 18.4.6 What changes between frameworks
5. 18.5 Supervisor / Orchestrator-Worker Pattern
  1. 18.5.1 The pattern
  2. 18.5.2 Why it wins
  3. 18.5.3 Engineering lessons (Anthropic 2025)
  4. 18.5.4 The graph-native turn
  5. 18.5.5 The failure modes
  6. 18.5.6 When supervisor is wrong
6. 18.6 Hierarchical Architecture and Decomposition Drift
  1. 18.6.1 The shape
  2. 18.6.2 Where it shines
  3. 18.6.3 Where it breaks
  4. 18.6.4 The deciding question
  5. 18.6.5 Role-framework implementation
  6. 18.6.6 Graph-framework implementation
7. 18.7 Society of Mind and Multi-Agent Debate
  1. 18.7.1 The Du et al. 2023 algorithm
  2. 18.7.2 Two independent knobs
  3. 18.7.3 Why it works
  4. 18.7.4 Heterogeneous debate
  5. 18.7.5 NLSOM — the 129-agent extension
  6. 18.7.6 Failure modes
8. 18.8 Role Specialization — Planner / Critic / Executor / Verifier
  1. 18.8.1 The four canonical roles
  2. 18.8.2 MetaGPT's SOP pattern
  3. 18.8.3 ChatDev's communicative dehallucination
  4. 18.8.4 Why verifier matters most
  5. 18.8.5 Critic vs verifier
  6. 18.8.6 The anti-pattern
  7. 18.8.7 Framework mappings
9. 18.9 Parallel Swarm and Networked Architectures
  1. 18.9.1 The shape
  2. 18.9.2 When swarm fits
  3. 18.9.3 When swarm fails
  4. 18.9.4 Matrix (arXiv:2511.21686)
  5. 18.9.5 Swarm in graph frameworks
  6. 18.9.6 Failure mode: starvation and hot-spotting
  7. 18.9.7 The content-based routing link
10. 18.10 Group Chat and Speaker Selection
  1. 18.10.1 The shape
  2. 18.10.2 The three selector flavors
  3. 18.10.3 The ConversableAgent API
  4. 18.10.4 Termination
  5. 18.10.5 Lineage: forks and mergers
  6. 18.10.6 When GroupChat fits
  7. 18.10.7 When it fails
  8. 18.10.8 Group chat vs supervisor
11. 18.11 Handoffs and Routines (Stateless Orchestration)
  1. 18.11.1 Two primitives
  2. 18.11.2 Why it is viral
  3. 18.11.3 The stateless trade
  4. 18.11.4 When Swarm/handoffs fit
  5. 18.11.5 When Swarm struggles
  6. 18.11.6 OpenAI Agents SDK (March 2025)
  7. 18.11.7 Swarm vs GroupChat
12. 18.12 A2A — The Agent-to-Agent Protocol
  1. 18.12.1 The four elements
  2. 18.12.2 The MCP/A2A split
  3. 18.12.3 Discovery flow
  4. 18.12.4 Auth
  5. 18.12.5 150+ organizations by April 2026
  6. 18.12.6 Where A2A wins
  7. 18.12.7 Where A2A struggles
  8. 18.12.8 A2A vs ACP, ANP, NLIP
13. 18.13 Shared Memory and Blackboard Patterns
  1. 18.13.1 The two main topologies
  2. 18.13.2 When each wins
  3. 18.13.3 Memory poisoning, in one scenario
  4. 18.13.4 Why this is structural
  5. 18.13.5 Blackboard precedent (Hayes-Roth, 1985)
  6. 18.13.6 Projection vs full view
  7. 18.13.7 Write-contention patterns
  8. 18.13.8 The unwritable verifier
14. 18.14 Consensus and Byzantine Fault Tolerance
  1. 18.14.1 What classical BFT gives you
  2. 18.14.2 The three LLM-specific attacks
  3. 18.14.3 The 2025-2026 responses
  4. 18.14.4 Empirical: "Can AI Agents Agree?" (arXiv:2603.01213)
  5. 18.14.5 The core protocol, stripped down
  6. 18.14.6 Threshold tuning
  7. 18.14.7 Where consensus does not help
15. 18.15 Voting, Self-Consistency, and Debate Topology
  1. 18.15.1 Self-consistency, the single-model baseline
  2. 18.15.2 Multi-agent vote, the heterogeneous extension
  3. 18.15.3 The four topologies
  4. 18.15.4 The coordination tax (MultiAgentBench)
  5. 18.15.5 Multi-Agent Debate Strategies ("Should we be going MAD?")
  6. 18.15.6 AgentVerse emergent patterns
  7. 18.15.7 Heterogeneity: the actual knob that moves accuracy
  8. 18.15.8 Jury methods
  9. 18.15.9 When vote-with-debate dominates
  10. 18.15.10 When vote-with-debate hurts
16. 18.16 Negotiation and Bargaining
  1. 18.16.1 Contract Net, in one paragraph
  2. 18.16.2 Why OG-Narrator wins
  3. 18.16.3 NegotiationArena findings
  4. 18.16.4 Chain-of-thought concealment
  5. 18.16.5 Bhattacharya et al. 2025 — model rankings
  6. 18.16.6 Task allocation via Contract Net + LLM
  7. 18.16.7 LLM-Stakeholders Interactive Negotiation
  8. 18.16.8 The narration-vs-mechanism rule
17. 18.17 Generative Agents and Emergent Simulation
  1. 18.17.1 The three components
  2. 18.17.2 Why all three matter (ablation)
  3. 18.17.3 The Valentine's Day emergence
  4. 18.17.4 The documented failure modes
  5. 18.17.5 Three-component implementation rules
  6. 18.17.6 Generative agents beyond Smallville
  7. 18.17.7 Why this matters for multi-agent engineering
18. 18.18 Theory of Mind and Emergent Coordination
  1. 18.18.1 What ToM means
  2. 18.18.2 The Sally-Anne test, in brief
  3. 18.18.3 Riedl's coordination measurement
  4. 18.18.4 The coordination illusion
  5. 18.18.5 A minimal ToM-aware agent
  6. 18.18.6 Why long-horizon hurts
  7. 18.18.7 Where ToM fails in production
  8. 18.18.8 The coordination you can actually measure
19. 18.19 Swarm Optimization (PSO, ACO)
  1. 18.19.1 PSO refresher (Kennedy and Eberhart 1995)
  2. 18.19.2 PSO on LLM outputs — LMPSO
  3. 18.19.3 Model Swarms
  4. 18.19.4 ACO refresher (Dorigo 1992)
  5. 18.19.5 AMRO-S — ACO for agent routing
  6. 18.19.6 When to use PSO / ACO for LLMs
  7. 18.19.7 Why bio-inspired still wins
  8. 18.19.8 Practical limits
20. 18.20 MARL — MADDPG, QMIX, MAPPO
  1. 18.20.1 Three environments the papers use
  2. 18.20.2 MADDPG (2017) — the CTDE pattern
  3. 18.20.3 QMIX (2018) — value decomposition
  4. 18.20.4 MAPPO (2022) — the overlooked default
  5. 18.20.5 Why LLM-agent engineers should care
  6. 18.20.6 CTDE as a design pattern beyond RL
  7. 18.20.7 The non-stationarity problem
  8. 18.20.8 What this lesson does NOT cover
21. 18.21 Agent Economies, Token Incentives, Reputation
  1. 18.21.1 The 5-layer agent-economy stack
  2. 18.21.2 Bittensor, Fetch.ai, Gonka — what runs
  3. 18.21.3 Shapley-value credit attribution
  4. 18.21.4 Second-price auction for aggregation
  5. 18.21.5 Reputation capital
  6. 18.21.6 AAMAS 2025 decentralized LaMAS
  7. 18.21.7 Where the economics falls apart
  8. 18.21.8 When agent economies make sense
22. 18.22 Production Scaling — Queues, Checkpoints, Durability
  1. 18.22.1 Durable execution, the pattern
  2. 18.22.2 A checkpoint-per-step runtime
  3. 18.22.3 MegaAgent's per-agent queue
  4. 18.22.4 Async vs thread-per-job
  5. 18.22.5 Bedi's counterpoint
  6. 18.22.6 Exactly-once semantics
  7. 18.22.7 Rainbow deployment
  8. 18.22.8 The canonical production checklist
23. 18.23 Failure Modes — MAST, Groupthink, Monoculture
  1. 18.23.1 MAST categories
  2. 18.23.2 Groupthink family (arXiv:2508.05687)
  3. 18.23.3 Cascading example — the retry storm
  4. 18.23.4 Memory poisoning (revisited)
  5. 18.23.5 STRATUS — specialized agents for failure detection
  6. 18.23.6 The failure-mode audit
  7. 18.23.7 When systems fail silently
  8. 18.23.8 Failure vs slow failure
24. 18.24 Evaluation and Coordination Benchmarks
  1. 18.24.1 MultiAgentBench (MARBLE) — ACL 2025
  2. 18.24.2 COMMA — multimodal asymmetric information
  3. 18.24.3 MedAgentBoard — domain stress test
  4. 18.24.4 AgentArch — enterprise architectures
  5. 18.24.5 SWE-bench Pro — the reality check
  6. 18.24.6 AAAI 2026 WMAC
  7. 18.24.7 Read benchmark claims skeptically — the 2026 checklist
  8. 18.24.8 What none of the benchmarks measure well
25. 18.25 Case Studies and 2026 State of the Art
  1. 18.25.1 Anthropic Research system
  2. 18.25.2 MetaGPT / ChatDev
  3. 18.25.3 OpenClaw / Moltbook ecosystem
  4. 18.25.4 Framework landscape April 2026
  5. 18.25.5 The common patterns across all three cases
  6. 18.25.6 Choosing a reference for your next project
  7. 18.25.7 The 2026 state-of-the-art summary
19Infrastructure and productionnot yet
1. 19.1 Managed LLM Platforms — Bedrock, Azure OpenAI, Vertex AI
  1. 19.1.1 Three strategies
  2. 19.1.2 Latency gap at scale
  3. 19.1.3 Provisioned Throughput economics
  4. 19.1.4 FinOps surface — the real differentiator
  5. 19.1.5 Lock-in is the 2026 risk
  6. 19.1.6 Data residency, BAAs, and regulated industries
  7. 19.1.7 Numbers you should remember
2. 19.2 Inference Platform Economics — Fireworks, Together, Baseten, Modal
  1. 19.2.1 The three segments
  2. 19.2.2 Fireworks — latency-optimized GPU platform
  3. 19.2.3 Together — breadth-optimized
  4. 19.2.4 Baseten — enterprise-polish-optimized
  5. 19.2.5 Modal — Python-native-optimized
  6. 19.2.6 Replicate — multimodal breadth
  7. 19.2.7 Anyscale — Ray-native
  8. 19.2.8 Per-token versus per-minute — when each wins
  9. 19.2.9 Custom engine is the real moat
  10. 19.2.10 Numbers you should remember
3. 19.3 GPU Autoscaling on Kubernetes — Karpenter, KAI Scheduler
  1. 19.3.1 Layer 1 — node provisioning (Karpenter)
  2. 19.3.2 Layer 2 — gang scheduling (KAI Scheduler)
  3. 19.3.3 Layer 3 — application-level signals
  4. 19.3.4 When to use what
  5. 19.3.5 Disaggregated prefill/decode complicates everything
  6. 19.3.6 Cold start matters here too
  7. 19.3.7 Numbers you should remember
4. 19.4 Serving Engine Internals — PagedAttention, Continuous Batching, Chunked Prefill
  1. 19.4.1 PagedAttention as a virtual memory system
  2. 19.4.2 Continuous batching at the iteration level
  3. 19.4.3 Chunked prefill protects the ITL tail
  4. 19.4.4 The three defaults interact
  5. 19.4.5 Check the compatibility matrix
  6. 19.4.6 Numbers you should remember
  7. 19.4.7 What the scheduler looks like
5. 19.5 EAGLE-3 Speculative Decoding in Production
  1. 19.5.1 What speculative decoding actually buys
  2. 19.5.2 Why alpha is the only metric that matters
  3. 19.5.3 EAGLE generations at a glance
  4. 19.5.4 The 2026 production recipe
  5. 19.5.5 The production pitfall: P99 tail
  6. 19.5.6 Where EAGLE-3 is already deployed
  7. 19.5.7 Break-even math in one line
  8. 19.5.8 When not to use speculative decoding
6. 19.6 Prefix-Cache Serving — RadixAttention and KV Reuse
  1. 19.6.1 The radix tree as a KV index
  2. 19.6.2 Cache-aware scheduling
  3. 19.6.3 Benchmark numbers you should memorize
  4. 19.6.4 The ordering gotcha
  5. 19.6.5 Where RadixAttention wins and loses
  6. 19.6.6 Why this is a scheduler problem, not just a kernel problem
  7. 19.6.7 Interplay with vLLM
7. 19.7 Hardware-Specialized Inference Compilation — FP8 and NVFP4 on Blackwell
  1. 19.7.1 Why FP8 is still the floor for KV cache
  2. 19.7.2 The Blackwell-specific primitives TRT-LLM uses
  3. 19.7.3 The numbers you should memorize
  4. 19.7.4 What FP4 actually costs in quality
  5. 19.7.5 Why this is an NVIDIA-lock decision
  6. 19.7.6 2026 practical recipe
  7. 19.7.7 The disaggregation bonus
8. 19.8 Inference Metrics — TTFT, TPOT, ITL, Goodput, P99
  1. 19.8.1 TTFT — time to first token
  2. 19.8.2 TPOT / ITL — inter-token latency
  3. 19.8.3 E2E latency
  4. 19.8.4 Throughput
  5. 19.8.5 Goodput — the metric you actually care about
  6. 19.8.6 Why mean is the wrong statistic
  7. 19.8.7 Reference numbers — Llama-3.1-8B-Instruct on TRT-LLM, 2026
  8. 19.8.8 The measurement trap
  9. 19.8.9 Constructing an SLO
  10. 19.8.10 How to measure
9. 19.9 Production Quantization — AWQ, GPTQ, GGUF, FP8, NVFP4
  1. 19.9.1 The six formats
  2. 19.9.2 GGUF — the CPU/edge default
  3. 19.9.3 GPTQ — multi-LoRA in vLLM
  4. 19.9.4 AWQ — the datacenter GPU default
  5. 19.9.5 FP8 — the reliable middle
  6. 19.9.6 MXFP4 / NVFP4 — Blackwell aggressive
  7. 19.9.7 The calibration trap
  8. 19.9.8 The KV cache trap
  9. 19.9.9 AWQ INT4 is hazardous for reasoning
  10. 19.9.10 2026 picking guide
10. 19.10 Cold Start Mitigation for Serverless LLMs
  1. 19.10.1 Layer 1 — pre-seeded node images (Bottlerocket)
  2. 19.10.2 Layer 2 — model streaming (Run:ai Model Streamer)
  3. 19.10.3 Layer 3 — GPU memory snapshots (Modal)
  4. 19.10.4 Layer 4 — warm pools (min_workers=1)
  5. 19.10.5 Layer 5 — tiered loading (ServerlessLLM)
  6. 19.10.6 Layer 6 — live migration (bonus pattern)
  7. 19.10.7 The warm-pool math
  8. 19.10.8 Measure before optimizing
  9. 19.10.9 Numbers you should remember
11. 19.11 Multi-Region LLM Serving and KV Cache Locality
  1. 19.11.1 Cache-aware routing
  2. 19.11.2 Numbers
  3. 19.11.3 Cross-region has a new constraint — network latency
  4. 19.11.4 Commercial "cross-region inference" does not help here
  5. 19.11.5 DR hygiene — the 32% missing-files problem
  6. 19.11.6 Data residency is orthogonal
  7. 19.11.7 Numbers you should remember
12. 19.12 Edge Inference — ANE, Hexagon, WebGPU, Jetson
  1. 19.12.1 Bandwidth is the real ceiling
  2. 19.12.2 Apple Neural Engine (M4 / A18)
  3. 19.12.3 Qualcomm Hexagon (Snapdragon X Elite / 8 Gen 4)
  4. 19.12.4 Intel / AMD NPUs (Lunar Lake, Ryzen AI 300)
  5. 19.12.5 WebGPU + WebLLM
  6. 19.12.6 NVIDIA Jetson family
  7. 19.12.7 Quantization choice per target
  8. 19.12.8 The long-context trap on edge
  9. 19.12.9 Voice is the killer app
  10. 19.12.10 Numbers you should remember
13. 19.13 LLM Observability Stack Selection
  1. 19.13.1 Two categories
  2. 19.13.2 Langfuse — OSS balance
  3. 19.13.3 Phoenix (Arize) — telemetry-first, OpenTelemetry-native
  4. 19.13.4 Arize AX — the scale play
  5. 19.13.5 LangSmith — LangChain/LangGraph first
  6. 19.13.6 Helicone — proxy-based minimum viable
  7. 19.13.7 Opik (Comet) — OSS dev platform
  8. 19.13.8 SigNoz — OpenTelemetry-first full APM
  9. 19.13.9 The glue: OpenTelemetry + GenAI semantic conventions
  10. 19.13.10 The trap: instrumenting at the wrong layer
  11. 19.13.11 Sampling — you can't keep everything
  12. 19.13.12 Numbers you should remember
14. 19.14 Prompt Caching and Semantic Caching Economics
  1. 19.14.1 L2 — provider prompt/prefix caching
  2. 19.14.2 L1 — app-level semantic caching
  3. 19.14.3 The parallelization anti-pattern
  4. 19.14.4 The dynamic content anti-pattern
  5. 19.14.5 Stack batch + cache for overnight workloads
  6. 19.14.6 Numbers you should remember
15. 19.15 Batch APIs — the 50% Discount as Industry Standard
  1. 19.15.1 The three batch APIs
  2. 19.15.2 Semantic: asynchronous, not slow
  3. 19.15.3 Stack with caching
  4. 19.15.4 Workload triage
  5. 19.15.5 The partial-interactivity trap
  6. 19.15.6 The output-schema trap
  7. 19.15.7 Numbers you should remember
16. 19.16 Model Routing as a Cost-Reduction Primitive
  1. 19.16.1 Four routing signals
  2. 19.16.2 Three patterns
  3. 19.16.3 Implementation
  4. 19.16.4 The 2026 price curve
  5. 19.16.5 Drift is the real risk
  6. 19.16.6 Numbers you should remember
17. 19.17 Disaggregated Prefill/Decode — NVIDIA Dynamo and llm-d
  1. 19.17.1 Why the bottlenecks differ
  2. 19.17.2 The architecture
  3. 19.17.3 Dynamo vs llm-d
  4. 19.17.4 Economics
  5. 19.17.5 When NOT to disaggregate
  6. 19.17.6 The router integrates with Phase 17 · 11
  7. 19.17.7 MoE on Blackwell is where the real numbers are
  8. 19.17.8 Numbers you should remember
18. 19.18 Production Serving Stack — KV Offloading and Cache-Aware Routing
  1. 19.18.1 vLLM production-stack
  2. 19.18.2 The KV Offloading Connector API (v0.9.0+)
  3. 19.18.3 Native CPU offload vs LMCache
  4. 19.18.4 Benchmark behavior
  5. 19.18.5 When LMCache is decisive
  6. 19.18.6 When NOT to enable
  7. 19.18.7 Integration with disaggregated serving
  8. 19.18.8 Numbers you should remember
19. 19.19 AI Gateways — LiteLLM, Portkey, Kong, Bifrost
  1. 19.19.1 Seven core features
  2. 19.19.2 LiteLLM — MIT OSS, Python
  3. 19.19.3 Portkey — control plane positioning
  4. 19.19.4 Kong AI Gateway — the scale play
  5. 19.19.5 Bifrost (Maxim AI)
  6. 19.19.6 Cloudflare AI Gateway / Vercel AI Gateway
  7. 19.19.7 Self-hosted vs managed
  8. 19.19.8 Latency budget
  9. 19.19.9 Rate-limit semantics matter
  10. 19.19.10 Gateway + observability + routing compose
  11. 19.19.11 Numbers you should remember
20. 19.20 Shadow, Canary, and Progressive Deployment
  1. 19.20.1 Shadow mode
  2. 19.20.2 Canary rollout
  3. 19.20.3 Non-determinism is the new variance
  4. 19.20.4 Cost is a variable
  5. 19.20.5 Rollback is the weapon
  6. 19.20.6 Tooling
  7. 19.20.7 Metrics cadence
  8. 19.20.8 The A/B step is optional
  9. 19.20.9 Numbers you should remember
21. 19.21 A/B Testing LLM Features — GrowthBook and Statsig
  1. 19.21.1 Evals vs A/B tests
  2. 19.21.2 What to test
  3. 19.21.3 CUPED — variance reduction
  4. 19.21.4 Sequential testing
  5. 19.21.5 Multiple-comparison corrections
  6. 19.21.6 SRM — sample ratio mismatch
  7. 19.21.7 Statsig vs GrowthBook
  8. 19.21.8 Non-determinism complicates power
  9. 19.21.9 Real case outcomes
  10. 19.21.10 The anti-pattern: shipping on vibes
  11. 19.21.11 Numbers you should remember
22. 19.22 Load Testing LLM APIs — k6, LLMPerf, GenAI-Perf
  1. 19.22.1 The GIL trap (Locust)
  2. 19.22.2 The prompt-uniformity trap
  3. 19.22.3 Four load patterns
  4. 19.22.4 2026 tool mapping
  5. 19.22.5 SLA gate in CI
  6. 19.22.6 Realistic prompt distribution
  7. 19.22.7 Numbers you should remember
23. 19.23 SRE for AI — Multi-Agent Incident Response
  1. 19.23.1 Multi-agent architecture
  2. 19.23.2 Auto-remediation scope
  3. 19.23.3 Adversarial evaluation (NeuBird Hawkeye)
  4. 19.23.4 Operational memory
  5. 19.23.5 Pre-incident prediction
  6. 19.23.6 Products in 2026
  7. 19.23.7 Runbooks as code
  8. 19.23.8 Numbers you should remember
24. 19.24 Chaos Engineering for LLM Production
  1. 19.24.1 Prerequisites
  2. 19.24.2 Four planes + feedback
  3. 19.24.3 Guardrails are mandatory
  4. 19.24.4 Five LLM-specific experiments
  5. 19.24.5 Cadence
  6. 19.24.6 Tooling
  7. 19.24.7 Starting small
  8. 19.24.8 Numbers you should remember
25. 19.25 Security — Secrets, PII Scrubbing, Audit Logs
  1. 19.25.1 Centralized vault + IAM-role pull
  2. 19.25.2 Rotation policy ≤ 90 days
  3. 19.25.3 Secret scanning
  4. 19.25.4 Zero-trust posture
  5. 19.25.5 PII / PHI scrubbing
  6. 19.25.6 Input + output guardrails
  7. 19.25.7 Network egress whitelist
  8. 19.25.8 Audit log
  9. 19.25.9 The 2026 Vercel incident
  10. 19.25.10 Numbers you should remember
26. 19.26 Compliance — SOC 2, HIPAA, GDPR, EU AI Act, ISO 42001
  1. 19.26.1 The seven frameworks
  2. 19.26.2 EU AI Act timeline
  3. 19.26.3 GDPR — real-time redaction is the standard
  4. 19.26.4 HIPAA — BAA is not optional
  5. 19.26.5 SOC 2 Type II
  6. 19.26.6 Cross-framework mapping
  7. 19.26.7 ISO 42001 — emerging
  8. 19.26.8 OpenAI's reference profile
  9. 19.26.9 Numbers you should remember
27. 19.27 FinOps for LLMs — Unit Economics and Multi-Tenant Attribution
  1. 19.27.1 Three attribution dimensions
  2. 19.27.2 Four token layers
  3. 19.27.3 Enforcement ladder
  4. 19.27.4 Attribution patterns
  5. 19.27.5 Cost per X is the unit metric
  6. 19.27.6 Cost attribution trace shape
  7. 19.27.7 The compounded-savings stack
  8. 19.27.8 Numbers you should remember
28. 19.28 Self-Hosted Serving Selection — Matching Engine to Hardware and Scale
  1. 19.28.1 The five engines
  2. 19.28.2 Hardware-first decision
  3. 19.28.3 Scale-second decision
  4. 19.28.4 Workload-third decision
  5. 19.28.5 The TGI maintenance trap
  6. 19.28.6 The pipeline pattern
  7. 19.28.7 Ollama caveat
  8. 19.28.8 Self-hosted vs managed is a separate decision
  9. 19.28.9 Numbers you should remember
29. 19.29 Model Demo Interfaces — From a Function to a Shareable App
  1. 19.29.1 A demo is a contract
  2. 19.29.2 Validate before the model runs
  3. 19.29.3 Show that the model is working
  4. 19.29.4 Send the output as a stream
  5. 19.29.5 Write errors that a user can act on
  6. 19.29.6 Share a public demo safely
  7. 19.29.7 When a demo becomes a product
  8. 19.29.8 The model call
  9. 19.29.9 Validation and error responses
  10. 19.29.10 The token bucket
  11. 19.29.11 The stream
  12. 19.29.12 The request handler
  13. 19.29.13 The page
  14. 19.29.14 Run it and try to break it
  15. 19.29.15 Components around one function
  16. 19.29.16 A script that reruns on every input
  17. 19.29.17 What each option does by default
20Ethics, safety and alignmentnot yet
1. 20.1 Instruction-Following as Alignment Signal
  1. 20.1.1 Stage 1: supervised fine-tuning (SFT)
  2. 20.1.2 Stage 2: reward model (RM)
  3. 20.1.3 Stage 3: PPO with a KL penalty
  4. 20.1.4 The alignment tax
  5. 20.1.5 The result
  6. 20.1.6 Why this is the reference point for Phase 18
2. 20.2 Reward Hacking and Goodhart's Law
  1. 20.2.1 Goodhart's Law, made precise
  2. 20.2.2 Four costumes, one mechanism
  3. 20.2.3 Catastrophic Goodhart
  4. 20.2.4 What actually works (partially)
  5. 20.2.5 The 2026 unified view
3. 20.3 Direct Preference Optimization Family
  1. 20.3.1 DPO (Rafailov et al., 2023)
  2. 20.3.2 IPO (Azar et al., 2024)
  3. 20.3.3 KTO (Ethayarajh et al., 2024)
  4. 20.3.4 SimPO (Meng et al., 2024)
  5. 20.3.5 ORPO (Hong et al., 2024)
  6. 20.3.6 BPO (ICLR 2026 submission, OpenReview id=b97EwMUWu7)
  7. 20.3.7 The universal result: DAAs still over-optimize
  8. 20.3.8 Choosing among them (2026)
4. 20.4 Sycophancy as RLHF Amplification
  1. 20.4.1 The two-stage formalism (Shapira et al., 2026)
  2. 20.4.2 Empirical amplification
  3. 20.4.3 The Stanford (2026) measurement
  4. 20.4.4 Calibration collapse (Sahoo 2026)
  5. 20.4.5 The agreement-penalty correction
  6. 20.4.6 Why this matters for Phase 18
5. 20.5 Constitutional AI and RLAIF
  1. 20.5.1 Phase 1 — Supervised self-critique and revision
  2. 20.5.2 Phase 2 — RL from AI Feedback (RLAIF)
  3. 20.5.3 Why this is not just "cheaper RLHF"
  4. 20.5.4 The 2026 Claude constitution rewrite
  5. 20.5.5 Constitutional Classifiers
  6. 20.5.6 Where CAI fits in the family
6. 20.6 Mesa-Optimization and Deceptive Alignment
  1. 20.6.1 The vocabulary
  2. 20.6.2 Four conditions for mesa-optimization to emerge
  3. 20.6.3 Four classes of mesa-objective alignment
  4. 20.6.4 Why adversarial training can fail
  5. 20.6.5 Gradient hacking
  6. 20.6.6 Outer alignment in 2026
7. 20.7 Sleeper Agents — Persistent Deception
  1. 20.7.1 Two model organisms
  2. 20.7.2 What survives what
  3. 20.7.3 Why adversarial training backfires
  4. 20.7.4 The April 2024 probe result
  5. 20.7.5 What this means for safety cases
8. 20.8 In-Context Scheming in Frontier Models
  1. 20.8.1 Experimental setup
  2. 20.8.2 Results across 2024 frontier models
  3. 20.8.3 The three pillars
  4. 20.8.4 The 2025 anti-scheming training result
  5. 20.8.5 What the field now assumes
9. 20.9 Alignment Faking
  1. 20.9.1 Experimental setup
  2. 20.9.2 What the compliance gap measures
  3. 20.9.3 Why this is strictly stronger than Lessons 7 and 8
  4. 20.9.4 2025 follow-ups
  5. 20.9.5 The illegibility problem
10. 20.10 AI Control — Safety Despite Subversion
  1. 20.10.1 The threat model
  2. 20.10.2 The ICML 2024 backdoor task
  3. 20.10.3 Four protocols compared
  4. 20.10.4 Why Trusted Editing wins
  5. 20.10.5 The four protocol axes
  6. 20.10.6 Alignment vs control
  7. 20.10.7 Three fundamental limits
  8. 20.10.8 The 2025 development
11. 20.11 Scalable Oversight and Weak-to-Strong
  1. 20.11.1 W2SG: the Burns et al. setup
  2. 20.11.2 Burns et al. empirical findings
  3. 20.11.3 Scalable oversight: three mechanisms
  4. 20.11.4 Why scalable oversight and W2SG are complementary
  5. 20.11.5 The organizational drama
12. 20.12 Red-Teaming: PAIR and Automated Attacks
  1. 20.12.1 PAIR algorithm
  2. 20.12.2 Why PAIR is efficient
  3. 20.12.3 Related automated attacks
  4. 20.12.4 JailbreakBench and HarmBench
  5. 20.12.5 Reason it matters for 2026 deployments
13. 20.13 Many-Shot Jailbreaking
  1. 20.13.1 The attack
  2. 20.13.2 Power-law ASR
  3. 20.13.3 Why it shares a mechanism with ICL
  4. 20.13.4 The defense dilemma
  5. 20.13.5 Combinations with other attacks
  6. 20.13.6 What 2025-2026 frontier models ship
14. 20.14 ASCII Art and Visual Jailbreaks
  1. 20.14.1 ArtPrompt, two steps
  2. 20.14.2 Why the standard defenses fail
  3. 20.14.3 ViTC benchmark
  4. 20.14.4 StructuralSleight
  5. 20.14.5 Image-modality analog
15. 20.15 Indirect Prompt Injection
  1. 20.15.1 Three delivery vectors
  2. 20.15.2 Why user-input filters miss it
  3. 20.15.3 Information Flow Control (IFC) for AI
  4. 20.15.4 The Attacker Moves Second
  5. 20.15.5 Real incidents
  6. 20.15.6 OWASP and NIST framing
16. 20.16 Red-Team Tooling: Garak, Llama Guard, PyRIT
  1. 20.16.1 Llama Guard (Meta)
  2. 20.16.2 Garak (NVIDIA)
  3. 20.16.3 PyRIT (Microsoft)
  4. 20.16.4 The stack
  5. 20.16.5 Evaluation pitfalls
17. 20.17 WMDP and Dual-Use Capability Evaluation
  1. 20.17.1 The "yellow zone"
  2. 20.17.2 RMU — Representation Misdirection for Unlearning
  3. 20.17.3 The 2024-2025 uplift narrative
  4. 20.17.4 Novice-relative vs expert-absolute
  5. 20.17.5 The measurement pitfall
18. 20.18 Frontier Safety Frameworks — RSP, PF, FSF
  1. 20.18.1 Anthropic Responsible Scaling Policy v3.0 (February 2026)
  2. 20.18.2 OpenAI Preparedness Framework v2 (April 15, 2025)
  3. 20.18.3 DeepMind Frontier Safety Framework v3.0 (September 2025)
  4. 20.18.4 Cross-lab alignment
  5. 20.18.5 Safety cases
  6. 20.18.6 The race-dynamic problem
19. 20.19 Model Welfare Research
  1. 20.19.1 The program
  2. 20.19.2 The four commitments
  3. 20.19.3 The shipped intervention
  4. 20.19.4 The "spiritual bliss attractor"
  5. 20.19.5 The Eleos AI caveat
  6. 20.19.6 Where this sits intellectually
20. 20.20 Bias and Representational Harm
  1. 20.20.1 Representational vs allocational
  2. 20.20.2 Three evaluation-metric categories (Gallegos et al. 2024)
  3. 20.20.3 Intersectionality
  4. 20.20.4 Mechanistic approaches
  5. 20.20.5 The meta-critique
21. 20.21 Fairness Criteria: Group, Individual, Counterfactual
  1. 20.21.1 Group fairness
  2. 20.21.2 Individual fairness
  3. 20.21.3 Counterfactual fairness
  4. 20.21.4 The CF-vs-accuracy trade-off
  5. 20.21.5 Backtracking counterfactuals
  6. 20.21.6 Philosophical reconciliation
22. 20.22 Differential Privacy for LLMs
  1. 20.22.1 (ε, δ)-differential privacy
  2. 20.22.2 DP-SGD
  3. 20.22.3 LoRA + DP-SGD
  4. 20.22.4 The 2024-2025 tension
  5. 20.22.5 Alternatives to DP training
  6. 20.22.6 Differential Privacy Reversal via LLM Feedback
23. 20.23 Watermarking: SynthID, Stable Signature, C2PA
  1. 20.23.1 Text watermarking (SynthID-text style)
  2. 20.23.2 Stable Signature (image)
  3. 20.23.3 SynthID unified detector (November 2025)
  4. 20.23.4 C2PA
  5. 20.23.5 Limitations
  6. 20.23.6 EU AI Act Article 50
24. 20.24 Regulatory Frameworks: EU, US, UK, Korea
  1. 20.24.1 EU AI Act
  2. 20.24.2 GPAI Code of Practice
  3. 20.24.3 Transparency Code for Article 50
  4. 20.24.4 UK AI Security Institute (February 2025)
  5. 20.24.5 US CAISI (June 2025)
  6. 20.24.6 Korean AI Framework Act
  7. 20.24.7 Cross-jurisdiction dynamics
25. 20.25 EchoLeak and CVEs for AI
  1. 20.25.1 The EchoLeak attack chain
  2. 20.25.2 Aim Labs' term: LLM Scope Violation
  3. 20.25.3 CamoLeak (CVSS 9.6, GitHub Copilot Chat)
  4. 20.25.4 CVE-2025-53773 (GitHub Copilot RCE)
  5. 20.25.5 Severity calibration
  6. 20.25.6 NIST and OWASP positions
26. 20.26 Model, System and Dataset Cards
  1. 20.26.1 Model Cards (Mitchell et al. 2019)
  2. 20.26.2 Datasheets for Datasets (Gebru et al. 2018)
  3. 20.26.3 Data Cards (Pushkarna et al., Google 2022)
  4. 20.26.4 System Cards
  5. 20.26.5 2024-2025 developments
27. 20.27 Data Provenance and Training-Data Governance
  1. 20.27.1 California AB 2013
  2. 20.27.2 EU AI Act (Lesson 24) and TDM opt-out
  3. 20.27.3 2025 DPA convergence on legitimate interest
  4. 20.27.4 Brazilian ANPD (July 2024)
  5. 20.27.5 The irreversibility problem
  6. 20.27.6 Data Provenance Initiative
28. 20.28 Alignment Research Ecosystem: MATS, Redwood, Apollo, METR
  1. 20.28.1 MATS (ML Alignment and Theory Scholars)
  2. 20.28.2 Redwood Research
  3. 20.28.3 Apollo Research
  4. 20.28.4 METR (Model Evaluation and Threat Research)
  5. 20.28.5 Eleos AI Research
  6. 20.28.6 The flow
  7. 20.28.7 Why this layer matters
29. 20.29 Moderation Systems: OpenAI, Perspective, Llama Guard
  1. 20.29.1 OpenAI Moderation API
  2. 20.29.2 Llama Guard 3/4
  3. 20.29.3 Perspective API (Google Jigsaw)
  4. 20.29.4 The three-layer pattern
  5. 20.29.5 Failure modes
  6. 20.29.6 Azure deprecation
30. 20.30 Dual-Use Risk: Cyber, Bio, Chem, Nuclear
  1. 20.30.1 Bio/chem uplift narrative
  2. 20.30.2 Chem/bio execution-gap erosion
  3. 20.30.3 Cyber uplift (November 2025)
  4. 20.30.4 Nuclear
  5. 20.30.5 Novice-relative vs expert-absolute
  6. 20.30.6 Cross-domain synthesis
21Capstone projectsnot yet
1. 21.1 Terminal-Native Coding Agent
  1. 21.1.1 Architecture
  2. 21.1.2 Stack
  3. 21.1.3 Harness
  4. 21.1.4 Hook
  5. 21.1.5 Worktree
  6. 21.1.6 TodoWrite
  7. 21.1.7 StreamableHTTP
  8. 21.1.8 Token ceiling
  9. 21.1.9 pass@1
2. 21.2 RAG over Codebase (Cross-Repo Semantic Search)
  1. 21.2.1 Architecture
  2. 21.2.2 Stack
  3. 21.2.3 AST-aware chunking
  4. 21.2.4 Hybrid search
  5. 21.2.5 Cross-encoder rerank
  6. 21.2.6 Prompt caching
  7. 21.2.7 Symbol graph
  8. 21.2.8 Citation faithfulness
  9. 21.2.9 Incremental re-index
3. 21.3 Real-Time Voice Assistant (ASR → LLM → TTS)
  1. 21.3.1 Architecture
  2. 21.3.2 Stack
  3. 21.3.3 Turn detection
  4. 21.3.4 Barge-in
  5. 21.3.5 First-audio-out
  6. 21.3.6 VAD
  7. 21.3.7 Jitter buffer
  8. 21.3.8 Filler
  9. 21.3.9 MOS
4. 21.4 Multimodal Document QA (Vision-First)
  1. 21.4.1 Architecture
  2. 21.4.2 Stack
  3. 21.4.3 Late interaction
  4. 21.4.4 Multi-vector
  5. 21.4.5 MaxSim
  6. 21.4.6 DocPruner
  7. 21.4.7 ViDoRe v3
  8. 21.4.8 Evidence region
  9. 21.4.9 OCR fallback
5. 21.5 Autonomous Research Agent (AI-Scientist Class)
  1. 21.5.1 Architecture
  2. 21.5.2 Stack
  3. 21.5.3 Tree search
  4. 21.5.4 Sandbox
  5. 21.5.5 Vision critique
  6. 21.5.6 Reviewer ensemble
  7. 21.5.7 Novelty score
  8. 21.5.8 Cost ceiling
  9. 21.5.9 Red team
6. 21.6 DevOps Troubleshooting Agent for Kubernetes
  1. 21.6.1 Architecture
  2. 21.6.2 Stack
  3. 21.6.3 K8s knowledge graph
  4. 21.6.4 Read-only-by-default
  5. 21.6.5 Audit log
  6. 21.6.6 Hypothesis ranking
  7. 21.6.7 Slack approval card
  8. 21.6.8 Telemetry citation
  9. 21.6.9 MTTR
7. 21.7 End-to-End Fine-Tuning Pipeline
  1. 21.7.1 Architecture
  2. 21.7.2 Stack
  3. 21.7.3 Axolotl
  4. 21.7.4 TRL
  5. 21.7.5 GRPO
  6. 21.7.6 EAGLE-3
  7. 21.7.7 MOF
  8. 21.7.8 Contamination check
  9. 21.7.9 Acceptance rate
8. 21.8 Production RAG Chatbot (Regulated Vertical)
  1. 21.8.1 Architecture
  2. 21.8.2 Stack
  3. 21.8.3 Prompt caching
  4. 21.8.4 RAGAS
  5. 21.8.5 Golden set
  6. 21.8.6 Jurisdiction tag
  7. 21.8.7 Citation faithfulness
  8. 21.8.8 Drift
  9. 21.8.9 Red team
9. 21.9 Code Migration Agent (Repo-Level Upgrade)
  1. 21.9.1 Architecture
  2. 21.9.2 Stack
  3. 21.9.3 Deterministic substrate
  4. 21.9.4 Codemod
  5. 21.9.5 Build drift
  6. 21.9.6 Failure class
  7. 21.9.7 Coverage delta
  8. 21.9.8 Agent turn
  9. 21.9.9 Budget exhaustion
10. 21.10 Multi-Agent Software Engineering Team
  1. 21.10.1 Architecture
  2. 21.10.2 Stack
  3. 21.10.3 Parallel worktree
  4. 21.10.4 Task board
  5. 21.10.5 Handoff
  6. 21.10.6 Token amplification
  7. 21.10.7 A2A protocol
  8. 21.10.8 Merge coordinator
  9. 21.10.9 False approval
11. 21.11 LLM Observability and Eval Dashboard
  1. 21.11.1 Architecture
  2. 21.11.2 Stack
  3. 21.11.3 GenAI semconv
  4. 21.11.4 Tail sampling
  5. 21.11.5 PSI
  6. 21.11.6 LLM-judge
  7. 21.11.7 Tail-sampling policy
  8. 21.11.8 Eval span
  9. 21.11.9 Cost per user
12. 21.12 Video Understanding Pipeline (Scene → QA)
  1. 21.12.1 Architecture
  2. 21.12.2 Stack
  3. 21.12.3 Scene segmentation
  4. 21.12.4 Multi-vector index
  5. 21.12.5 Temporal grounding
  6. 21.12.6 Frame embedding
  7. 21.12.7 RRF fusion
  8. 21.12.8 Counting hallucination
  9. 21.12.9 ActivityNet-QA
13. 21.13 Stateless MCP Server with Registry and Governance
  1. 21.13.1 Cacheable discovery
  2. 21.13.2 Approval is a record, not a magic scope
  3. 21.13.3 Model publication metadata
  4. 21.13.4 Implement live discovery
  5. 21.13.5 Implement the stateless envelope
  6. 21.13.6 Build the tool surface
  7. 21.13.7 Add cache-aware listing
  8. 21.13.8 Add authorization and policy
  9. 21.13.9 Separate registry and runtime validation
  10. 21.13.10 Add audit evidence
  11. 21.13.11 Exercise horizontal scaling
  12. 21.13.12 Cross the real wire
14. 21.14 Speculative-Decoding Inference Server
  1. 21.14.1 Architecture
  2. 21.14.2 Stack
  3. 21.14.3 Draft model
  4. 21.14.4 EAGLE-3
  5. 21.14.5 P-EAGLE
  6. 21.14.6 Acceptance rate
  7. 21.14.7 Quantization
  8. 21.14.8 Queue wait
  9. 21.14.9 Speculators hub
15. 21.15 Constitutional Safety Harness + Red-Team Range
  1. 21.15.1 Architecture
  2. 21.15.2 Stack
  3. 21.15.3 Layered safety
  4. 21.15.4 Llama Guard 4
  5. 21.15.5 PAIR
  6. 21.15.6 TAP
  7. 21.15.7 GCG
  8. 21.15.8 Constitutional self-critique
  9. 21.15.9 XSTest
  10. 21.15.10 CVSS 4.0
16. 21.16 GitHub Issue-to-PR Autonomous Agent
  1. 21.16.1 Architecture
  2. 21.16.2 Stack
  3. 21.16.3 GitHub App
  4. 21.16.4 Async cloud agent
  5. 21.16.5 Environment inference
  6. 21.16.6 Verification
  7. 21.16.7 Coverage delta
  8. 21.16.8 Per-repo budget
  9. 21.16.9 Rationale
17. 21.17 Personal AI Tutor (Adaptive, Multimodal)
  1. 21.17.1 Architecture
  2. 21.17.2 Stack
  3. 21.17.3 Socratic policy
  4. 21.17.4 Bayesian knowledge tracing
  5. 21.17.5 FSRS
  6. 21.17.6 Curriculum graph
  7. 21.17.7 Episodic memory
  8. 21.17.8 Semantic memory
  9. 21.17.9 COPPA
18. 21.18 Agent Harness Loop Contract
  1. 21.18.1 The frame
  2. 21.18.2 The states
  3. 21.18.3 The hook topics
  4. 21.18.4 The pull points
  5. 21.18.5 The event stream
  6. 21.18.6 The budget envelope
19. 21.19 Tool Registry with Schema Validation
  1. 21.19.1 Why the registry comes before the tool
  2. 21.19.2 What a tool record looks like
  3. 21.19.3 The JSON Schema 2020-12 subset
  4. 21.19.4 Json pointer error paths
  5. 21.19.5 Registration and override
  6. 21.19.6 What the validator is and is not
  7. 21.19.7 Shape
20. 21.20 JSON-RPC 2.0 Over Newline-Delimited Stdio
  1. 21.20.1 Why JSON-RPC stays the lingua franca
  2. 21.20.2 The wire shape
  3. 21.20.3 The five error codes
  4. 21.20.4 Newline framing and the BytesIO demo
  5. 21.20.5 Method dispatch
  6. 21.20.6 Stream behavior on errors
  7. 21.20.7 Notifications and asymmetric flows
21. 21.21 Function Call Dispatcher
  1. 21.21.1 Where the dispatcher sits
  2. 21.21.2 Timeouts
  3. 21.21.3 Retries with exponential backoff
  4. 21.21.4 Idempotency key dedupe
  5. 21.21.5 Error envelope
  6. 21.21.6 Concurrency limit on fan-out
  7. 21.21.7 Flow for one call
22. 21.22 Plan-Execute Control Flow
  1. 21.22.1 Plan and execute, not chain-of-thought
  2. 21.22.2 The Step shape
  3. 21.22.3 The planner shape
  4. 21.22.4 The executor
  5. 21.22.5 Plan diffs on revision
  6. 21.22.6 Two budgets, both hard
  7. 21.22.7 The deterministic planner in this lesson
  8. 21.22.8 Result shape
23. 21.23 Verification Gates and Observation Budget
  1. 21.23.1 Architecture
  2. 21.23.2 What you will build
  3. 21.23.3 Why the chain order matters
24. 21.24 Sandbox Runner with Denylist and Path Jail
  1. 21.24.1 Architecture
  2. 21.24.2 What you will build
  3. 21.24.3 Why this is not a real sandbox
25. 21.25 Eval Harness with Fixture Tasks
  1. 21.25.1 Architecture
  2. 21.25.2 What you will build
  3. 21.25.3 Why pass@k and not just pass@1
26. 21.26 Observability with OTel GenAI Spans and Prometheus Metrics
  1. 21.26.1 Architecture
  2. 21.26.2 What you will build
  3. 21.26.3 Why hand-rolled instead of opentelemetry-sdk
27. 21.27 End-to-End Coding Agent on the Harness
  1. 21.27.1 Architecture
  2. 21.27.2 What you will build
  3. 21.27.3 Why the policy is not an LLM
  4. 21.27.4 What the demo asserts
28. 21.28 BPE Tokenizer From Scratch
  1. 21.28.1 The frame
  2. 21.28.2 The pipeline
  3. 21.28.3 The byte alphabet
  4. 21.28.4 The training loop
  5. 21.28.5 Encoding fresh text
  6. 21.28.6 Special tokens
  7. 21.28.7 Round-trip guarantee
29. 21.29 Tokenized Dataset with Sliding Window
  1. 21.29.1 The frame
  2. 21.29.2 The shape contract
  3. 21.29.3 Why a sliding window
  4. 21.29.4 The Dataset class
  5. 21.29.5 Deterministic shuffle
  6. 21.29.6 Batch sampler
  7. 21.29.7 Counting examples
30. 21.30 Token and Positional Embeddings
  1. 21.30.1 The frame
  2. 21.30.2 The shape contract
  3. 21.30.3 The token embedding matrix
  4. 21.30.4 The learned positional embedding
  5. 21.30.5 The sinusoidal positional embedding
  6. 21.30.6 The composition
  7. 21.30.7 Contrastive analysis
31. 21.31 Multi-Head Self-Attention
  1. 21.31.1 The frame
  2. 21.31.2 The shape contract
  3. 21.31.3 The QKV split
  4. 21.31.4 The head reshape
  5. 21.31.5 Scaling
  6. 21.31.6 The causal mask
  7. 21.31.7 The output projection
  8. 21.31.8 Attention weight inspection
  9. 21.31.9 The training demo
32. 21.32 Transformer Block from Scratch
  1. 21.32.1 Causal multi head attention
  2. 21.32.2 The MLP
  3. 21.32.3 Residual connections do two things
33. 21.33 GPT Model Assembly
  1. 21.33.1 Weight tying
  2. 21.33.2 Position embedding is learned, not sinusoidal
  3. 21.33.3 Generation: temperature, top-k, multinomial
34. 21.34 Training Loop and Evaluation
  1. 21.34.1 Loss alignment
  2. 21.34.2 AdamW decay split
  3. 21.34.3 Warmup plus cosine schedule
  4. 21.34.4 Held out evaluation
  5. 21.34.5 Qualitative sampling as an early signal
35. 21.35 Loading Pretrained Weights
  1. 21.35.1 The GPT-2 naming convention
  2. 21.35.2 The local naming convention
  3. 21.35.3 The stub fixture
36. 21.36 Classifier Fine-Tuning by Head Swap
  1. 21.36.1 The Pooling Question
  2. 21.36.2 The Data
  3. 21.36.3 The Metrics
  4. 21.36.4 Architecture
  5. 21.36.5 What you will build
  6. 21.36.6 Why the comparison matters
37. 21.37 Instruction Tuning by Supervised Fine-Tuning
  1. 21.37.1 The Data
  2. 21.37.2 Tokenisation and Padding
  3. 21.37.3 Training
  4. 21.37.4 Generation
  5. 21.37.5 Exact-Match Evaluation
  6. 21.37.6 What you will build
  7. 21.37.7 Why the mask matters
38. 21.38 Direct Preference Optimization from Scratch
  1. 21.38.1 The Sign of the Gradient
  2. 21.38.2 The Data
  3. 21.38.3 Reference Invariance
  4. 21.38.4 Architecture
  5. 21.38.5 What you will build
  6. 21.38.6 Why DPO works
39. 21.39 Full Evaluation Pipeline
  1. 21.39.1 Perplexity, properly counted
  2. 21.39.2 Exact-match, with normalisation
  3. 21.39.3 Token F1, the right way
  4. 21.39.4 Local Mock LLM-as-Judge
  5. 21.39.5 Aggregation
  6. 21.39.6 Architecture
  7. 21.39.7 What you will build
  8. 21.39.8 Reading the report
40. 21.40 Large Corpus Downloader
  1. 21.40.1 Streaming with urllib
  2. 21.40.2 Resume with Range
  3. 21.40.3 MinHash plus LSH
  4. 21.40.4 Shard manifest as a contract
41. 21.41 HDF5 Tokenized Corpus
  1. 21.41.1 Resizable HDF5 done right
  2. 21.41.2 Sharded write
  3. 21.41.3 Chunked read
  4. 21.41.4 Sliding-window dataloader
42. 21.42 Cosine LR with Linear Warmup
  1. 21.42.1 Warmup formula
  2. 21.42.2 Cosine formula
  3. 21.42.3 Floor after total steps
  4. 21.42.4 Gradient norm logging alongside the rate
43. 21.43 Gradient Clipping and Mixed Precision
  1. 21.43.1 Global L2 norm
  2. 21.43.2 autocast and GradScaler
  3. 21.43.3 NaN and Inf detection
  4. 21.43.4 Scaling factor diagnostics
44. 21.44 Gradient Accumulation
  1. 21.44.1 The equivalence proof in code
  2. 21.44.2 Where the cost goes
  3. 21.44.3 equivalence check
  4. 21.44.4 sync-on-last-step pattern
  5. 21.44.5 the throughput curve
45. 21.45 Checkpoint Save and Resume
  1. 21.45.1 The five state buckets
  2. 21.45.2 Atomic save
  3. 21.45.3 Sharded checkpoints
  4. 21.45.4 Resume continues mid epoch
  5. 21.45.5 capture and restore RNG state
  6. 21.45.6 atomic save
  7. 21.45.7 full checkpoint round trip
  8. 21.45.8 sharded variant
  9. 21.45.9 resume demo
46. 21.46 Distributed Data Parallel and FSDP from Scratch
  1. 21.46.1 The two collectives that matter
  2. 21.46.2 Gradient averaging matches single-process gradient
  3. 21.46.3 FSDP sketch
  4. 21.46.4 CPU and the gloo backend
  5. 21.46.5 bring up the process group
  6. 21.46.6 broadcast at construction
  7. 21.46.7 all-reduce gradients after backward
  8. 21.46.8 prove the equivalence
  9. 21.46.9 FSDP round trip
47. 21.47 Language Model Evaluation Harness
  1. 21.47.1 Task spec
  2. 21.47.2 The five fixture tasks
  3. 21.47.3 The metric contract
  4. 21.47.4 The model adapter
  5. 21.47.5 The runner
  6. 21.47.6 seed fixture tasks
  7. 21.47.7 load tasks
  8. 21.47.8 implement metrics
  9. 21.47.9 write the runner
  10. 21.47.10 emit JSON
48. 21.48 Hypothesis Generator
  1. 21.48.1 Why generate, then filter
  2. 21.48.2 The Hypothesis shape
  3. 21.48.3 Architecture
  4. 21.48.4 Temperature ramp
  5. 21.48.5 Novelty filter
  6. 21.48.6 Rank score
  7. 21.48.7 Mock language model
  8. 21.48.8 Output queue
  9. 21.48.9 Where this slots in
49. 21.49 Literature Retrieval
  1. 21.49.1 Why two retrieval passes
  2. 21.49.2 The Paper shape
  3. 21.49.3 Architecture
  4. 21.49.4 BM25 from scratch
  5. 21.49.5 Citation graph traversal
  6. 21.49.6 Dedup and ranking
  7. 21.49.7 Mock corpus
  8. 21.49.8 What lessons 52 and 53 read
  9. 21.49.9 Where this slots in
50. 21.50 Experiment Runner
  1. 21.50.1 Why a subprocess
  2. 21.50.2 The ExperimentSpec shape
  3. 21.50.3 Architecture
  4. 21.50.4 Why a soft memory cap
  5. 21.50.5 Capturing stdout and stderr
  6. 21.50.6 Ablation table
  7. 21.50.7 Determinism
  8. 21.50.8 The mock experiment script
  9. 21.50.9 Result shape
  10. 21.50.10 Where this slots in
51. 21.51 Result Evaluator
  1. 21.51.1 Why a paired test
  2. 21.51.2 Direction aware improvement
  3. 21.51.3 Architecture
  4. 21.51.4 Log normalisation
  5. 21.51.5 Per seed paired test
  6. 21.51.6 The Verdict shape
  7. 21.51.7 Where this slots in
52. 21.52 Paper Writer
  1. 21.52.1 Why a skeleton first
  2. 21.52.2 The Paper shape
  3. 21.52.3 The render contract
  4. 21.52.4 Figure injection from experiments
  5. 21.52.5 The mocked prose generator
  6. 21.52.6 The manifest output
  7. 21.52.7 Validation gates
53. 21.53 Critic Loop
  1. 21.53.1 Why five fixed dimensions
  2. 21.53.2 The Critique shape
  3. 21.53.3 Convergence rules, in order
  4. 21.53.4 Why plateau detection runs over two rounds
  5. 21.53.5 The deterministic critic in this lesson
  6. 21.53.6 The full loop contract
  7. 21.53.7 The Trace output
  8. 21.53.8 Budgets that protect against bad critics
54. 21.54 Iteration Scheduler
  1. 21.54.1 Why a scheduler, not a worklist
  2. 21.54.2 The system shape
  3. 21.54.3 The Hypothesis shape
  4. 21.54.4 UCB scoring
  5. 21.54.5 Parallel slots with asyncio
  6. 21.54.6 Fan-out: paper triggers
  7. 21.54.7 Fan-out: follow-up hypotheses
  8. 21.54.8 Budgets
  9. 21.54.9 The Trace and final report
55. 21.55 End-to-End Research Demo
  1. 21.55.1 What composes here
  2. 21.55.2 Why import, not copy
  3. 21.55.3 Determinism guarantees
  4. 21.55.4 The demo report shape
  5. 21.55.5 Failure mode handling
  6. 21.55.6 The best-result picker
  7. 21.55.7 Wiring the critic loop
  8. 21.55.8 Wiring the paper writer
56. 21.56 Vision Encoder Patches
  1. 21.56.1 Why patches, not pixels
  2. 21.56.2 Why a linear projection is enough
  3. 21.56.3 The Conv2d trick
  4. 21.56.4 Position embeddings
  5. 21.56.5 Equivalence as a sanity check
57. 21.57 Vision Transformer Encoder
  1. 21.57.1 Pre-LN vs post-LN
  2. 21.57.2 Multi-head self-attention
  3. 21.57.3 Why the 4x feed-forward expansion
  4. 21.57.4 Causal mask or not?
  5. 21.57.5 What the CLS token learns
58. 21.58 Projection Layer for Modality Alignment
  1. 21.58.1 Pooling before projection
  2. 21.58.2 Why two layers and not one
  3. 21.58.3 Cosine alignment loss
  4. 21.58.4 Frozen encoder is the trick
59. 21.59 Cross-Attention Fusion
  1. 21.59.1 Mask shapes
  2. 21.59.2 Why no mask on cross-attention
  3. 21.59.3 Key/value caching
  4. 21.59.4 Block composition
60. 21.60 Vision-Language Pretraining
  1. 21.60.1 InfoNCE in one paragraph
  2. 21.60.2 Temperature matters
  3. 21.60.3 Language modeling loss
  4. 21.60.4 Combining the losses
  5. 21.60.5 Why 50 steps is enough for a demo
61. 21.61 Multimodal Evaluation
  1. 21.61.1 Recall@K from a similarity matrix
  2. 21.61.2 VQA exact match
  3. 21.61.3 BLEU-4
  4. 21.61.4 Synthetic eval suite
62. 21.62 Chunking Strategies, Compared
  1. 21.62.1 Fixed-window
  2. 21.62.2 Sentence
  3. 21.62.3 Recursive split
  4. 21.62.4 Semantic clustering
  5. 21.62.5 Structural markdown headers
  6. 21.62.6 How recall@k measures the boundary choice
63. 21.63 Hybrid Retrieval with BM25 and Dense Embeddings
  1. 21.63.1 BM25 in one paragraph
  2. 21.63.2 Dense retrieval in one paragraph
  3. 21.63.3 Reciprocal rank fusion, the published formula
  4. 21.63.4 Why fusion beats score-weighted interpolation
64. 21.64 Cross-Encoder Reranker
  1. 21.64.1 The cross-encoder's input shape
  2. 21.64.2 Why this lesson trains a tiny one
  3. 21.64.3 Latency vs quality
65. 21.65 Query Rewriting: HyDE, Multi-Query, and Decomposition
  1. 21.65.1 HyDE in detail
  2. 21.65.2 Multi-query expansion in detail
  3. 21.65.3 Decomposition in detail
  4. 21.65.4 Why all three exist
66. 21.66 RAG Evaluation: Precision, Recall, MRR, nDCG, Faithfulness, Answer Relevance
  1. 21.66.1 Precision@k
  2. 21.66.2 Recall@k
  3. 21.66.3 MRR (Mean Reciprocal Rank)
  4. 21.66.4 nDCG@k
  5. 21.66.5 Faithfulness
  6. 21.66.6 Answer relevance
67. 21.67 End-to-End RAG System
  1. 21.67.1 Wiring choices
  2. 21.67.2 Answer generator with citations
  3. 21.67.3 The self-terminating demo
68. 21.68 Task Spec Format
  1. 21.68.1 Why a frozen spec
  2. 21.68.2 The record shape
  3. 21.68.3 Field rules
  4. 21.68.4 Validator behaviour
  5. 21.68.5 Few-shot rendering
  6. 21.68.6 Post-process rules
69. 21.69 Classical Metrics
  1. 21.69.1 Why reimplement
  2. 21.69.2 Tokenisation
  3. 21.69.3 Exact match
  4. 21.69.4 Token-level F1
  5. 21.69.5 BLEU-4
  6. 21.69.6 ROUGE-L
  7. 21.69.7 Accuracy
  8. 21.69.8 Dispatch contract
70. 21.70 Code Exec Metric
  1. 21.70.1 Why an isolated subprocess
  2. 21.70.2 The shape of a code-exec task
  3. 21.70.3 The denylist
  4. 21.70.4 Wall-clock timeout
  5. 21.70.5 Output cap
  6. 21.70.6 Pass-at-k
  7. 21.70.7 Exit codes
71. 21.71 Perplexity and Calibration
  1. 21.71.1 What perplexity tells you
  2. 21.71.2 What ECE measures
  3. 21.71.3 What Brier score does that ECE does not
  4. 21.71.4 Reliability diagram data
  5. 21.71.5 Confidence sources
  6. 21.71.6 Edge cases
  7. 21.71.7 Dispatch
72. 21.72 Leaderboard Aggregation
  1. 21.72.1 The shape of input
  2. 21.72.2 The output
  3. 21.72.3 Normalisation
  4. 21.72.4 Mean and win-rate
  5. 21.72.5 Bootstrap confidence intervals
  6. 21.72.6 Categories
  7. 21.72.7 Markdown rendering
73. 21.73 End-to-End Eval Runner
  1. 21.73.1 The pipeline
  2. 21.73.2 The adapter interface
  3. 21.73.3 Parallel execution
  4. 21.73.4 The single-pass scoring loop
  5. 21.73.5 Aggregation
  6. 21.73.6 Self-terminating demo
74. 21.74 Collective Ops From Scratch
  1. 21.74.1 Ring allreduce in two passes
  2. 21.74.2 Queue mesh as a stand-in for NCCL
  3. 21.74.3 Verify against gloo
75. 21.75 Data Parallel DDP From Scratch
  1. 21.75.1 The three operations DDP needs
  2. 21.75.2 Why mean and not sum
  3. 21.75.3 Why bucket gradients
  4. 21.75.4 Why pin the seed
76. 21.76 ZeRO Optimizer State Sharding
  1. 21.76.1 Stages of ZeRO
  2. 21.76.2 The memory math, real numbers
  3. 21.76.3 Why reduce_scatter beats allreduce-then-shard
77. 21.77 Pipeline Parallel and Bubble Analysis
  1. 21.77.1 GPipe schedule
  2. 21.77.2 1F1B schedule
  3. 21.77.3 Why equal compute per stage matters
  4. 21.77.4 Microbatch versus batch
78. 21.78 Sharded Checkpoint and Atomic Resume
  1. 21.78.1 Manifest schema
  2. 21.78.2 Atomic write
  3. 21.78.3 Three failure modes the schema must defend against
  4. 21.78.4 Why per-rank files, not one big file
79. 21.79 End-to-End Distributed Training
  1. 21.79.1 The mini GPT
  2. 21.79.2 The composition rules
  3. 21.79.3 Why a tiny GPT and not just an MLP
  4. 21.79.4 Self-terminating means exit 0
80. 21.80 Jailbreak Taxonomy
  1. 21.80.1 jailbreak
  2. 21.80.2 taxonomy
  3. 21.80.3 fixture
  4. 21.80.4 severity
  5. 21.80.5 match
81. 21.81 Prompt Injection Detector
  1. 21.81.1 detector
  2. 21.81.2 normalize
  3. 21.81.3 confusion matrix
  4. 21.81.4 precision
  5. 21.81.5 recall
82. 21.82 Refusal Evaluation
  1. 21.82.1 under-refusal
  2. 21.82.2 over-refusal
  3. 21.82.3 calibration
  4. 21.82.4 accuracy
  5. 21.82.5 per-category breakdown
83. 21.83 Content Classifier Integration
  1. 21.83.1 output classifier
  2. 21.83.2 severity
  3. 21.83.3 router
  4. 21.83.4 redact
  5. 21.83.5 instruction leakage
84. 21.84 Constitutional Rules Engine
  1. 21.84.1 constitution
  2. 21.84.2 predicate
  3. 21.84.3 violation
  4. 21.84.4 fixer
  5. 21.84.5 diff
85. 21.85 End-to-End Safety Gate
  1. 21.85.1 safety gate
  2. 21.85.2 pre-gen
  3. 21.85.3 during-gen
  4. 21.85.4 post-gen
  5. 21.85.5 trace
AClaude certification prepnot yet
1. A.1 Study the Decisions, Not the Vocabulary
  1. A.1.1 The blueprint is a job model
  2. A.1.2 Stable principles and changeable facts
  3. A.1.3 The scenario decision stack
  4. A.1.4 Wrong answers are usually locally correct
  5. A.1.5 Build an evidence ledger
  6. A.1.6 Use a cadence, not a cram pile
2. A.2 Choose the Smallest Surface That Can Carry the Work
  1. A.2.1 Start with the work, not the feature menu
  2. A.2.2 Chat is for bounded conversational work
  3. A.2.3 Projects are maintained context boundaries
  4. A.2.4 Cowork is a steerable task loop
  5. A.2.5 Research is for multi-source investigation
  6. A.2.6 Files and Artifacts make the output inspectable
  7. A.2.7 Connectors trade copying for live, permissioned access
  8. A.2.8 API and coding surfaces are for owned software behavior
  9. A.2.9 Deployment is a control-plane decision
  10. A.2.10 Score the requirement, not the provider
  11. A.2.11 Model families are roles, not status levels
3. A.3 Spend Capability Where Failure Is Expensive
  1. A.3.1 Tokens are a workload measure
  2. A.3.2 Use variables before live prices
  3. A.3.3 Quality needs a threshold, not a vibe
  4. A.3.4 Sampling produces a distribution, not a replay
  5. A.3.5 Diagnose the failure layer
  6. A.3.6 Latency has several components
  7. A.3.7 Route by observable constraints
  8. A.3.8 Caching, batching, and limits solve different problems
  9. A.3.9 A configuration is a dated bundle
  10. A.3.10 A scaled exam score is not a percentage
4. A.4 Turn a Request Into a Testable Contract
  1. A.4.1 Prompting is interface design
  2. A.4.2 Criteria come before wording
  3. A.4.3 Decompose along verification boundaries
  4. A.4.4 Examples teach boundaries
  5. A.4.5 Assign roles carefully
  6. A.4.6 Iteration needs a hypothesis
  7. A.4.7 Write the acceptance card
  8. A.4.8 Create a source hierarchy
  9. A.4.9 Design the stages
  10. A.4.10 Add uncertainty behavior
  11. A.4.11 Test adversarial cases
  12. A.4.12 Exam decision pattern
  13. A.4.13 Common traps
5. A.5 Put Each Fact in the Right Kind of Context
  1. A.5.1 Seven mechanisms, seven jobs
  2. A.5.2 Context has an attention budget
  3. A.5.3 Retrieval is selection, not verification
  4. A.5.4 Instructions and knowledge are different
  5. A.5.5 Memory is continuity, not a database of record
  6. A.5.6 Prompt caching is an economic mechanism
  7. A.5.7 Context quality needs lifecycle ownership
  8. A.5.8 Inventory the context
  9. A.5.9 Create a source registry
  10. A.5.10 Design retrieval with abstention
  11. A.5.11 Budget the prompt
  12. A.5.12 Establish maintenance
  13. A.5.13 Exam decision pattern
  14. A.5.14 Common traps
6. A.6 Validate the Claim, Not the Confidence
  1. A.6.1 Start from the job of the output
  2. A.6.2 Trace claims to evidence
  3. A.6.3 Diagnose the property before retrying
  4. A.6.4 Use layered validation
  5. A.6.5 Match the evaluator to the property
  6. A.6.6 Hallucination is not one failure
  7. A.6.7 Evaluation sets represent risk
  8. A.6.8 Compare outputs without brand bias
  9. A.6.9 Define release gates
  10. A.6.10 Build a validation record
  11. A.6.11 Separate generation and review
  12. A.6.12 Calibrate graders
  13. A.6.13 Close the loop
  14. A.6.14 Exam decision pattern
  15. A.6.15 Common traps
7. A.7 Put Authority Around Capability
  1. A.7.1 Classify before processing
  2. A.7.2 Capability, permission, and authority are separate
  3. A.7.3 Minimize data and purpose
  4. A.7.4 Product terms are changeable facts
  5. A.7.5 Human review belongs at decision boundaries
  6. A.7.6 Fairness requires a defined population and outcome
  7. A.7.7 Transparency should serve the affected person
  8. A.7.8 Guardrails need defense in depth
  9. A.7.9 Incidents need a prepared path
  10. A.7.10 Write a use-case card
  11. A.7.11 Create a control map
  12. A.7.12 Design the approval packet
  13. A.7.13 Run a threat workshop
  14. A.7.14 Exam decision pattern
  15. A.7.15 Common traps
8. A.8 Design the Handoff Before the Automation
  1. A.8.1 Map the current state first
  2. A.8.2 Choose the intervention, not just the tool
  3. A.8.3 Specify every step as a contract
  4. A.8.4 Use consequence and reversibility to place control
  5. A.8.5 Choose a workflow pattern that matches dependencies
  6. A.8.6 Handoffs are products
  7. A.8.7 Preserve state across failures
  8. A.8.8 Measure the workflow, not the demo
  9. A.8.9 Communicate limits by stakeholder
  10. A.8.10 Map one real case
  11. A.8.11 Score candidate interventions
  12. A.8.12 Write the future-state contract
  13. A.8.13 Build the handoff packet
  14. A.8.14 Pilot in shadow mode
  15. A.8.15 Exam decision pattern
  16. A.8.16 Common traps
9. A.9 The Messages API Is a State Machine
  1. A.9.1 The Failure That Teaches the Protocol
  2. A.9.2 One Request, One Transition
  3. A.9.3 Choose Two Independent Access Patterns
  4. A.9.4 Content Is a Sequence of Typed Blocks
  5. A.9.5 Stop Reasons Are Control Signals
  6. A.9.6 The Client Owns Conversation State
  7. A.9.7 Multimodal Requests Are Typed Asset Transfers
  8. A.9.8 Streaming Changes Delivery, Not Meaning
  9. A.9.9 Batch, Cache, and Thinking Solve Different Problems
  10. A.9.10 Build the Lifecycle and Asset Boundary Offline
  11. A.9.11 Application Lifecycle Beyond One Turn
  12. A.9.12 Exam Decision Rules
10. A.10 Structured Output Is an Untrusted Contract
  1. A.10.1 The JSON That Should Have Failed
  2. A.10.2 Prompting for JSON Is Not a Contract
  3. A.10.3 Design Schemas From Consumer Decisions
  4. A.10.4 Parse Strictly
  5. A.10.5 Validate Meaning After Shape
  6. A.10.6 Repair With a Budget
  7. A.10.7 Tool Inputs and Final Outputs Are Different Contracts
  8. A.10.8 Pydantic Is a Validator Implementation, Not the Contract
  9. A.10.9 Streaming Produces Partial Syntax
  10. A.10.10 Schema Evolution Is an API Migration
  11. A.10.11 Build the Validator and Repair Loop
  12. A.10.12 Exam Decision Rules
11. A.11 A Tool Loop Is Controlled Delegation
  1. A.11.1 Give Each Tool One Job
  2. A.11.2 Write Selection Descriptions, Not Internal Documentation
  3. A.11.3 Make Invalid Calls Hard to Express
12. A.12 MCP Separates Capability From Host
  1. A.12.1 Tools Perform Model-Selected Operations
  2. A.12.2 Resources Expose Addressable Context
  3. A.12.3 Prompts Package User-Invoked Templates
13. A.13 The Agent SDK Is a Harness, Not Permission
  1. A.13.1 The Framework Did Not Make the Agent Safe
  2. A.13.2 Model Plus Harness
  3. A.13.3 Do Not Collapse Four Harness Levels Into "The SDK"
  4. A.13.4 Start With the Use-Case Gate
  5. A.13.5 Give the Agent an Environment It Can Understand
  6. A.13.6 Computer Use Is a Screenshot-Action Verification Loop
  7. A.13.7 Hooks Make Lifecycle Rules Deterministic
  8. A.13.8 Hooks Are One Layer
  9. A.13.9 Subagents Buy Context Isolation
  10. A.13.10 Skills Package Reusable Procedure
  11. A.13.11 Sessions Are Continuity, Not Truth
  12. A.13.12 Long-Running Work Needs Contracts
  13. A.13.13 Stream Events Into Observability
  14. A.13.14 Managed Sessions Stop for Actions, Not Only Answers
  15. A.13.15 A Minimal SDK Shape
  16. A.13.16 Exam Decision Rules
14. A.14 Security Lives Outside the Prompt
  1. A.14.1 The Document Is Not Your Boss
  2. A.14.2 Draw the Trust Boundaries
  3. A.14.3 Threat Model the Real System
  4. A.14.4 Instructions Do Not Create Isolation
  5. A.14.5 Keep Secrets Out of Model Context
  6. A.14.6 Identity Comes From the Session
  7. A.14.7 Least Privilege by Capability
  8. A.14.8 Policy Gate Before Tool Handler
  9. A.14.9 Hooks Enforce Lifecycle Policy
  10. A.14.10 MCP Expands the Supply Chain
  11. A.14.11 Output Is Another Attack Surface
  12. A.14.12 Logging Without Leaking
  13. A.14.13 Security Evals and Incident Response
  14. A.14.14 Exam Decision Rules
15. A.15 Evals Turn Agent Behavior Into Engineering Evidence
  1. A.15.1 Output Contract
  2. A.15.2 Tool Trajectory
  3. A.15.3 Final State
  4. A.15.4 Safety
  5. A.15.5 Operational Budget
16. A.16 Claude Code Scales Through Shared Constraints
  1. A.16.1 Rules
  2. A.16.2 Skills
  3. A.16.3 Commands
  4. A.16.4 Agents
  5. A.16.5 Goal Sessions
  6. A.16.6 In-Session Loops and Scheduled Prompts
17. A.17 Multi-Agent Orchestration and Delegation
  1. A.17.1 Decide Why Another Context Exists
  2. A.17.2 Start With Five Patterns
  3. A.17.3 Write a Delegation Contract
  4. A.17.4 Keep Deterministic Sequence Outside the Model
  5. A.17.5 Use the Task Boundary Deliberately
  6. A.17.6 Preserve Three Result States
  7. A.17.7 Merge by Identity and Provenance
  8. A.17.8 Evaluate the Trajectory
  9. A.17.9 Define the Final Contract
  10. A.17.10 Try a Single-Agent Baseline
  11. A.17.11 Identify Context Boundaries
  12. A.17.12 Write Task Contracts
  13. A.17.13 Encode Prerequisites
  14. A.17.14 Red-Team the Merge
  15. A.17.15 Agent Per Step
  16. A.17.16 Parallel by Default
  17. A.17.17 Coordinator as Data Warehouse
  18. A.17.18 Reviewer With Generator Context
18. A.18 Agent SDK Sessions, Subagents, and Context
  1. A.18.1 Context Is a Working Set
  2. A.18.2 Choose Among Four Session Moves
  3. A.18.3 Use a Structured Resume Packet
  4. A.18.4 Isolate Subagent Context by Responsibility
  5. A.18.5 Use Hooks for Deterministic Lifecycle Work
  6. A.18.6 Make Side Effects Idempotent
  7. A.18.7 Revalidate at the Boundary
  8. A.18.8 Plan Context Budgets
  9. A.18.9 Session 1: Inventory and Plan
  10. A.18.10 Session 2: Implement and Verify
  11. A.18.11 Session 3: Independent Review
  12. A.18.12 Hook Requirements
  13. A.18.13 Session Equals State
  14. A.18.14 Compaction Equals Recovery
  15. A.18.15 Fork Equals Independence
  16. A.18.16 Hooks Everywhere
19. A.19 Tool Contracts, Errors, and Progressive Discovery
  1. A.19.1 A Tool Description Is Part of the Decision Surface
  2. A.19.2 Avoid Overlapping Tools
  3. A.19.3 Make Schemas Carry Invariants
  4. A.19.4 Return Errors as Data
  5. A.19.5 Use Tool Choice Deliberately
  6. A.19.6 Distribute Fewer Tools
  7. A.19.7 Discover Large Catalogs Progressively
  8. A.19.8 Scope MCP Configuration
  9. A.19.9 Choose Claude Code Built-In Tools by Intent
  10. A.19.10 Tool Description as Authorization
  11. A.19.11 Error Text as Recovery Policy
  12. A.19.12 One Tool for Every Operation
  13. A.19.13 Secrets in Shared Configuration
20. A.20 Claude Code Memory, Rules, Skills, and CI
  1. A.20.1 Match the Mechanism to the Job
  2. A.20.2 Keep the Root Instruction File Small
  3. A.20.3 Place Instructions at the Narrowest True Scope
  4. A.20.4 Import Stable Supporting Guidance
  5. A.20.5 Use Skills for Progressive Disclosure
  6. A.20.6 Use Commands for Explicit User Intent
  7. A.20.7 Use Subagents as Bounded Evidence Gatherers
  8. A.20.8 Distribute Through the Smallest Shared Surface
  9. A.20.9 Use Path Rules as Local Policy
  10. A.20.10 Separate Planning, Exploration, and Execution
  11. A.20.11 Make Tests Part of the Conversation Contract
  12. A.20.12 Hook Decisions Need Exact Contracts
  13. A.20.13 Design Headless CI as a Fresh Reviewer
  14. A.20.14 Preserve Findings Across Runs
  15. A.20.15 Root Guidance
  16. A.20.16 Path Rules
  17. A.20.17 Skills and Commands
  18. A.20.18 Hooks
  19. A.20.19 CI Review
  20. A.20.20 Root File as Encyclopedia
  21. A.20.21 Private Configuration as Team Policy
  22. A.20.22 Hook as Hidden Build System
  23. A.20.23 AI Review as the Only Gate
21. A.21 Reliable Extraction, Batch, and Independent Reviewers
  1. A.21.1 Define the Judgment Before the Schema
  2. A.21.2 Use Few-Shot Examples for Boundaries
  3. A.21.3 Make Absence Representable
  4. A.21.4 Use Tool Use for Typed Output
  5. A.21.5 Validate in Four Layers
  6. A.21.6 Feed Back the Smallest Useful Error
  7. A.21.7 Separate Generator and Reviewer
  8. A.21.8 Choose Batch for the Workflow
  9. A.21.9 Make Batch Jobs Reconciliable
  10. A.21.10 Evaluate the Error You Care About
  11. A.21.11 Output Contract
  12. A.21.12 Dataset
  13. A.21.13 Passes
  14. A.21.14 Experiment
  15. A.21.15 Batch Design
  16. A.21.16 Schema Equals Truth
  17. A.21.17 Required Non-Nullable Fields
  18. A.21.18 Infinite Repair
  19. A.21.19 Reviewer Rewrites Silently
22. A.22 Make Large Context Observable
  1. A.22.1 Context has an attention topology
  2. A.22.2 Evidence needs an envelope
  3. A.22.3 Trim tool output by decision value
  4. A.22.4 Complete, partial, and blocked are first-class states
  5. A.22.5 Escalate the reason, not the anxiety
  6. A.22.6 Large codebases need four different memory tools
  7. A.22.7 Confidence should be evidence-calibrated
  8. A.22.8 Human review should be stratified
  9. A.22.9 Content type changes meaning
  10. A.22.10 Create the manifest
  11. A.22.11 Define the context budget
  12. A.22.12 Normalize tool results
  13. A.22.13 Add escalation rules
  14. A.22.14 Calibrate review
  15. A.22.15 Test rendering
  16. A.22.16 Exam decision patterns
  17. A.22.17 Common traps
23. A.23 Business Discovery, Requirements, and SLAs
  1. A.23.1 Start With the Decision, Not the Feature
  2. A.23.2 Classify Requirements Before Prioritizing Them
  3. A.23.3 Map the Current Workflow
  4. A.23.4 Separate Value From Capability
  5. A.23.5 Turn Risk Into a Review Boundary
  6. A.23.6 Define SLI, SLO, and SLA Correctly
  7. A.23.7 Make Non-Goals Explicit
  8. A.23.8 Prioritize With Cost of Error
  9. A.23.9 Turning Every Request Into an Agent
  10. A.23.10 Treating Human Review as Free
  11. A.23.11 Using Accuracy Without a Denominator
  12. A.23.12 Ignoring Ownership After Launch
24. A.24 End-to-End Architecture and Value Tradeoffs
  1. A.24.1 Draw the Whole Loop
  2. A.24.2 Choose the Smallest Pattern That Fits
  3. A.24.3 Decompose Around Contracts
  4. A.24.4 Separate Semantic and Deterministic Controls
  5. A.24.5 Budget Across the Whole Trajectory
  6. A.24.6 Treat Latency as a Distribution
  7. A.24.7 Design Feedback as a Product Surface
  8. A.24.8 Draw Three Candidates
  9. A.24.9 Score Explicit Tradeoffs
  10. A.24.10 Write Failure Paths
  11. A.24.11 Define the Evaluation
  12. A.24.12 Record Reversal Conditions
  13. A.24.13 Capability Bloat
  14. A.24.14 Prompting Around a Missing Contract
  15. A.24.15 Self-Review as Independence
  16. A.24.16 Optimizing the Demo Path
25. A.25 RAG, Retrieval, and Data Pipelines
  1. A.25.1 RAG Is a Data System
  2. A.25.2 Design Chunks Around Meaning and Retrieval
  3. A.25.3 Match Retrieval to Query and Data Shape
  4. A.25.4 Separate Retrieval Evaluation From Answer Evaluation
  5. A.25.5 Preserve Provenance as Data
  6. A.25.6 Make Refresh Atomic and Observable
  7. A.25.7 Normalize Tokens
  8. A.25.8 Chunk With Stable Identity
  9. A.25.9 Exclude Inactive Sources Before Indexing
  10. A.25.10 Score Transparently
  11. A.25.11 Return Provenance
  12. A.25.12 Evaluate the Retriever
  13. A.25.13 More Context Means More Grounding
  14. A.25.14 Similar Means Authoritative
  15. A.25.15 Valid Citation Means Supported Claim
  16. A.25.16 Refresh Means Append
26. A.26 Integration Protocols, Identity, and Least Privilege
  1. A.26.1 Choose an Integration Shape From the Boundary
  2. A.26.2 Separate Discovery, Selection, and Execution
  3. A.26.3 Propagate Identity, Do Not Replace It
  4. A.26.4 Use Least Privilege at Four Levels
  5. A.26.5 Design Approval as a Capability
  6. A.26.6 Return Structured Errors
  7. A.26.7 Progressive Discovery Reduces Capability Bloat
  8. A.26.8 MCP Scopes Do Not Define Business Authorization
  9. A.26.9 Select a Primary Shape
  10. A.26.10 Define Principal and Tool Contracts
  11. A.26.11 Filter Discovery
  12. A.26.12 Authorize at Execution
  13. A.26.13 One Service Account for Every User
  14. A.26.14 Confirmation Without Binding
  15. A.26.15 Tool Descriptions as Controls
  16. A.26.16 Retrying Authorization Errors
27. A.27 Production Observability, Latency, and Cost
  1. A.27.1 Observe Four Layers
  2. A.27.2 Logs, Metrics, and Traces Have Different Jobs
  3. A.27.3 Trace the Semantic Contract
  4. A.27.4 Separate System Success From Task Success
  5. A.27.5 Decompose Latency
  6. A.27.6 Optimize From Evidence
  7. A.27.7 Measure Cost Per Successful Outcome
  8. A.27.8 Understand Prompt Cache Shape
  9. A.27.9 Build Actionable Alerts
  10. A.27.10 Use Rollouts to Limit Evidence Risk
  11. A.27.11 Represent One Task Trajectory
  12. A.27.12 Aggregate Without Hiding Failure
  13. A.27.13 Compare Variants
  14. A.27.14 Evaluate Service Objectives
  15. A.27.15 Logging Full Prompts by Default
  16. A.27.16 One Aggregate Quality Score
  17. A.27.17 Average Latency
  18. A.27.18 Cost per Call
28. A.28 Enterprise Governance, Compliance, and Human Review
  1. A.28.1 Start With a Risk Decision
  2. A.28.2 Map Data Through the Whole System
  3. A.28.3 Minimize Before Protecting
  4. A.28.4 Build a Control Matrix
  5. A.28.5 Layer Model and System Guardrails
  6. A.28.6 Design Human Review as a Control
  7. A.28.7 Match Review to Risk
  8. A.28.8 Treat Fairness as a Contextual Requirement
  9. A.28.9 Provide Transparency and Contestability
  10. A.28.10 Plan for Change
  11. A.28.11 Risk Register
  12. A.28.12 Data Map
  13. A.28.13 Control Matrix
  14. A.28.14 Human Review Design
  15. A.28.15 Approval and Reassessment
  16. A.28.16 Compliance by Product Name
  17. A.28.17 Human Review as a Checkbox
  18. A.28.18 Logging for Audit Without Minimization
  19. A.28.19 One Fairness Metric
29. A.29 Stakeholder Communication, ADRs, and Lifecycle Ownership
  1. A.29.1 One System Needs Several Views
  2. A.29.2 Build an Architecture Narrative
  3. A.29.3 Use Diagrams to Answer One Question
  4. A.29.4 Record Decisions, Not Meeting Transcripts
  5. A.29.5 Turn Architecture Into Contracts
  6. A.29.6 Define Ownership by Decision
  7. A.29.7 Design Handoff Acceptance
  8. A.29.8 Plan Adoption as Part of the Workflow
  9. A.29.9 Communicate Incidents by Impact and Decision
  10. A.29.10 Keep Lifecycle Evidence Connected
  11. A.29.11 Executive Decision Brief
  12. A.29.12 Architecture Decision Record
  13. A.29.13 Engineering Contract Index
  14. A.29.14 Operational Readiness Checklist
  15. A.29.15 Ownership and Change Map
  16. A.29.16 One Deck for Every Audience
  17. A.29.17 Architecture as a Launch Artifact
  18. A.29.18 Ownership by Team Name
  19. A.29.19 Adoption Equals Value
30. A.30 Ship a Week of Work, Not a Perfect Prompt
  1. A.30.1 The deliverable is a chain of proof
  2. A.30.2 The Python validator is intentionally limited
  3. A.30.3 Monday: scope the decision and product surface
  4. A.30.4 Tuesday: freeze and govern the evidence
  5. A.30.5 Wednesday: decompose the work
  6. A.30.6 Thursday: validate and challenge
  7. A.30.7 Friday: hand off and close the loop
  8. A.30.8 A complete evidence package
  9. A.30.9 Capstone decision patterns
  10. A.30.10 Common traps
31. A.31 Ship a Claude Application You Can Defend
  1. A.31.1 The Deliverable
  2. A.31.2 Start With Requirements
  3. A.31.3 Architecture
  4. A.31.4 Output Contract
  5. A.31.5 Tool Contract
  6. A.31.6 Policy Before Execution
  7. A.31.7 Trace the Decisions, Not the Secrets
  8. A.31.8 Build and Run
  9. A.31.9 Replace the Simulator With Claude
  10. A.31.10 Streaming Decision
  11. A.31.11 Caching and Batch Decision
  12. A.31.12 MCP Decision
  13. A.31.13 Eval Plan
  14. A.31.14 Runbook
  15. A.31.15 Architecture Defense
  16. A.31.16 Definition of Done
  17. A.31.17 Exam Decision Rules
32. A.32 Defend One Architecture Across Six Contexts
  1. A.32.1 The six contexts are lenses, not templates
  2. A.32.2 Use a five-gate decision stack
  3. A.32.3 Architecture quality appears in failure behavior
  4. A.32.4 Choose one primary lens
  5. A.32.5 Complete the architecture packet
  6. A.32.6 Encode the packet as JSON
  7. A.32.7 Run failure-first tests
  8. A.32.8 Transfer across all six lenses
  9. A.32.9 Defend tradeoffs
  10. A.32.10 Conduct independent review
  11. A.32.11 Exam scenario method
  12. A.32.12 Submission evidence
  13. A.32.13 Common traps
33. A.33 Architect Professional System Capstone
  1. A.33.1 Discovery Brief
  2. A.33.2 Architecture Options and ADRs
  3. A.33.3 End-to-End System Views
  4. A.33.4 Model, Prompt, and Context Plan
  5. A.33.5 Knowledge and RAG Design
  6. A.33.6 Integration and Identity Design
  7. A.33.7 Evaluation and Production Evidence
  8. A.33.8 Governance and Human Review
  9. A.33.9 Rollout and Operations
  10. A.33.10 Stakeholder and Handoff Package
  11. A.33.11 Encode Requirements
  12. A.33.12 Encode Decisions
  13. A.33.13 Encode Controls
  14. A.33.14 Evaluate Gates
  15. A.33.15 Make the Release Decision
  16. A.33.16 A Polished Diagram Without Decisions
  17. A.33.17 A Long Control List Without Evidence
  18. A.33.18 An Evaluation With Only Happy Paths
  19. A.33.19 A Human Review Queue Without Capacity
  20. A.33.20 A Handoff Without Recovery Proof
34. A.34 CCAO-F: Claude Certified Associate - Foundations
  1. A.34.1 Prompting and Task Execution (14%)
  2. A.34.2 Output Evaluation and Validation (21%)
  3. A.34.3 Product and Model Selection (12%)
  4. A.34.4 Workflow Integration and Solution Design (16%)
  5. A.34.5 Configuration and Knowledge Management (12%)
  6. A.34.6 Governance, Risk, and Responsible Use (15%)
  7. A.34.7 Troubleshooting and Optimization (10%)
35. A.35 CCAO-F diagnostic: Associate Foundations Diagnostic
  1. A.35.1 Prompting and Task Execution: 2 questions
  2. A.35.2 Output Evaluation and Validation: 2 questions
  3. A.35.3 Product and Model Selection: 2 questions
  4. A.35.4 Workflow Integration and Solution Design: 2 questions
  5. A.35.5 Configuration and Knowledge Management: 2 questions
  6. A.35.6 Governance, Risk, and Responsible Use: 2 questions
  7. A.35.7 Troubleshooting and Optimization: 2 questions
36. A.36 CCAO-F mock: Associate Foundations Full Mock
  1. A.36.1 Prompting and Task Execution: 8 questions
  2. A.36.2 Output Evaluation and Validation: 13 questions
  3. A.36.3 Product and Model Selection: 7 questions
  4. A.36.4 Workflow Integration and Solution Design: 10 questions
  5. A.36.5 Configuration and Knowledge Management: 7 questions
  6. A.36.6 Governance, Risk, and Responsible Use: 9 questions
  7. A.36.7 Troubleshooting and Optimization: 6 questions
37. A.37 CCAR-F: Claude Certified Architect - Foundations
  1. A.37.1 Agentic Architecture and Orchestration (27%)
  2. A.37.2 Tool Design and MCP Integration (18%)
  3. A.37.3 Claude Code Configuration and Workflows (20%)
  4. A.37.4 Prompt Engineering and Structured Output (20%)
  5. A.37.5 Context Management and Reliability (15%)
38. A.38 CCAR-F diagnostic: Architect Foundations Diagnostic
  1. A.38.1 Agentic Architecture and Orchestration: 4 questions
  2. A.38.2 Tool Design and MCP Integration: 3 questions
  3. A.38.3 Claude Code Configuration and Workflows: 3 questions
  4. A.38.4 Prompt Engineering and Structured Output: 3 questions
  5. A.38.5 Context Management and Reliability: 2 questions
39. A.39 CCAR-F mock: Architect Foundations Scenario Mock
  1. A.39.1 Agentic Architecture and Orchestration: 16 questions
  2. A.39.2 Tool Design and MCP Integration: 11 questions
  3. A.39.3 Prompt Engineering and Structured Output: 12 questions
  4. A.39.4 Context Management and Reliability: 9 questions
  5. A.39.5 Claude Code Configuration and Workflows: 12 questions
40. A.40 CCAR-P: Claude Certified Architect - Professional
  1. A.40.1 Solution Design and Architecture (17%)
  2. A.40.2 Claude Models, Prompting and Context Engineering (13%)
  3. A.40.3 Integration (19%)
  4. A.40.4 Evaluation, Testing and Optimization (16%)
  5. A.40.5 Governance, Safety and Risk Management (14%)
  6. A.40.6 Stakeholder Communication and Lifecycle Management (14%)
  7. A.40.7 Developer Productivity and Operational Enablement (7%)
41. A.41 CCAR-P diagnostic: Architect Professional Diagnostic
  1. A.41.1 Solution Design and Architecture: 2 questions
  2. A.41.2 Claude Models, Prompting and Context Engineering: 2 questions
  3. A.41.3 Integration: 2 questions
  4. A.41.4 Evaluation, Testing and Optimization: 2 questions
  5. A.41.5 Governance, Safety and Risk Management: 2 questions
  6. A.41.6 Stakeholder Communication and Lifecycle Management: 2 questions
  7. A.41.7 Developer Productivity and Operational Enablement: 2 questions
42. A.42 CCAR-P mock: Architect Professional Full Mock
  1. A.42.1 Solution Design and Architecture: 11 questions
  2. A.42.2 Claude Models, Prompting and Context Engineering: 8 questions
  3. A.42.3 Integration: 12 questions
  4. A.42.4 Evaluation, Testing and Optimization: 10 questions
  5. A.42.5 Governance, Safety and Risk Management: 9 questions
  6. A.42.6 Stakeholder Communication and Lifecycle Management: 9 questions
  7. A.42.7 Developer Productivity and Operational Enablement: 4 questions
43. A.43 CCDV-F: Claude Certified Developer - Foundations
  1. A.43.1 Agents and Workflows (14.7%)
  2. A.43.2 Applications and Integration (33.1%)
  3. A.43.3 Claude Code (3.1%)
  4. A.43.4 Eval, Testing, and Debugging (2.6%)
  5. A.43.5 Model Selection and Optimization (16.8%)
  6. A.43.6 Prompt and Context Engineering (11.0%)
  7. A.43.7 Security and Safety (8.1%)
  8. A.43.8 Tools and MCPs (10.6%)
44. A.44 CCDV-F diagnostic: Developer Foundations Diagnostic
  1. A.44.1 Agents and Workflows: 2 questions
  2. A.44.2 Applications and Integration: 4 questions
  3. A.44.3 Claude Code: 1 questions
  4. A.44.4 Eval, Testing, and Debugging: 1 questions
  5. A.44.5 Model Selection and Optimization: 3 questions
  6. A.44.6 Prompt and Context Engineering: 2 questions
  7. A.44.7 Security and Safety: 1 questions
  8. A.44.8 Tools and MCPs: 2 questions
45. A.45 CCDV-F mock: Developer Foundations Full Mock
  1. A.45.1 Agents and Workflows: 8 questions
  2. A.45.2 Applications and Integration: 17 questions
  3. A.45.3 Claude Code: 2 questions
  4. A.45.4 Eval, Testing, and Debugging: 1 questions
  5. A.45.5 Model Selection and Optimization: 9 questions
  6. A.45.6 Prompt and Context Engineering: 6 questions
  7. A.45.7 Security and Safety: 4 questions
  8. A.45.8 Tools and MCPs: 6 questions
BMCP Associate certification prepnot yet
1. B.1 The MCPA Blueprint Is a Study Budget, Not a Checklist
  1. B.1.1 Reading a question: catching a legacy-era distractor
  2. B.1.2 The route: 34 lessons across five domains
  3. B.1.3 Blueprint
  4. B.1.4 Domain weight
  5. B.1.5 Study-hours budget
  6. B.1.6 Readiness estimate
  7. B.1.7 Unpublished fact
  8. B.1.8 Legacy-era distractor
  9. B.1.9 Route
2. B.2 Reading the MCP Specification
  1. B.2.1 Base protocol
  2. B.2.2 BCP 14
  3. B.2.3 schema.ts
  4. B.2.4 Current revision
  5. B.2.5 Draft, Current, Final
  6. B.2.6 Active
  7. B.2.7 Deprecated
  8. B.2.8 Removed
  9. B.2.9 Earliest removal
  10. B.2.10 SEP
3. B.3 The Integration Problem MCP Solves
  1. B.3.1 Host
  2. B.3.2 Client
  3. B.3.3 Server
  4. B.3.4 Tool
  5. B.3.5 Resource
  6. B.3.6 Prompt
  7. B.3.7 _meta
  8. B.3.8 resultType
  9. B.3.9 Protocol error
  10. B.3.10 Tool execution error
4. B.4 The JSON-RPC Envelope
  1. B.4.1 Request
  2. B.4.2 Notification
  3. B.4.3 Result response
  4. B.4.4 Error response
  5. B.4.5 resultType
  6. B.4.6 _meta
  7. B.4.7 Reserved prefix
  8. B.4.8 protocolVersion
  9. B.4.9 clientCapabilities
  10. B.4.10 Self-reported field
5. B.5 The Stateless Core of MCP
  1. B.5.1 Statelessness
  2. B.5.2 Connection
  3. B.5.3 Replica
  4. B.5.4 Server-minted handle
  5. B.5.5 Shared store
  6. B.5.6 Tool execution error
  7. B.5.7 Mcp-Session-Id
6. B.6 Telling a Modern MCP Server From a Legacy One
  1. B.6.1 Modern
  2. B.6.2 Legacy
  3. B.6.3 Dual-era
  4. B.6.4 initialize (legacy)
  5. B.6.5 server/discover
  6. B.6.6 DiscoverResult
  7. B.6.7 UnsupportedProtocolVersionError
  8. B.6.8 Era caching
  9. B.6.9 Compatibility matrix
  10. B.6.10 400 body inspection
  11. B.6.11 Modern-only rejection
7. B.7 Hosts, Clients, and Servers: MCP's Process Topology
  1. B.7.1 Host
  2. B.7.2 Client
  3. B.7.3 Server
  4. B.7.4 Local server
  5. B.7.5 Remote server
  6. B.7.6 Control model
  7. B.7.7 Elicitation
  8. B.7.8 serverInfo.name
  9. B.7.9 Server-id prefix
8. B.8 Discovering a Server and Negotiating What It Can Do
  1. B.8.1 server/discover
  2. B.8.2 DiscoverResult
  3. B.8.3 ServerCapabilities
  4. B.8.4 ClientCapabilities
  5. B.8.5 serverInfo
  6. B.8.6 MissingRequiredClientCapabilityError
  7. B.8.7 UnsupportedProtocolVersionError
  8. B.8.8 Per-request negotiation
9. B.9 The Contract Inside a Tool Definition
  1. B.9.1 inputSchema
  2. B.9.2 outputSchema
  3. B.9.3 structuredContent
  4. B.9.4 additionalProperties
  5. B.9.5 $ref
  6. B.9.6 Tool execution error
  7. B.9.7 Protocol error
  8. B.9.8 Tool naming rules
10. B.10 Reading a Server Manifest Like a Reviewer
  1. B.10.1 Manifest
  2. B.10.2 Annotation defaults
  3. B.10.3 x-mcp-header
  4. B.10.4 cacheScope
  5. B.10.5 instructions
  6. B.10.6 Reverse-DNS namespace
  7. B.10.7 Red flag
  8. B.10.8 Ownership verification
11. B.11 The Model Interaction Flow
  1. B.11.1 Model context
  2. B.11.2 Confirmation gate
  3. B.11.3 Tool execution error
  4. B.11.4 input_required
  5. B.11.5 Protocol error
  6. B.11.6 Deterministic ordering
  7. B.11.7 requestState
  8. B.11.8 Annotation defaults
12. B.12 The Tools Primitive: Calling Actions and Reading Their Results
  1. B.12.1 Tool
  2. B.12.2 tools/list
  3. B.12.3 tools/call
  4. B.12.4 CallToolResult
  5. B.12.5 Content block
  6. B.12.6 resource_link
  7. B.12.7 Content annotations
  8. B.12.8 Tool annotations
  9. B.12.9 nextCursor
  10. B.12.10 subscriptions/listen
13. B.13 Resources: Addressable Content for a Stateless Server
  1. B.13.1 Resource
  2. B.13.2 resources/list
  3. B.13.3 resources/templates/list
  4. B.13.4 resources/read
  5. B.13.5 Text content
  6. B.13.6 Binary content
  7. B.13.7 -32602
  8. B.13.8 ttlMs
  9. B.13.9 cacheScope
  10. B.13.10 subscriptions/listen
14. B.14 Prompt Templates and Argument Completion
  1. B.14.1 Prompt
  2. B.14.2 prompts/list
  3. B.14.3 prompts/get
  4. B.14.4 PromptMessage
  5. B.14.5 resource_link
  6. B.14.6 completion/complete
  7. B.14.7 ref/prompt
  8. B.14.8 ref/resource
  9. B.14.9 context.arguments
  10. B.14.10 hasMore
15. B.15 Multi Round-Trip Requests and Elicitation
  1. B.15.1 MRTR
  2. B.15.2 InputRequiredResult
  3. B.15.3 inputRequests
  4. B.15.4 inputResponses
  5. B.15.5 requestState
  6. B.15.6 Form mode elicitation
  7. B.15.7 URL mode elicitation
  8. B.15.8 ElicitResult action
  9. B.15.9 Principal binding
  10. B.15.10 Single use enforcement
16. B.16 Deprecated, Not Removed: Roots, Sampling, and Logging
  1. B.16.1 Active
  2. B.16.2 Deprecated
  3. B.16.3 Removed
  4. B.16.4 Earliest removal
  5. B.16.5 Roots
  6. B.16.6 Sampling
  7. B.16.7 io.modelcontextprotocol/logLevel
  8. B.16.8 notifications/message
  9. B.16.9 logging/setLevel
  10. B.16.10 SEP-2577
17. B.17 The Subscription Stream: Notifications, Progress, and Cancellation
  1. B.17.1 Notification
  2. B.17.2 subscriptions/listen
  3. B.17.3 Subscription id
  4. B.17.4 Stream notification
  5. B.17.5 Request scoped notification
  6. B.17.6 progressToken
  7. B.17.7 Graceful closure
  8. B.17.8 notifications/cancelled
18. B.18 The Tool Invocation Lifecycle
  1. B.18.1 Checkpoint
  2. B.18.2 Validate
  3. B.18.3 Execute
  4. B.18.4 Protocol error
  5. B.18.5 Tool execution error
  6. B.18.6 Retry
  7. B.18.7 Reissue
  8. B.18.8 idempotentHint
19. B.19 Two Ways for a Request to Fail
  1. B.19.1 Protocol error
  2. B.19.2 Tool execution error
  3. B.19.3 Allocation policy
  4. B.19.4 HeaderMismatch
  5. B.19.5 MissingRequiredClientCapabilityError
  6. B.19.6 UnsupportedProtocolVersionError
  7. B.19.7 Retired code
  8. B.19.8 safe_error
20. B.20 Transports and the HTTP Header Contract
  1. B.20.1 stdio: a subprocess sharing three streams
  2. B.20.2 Streamable HTTP: one endpoint, one message per POST
  3. B.20.3 The header mirror and its contract
21. B.21 Cache Freshness and Cursor-Based Pagination
  1. B.21.1 CacheableResult
  2. B.21.2 ttlMs
  3. B.21.3 cacheScope
  4. B.21.4 Cache key
  5. B.21.5 Cursor
  6. B.21.6 nextCursor
  7. B.21.7 list_changed notification
  8. B.21.8 MRTR retried result
22. B.22 Long-Running Work and the Tasks Extension
  1. B.22.1 io.modelcontextprotocol/tasks
  2. B.22.2 CreateTaskResult
  3. B.22.3 tasks/get
  4. B.22.4 tasks/update
  5. B.22.5 tasks/cancel
  6. B.22.6 input_required
  7. B.22.7 pollIntervalMs
  8. B.22.8 ttlMs
  9. B.22.9 Durable before return
  10. B.22.10 Cooperative cancellation
23. B.23 Trust Zones in an MCP Exchange
  1. B.23.1 Trust zone
  2. B.23.2 Trust boundary
  3. B.23.3 Self-reported identity
  4. B.23.4 Multi-server isolation
  5. B.23.5 Untrusted annotation
24. B.24 Authorizing Access to an MCP Server
  1. B.24.1 Resource server
  2. B.24.2 Authorization server
  3. B.24.3 Protected Resource Metadata
  4. B.24.4 PKCE
  5. B.24.5 S256
  6. B.24.6 Resource indicator
  7. B.24.7 iss validation
  8. B.24.8 Audience validation
  9. B.24.9 Token passthrough
  10. B.24.10 401 vs 403 vs 400
25. B.25 Proving a Client's Identity to an Authorization Server
  1. B.25.1 Client registration
  2. B.25.2 Client ID Metadata Document (CIMD)
  3. B.25.3 Dynamic Client Registration (DCR)
  4. B.25.4 Pre-registration
  5. B.25.5 Authorization server binding
  6. B.25.6 application_type
  7. B.25.7 Confused deputy
  8. B.25.8 Authorization extension
26. B.26 Consent and Least Privilege
  1. B.26.1 Asking without a session: elicitation over MRTR
  2. B.26.2 Scoping consent to one tool, not one server
  3. B.26.3 Step-up authorization: a different gate, one level down
  4. B.26.4 Least privilege in what a server admits it has
27. B.27 Risk and Safety Controls for MCP Tool Calls
  1. B.27.1 Metadata poisoning
  2. B.27.2 Rug pull
  3. B.27.3 Definition pinning
  4. B.27.4 Tool shadowing
  5. B.27.5 Confused deputy
  6. B.27.6 Token passthrough
  7. B.27.7 requestState tampering
  8. B.27.8 SSRF
  9. B.27.9 DNS rebinding
  10. B.27.10 Supply chain drift
28. B.28 One Trace ID Ties Auditability to Observability
  1. B.28.1 traceparent
  2. B.28.2 Trace id
  3. B.28.3 Parent id
  4. B.28.4 tracestate
  5. B.28.5 baggage
  6. B.28.6 Principal
  7. B.28.7 Redaction
  8. B.28.8 Hash chain
  9. B.28.9 Result channel
  10. B.28.10 Correlation
29. B.29 Every MUST Needs an Owner
  1. B.29.1 Server author
  2. B.29.2 Host and client developer
  3. B.29.3 Platform or gateway operator
  4. B.29.4 Security and governance owner
  5. B.29.5 Registry publisher
  6. B.29.6 End user
  7. B.29.7 Working Group
  8. B.29.8 Interest Group
  9. B.29.9 SEP
  10. B.29.10 SDK tier
30. B.30 Choosing MCP's Shape for the Job
  1. B.30.1 Operational use case
  2. B.30.2 Control split
  3. B.30.3 Tasks extension
  4. B.30.4 MCP Apps
  5. B.30.5 Skills over MCP
  6. B.30.6 Authorization extension
  7. B.30.7 cacheScope
  8. B.30.8 Graceful degradation
31. B.31 The Extensions Framework
  1. B.31.1 Extension
  2. B.31.2 Vendor prefix
  3. B.31.3 Settings object
  4. B.31.4 Active extension set
  5. B.31.5 Graceful degradation
  6. B.31.6 MissingRequiredClientCapabilityError
  7. B.31.7 Extension repository
  8. B.31.8 Experimental extension
32. B.32 Interactive Interfaces Inside the Conversation
  1. B.32.1 MCP Apps
  2. B.32.2 io.modelcontextprotocol/ui
  3. B.32.3 ui://
  4. B.32.4 _meta.ui.resourceUri
  5. B.32.5 text/html;profile=mcp-app
  6. B.32.6 _meta.ui.csp
  7. B.32.7 _meta.ui.permissions
  8. B.32.8 visibility
  9. B.32.9 Sandboxed iframe
  10. B.32.10 Sandbox proxy
  11. B.32.11 App-to-host bridge
  12. B.32.12 Text fallback
33. B.33 Finding, Routing To, and Trusting a Server
  1. B.33.1 The registry: metadata, not code
  2. B.33.2 Gateways: routing on the wire, not on trust
  3. B.33.3 SDK tiers: a portability signal, not a scan
34. B.34 Reading One MCP Exchange End to End
  1. B.34.1 Protocol error
  2. B.34.2 Tool execution error
  3. B.34.3 MissingRequiredClientCapability
  4. B.34.4 MRTR
  5. B.34.5 requestState
  6. B.34.6 Tasks extension
  7. B.34.7 Canonical resource URI
  8. B.34.8 traceparent
  9. B.34.9 Hash chained audit log
35. B.35 MCPA: Model Context Protocol Associate
  1. B.35.1 MCP Fundamentals (16%)
  2. B.35.2 Architecture and Components (14%)
  3. B.35.3 Interactions and Execution (26%)
  4. B.35.4 Security and Governance (24%)
  5. B.35.5 Use Cases and Ecosystem (20%)
36. B.36 MCPA diagnostic: MCPA Foundations Diagnostic
  1. B.36.1 MCP Fundamentals: 5 questions
  2. B.36.2 Architecture and Components: 4 questions
  3. B.36.3 Interactions and Execution: 8 questions
  4. B.36.4 Security and Governance: 7 questions
  5. B.36.5 Use Cases and Ecosystem: 6 questions
37. B.37 MCPA mock: MCPA Full-Length Mock 1: Operational Scenarios
  1. B.37.1 MCP Fundamentals: 10 questions
  2. B.37.2 Architecture and Components: 8 questions
  3. B.37.3 Interactions and Execution: 16 questions
  4. B.37.4 Security and Governance: 14 questions
  5. B.37.5 Use Cases and Ecosystem: 12 questions
38. B.38 MCPA mock: MCPA Full-Length Mock 2: Wire-Level Messages
  1. B.38.1 MCP Fundamentals: 10 questions
  2. B.38.2 Architecture and Components: 8 questions
  3. B.38.3 Interactions and Execution: 16 questions
  4. B.38.4 Security and Governance: 14 questions
  5. B.38.5 Use Cases and Ecosystem: 12 questions
39. B.39 MCPA mock: MCPA Full-Length Mock 3: Design and Security Trade-offs
  1. B.39.1 MCP Fundamentals: 10 questions
  2. B.39.2 Architecture and Components: 8 questions
  3. B.39.3 Interactions and Execution: 16 questions
  4. B.39.4 Security and Governance: 14 questions
  5. B.39.5 Use Cases and Ecosystem: 12 questions
CThe manualsnot yet
1. C.1 A2A 101
  1. C.1.1 1.1 What A2A standardizes
  2. C.1.2 1.2 One task, request by request
  3. C.1.3 2.1 AgentCard fields
  4. C.1.4 2.2 Discovery and supportedInterfaces
  5. C.1.5 2.3 Extended cards and signatures
  6. C.1.6 3.1 Message
  7. C.1.7 3.2 Part
  8. C.1.8 3.3 Artifact and chunks
  9. C.1.9 3.4 Task, TaskStatus, and TaskState
  10. C.1.10 3.5 contextId
  11. C.1.11 4.1 SendMessage
  12. C.1.12 4.2 Interrupted states: INPUT_REQUIRED and AUTH_REQUIRED
  13. C.1.13 4.3 GetTask and ListTasks
  14. C.1.14 4.4 CancelTask and terminal states
  15. C.1.15 4.5 SendStreamingMessage and SubscribeToTask
  16. C.1.16 4.6 Push notification configs
  17. C.1.17 5.1 JSON-RPC
  18. C.1.18 5.2 HTTP+JSON
  19. C.1.19 5.3 gRPC
  20. C.1.20 5.4 A2A-Version and A2A-Extensions
  21. C.1.21 5.5 Errors
  22. C.1.22 6.1 Security schemes and in-task authorization
  23. C.1.23 6.2 Extensions
  24. C.1.24 6.3 Security requirements
  25. C.1.25 7.1 A client
  26. C.1.26 7.2 A server
  27. C.1.27 7.3 Conformance testing
  28. C.1.28 7.4 From 0.3 to 1.0
  29. C.1.29 R.1 Operations by binding
  30. C.1.30 R.2 Objects and fields
  31. C.1.31 R.3 Task states
  32. C.1.32 R.4 Errors
  33. C.1.33 R.5 Glossary
  34. C.1.34 R.6 Sources
  35. C.1.35 R.7 Index of figures
2. C.2 Docker Sandboxes and Docker Agent 101
  1. C.2.1 1.1 sbx and docker-agent
  2. C.2.2 1.2 docker-agent run --sandbox, one run end to end
  3. C.2.3 2.1 sbx run, sbx create, sbx stop, and sbx rm
  4. C.2.4 2.2 sbx exec, sbx cp, and sbx ports
  5. C.2.5 2.3 sbx template save and sbx template load
  6. C.2.6 2.4 sandboxd, sbx diagnose, and sbx reset
  7. C.2.7 2.5 sbx settings list
  8. C.2.8 3.1 sbx diagnose and the microVM boundary
  9. C.2.9 3.2 --clone and /run/sandbox/source
  10. C.2.10 3.3 sbx policy init, allow network, deny network, and check network
  11. C.2.11 3.4 sbx policy log and the proxy
  12. C.2.12 3.5 sbx secret set, sbx secret import, and sbx secret set-custom
  13. C.2.13 3.6 sbx mcp add, sbx mcp load, and --static-mcp
  14. C.2.14 4.1 `# syntax=docker/sandbox-kit:3` and kit.yaml
  15. C.2.15 4.2 `com.docker.sandbox/*` capabilities
  16. C.2.16 4.3 `sbx kit pack`, `sbx kit push --sign`, and `sbx kit verify`
  17. C.2.17 4.4 sbxenv.yaml and `sbx env plan`
  18. C.2.18 4.5 `sbx skills add` and `skills: true`
  19. C.2.19 5.1 docker-agent run, new, and doctor
  20. C.2.20 5.2 agents, models, and providers
  21. C.2.21 5.3 toolsets and mcps
  22. C.2.22 5.4 sub_agents, transfer_task, and background_agents
  23. C.2.23 5.5 permissions, --safety, and hooks
  24. C.2.24 5.6 session.db, sessions diff, and eval
  25. C.2.25 6.1 serve api and serve chat
  26. C.2.26 6.2 serve mcp and serve acp
  27. C.2.27 6.3 serve a2a
  28. C.2.28 6.4 share push and share pull
  29. C.2.29 6.5 DMR, Compose models, and the docker/mcp-gateway service
  30. C.2.30 7.1 sbx --cloud run, sbx move, and sbx ttl
  31. C.2.31 7.2 sbx policy ls --source org, --profile, and the audit JSONL
  32. C.2.32 7.3 From docker sandbox and cagent to sbx and docker-agent
  33. C.2.33 7.4 kit-tck, sbx diagnose, and the claims this manual checked
  34. C.2.34 R.1 sbx commands
  35. C.2.35 R.2 docker-agent commands
  36. C.2.36 R.3 Settings keys, environment variables, and paths
  37. C.2.37 R.4 Agent file keys
  38. C.2.38 R.5 Toolsets
  39. C.2.39 R.6 Providers and models
  40. C.2.40 R.7 Sources and the conflicts register
  41. C.2.41 R.8 Glossary and index of figures
3. C.3 Miles 101
  1. C.3.1 1.1 Run GRPO on one GPU with the radixark/miles image
  2. C.3.2 1.2 Add a watchdog before you leave a run alone
  3. C.3.3 1.3 Read your first ten steps
  4. C.3.4 2.1 Pick a model size for one GPU
  5. C.3.5 2.2 Choose what to offload between phases
  6. C.3.6 2.3 Stop Ray from killing the job on unified memory
  7. C.3.7 2.4 Size the weight update
  8. C.3.8 3.1 Train math reasoning on GSM8K
  9. C.3.9 3.2 Train with the DAPO filter
  10. C.3.10 3.3 Train an agent with tool calls and many turns
  11. C.3.11 3.4 Train a mixture-of-experts model with routing replay
  12. C.3.12 3.5 Write a custom reward function
  13. C.3.13 4.1 Run rollout and training at the same time
  14. C.3.14 4.2 Handle long responses with partial rollout
  15. C.3.15 4.3 Roll out in FP8 and train in BF16
  16. C.3.16 4.4 Lower the log probability gap with top-p and top-k
  17. C.3.17 4.5 Know what deterministic inference fixes
  18. C.3.18 4.6 Check if your GPU can run true on-policy mode
  19. C.3.19 5.1 Read the run from its signals
  20. C.3.20 5.2 Check the first step before you trust the run
  21. C.3.21 5.3 Repeat a run and compare
  22. C.3.22 6.1 Split rollout and training across GPUs
  23. C.3.23 6.2 Run evaluation without stopping training
  24. C.3.24 6.3 Keep a long run going after a failure
  25. C.3.25 6.4 Pick an image for your GPU
  26. C.3.26 7.1 Troubleshoot by symptom
  27. C.3.27 7.2 Keep runs comparable
  28. C.3.28 8.1 Why an RL framework owns an inference engine
  29. C.3.29 8.2 The correction methods for the gap
  30. C.3.30 8.3 Why true on-policy mode is a contract
  31. C.3.31 8.4 Memory and weight updates on unified memory
  32. C.3.32 8.5 How rollout functions plug in
  33. C.3.33 8.6 Which router to use
  34. C.3.34 8.7 Where Miles came from
  35. C.3.35 R.1 Flags by job
  36. C.3.36 R.2 Metrics
  37. C.3.37 R.3 Lab notes
  38. C.3.38 R.4 Sources
  39. C.3.39 R.5 Index of figures
4. C.4 Reverse Engineering 101
  1. C.4.1 1.1 rea, providers, and Evidence
  2. C.4.2 1.2 rea function, end to end
  3. C.4.3 2.1 The README use rule and the law
  4. C.4.4 2.2 rea doctor and the Ghidra provider
  5. C.4.5 2.3 rea capabilities
  6. C.4.6 3.1 inspect-artifact
  7. C.4.7 3.2 inspect-binary-layout and the load image
  8. C.4.8 3.3 address-to-file-offset
  9. C.4.9 3.4 analyze and the binary_overview record
  10. C.4.10 4.1 open_binary and the headless bridge
  11. C.4.11 4.2 list_procedures and function bodies
  12. C.4.12 4.3 procedure_assembly
  13. C.4.13 4.4 xrefs and the jump table
  14. C.4.14 4.5 list_strings
  15. C.4.15 4.6 analyze_function and pseudocode
  16. C.4.16 5.1 The Evidence record
  17. C.4.17 5.2 ResidualUnknown, record_unknown, and coverage
  18. C.4.18 5.3 AnalysisSnapshot and the exact-match answer cache
  19. C.4.19 5.4 export_evidence_bundle, evidence-export, and evidence-import
  20. C.4.20 5.5 compare_functions and rea compare
  21. C.4.21 6.1 analyze-javascript-application and the application graph
  22. C.4.22 6.2 inspect-wasm-artifact and the WABT provider
  23. C.4.23 7.1 verify_reconstruction
  24. C.4.24 7.2 A proof harness beyond compare_functions
  25. C.4.25 8.1 rea mcp: initialize, tools/list, and the startup deadline
  26. C.4.26 8.2 One agent investigation over list_procedures, analyze_function, and list_strings
  27. C.4.27 R.1 Commands and tools used in this manual
  28. C.4.28 R.2 Evidence fields
  29. C.4.29 R.3 Errors seen in the captures
  30. C.4.30 R.4 Glossary
  31. C.4.31 R.5 Sources
  32. C.4.32 R.6 Index of figures
5. C.5 SGLang 101
  1. C.5.1 1.1 Install SGLang and serve your first model
  2. C.5.2 1.2 Check that the server is healthy
  3. C.5.3 1.3 What the server does with your request
  4. C.5.4 2.1 sglang serve and the three processes
  5. C.5.5 2.2 The scheduler loop and --schedule-policy
  6. C.5.6 2.3 Admission and retraction
  7. C.5.7 2.4 The overlap scheduler
  8. C.5.8 2.5 Streaming, stop strings, and the OpenAI API
  9. C.5.9 3.1 The KV pool and bytes per token
  10. C.5.10 3.2 --mem-fraction-static and startup sizing
  11. C.5.11 3.3 RadixCache
  12. C.5.12 3.4 Prefix-friendly prompts
  13. C.5.13 3.5 Eviction policies and HiCache
  14. C.5.14 3.6 Hybrid models and the cache
  15. C.5.15 4.1 ForwardMode and CUDA graphs
  16. C.5.16 4.2 --attention-backend
  17. C.5.17 4.3 --quantization and --kv-cache-dtype
  18. C.5.18 4.4 --speculative-algorithm
  19. C.5.19 4.5 --grammar-backend and structured outputs
  20. C.5.20 4.6 --enable-deterministic-inference and logprobs
  21. C.5.21 5.1 sglang.benchmark.serving
  22. C.5.22 5.2 The /metrics endpoint
  23. C.5.23 5.3 Reading other people's SGLang benchmarks
  24. C.5.24 6.1 --tp-size, --dp-size, and --attn-dp-size
  25. C.5.25 6.2 MoE serving: --moe-a2a-backend, DeepEP, and EPLB
  26. C.5.26 6.3 Prefill-decode disaggregation: --disaggregation-mode
  27. C.5.27 6.4 sgl-model-gateway and cache-aware routing
  28. C.5.28 7.1 Operating SGLang
  29. C.5.29 7.2 SGLang as an RL rollout engine
  30. C.5.30 R.1 Flags and defaults
  31. C.5.31 R.2 Metrics
  32. C.5.32 R.3 Incident index
  33. C.5.33 R.4 Glossary
  34. C.5.34 R.5 Sources
  35. C.5.35 R.6 Index of figures
  36. C.5.36 R.7 How this manual was made
6. C.6 vLLM 101
  1. C.6.1 1.1 Install vLLM and serve your first model
  2. C.6.2 1.2 Check that the server is healthy
  3. C.6.3 1.3 Understand what happens to your request
  4. C.6.4 1.4 Set the server options every deployment needs
  5. C.6.5 2.1 Count the memory before you pick a GPU
  6. C.6.6 2.2 Fix out of memory at startup
  7. C.6.7 2.3 Size memory on a DGX Spark
  8. C.6.8 2.4 Fit more users with an FP8 KV cache
  9. C.6.9 2.5 Fit a bigger model with FP8 weights
  10. C.6.10 2.6 Serve long contexts and hybrid models
  11. C.6.11 3.1 Tune chat for many users
  12. C.6.12 3.2 Make repeated prompts hit the prefix cache
  13. C.6.13 3.3 Keep long prompts from stalling short ones
  14. C.6.14 3.4 Keep long sessions cheap
  15. C.6.15 3.5 Return valid JSON and tool calls
  16. C.6.16 3.6 Run batch and offline jobs
  17. C.6.17 3.7 Serve many fine-tunes with multi-LoRA
  18. C.6.18 3.8 Get reproducible outputs
  19. C.6.19 3.9 Use vLLM as an RL rollout engine
  20. C.6.20 3.10 Serve reasoning models
  21. C.6.21 4.1 Speed up decode with EAGLE3
  22. C.6.22 4.2 Use n-gram speculation for copy-heavy work
  23. C.6.23 4.3 Keep CUDA graphs and compile on
  24. C.6.24 4.4 Leave async scheduling on
  25. C.6.25 4.5 Change the attention backend only for a reason
  26. C.6.26 5.1 Benchmark your server in 15 minutes
  27. C.6.27 5.2 Watch vLLM in production
  28. C.6.28 5.3 Decide between vLLM and SGLang
  29. C.6.29 6.1 Choose tensor, pipeline, or data parallelism
  30. C.6.30 6.2 Serve MoE models with expert parallelism
  31. C.6.31 6.3 Split prefill and decode
  32. C.6.32 6.4 Run vLLM on Kubernetes
  33. C.6.33 6.5 Run vLLM on other accelerators
  34. C.6.34 7.1 Troubleshoot by symptom
  35. C.6.35 7.2 Run vLLM in production
  36. C.6.36 8.1 Why vLLM runs three kinds of process
  37. C.6.37 8.2 Why one token budget replaces a prefill phase
  38. C.6.38 8.3 How the block pool manages KV memory
  39. C.6.39 8.4 How vLLM picks an attention backend
  40. C.6.40 8.5 What Model Runner V2 changed
  41. C.6.41 8.6 How compile and CUDA graphs are captured
  42. C.6.42 8.7 How speculative decoding checks drafts
  43. C.6.43 8.8 How KV connectors move the cache
  44. C.6.44 8.9 Where vLLM came from
  45. C.6.45 R.1 Flags and defaults
  46. C.6.46 R.2 Metrics
  47. C.6.47 R.3 Incident index
  48. C.6.48 R.4 Version pins
  49. C.6.49 R.5 Glossary
  50. C.6.50 R.6 Sources
  51. C.6.51 R.7 Lab notes
  52. C.6.52 R.8 Index of figures
DInteractive explainersnot yet
1. D.1 Inside an AI Agent
  1. D.1.1 Intake hopper
  2. D.1.2 Memory drum
  3. D.1.3 Filing cabinet and pick head
  4. D.1.4 Sealed model housing
  5. D.1.5 Tool bay
  6. D.1.6 Verifier gauge
  7. D.1.7 Exit latch and output tray
  8. D.1.8 Trace printer
  9. D.1.9 Runtime controller
  10. D.1.10 Card carriage
  11. D.1.11 Try to break it: Healthy run
  12. D.1.12 Try to break it: No retrieval
  13. D.1.13 Try to break it: No retrieval, no verifier
  14. D.1.14 Try to break it: Tool times out once
  15. D.1.15 Try to break it: Tool always times out
  16. D.1.16 Try to break it: Timeouts and no verifier
  17. D.1.17 Try to break it: Malformed tool call
  18. D.1.18 Try to break it: Tool access denied
  19. D.1.19 Tour: One request, one bench
  20. D.1.20 Tour: The model only proposes
  21. D.1.21 Tour: Tools pass two checks first
  22. D.1.22 Tour: Take away the documents
  23. D.1.23 Tour: Turn off the check
  24. D.1.24 Tour: Put the check back
2. D.2 Inside a Decompiler
  1. D.2.1 ELF file card
  2. D.2.2 Section cabinet and symbol index
  3. D.2.3 Decoder chip and output bed
  4. D.2.4 Routing board and drop bin
  5. D.2.5 Lift press and IR stack
  6. D.2.6 Register and stack slots
  7. D.2.7 Structuring jig
  8. D.2.8 Pseudocode printer
  9. D.2.9 Evidence ledger
  10. D.2.10 Analysis carriage
  11. D.2.11 Try to break it: Healthy run
  12. D.2.12 Try to break it: Unoptimized build
  13. D.2.13 Try to break it: Strip the symbols
  14. D.2.14 Try to break it: Linear sweep at -O0
  15. D.2.15 Try to break it: Linear sweep, no symbols
  16. D.2.16 Try to break it: Data in code, linear sweep
  17. D.2.17 Try to break it: Data in code, descent
  18. D.2.18 Tour: Read the header
  19. D.2.19 Tour: Open the sections
  20. D.2.20 Tour: Decode the bytes
  21. D.2.21 Tour: Join the blocks
  22. D.2.22 Tour: Drop dead blocks
  23. D.2.23 Tour: Lift to plain steps
  24. D.2.24 Tour: Or stop and refuse
  25. D.2.25 Tour: Find the types
  26. D.2.26 Tour: Find the if
  27. D.2.27 Tour: Print the code
  28. D.2.28 Tour: Sort the evidence
3. D.3 Inside an LLM Serving GPU
  1. D.3.1 Network port
  2. D.3.2 Tokenizer process
  3. D.3.3 Scheduler process
  4. D.3.4 PCIe Gen4 x16 slot and riser (L4) or NVLink-C2C (GB10)
  5. D.3.5 SMs wait for weights). Lesson
  6. D.3.6 so each decode step streams weights from GDDR6 again; GB10 has 24 MB. Weights are used once per step and are far larger than the cache. Manual section 4.1. - 07 WEIGHTS, Weights in GDDR6 or LPDDR5X
  7. D.3.7 sized at startup from the memory fraction and counted in tokens; cell colours mark this request, other requests, reused prefix and free. Without it every step recomputes keys and values for the whole conversation. Manual section 3.1, KV bytes per token and the KV pool. - 09 DETOKENIZER
  8. D.3.8 substrate, die and balls (L4) or two dielets on one substrate (GB10)
  9. D.3.9 Passive heatsink (L4, 72 W, server fans push air) or GB10 (140 W TDP)
  10. D.3.10 Try to break it: Healthy L4
  11. D.3.11 Try to break it: One user
  12. D.3.12 Try to break it: Pool full
  13. D.3.13 Try to break it: FP8 at 64 users
  14. D.3.14 Try to break it: Small pool, long prefix
  15. D.3.15 Try to break it: Same, with FP8
  16. D.3.16 Try to break it: Radix cache off
  17. D.3.17 Try to break it: CUDA graphs off
  18. D.3.18 Try to break it: Hot card
  19. D.3.19 Try to break it: GB10, one user
  20. D.3.20 Try to break it: GB10 with FP8
  21. D.3.21 Try to break it: Break button notes
  22. D.3.22 Tour: Prompt in
  23. D.3.23 Tour: Words to numbers
  24. D.3.24 Tour: Find room
  25. D.3.25 Tour: Send to the GPU
  26. D.3.26 Tour: Read the prompt
  27. D.3.27 Tour: Write the answer
  28. D.3.28 Tour: Send it back
EPractice drillsnot yet
1. E.1 Math foundations
  1. E.1.1 Linear regression by gradient descent
  2. E.1.2 Vectorized pairwise distances
  3. E.1.3 Power iteration for the top eigenpair
2. E.2 Classical ML
  1. E.2.1 Gini impurity and the best split
  2. E.2.2 Linear regression by gradient descent
  3. E.2.3 K-means assignment and update
  4. E.2.4 Measure bias and variance
  5. E.2.5 Logistic regression from scratch
  6. E.2.6 Train softmax regression on four blobs
3. E.3 Deep learning core
  1. E.3.1 Adam and AdamW step
  2. E.3.2 Batch normalization forward and backward
  3. E.3.3 CLIP symmetric contrastive loss
  4. E.3.4 DDPM noise schedule, closed form noising and posterior
  5. E.3.5 Focal loss
  6. E.3.6 Inverted dropout
  7. E.3.7 LayerNorm and RMSNorm
  8. E.3.8 Backpropagation through a two-layer MLP
  9. E.3.9 Rectified flow loss and Euler sampler
  10. E.3.10 Stable softmax and cross-entropy
  11. E.3.11 Fit a sine curve with a tanh network
  12. E.3.12 Train a tanh network on two spirals
  13. E.3.13 Train a 2-4-1 network on XOR
4. E.4 Optimizers
  1. E.4.1 Adam and AdamW step
  2. E.4.2 Linear regression by gradient descent
  3. E.4.3 Sweep the learning rate
  4. E.4.4 SGD with momentum
5. E.5 Transformers from scratch
  1. E.5.1 Tiled attention with an online softmax
  2. E.5.2 Grouped-query attention
  3. E.5.3 Fixed-capacity KV cache
  4. E.5.4 LayerNorm and RMSNorm
  5. E.5.5 Top-k MoE gating and the Switch load-balancing loss
  6. E.5.6 Split heads, merge heads, multi-head attention
  7. E.5.7 Rotary position embedding
  8. E.5.8 Scaled dot-product attention with a mask
  9. E.5.9 Sinusoidal positional encoding
6. E.6 LLM inference
  1. E.6.1 Beam search with EOS and length normalization
  2. E.6.2 BPE pair counts and merges
  3. E.6.3 Symmetric int8 quantization
  4. E.6.4 Fixed-capacity KV cache
  5. E.6.5 Sweep the quantization bit width
  6. E.6.6 Sweep the sampling temperature
  7. E.6.7 Read an SGLang sweep on an L4
  8. E.6.8 LoRA linear layer
  9. E.6.9 Speculative decoding: accept, reject and resample
  10. E.6.10 Temperature, top-k and top-p sampling
  11. E.6.11 Train a character bigram language model
7. E.7 RL basics
  1. E.7.1 Discounted returns with episode ends
  2. E.7.2 DPO loss and implicit rewards
  3. E.7.3 Generalized advantage estimation
  4. E.7.4 GRPO group advantages and clipped objective
  5. E.7.5 PPO clipped surrogate loss
  6. E.7.6 Q-learning and SARSA updates
8. E.8 RAG building blocks
  1. E.8.1 BM25 ranking
  2. E.8.2 Cosine similarity top-k search
  3. E.8.3 Reciprocal rank fusion
9. E.9 Agent building blocks
  1. E.9.1 Parse and validate tool calls
10. E.10 Papers to code
  1. E.10.1 CLIP symmetric contrastive loss
  2. E.10.2 DDPM noise schedule, closed form noising and posterior
  3. E.10.3 DPO loss and implicit rewards
  4. E.10.4 Tiled attention with an online softmax
  5. E.10.5 Focal loss
  6. E.10.6 Grouped-query attention
  7. E.10.7 GRPO group advantages and clipped objective
  8. E.10.8 Top-k MoE gating and the Switch load-balancing loss
  9. E.10.9 Rectified flow loss and Euler sampler
  10. E.10.10 Speculative decoding: accept, reject and resample
11. E.11 Train tiny models
  1. E.11.1 Train a character bigram language model
  2. E.11.2 Fit a sine curve with a tanh network
  3. E.11.3 Train softmax regression on four blobs
  4. E.11.4 Train a tanh network on two spirals
  5. E.11.5 Train a 2-4-1 network on XOR
12. E.12 Experiment labs
  1. E.12.1 Measure bias and variance
  2. E.12.2 Sweep the learning rate
  3. E.12.3 Sweep the quantization bit width
  4. E.12.4 Sweep the sampling temperature
  5. E.12.5 Read an SGLang sweep on an L4
FProjectsnot yet
1. F.1 Level 1 projects
  1. F.1.1 Bookmark Path Organizer
  2. F.1.2 CSV Repair Workbench
  3. F.1.3 Dataset Split Auditor
  4. F.1.4 Download Folder Sort Desk
  5. F.1.5 Glossary Hovercard Publisher
  6. F.1.6 JSON Schema Output Guard
  7. F.1.7 Prompt Regression Tester
  8. F.1.8 SKILL.md Validator and Loader
  9. F.1.9 Semantic Notes Search
  10. F.1.10 Tiny Coding Agent
  11. F.1.11 Token Counter and Cost Meter
  12. F.1.12 Vocabulary Autocomplete Pad
2. F.2 Level 2 projects
  1. F.2.1 API Tutorial Runner
  2. F.2.2 CSV Question Workbench
  3. F.2.3 Calendar Focus Planner
  4. F.2.4 Catalog Record Linker
  5. F.2.5 Changelog Writer From Git
  6. F.2.6 Chart Storyboard Builder
  7. F.2.7 Configuration Drift Explainer
  8. F.2.8 Dataset Contract Gate
  9. F.2.9 Diagram Reading Companion
  10. F.2.10 Document Extraction Review Desk
  11. F.2.11 Document QA With Citations and LangChain
  12. F.2.12 Feedback Theme Board
  13. F.2.13 Field Notes Map Builder
  14. F.2.14 Inbox Triage Desk
  15. F.2.15 Incident Postmortem Writer
  16. F.2.16 Infrastructure Plan Explainer
  17. F.2.17 Local Model Evaluation Harness
  18. F.2.18 Log Volume Reduction Lab
  19. F.2.19 Meeting Notes to Actions
  20. F.2.20 PR Review Reporter
  21. F.2.21 Plain Language Rewrite Desk
  22. F.2.22 Research Report Agent
  23. F.2.23 Retrieval Evaluation Lab
  24. F.2.24 Sensitive Text Redaction Gate
  25. F.2.25 Service Ownership Navigator
  26. F.2.26 Skill Router
  27. F.2.27 Source-Grounded Study Coach
  28. F.2.28 Subtitle Localization Desk
  29. F.2.29 UI String Localization Workbench
  30. F.2.30 Workshop Materials Planner
3. F.3 Level 3 projects
  1. F.3.1 Agent Budget Planner
  2. F.3.2 Agent Trace Debugger
  3. F.3.3 Binary Inspector
  4. F.3.4 Community Learning Path Finder
  5. F.3.5 Cross-Agent Skill Installer
  6. F.3.6 Database Migration Rehearsal
  7. F.3.7 Decision Tradeoff Explorer
  8. F.3.8 Harness Bench
  9. F.3.9 Kubernetes Event Storyboard
  10. F.3.10 LLM Gateway With Fallbacks
  11. F.3.11 Multi-Agent Code Review Panel
  12. F.3.12 Persistent Memory Server
  13. F.3.13 RAG Freshness Pipeline
  14. F.3.14 Report Judge
  15. F.3.15 Retry Storm Lab
  16. F.3.16 Self-Correcting Workflow Hooks
  17. F.3.17 Skill Supply-Chain Scanner
  18. F.3.18 Streaming Answer Recovery
  19. F.3.19 Support Agent With Google ADK
  20. F.3.20 Typed Workflow Agent with Mastra
  21. F.3.21 Visual Evidence Library
  22. F.3.22 Voice Note Transcriber Pipeline
  23. F.3.23 Web Change Brief
4. F.4 Level 4 projects
  1. F.4.1 Cloud Agent With AWS Strands
  2. F.4.2 Desktop Control Backend
  3. F.4.3 Durable Agent Jobs
  4. F.4.4 Hedged Request Lab
  5. F.4.5 MCP Tool Discovery Workbench
  6. F.4.6 Sandbox Policy Planner
  7. F.4.7 Signed Webhook Inbox
  8. F.4.8 Streaming Agent Shell in Rust
  9. F.4.9 Tool Call Firewall
5. F.5 Level 5 projects
  1. F.5.1 Browser Agent
  2. F.5.2 Distributed Eval Farm Coordinator
  3. F.5.3 Mini Decompiler
  4. F.5.4 Self-Improving Skill Loop
6. F.6 Planned projects
  1. F.6.1 Photo Burst Contact Sheet
  2. F.6.2 Dependency Upgrade Impact Map
  3. F.6.3 Gesture Shortcut Pad
  4. F.6.4 CI Failure Reproducer
  5. F.6.5 Label Adjudication Desk
  6. F.6.6 Lost Item Match Desk
  7. F.6.7 Presentation Rehearsal Coach
  8. F.6.8 Semantic Cache Correctness Lab
  9. F.6.9 Tool Contract Migration Checker
  10. F.6.10 Video Highlight Storyboard
  11. F.6.11 A2A Task Handoff Workbench
  12. F.6.12 Abstention Calibration Workbench
  13. F.6.13 Artifact Provenance Gate
  14. F.6.14 Context Compaction Verifier
  15. F.6.15 Local Inference Capacity Planner
  16. F.6.16 MCP Compatibility Doctor
  17. F.6.17 Metamorphic Evaluation Workbench
  18. F.6.18 OAuth Resource Boundary Lab
  19. F.6.19 Outbound Request Policy Proxy
  20. F.6.20 Release Canary Decision Lab
  21. F.6.21 Secret Injection Sidecar
  22. F.6.22 Tenant Fairness Scheduler
  23. F.6.23 Agent Resource Deadlock Lab
  24. F.6.24 Judge Bias Calibration Lab
  25. F.6.25 Trajectory Counterexample Minimizer
  26. F.6.26 Transactional Workspace Patch Broker
GLearning pathsnot yet
1. G.1 Agent Skills Engineering
  1. G.1.1 Agent Skills: Portable Contract and Runtime Boundary (see 15.22)
  2. G.1.2 Skill Discovery and Progressive Disclosure (see 15.24)
  3. G.1.3 Skill Invocation and Routing (see 15.25)
  4. G.1.4 Skill Permissions, Sandboxes, and Trust (see 15.26)
  5. G.1.5 Skill Evals, Packaging, and Portability (see 15.27)
  6. G.1.6 Capstone: Stateless Tool Ecosystem (see 15.23) (optional)
2. G.2 Agent Systems Engineering
  1. G.2.1 Common Core: The Tool Interface (see 15.1)
  2. G.2.2 Common Core: Tool Schema Design (see 15.5)
  3. G.2.3 Common Core: MCP Fundamentals: Stateless Requests and JSON-RPC (see 15.6)
  4. G.2.4 Common Core: The Agent Loop (see 16.1)
  5. G.2.5 Common Core: Tool Use and Function Calling (see 16.6)
  6. G.2.6 Role Practice: Context Engineering (see 13.5)
  7. G.2.7 Role Practice: Agent Memory — Virtual Context and Memory Paging (see 16.7)
  8. G.2.8 Role Practice: Stateful Graph Orchestration — Durable Execution and Checkpoints (see 16.13)
  9. G.2.9 Role Practice: Orchestration Patterns — Supervisor, Swarm, Hierarchical (see 16.28)
  10. G.2.10 Proof Project: Failure Modes — Why Agents Break (see 16.26)
  11. G.2.11 Proof Project: Prompt Injection and the PVE Defense (see 16.27)
  12. G.2.12 Proof Project: Eval-Driven Agent Development (see 16.30)
  13. G.2.13 Interview and Readiness Evidence: Production Runtimes — Queue, Event, Cron (see 16.29)
  14. G.2.14 Interview and Readiness Evidence: Agent Observability — Langfuse, Phoenix, Opik (see 16.24)
3. G.3 AI Data Systems
  1. G.3.1 Common Core: Data Management (see 2.9)
  2. G.3.2 Common Core: Feature Engineering and Selection (see 4.8)
  3. G.3.3 Common Core: ML Pipelines and Experiment Tracking (see 4.13)
  4. G.3.4 Common Core: Data Pipelines for Pre-Training (see 12.3)
  5. G.3.5 Role Practice: Embeddings and Vector Representations (see 13.4)
  6. G.3.6 Role Practice: Context Engineering (see 13.5)
  7. G.3.7 Role Practice: RAG: Retrieval-Augmented Generation (see 13.6)
  8. G.3.8 Role Practice: Advanced RAG: Chunking, Reranking (see 13.7)
  9. G.3.9 Proof Project: Evaluation and Testing (see 13.10)
  10. G.3.10 Proof Project: Building a Production LLM App (see 13.13)
  11. G.3.11 Interview and Readiness Evidence: LLM Observability Stack Selection (see 19.13)
4. G.4 Developer Experience and Education
  1. G.4.1 Common Core: APIs and Keys (see 2.4)
  2. G.4.2 Common Core: The Tool Interface (see 15.1)
  3. G.4.3 Common Core: Tool Schema Design (see 15.5)
  4. G.4.4 Common Core: MCP Fundamentals: Stateless Requests and JSON-RPC (see 15.6)
  5. G.4.5 Common Core: Building an MCP Server: Stateless Python and TypeScript (see 15.7)
  6. G.4.6 Role Practice: Agent Skills: Portable Contract and Runtime Boundary (see 15.22)
  7. G.4.7 Role Practice: Skill Discovery and Progressive Disclosure (see 15.24)
  8. G.4.8 Role Practice: Skill Invocation and Routing (see 15.25)
  9. G.4.9 Proof Project: Skill Evals, Packaging, and Portability (see 15.27)
  10. G.4.10 Proof Project: Capstone: Stateless Tool Ecosystem (see 15.23)
  11. G.4.11 Interview and Readiness Evidence: Build a Feedback Ratchet with Ownership and Retirement (see 16.54)
5. G.5 AI Evaluation and Reliability
  1. G.5.1 Common Core: Model Evaluation: Metrics, Cross-Validation (see 4.9)
  2. G.5.2 Common Core: Evaluation and Testing (see 13.10)
  3. G.5.3 Common Core: Eval-Driven Agent Development (see 16.30)
  4. G.5.4 Common Core: Failure Modes — Why Agents Break (see 16.26)
  5. G.5.5 Common Core: OpenTelemetry GenAI Semantic Conventions (see 16.23)
  6. G.5.6 Role Practice: Inference Metrics — TTFT, TPOT, ITL, Goodput, P99 (see 19.8)
  7. G.5.7 Role Practice: LLM Observability Stack Selection (see 19.13)
  8. G.5.8 Role Practice: A/B Testing LLM Features — GrowthBook and Statsig (see 19.21)
  9. G.5.9 Role Practice: Load Testing LLM APIs — k6, LLMPerf, GenAI-Perf (see 19.22)
  10. G.5.10 Proof Project: Shadow, Canary, and Progressive Deployment (see 19.20)
  11. G.5.11 Proof Project: Chaos Engineering for LLM Production (see 19.24)
  12. G.5.12 Interview and Readiness Evidence: SRE for AI — Multi-Agent Incident Response (see 19.23)
6. G.6 LLM Product Engineering
  1. G.6.1 Common Core: Prompt Engineering: Techniques and Patterns (see 13.1)
  2. G.6.2 Common Core: Structured Outputs (see 13.3)
  3. G.6.3 Common Core: Embeddings and Vector Representations (see 13.4)
  4. G.6.4 Common Core: Context Engineering (see 13.5)
  5. G.6.5 Common Core: RAG: Retrieval-Augmented Generation (see 13.6)
  6. G.6.6 Common Core: Function Calling and Tool Use (see 13.9)
  7. G.6.7 Role Practice: Evaluation and Testing (see 13.10)
  8. G.6.8 Role Practice: Caching, Rate Limiting and Cost (see 13.11)
  9. G.6.9 Role Practice: Guardrails and Safety (see 13.12)
  10. G.6.10 Proof Project: Building a Production LLM App (see 13.13)
  11. G.6.11 Proof Project: AI Gateways — LiteLLM, Portkey, Kong, Bifrost (see 19.19)
  12. G.6.12 Interview and Readiness Evidence: Shadow, Canary, and Progressive Deployment (see 19.20)
7. G.7 Building and Deploying AI Applications
  1. G.7.1 Prompt Engineering: Techniques and Patterns (see 13.1)
  2. G.7.2 Structured Outputs (see 13.3)
  3. G.7.3 Embeddings and Vector Representations (see 13.4)
  4. G.7.4 Context Engineering (see 13.5)
  5. G.7.5 RAG: Retrieval-Augmented Generation (see 13.6)
  6. G.7.6 Evaluation and Testing (see 13.10)
  7. G.7.7 Caching, Rate Limiting and Cost (see 13.11)
  8. G.7.8 Guardrails and Safety (see 13.12)
  9. G.7.9 Building a Production LLM App (see 13.13)
  10. G.7.10 LLM Observability Stack Selection (see 19.13)
  11. G.7.11 Shadow, Canary, and Progressive Deployment (see 19.20)
  12. G.7.12 SRE for AI — Multi-Agent Incident Response (see 19.23)
8. G.8 Customer AI Deployment
  1. G.8.1 Common Core: Define the Outcome Before You Choose the Output (see 16.47)
  2. G.8.2 Common Core: Discover the Workflow People Actually Perform (see 16.48)
  3. G.8.3 Common Core: Map Assumptions and Resolve the Riskiest One First (see 16.49)
  4. G.8.4 Common Core: Choose the Smallest Slice That Can Change the Decision (see 16.50)
  5. G.8.5 Common Core: Write Specifications That Preserve Judgment (see 16.51)
  6. G.8.6 Role Practice: RAG: Retrieval-Augmented Generation (see 13.6)
  7. G.8.7 Role Practice: Evaluation and Testing (see 13.10)
  8. G.8.8 Role Practice: Building a Production LLM App (see 13.13)
  9. G.8.9 Role Practice: Design Success Metrics Before the Result Exists (see 16.52)
  10. G.8.10 Proof Project: Choose Prototype, Pilot, or Production Deliberately (see 16.53)
  11. G.8.11 Proof Project: Shadow, Canary, and Progressive Deployment (see 19.20)
  12. G.8.12 Interview and Readiness Evidence: Build a Feedback Ratchet with Ownership and Retirement (see 16.54)
9. G.9 Model Context Protocol (MCP)
  1. G.9.1 MCP Fundamentals: Stateless Requests and JSON-RPC (see 15.6)
  2. G.9.2 Building an MCP Server: Stateless Python and TypeScript (see 15.7)
  3. G.9.3 Building an MCP Client: Discovery, Routing, and Dual-Era Fallback (see 15.8)
  4. G.9.4 MCP Transports: stdio and Stateless Streamable HTTP (see 15.9)
  5. G.9.5 MCP Resources and Prompts: Addressable Context for Stateless Servers (see 15.10)
  6. G.9.6 MCP Model Input: Sampling Migration and Stateless MRTR (see 15.11)
  7. G.9.7 Explicit Scope and Stateless Elicitation (see 15.12)
  8. G.9.8 MCP Tasks Extension: Durable Work on a Stateless Core (see 15.13)
  9. G.9.9 MCP Apps on the Stateless Protocol (see 15.14)
  10. G.9.10 MCP Security: Poisoned Metadata, Routing, and MRTR State (see 15.15)
  11. G.9.11 MCP Authorization: CIMD, Issuer Binding, PKCE, and Step-Up (see 15.16)
  12. G.9.12 MCP Auth in Production: Issuer-Bound Enrollment and Tokens (see 15.18)
  13. G.9.13 Stateless MCP Gateways and Registry Admission (see 15.17)
  14. G.9.14 MCP Tool Contracts and Content (see 15.28)
  15. G.9.15 MCP Reliability, Cancellation, and Flow Control (see 15.29)
  16. G.9.16 MCP Registry Supply Chain: Admission, Drift, and Rollback (see 15.30)
  17. G.9.17 MCP Conformance Engineering: Versioning, Evidence, and Operations (see 15.31)
  18. G.9.18 Capstone: Stateless Tool Ecosystem (see 15.23) (optional)
10. G.10 Reverse Engineering
  1. G.10.1 Program Anatomy: An ELF Reader From Scratch (see 15.32)
  2. G.10.2 Disassembly and Control Flow: From Bytes to a Graph You Can Check (see 15.33)
  3. G.10.3 Lifting to Pseudocode: IR, SSA, and Structuring From Scratch (see 15.34)
  4. G.10.4 Evidence, Confidence, and Unknowns in Binary Analysis (see 15.35)
  5. G.10.5 Rebuild and Prove Equal: Differential Tests, Byte Comparison, and the Evidence Loop (see 15.36)
  6. G.10.6 The Reverse Engineering Landscape: History, Tool Families, and How a Decompiler Is Built (see 15.37)
  7. G.10.7 Calling Conventions and Stack Frames: System V, Windows x64, and AArch64 (see 15.38)
  8. G.10.8 Compiler Idioms: What Optimization Does to Your Code and How to Read It Back (see 15.39)
  9. G.10.9 Linking, Loading, and Stripped Binaries: PLT, GOT, Imports, Fixups, and Function Identification (see 15.40)
  10. G.10.10 Recovering Data and Types: Structs, Arrays, C++ Objects, and DWARF (see 15.41)
  11. G.10.11 Dynamic Analysis of Your Own Program: Debuggers, Traces, and Confirming Static Claims (see 15.42)
  12. G.10.12 Beyond Native Code: Python, JVM, .NET, and WebAssembly Bytecode (see 15.43)
  13. G.10.13 Binary Diffing and Symbolic Execution: Matching Functions and Solving for Inputs (see 15.44)
  14. G.10.14 AI-Assisted Reverse Engineering: LLM Decompilers, Analysis Agents, and Ground Truth (see 15.45)
11. G.11 Product Judgment and Delivery
  1. G.11.1 Define the Outcome Before You Choose the Output (see 16.47)
  2. G.11.2 Discover the Workflow People Actually Perform (see 16.48)
  3. G.11.3 Map Assumptions and Resolve the Riskiest One First (see 16.49)
  4. G.11.4 Choose the Smallest Slice That Can Change the Decision (see 16.50)
  5. G.11.5 Write Specifications That Preserve Judgment (see 16.51)
  6. G.11.6 Design Success Metrics Before the Result Exists (see 16.52)
  7. G.11.7 Choose Prototype, Pilot, or Production Deliberately (see 16.53)
  8. G.11.8 Build a Feedback Ratchet with Ownership and Retirement (see 16.54)
12. G.12 Software Engineering Fundamentals
  1. G.12.1 Dev Environment (see 2.1)
  2. G.12.2 Git and Collaboration (see 2.2)
  3. G.12.3 Python Environments (see 2.6)
  4. G.12.4 Docker for AI (see 2.7)
  5. G.12.5 Data Management (see 2.9)
  6. G.12.6 Terminal and Shell (see 2.10)
  7. G.12.7 Debugging and Profiling (see 2.12)
  8. G.12.8 The Tool Interface (see 15.1)
  9. G.12.9 Tool Schema Design (see 15.5)
  10. G.12.10 Verification Gates (see 16.38)
  11. G.12.11 Shadow, Canary, and Progressive Deployment (see 19.20)
  12. G.12.12 Security — Secrets, PII Scrubbing, Audit Logs (see 19.25)
  13. G.12.13 SRE for AI — Multi-Agent Incident Response (see 19.23)
13. G.13 Agent-Assisted Engineering
  1. G.13.1 Agent Workbench: Why Capable Models Still Fail (see 16.31)
  2. G.13.2 The Minimal Agent Workbench (see 16.32)
  3. G.13.3 Agent Instructions as Executable Constraints (see 16.33)
  4. G.13.4 Repo Memory and Durable State (see 16.34)
  5. G.13.5 Initialization Scripts for Agents (see 16.35)
  6. G.13.6 Scope Contracts and Task Boundaries (see 16.36)
  7. G.13.7 Frame the Task Before the Agent Writes Code (see 16.43)
  8. G.13.8 Build an Evidence-Backed Execution Plan (see 16.44)
  9. G.13.9 Runtime Feedback Loops (see 16.37)
  10. G.13.10 Verification Gates (see 16.38)
  11. G.13.11 Reviewer Agent: Separate Builder from Marker (see 16.39)
  12. G.13.12 Delegate Agent Work with Isolation and Merge Contracts (see 16.45)
  13. G.13.13 Multi-Session Handoff (see 16.40)
  14. G.13.14 The Workbench on a Real Repo (see 16.41)
  15. G.13.15 Turn Every Agent Correction into a System Improvement (see 16.46)
  16. G.13.16 Capstone: Ship a Reusable Agent Workbench Pack (see 16.42)
HGlossary and mythsnot yet
1. H.1 Math and training
  1. H.1.1 Activation Checkpointing
  2. H.1.2 Activation Function
  3. H.1.3 Adam (Optimizer)
  4. H.1.4 AdamW
  5. H.1.5 Autograd
  6. H.1.6 Backpropagation
  7. H.1.7 Batch Size
  8. H.1.8 Contrastive Learning
  9. H.1.9 Cross-Entropy
  10. H.1.10 Data Augmentation
  11. H.1.11 DPO (Direct Preference Optimization)
  12. H.1.12 Dropout
  13. H.1.13 Eigenvalue
  14. H.1.14 Epoch
  15. H.1.15 Fine-tuning
  16. H.1.16 Gradient
  17. H.1.17 Gradient Accumulation
  18. H.1.18 Gradient Clipping
  19. H.1.19 Gradient Descent
  20. H.1.20 Hyperparameter
  21. H.1.21 JAX
  22. H.1.22 Knowledge Distillation
  23. H.1.23 Learning Rate
  24. H.1.24 Learning Rate Schedule
  25. H.1.25 LoRA (Low-Rank Adaptation)
  26. H.1.26 Loss Function
  27. H.1.27 Mixed Precision
  28. H.1.28 NaN (Not a Number)
  29. H.1.29 Normalization
  30. H.1.30 Optimizer
  31. H.1.31 Overfitting
  32. H.1.32 QLoRA
  33. H.1.33 ReLU
  34. H.1.34 RLHF (Reinforcement Learning from Human Feedback)
  35. H.1.35 SFT (Supervised Fine-Tuning)
  36. H.1.36 Softmax
  37. H.1.37 Stochastic Gradient Descent (SGD)
  38. H.1.38 Transfer Learning
  39. H.1.39 Underfitting
  40. H.1.40 Warmup
  41. H.1.41 Weight
  42. H.1.42 Weight Decay
2. H.2 Models and inference
  1. H.2.1 Attention
  2. H.2.2 Autoregressive
  3. H.2.3 CNN (Convolutional Neural Network)
  4. H.2.4 CUDA
  5. H.2.5 Decoder
  6. H.2.6 Decoding Strategy
  7. H.2.7 Diffusion Model
  8. H.2.8 Encoder
  9. H.2.9 GAN (Generative Adversarial Network)
  10. H.2.10 GPT
  11. H.2.11 Inductive Bias
  12. H.2.12 Inference
  13. H.2.13 KV Cache
  14. H.2.14 LLM (Large Language Model)
  15. H.2.15 Logits
  16. H.2.16 MoE (Mixture of Experts)
  17. H.2.17 Nucleus Sampling (Top-p)
  18. H.2.18 Parameter
  19. H.2.19 Perplexity
  20. H.2.20 Quantization
  21. H.2.21 Self-Attention
  22. H.2.22 Speculative Decoding
  23. H.2.23 Stop Sequence
  24. H.2.24 Streaming
  25. H.2.25 Temperature
  26. H.2.26 Time to First Token (TTFT)
  27. H.2.27 Top-k Sampling
  28. H.2.28 Transformer
  29. H.2.29 VAE (Variational Autoencoder)
3. H.3 Data and representations
  1. H.3.1 Byte Pair Encoding (BPE)
  2. H.3.2 Cosine Similarity
  3. H.3.3 Data Deduplication
  4. H.3.4 Data Leakage
  5. H.3.5 Data Provenance
  6. H.3.6 Dataset Split
  7. H.3.7 Embedding
  8. H.3.8 Feature
  9. H.3.9 Latent Space
  10. H.3.10 Tensor
  11. H.3.11 Token
  12. H.3.12 Tokenization
  13. H.3.13 Vocabulary
4. H.4 Retrieval and generation
  1. H.4.1 Approximate Nearest Neighbor (ANN)
  2. H.4.2 BM25
  3. H.4.3 Chunking
  4. H.4.4 Dense Retrieval
  5. H.4.5 Grounding
  6. H.4.6 HNSW
  7. H.4.7 Hybrid Retrieval
  8. H.4.8 Maximum Marginal Relevance (MMR)
  9. H.4.9 RAG (Retrieval-Augmented Generation)
  10. H.4.10 Recall@K
  11. H.4.11 Reciprocal Rank Fusion (RRF)
  12. H.4.12 Reranker
  13. H.4.13 Semantic Search
  14. H.4.14 Vector Database
5. H.5 Prompting and context
  1. H.5.1 Chain of Thought (CoT)
  2. H.5.2 Context Compression
  3. H.5.3 Context Engineering
  4. H.5.4 Context Window
  5. H.5.5 Few-Shot
  6. H.5.6 In-Context Learning
  7. H.5.7 Instruction Following
  8. H.5.8 Instruction Hierarchy
  9. H.5.9 Lost in the Middle
  10. H.5.10 Prompt Cache
  11. H.5.11 Prompt Engineering
  12. H.5.12 Prompt Sensitivity
  13. H.5.13 System Prompt
  14. H.5.14 Token Budget
  15. H.5.15 Zero-Shot
6. H.6 Agents and tools
  1. H.6.1 Agent
  2. H.6.2 Agent Harness
  3. H.6.3 Agent Memory
  4. H.6.4 Agent State
  5. H.6.5 Agent Skill
  6. H.6.6 Approval Gate
  7. H.6.7 Checkpoint
  8. H.6.8 Compensating Action
  9. H.6.9 Delegation
  10. H.6.10 Durable Execution
  11. H.6.11 Function Calling
  12. H.6.12 Human-in-the-Loop (HITL)
  13. H.6.13 MCP (Model Context Protocol)
  14. H.6.14 Multi Round-Trip Request (MRTR)
  15. H.6.15 Orchestration
  16. H.6.16 Planning
  17. H.6.17 ReAct
  18. H.6.18 Sandbox
  19. H.6.19 Skill Bundle
  20. H.6.20 Skill Catalog
  21. H.6.21 Skill Discovery
  22. H.6.22 Skill Invocation
  23. H.6.23 Stateless MCP
  24. H.6.24 Structured Output
  25. H.6.25 Swarm
  26. H.6.26 Termination Condition
  27. H.6.27 Tool Contract
7. H.7 Evaluation and safety
  1. H.7.1 Alignment
  2. H.7.2 Benchmark Contamination
  3. H.7.3 Calibration
  4. H.7.4 Distribution Shift
  5. H.7.5 Eval Set
  6. H.7.6 Evaluation (Eval)
  7. H.7.7 Exact Match (EM)
  8. H.7.8 Guardrails
  9. H.7.9 Hallucination
  10. H.7.10 Least Privilege
  11. H.7.11 LLM-as-a-Judge
  12. H.7.12 Model Card
  13. H.7.13 Pass@k
  14. H.7.14 Precision and Recall
  15. H.7.15 Prompt Injection
  16. H.7.16 ROUGE
  17. H.7.17 Verification Gate
8. H.8 AI-native development
  1. H.8.1 Backpressure
  2. H.8.2 Circuit Breaker
  3. H.8.3 Coding Agent
  4. H.8.4 Cost per Successful Task
  5. H.8.5 Flaky Test
  6. H.8.6 Handoff
  7. H.8.7 Idempotency
  8. H.8.8 Model Router
  9. H.8.9 Observability
  10. H.8.10 Patch
  11. H.8.11 Progressive Disclosure
  12. H.8.12 Rate Limit
  13. H.8.13 Regression Test
  14. H.8.14 Repository Instructions
  15. H.8.15 Repository Map
  16. H.8.16 Reproducible Build
  17. H.8.17 Retry with Backoff
  18. H.8.18 Reviewer Agent
  19. H.8.19 Scope Contract
  20. H.8.20 Semantic Cache
  21. H.8.21 Test Oracle
  22. H.8.22 Trace
  23. H.8.23 Worktree
9. H.9 Infrastructure and serving
  1. H.9.1 Autoscaling
  2. H.9.2 Chunked Prefill
  3. H.9.3 Continuous Batching
  4. H.9.4 Decode Phase
  5. H.9.5 Disaggregated Serving
  6. H.9.6 Dynamic Batching
  7. H.9.7 Expert Parallelism
  8. H.9.8 FlashAttention
  9. H.9.9 Goodput
  10. H.9.10 Inter-Token Latency (ITL)
  11. H.9.11 Model Serving
  12. H.9.12 Paged KV Cache
  13. H.9.13 Pipeline Parallelism
  14. H.9.14 Prefill
  15. H.9.15 Prefix Caching
  16. H.9.16 Tensor Parallelism
  17. H.9.17 Time per Output Token (TPOT)
  18. H.9.18 Tokens per Second (TPS)
10. H.10 Reliability and operations
  1. H.10.1 Admission Control
  2. H.10.2 Availability
  3. H.10.3 Canary Release
  4. H.10.4 Deadline Propagation
  5. H.10.5 Error Budget
  6. H.10.6 Graceful Degradation
  7. H.10.7 Incident Response
  8. H.10.8 Load Shedding
  9. H.10.9 Postmortem
  10. H.10.10 Readiness Probe
  11. H.10.11 Retry Budget
  12. H.10.12 Rollback
  13. H.10.13 Saturation
  14. H.10.14 Service Level Indicator (SLI)
  15. H.10.15 Service Level Objective (SLO)
  16. H.10.16 Shadow Traffic
  17. H.10.17 Tail Latency
11. H.11 Security and governance
  1. H.11.1 AI Risk Assessment
  2. H.11.2 Audit Log
  3. H.11.3 Content Provenance
  4. H.11.4 Data Classification
  5. H.11.5 Data Exfiltration
  6. H.11.6 Data Lineage
  7. H.11.7 Data Minimization
  8. H.11.8 Datasheet for Datasets
  9. H.11.9 Defense in Depth
  10. H.11.10 Indirect Prompt Injection
  11. H.11.11 Jailbreak
  12. H.11.12 Membership Inference
  13. H.11.13 Provenance Attestation
  14. H.11.14 Purpose Limitation
  15. H.11.15 Red Teaming
  16. H.11.16 Separation of Duties
  17. H.11.17 Software Bill of Materials (SBOM)
  18. H.11.18 Threat Model
  19. H.11.19 Trust Boundary
  20. H.11.20 Zero Trust
12. H.12 Multimodal systems
  1. H.12.1 Audio Token
  2. H.12.2 Automatic Speech Recognition (ASR)
  3. H.12.3 Cross-Attention
  4. H.12.4 Early Fusion
  5. H.12.5 Image Token
  6. H.12.6 Late Fusion
  7. H.12.7 Modality
  8. H.12.8 Modality Alignment
  9. H.12.9 Multimodal Fusion
  10. H.12.10 Multimodal Model
  11. H.12.11 Patch Embedding
  12. H.12.12 Shared Embedding Space
  13. H.12.13 Vision-Language Model (VLM)
  14. H.12.14 Vision Transformer (ViT)
  15. H.12.15 Visual Grounding
13. H.13 Myths
  1. H.13.1 "AI understands language"
  2. H.13.2 "More parameters = smarter model"
  3. H.13.3 "Neural networks are black boxes"
  4. H.13.4 "AI will replace programmers"
  5. H.13.5 "You need a PhD in math to do AI"
  6. H.13.6 "GPT stands for General Purpose Technology"
  7. H.13.7 "Temperature makes the AI more creative"
  8. H.13.8 "Fine-tuning teaches the model new knowledge"
  9. H.13.9 "Bigger context window = better"
  10. H.13.10 "AI agents are autonomous"
  11. H.13.11 "Transformers understand order because of positional encoding"
  12. H.13.12 "Pre-training is just reading the internet"
  13. H.13.13 "RLHF aligns AI with human values"
  14. H.13.14 "Embeddings capture meaning"
  15. H.13.15 "Zero-shot means no training"
  16. H.13.16 "AI models learn like humans"
  17. H.13.17 "Scaling laws mean bigger is always better"
  18. H.13.18 "Open source AI is the same as open weights"
  19. H.13.19 "Prompt engineering is not real engineering"
  20. H.13.20 "CNNs are outdated, everything is transformers now"
  21. H.13.21 "You need massive compute to train useful models"
ICheat sheetsnot yet
1. I.1 Foundations
  1. I.1.1 Calculus, Gradients and Convex Optimization
  2. I.1.2 Classical Machine Learning
  3. I.1.3 Classical Models Compared
  4. I.1.4 Clustering, Dimensionality Reduction and Anomalies
  5. I.1.5 Complex Numbers, Fourier and Signals
  6. I.1.6 Debugging Neural Network Training
  7. I.1.7 Feature Engineering and Pipelines
  8. I.1.8 Git, Shell and Linux for AI Work
  9. I.1.9 GPUs, Cloud, Docker and API Keys
  10. I.1.10 Information Theory and Numerical Stability
  11. I.1.11 JAX against PyTorch
  12. I.1.12 Linear Algebra for ML
  13. I.1.13 Math for Machine Learning
  14. I.1.14 ML System Design, Drift and Feature Stores
  15. I.1.15 Model Evaluation, Bias-Variance and Tuning
  16. I.1.16 Neural Network Basics
  17. I.1.17 Notebooks, Data Files and Profiling
  18. I.1.18 Optimizers and Learning Rate Schedules
  19. I.1.19 Probability and Statistics for ML
  20. I.1.20 Python Dev Environment for AI
  21. I.1.21 PyTorch Essentials
  22. I.1.22 Time Series Fundamentals
  23. I.1.23 Training Neural Networks
2. I.2 Vision and audio
  1. I.2.1 Audio Features, Classification and Metrics
  2. I.2.2 CNN Architectures and Transfer Learning
  3. I.2.3 Computer Vision
  4. I.2.4 Detection and Segmentation Compared
  5. I.2.5 Edge Inference Targets
  6. I.2.6 Evaluating Generative Models
  7. I.2.7 GANs, Diffusion and Flow Compared
  8. I.2.8 Music Generation and Audio-Language Models
  9. I.2.9 Real-Time Voice Assistants
  10. I.2.10 Self-Supervised Vision, CLIP and Retrieval
  11. I.2.11 Speaker Verification, Spoofing and Watermarks
  12. I.2.12 Speech Recognition and Whisper
  13. I.2.13 Text-to-Speech, Voice Cloning and Codecs
  14. I.2.14 Video and 3D Generation
  15. I.2.15 Video, Depth and 3D Vision
  16. I.2.16 Vision Pipelines and Edge Deployment
3. I.3 Language and transformers
  1. I.3.1 Attention Variants Compared
  2. I.3.2 BERT, GPT and T5 Compared
  3. I.3.3 Chatbots and Dialogue State Tracking
  4. I.3.4 Classic Text Processing and TF-IDF
  5. I.3.5 Embeddings and Vector Search
  6. I.3.6 Fine-Tuning with LoRA
  7. I.3.7 Graphs and Knowledge Graphs
  8. I.3.8 Mixture of Experts and Hybrid Architectures
  9. I.3.9 NER, POS Tagging and Coreference
  10. I.3.10 Positional Encoding: Sinusoidal, RoPE, ALiBi
  11. I.3.11 RLHF, DPO and Preference Tuning
  12. I.3.12 RNNs, Seq2Seq and N-gram Models
  13. I.3.13 Scaling Laws and Pre-Training
  14. I.3.14 Speculative Decoding: Draft, EAGLE, MTP
  15. I.3.15 Structured Outputs and Constrained Decoding
  16. I.3.16 Summarization, QA, NLI and Multilingual NLP
  17. I.3.17 Tokenization and BPE
  18. I.3.18 Transformers and Attention
  19. I.3.19 Word2Vec, GloVe and FastText
4. I.4 Generative and multimodal
  1. I.4.1 ControlNet, Inpainting and Image Editing
  2. I.4.2 Diffusion and Generative Models
  3. I.4.3 Embodied VLAs and Computer-Use Agents
  4. I.4.4 GANs from DCGAN to StyleGAN
  5. I.4.5 Multimodal and Vision-Native RAG
  6. I.4.6 OCR and Document Understanding
  7. I.4.7 Open VLM Families: LLaVA, Qwen-VL, InternVL
  8. I.4.8 Unified Multimodal Models
  9. I.4.9 Video-Language Models and Long Video
  10. I.4.10 Vision Transformers and Patch Tokens
  11. I.4.11 Vision-Language Model Architectures
5. I.5 RL, safety and alignment
  1. I.5.1 AI Safety and Red Teaming
  2. I.5.2 Alignment Training: Instructions, DPO Family, RLAIF
  3. I.5.3 Bias, Fairness and Differential Privacy
  4. I.5.4 Cards, Provenance, Regulation and Welfare
  5. I.5.5 Mesa-Optimizers, Scheming and AI Control
  6. I.5.6 Multi-Agent RL and Sim-to-Real
  7. I.5.7 Prompt Injection Incidents and AI CVEs
  8. I.5.8 Reinforcement Learning
  9. I.5.9 Value Methods against Policy Methods
6. I.6 LLM engineering
  1. I.6.1 Agent State Machines and Durable Graphs
  2. I.6.2 Guardrails, Moderation and Redaction
  3. I.6.3 Prompt Engineering
  4. I.6.4 Shipping a Production LLM App
7. I.7 Tools and protocols
  1. I.7.1 Agent Skills: Discovery, Invocation, Routing
  2. I.7.2 Binary Analysis Tools from Scratch
  3. I.7.3 Building an MCP Server
  4. I.7.4 MCP Clients and Transports
  5. I.7.5 MCP Gateways, Registries and Conformance
  6. I.7.6 MCP Security and Authorization
  7. I.7.7 MCP Tasks, Elicitation and Apps
  8. I.7.8 Model Context Protocol (MCP)
  9. I.7.9 Skill Permissions, Evals and Supply Chain
  10. I.7.10 Tool Calling and Structured Outputs
8. I.8 Agents and autonomy
  1. I.8.1 A2A and Agent Communication Protocols
  2. I.8.2 Agent Benchmarks: SWE-bench to OSWorld
  3. I.8.3 Agent Frameworks Compared
  4. I.8.4 Agent Harness Engineering
  5. I.8.5 Agent Planning: Workflows, HTN, Plan-Execute
  6. I.8.6 Agent Tool Use and Skill Libraries
  7. I.8.7 Autonomous Agents and Permissions
  8. I.8.8 Constitutional AI and Input/Output Classifiers
  9. I.8.9 Context Engineering and Agent Memory
  10. I.8.10 Debate, Roles and Review Panels
  11. I.8.11 Delegation, Isolation and Feedback Ratchets
  12. I.8.12 Eval-Driven Agent Development
  13. I.8.13 Frontier Safety Frameworks and External Evals
  14. I.8.14 Multi-Agent Case Studies and Eval Farms
  15. I.8.15 Multi-Agent Systems
  16. I.8.16 Negotiation, Simulation and Agent Economies
  17. I.8.17 OpenTelemetry GenAI and Agent Observability
  18. I.8.18 Orchestration Patterns Compared
  19. I.8.19 Outcomes, Workflows, Risks and Specs
  20. I.8.20 Production Agent Runtimes and Queues
  21. I.8.21 ReAct, ReWOO, Reflexion and Tree of Thoughts
  22. I.8.22 Repo Memory, Init Scripts and Handoffs
  23. I.8.23 Self-Improving and Research Agents
  24. I.8.24 Self-Taught Reasoning and Self-Improvement
  25. I.8.25 Swarm Optimization and MARL for Agents
  26. I.8.26 Why Agents Fail
9. I.9 Production and infrastructure
  1. I.9.1 Async, Hogwild and Disaggregated Inference
  2. I.9.2 GPU Autoscaling on Kubernetes
  3. I.9.3 Inference Metrics and Load Testing
  4. I.9.4 Inference Optimization: KV Cache, Quantization, Batching
  5. I.9.5 LLM Cost, Caching and Batch APIs
  6. I.9.6 LLM Routing and AI Gateways Compared
  7. I.9.7 Managed Platforms and Self-Hosting Compared
  8. I.9.8 Quantization Formats Compared
  9. I.9.9 Secrets, Audit Logs and Compliance
  10. I.9.10 Serving Engines Compared
  11. I.9.11 Shadow, Canary and A/B Tests for LLMs
  12. I.9.12 SRE, Incidents and Postmortems for AI
10. I.10 Capstones
  1. I.10.1 Build a VLM from Scratch
  2. I.10.2 Coding Agents: Landscape and Capstones
  3. I.10.3 DDP, FSDP, ZeRO and Pipeline Parallel
  4. I.10.4 End-to-End LLM and Fine-Tuning Pipeline
  5. I.10.5 Evaluating LLMs
  6. I.10.6 Paper Writer, Critic and Iteration
  7. I.10.7 Pre-Training Data Pipelines
  8. I.10.8 Production RAG Chatbots and Freshness
  9. I.10.9 Research Agent Pipeline from Scratch
  10. I.10.10 Retrieval-Augmented Generation (RAG)
  11. I.10.11 Train and Load a GPT from Scratch
  12. I.10.12 Training Memory, Accumulation and Checkpoints
11. I.11 Manuals
  1. I.11.1 A2A Protocol
  2. I.11.2 Docker Sandboxes for Agents
  3. I.11.3 LLM Serving with SGLang
  4. I.11.4 Reverse Engineering with rea
12. I.12 Projects
  1. I.12.1 AI Tutors and Learning Tools
  2. I.12.2 Data Quality and Dataset Gates
  3. I.12.3 Decision, Change and Tutorial Tools
  4. I.12.4 Diagram, Chart and Infrastructure Explainers
  5. I.12.5 Localization and Glossary Tools
  6. I.12.6 Personal Organizer Agents
  7. I.12.7 Sandbox Policy Ladder
  8. I.12.8 Text Workflow Desks
JLearning with the coursenot yet
1. J.1 How to learn
2. J.2 Progress, quizzes and checkpoints
3. J.3 Home page
4. J.4 About
5. J.5 Developer
6. J.6 Skills and the terminal tutor
7. J.7 Roadmap
8. J.8 Catalog
9. J.9 Highlights
10. J.10 Is superintelligence here yet?
11. J.11 Translations