pageman/sutskever-30-implementations

Sutskever 30 - Complete Implementation Suite

Comprehensive toy implementations of the 30 foundational papers recommended by Ilya Sutskever

Buy the Google Colab Code for this Repo at Gumroad

Overview

This repository contains detailed, educational implementations of the papers from Ilya Sutskever's famous reading list - the collection he told John Carmack would teach you "90% of what matters" in deep learning.

Progress: 30/30 papers (100%) - COMPLETE! 🎉

Each implementation:

• ✅ Uses only NumPy (no deep learning frameworks) for educational clarity
• ✅ Includes synthetic/bootstrapped data for immediate execution
• ✅ Provides extensive visualizations and explanations
• ✅ Demonstrates core concepts from each paper
• ✅ Runs in Jupyter notebooks for interactive learning

Quick Start

# Navigate to the directory
cd sutskever-30-implementations

# Install dependencies
pip install numpy matplotlib scipy

# Run any notebook
jupyter notebook 02_char_rnn_karpathy.ipynb

The Sutskever 30 Papers

Foundational Concepts (Papers 1-5)

Architectures & Mechanisms (Papers 6-15)

Advanced Topics (Papers 16-22)

Theory & Meta-Learning (Papers 23-30)

Featured Implementations

🌟 Must-Read Notebooks

These implementations cover the most influential papers and demonstrate core deep learning concepts:

Foundations

1. 02_char_rnn_karpathy.ipynb - Character-level RNN

   • Build RNN from scratch
   • Understand backpropagation through time
   • Generate text

2. 03_lstm_understanding.ipynb - LSTM Networks

   • Implement forget/input/output gates
   • Visualize gate activations
   • Compare with vanilla RNN

3. 04_rnn_regularization.ipynb - RNN Regularization

   • Variational dropout for RNNs
   • Proper dropout placement
   • Training improvements

4. 05_neural_network_pruning.ipynb - Network Pruning & MDL

   • Magnitude-based pruning
   • Iterative pruning with fine-tuning
   • 90%+ sparsity with minimal loss
   • Minimum Description Length principle

Computer Vision

5. 07_alexnet_cnn.ipynb - CNNs & AlexNet

   • Convolutional layers from scratch
   • Max pooling and ReLU
   • Data augmentation techniques

6. 10_resnet_deep_residual.ipynb - ResNet

   • Skip connections solve degradation
   • Gradient flow visualization
   • Identity mapping intuition

7. 15_identity_mappings_resnet.ipynb - Pre-activation ResNet

   • Pre-activation vs post-activation
   • Better gradient flow
   • Training 1000+ layer networks

8. 11_dilated_convolutions.ipynb - Dilated Convolutions

   • Multi-scale receptive fields
   • No pooling required
   • Semantic segmentation

Attention & Transformers

9. 14_bahdanau_attention.ipynb - Neural Machine Translation

   • Original attention mechanism
   • Seq2seq with alignment
   • Attention visualization

10. 13_attention_is_all_you_need.ipynb - Transformers

    • Scaled dot-product attention
    • Multi-head attention
    • Positional encoding
    • Foundation of modern LLMs

11. 06_pointer_networks.ipynb - Pointer Networks

    • Attention as selection
    • Combinatorial optimization
    • Variable output size

12. 08_seq2seq_for_sets.ipynb - Seq2Seq for Sets

    • Permutation-invariant set encoder
    • Read-Process-Write architecture
    • Attention over unordered elements
    • Sorting and set operations
    • Comparison: order-sensitive vs order-invariant

13. 09_gpipe.ipynb - GPipe Pipeline Parallelism

    • Model partitioning across devices
    • Micro-batching for pipeline utilization
    • F-then-B schedule (forward all, backward all)
    • Re-materialization (gradient checkpointing)
    • Bubble time analysis
    • Training models larger than single-device memory

Advanced Topics

14. 12_graph_neural_networks.ipynb - Graph Neural Networks

    • Message passing framework
    • Graph convolutions
    • Molecular property prediction

15. 16_relational_reasoning.ipynb - Relation Networks

    • Pairwise relational reasoning
    • Visual QA
    • Permutation invariance

16. 18_relational_rnn.ipynb - Relational RNN

    • LSTM with relational memory
    • Multi-head self-attention across memory slots
    • Architecture demonstration (forward pass)
    • Sequential reasoning tasks
    • Section 11: Manual backpropagation implementation (~1100 lines)
    • Complete gradient computation for all components
    • Gradient checking with numerical verification

17. 20_neural_turing_machine.ipynb - Memory-Augmented Networks

    • Content & location addressing
    • Differentiable read/write
    • External memory

18. 21_ctc_speech.ipynb - CTC Loss & Speech Recognition

    • Connectionist Temporal Classification
    • Alignment-free training
    • Forward algorithm

Generative Models

19. 17_variational_autoencoder.ipynb - VAE
    • Generative modeling
    • ELBO loss
    • Latent space visualization

Modern Applications

20. 27_multi_token_prediction.ipynb - Multi-Token Prediction

    • Predict multiple future tokens
    • 2-3x sample efficiency
    • Speculative decoding
    • Faster training & inference

21. 28_dense_passage_retrieval.ipynb - Dense Retrieval

    • Dual encoder architecture
    • In-batch negatives
    • Semantic search

22. 29_rag.ipynb - Retrieval-Augmented Generation

    • RAG-Sequence vs RAG-Token
    • Combining retrieval + generation
    • Knowledge-grounded outputs

23. 30_lost_in_middle.ipynb - Long Context Analysis

    • Position bias in LLMs
    • U-shaped performance curve
    • Document ordering strategies

Scaling & Theory

24. 22_scaling_laws.ipynb - Scaling Laws

    • Power law relationships
    • Compute-optimal training
    • Performance prediction

25. 23_mdl_principle.ipynb - Minimum Description Length

    • Information-theoretic model selection
    • Compression = Understanding
    • MDL vs AIC/BIC comparison
    • Neural network architecture selection
    • MDL-based pruning (connects to Paper 5)
    • Kolmogorov complexity preview

26. 25_kolmogorov_complexity.ipynb - Kolmogorov Complexity

    • K(x) = shortest program generating x
    • Randomness = Incompressibility
    • Algorithmic probability (Solomonoff)
    • Universal prior for induction
    • Connection to Shannon entropy
    • Occam's Razor formalized
    • Theoretical foundation for ML

27. 24_machine_super_intelligence.ipynb - Universal Artificial Intelligence

    • Formal theory of intelligence (Legg & Hutter)
    • Psychometric g-factor and universal intelligence Υ(π)
    • Solomonoff induction for sequence prediction
    • AIXI: Theoretically optimal RL agent
    • Monte Carlo AIXI (MC-AIXI) approximation
    • Kolmogorov complexity estimation
    • Intelligence measurement across environments
    • Recursive self-improvement dynamics
    • Intelligence explosion scenarios
    • 6 sections: from psychometrics to superintelligence
    • Connects Papers #23 (MDL), #25 (Kolmogorov), #8 (DQN)

28. 01_complexity_dynamics.ipynb - Complexity & Entropy

    • Cellular automata (Rule 30)
    • Entropy growth
    • Irreversibility (basic introduction)

29. 19_coffee_automaton.ipynb - The Coffee Automaton (Deep Dive)

    • Comprehensive exploration of irreversibility
    • Coffee mixing and diffusion processes
    • Entropy growth and coarse-graining
    • Phase space and Liouville's theorem
    • Poincaré recurrence theorem (will unmix after e^N time!)
    • Maxwell's demon and Landauer's principle
    • Computational irreversibility (one-way functions, hashing)
    • Information bottleneck in machine learning
    • Biological irreversibility (life and the 2nd law)
    • Arrow of time: fundamental vs emergent
    • 10 comprehensive sections exploring irreversibility across all scales

30. 26_cs231n_cnn_fundamentals.ipynb - CS231n: Vision from First Principles

    • Complete vision pipeline in pure NumPy
    • k-Nearest Neighbors baseline
    • Linear classifiers (SVM and Softmax)
    • Optimization (SGD, Momentum, Adam, learning rate schedules)
    • 2-layer neural networks with backpropagation
    • Convolutional layers (conv, pool, ReLU)
    • Complete CNN architecture (Mini-AlexNet)
    • Visualization techniques (filters, saliency maps)
    • Transfer learning principles
    • Babysitting tips (sanity checks, hyperparameter tuning, monitoring)
    • 10 sections covering entire CS231n curriculum
    • Ties together Papers #7 (AlexNet), #10 (ResNet), #11 (Dilated Conv)

Repository Structure

sutskever-30-implementations/
├── README.md                           # This file
├── PROGRESS.md                         # Implementation progress tracking
├── IMPLEMENTATION_TRACKS.md            # Detailed tracks for all 30 papers
│
├── 01_complexity_dynamics.ipynb        # Entropy & complexity
├── 02_char_rnn_karpathy.ipynb         # Vanilla RNN
├── 03_lstm_understanding.ipynb         # LSTM gates
├── 04_rnn_regularization.ipynb         # Dropout for RNNs
├── 05_neural_network_pruning.ipynb     # Pruning & MDL
├── 06_pointer_networks.ipynb           # Attention pointers
├── 07_alexnet_cnn.ipynb               # CNNs & AlexNet
├── 08_seq2seq_for_sets.ipynb          # Permutation-invariant sets
├── 09_gpipe.ipynb                     # Pipeline parallelism
├── 10_resnet_deep_residual.ipynb      # Residual connections
├── 11_dilated_convolutions.ipynb       # Multi-scale convolutions
├── 12_graph_neural_networks.ipynb      # Message passing GNNs
├── 13_attention_is_all_you_need.ipynb # Transformer architecture
├── 14_bahdanau_attention.ipynb         # Original attention
├── 15_identity_mappings_resnet.ipynb   # Pre-activation ResNet
├── 16_relational_reasoning.ipynb       # Relation networks
├── 17_variational_autoencoder.ipynb   # VAE
├── 18_relational_rnn.ipynb             # Relational RNN
├── 19_coffee_automaton.ipynb           # Irreversibility deep dive
├── 20_neural_turing_machine.ipynb     # External memory
├── 21_ctc_speech.ipynb                # CTC loss
├── 22_scaling_laws.ipynb              # Empirical scaling
├── 23_mdl_principle.ipynb             # MDL & compression
├── 24_machine_super_intelligence.ipynb # Universal AI & AIXI
├── 25_kolmogorov_complexity.ipynb     # K(x) & randomness
├── 26_cs231n_cnn_fundamentals.ipynb    # Vision from first principles
├── 27_multi_token_prediction.ipynb     # Multi-token prediction
├── 28_dense_passage_retrieval.ipynb    # Dense retrieval
├── 29_rag.ipynb                       # RAG architecture
└── 30_lost_in_middle.ipynb            # Long context analysis

All 30 papers implemented! (100% complete!) 🎉

Learning Path

Beginner Track (Start here!)

1. Character RNN (02_char_rnn_karpathy.ipynb) - Learn basic RNNs
2. LSTM (03_lstm_understanding.ipynb) - Understand gating mechanisms
3. CNNs (07_alexnet_cnn.ipynb) - Computer vision fundamentals
4. ResNet (10_resnet_deep_residual.ipynb) - Skip connections
5. VAE (17_variational_autoencoder.ipynb) - Generative models

Intermediate Track

6. RNN Regularization (04_rnn_regularization.ipynb) - Better training
7. Bahdanau Attention (14_bahdanau_attention.ipynb) - Attention basics
8. Pointer Networks (06_pointer_networks.ipynb) - Attention as selection
9. Seq2Seq for Sets (08_seq2seq_for_sets.ipynb) - Permutation invariance
10. CS231n (26_cs231n_cnn_fundamentals.ipynb) - Complete vision pipeline (kNN → CNNs)
11. GPipe (09_gpipe.ipynb) - Pipeline parallelism for large models
12. Transformers (13_attention_is_all_you_need.ipynb) - Modern architecture
13. Dilated Convolutions (11_dilated_convolutions.ipynb) - Receptive fields
14. Scaling Laws (22_scaling_laws.ipynb) - Understanding scale

Advanced Track

15. Pre-activation ResNet (15_identity_mappings_resnet.ipynb) - Architecture details
16. Graph Neural Networks (12_graph_neural_networks.ipynb) - Graph learning
17. Relation Networks (16_relational_reasoning.ipynb) - Relational reasoning
18. Neural Turing Machines (20_neural_turing_machine.ipynb) - External memory
19. CTC Loss (21_ctc_speech.ipynb) - Speech recognition
20. Dense Retrieval (28_dense_passage_retrieval.ipynb) - Semantic search
21. RAG (29_rag.ipynb) - Retrieval-augmented generation
22. Lost in the Middle (30_lost_in_middle.ipynb) - Long context analysis

Theory & Fundamentals

23. MDL Principle (23_mdl_principle.ipynb) - Model selection via compression
24. Kolmogorov Complexity (25_kolmogorov_complexity.ipynb) - Randomness & information
25. Complexity Dynamics (01_complexity_dynamics.ipynb) - Entropy & emergence
26. Coffee Automaton (19_coffee_automaton.ipynb) - Deep dive into irreversibility

Key Insights from the Sutskever 30

Architecture Evolution

• RNN → LSTM: Gating solves vanishing gradients
• Plain Networks → ResNet: Skip connections enable depth
• RNN → Transformer: Attention enables parallelization
• Fixed vocab → Pointers: Output can reference input

Fundamental Mechanisms

• Attention: Differentiable selection mechanism
• Residual Connections: Gradient highways
• Gating: Learned information flow control
• External Memory: Separate storage from computation

Training Insights

• Scaling Laws: Performance predictably improves with scale
• Regularization: Dropout, weight decay, data augmentation
• Optimization: Gradient clipping, learning rate schedules
• Compute-Optimal: Balance model size and training data

Theoretical Foundations

• Information Theory: Compression, entropy, MDL
• Complexity: Kolmogorov complexity, power laws
• Generative Modeling: VAE, ELBO, latent spaces
• Memory: Differentiable data structures

Implementation Philosophy

Why NumPy-only?

These implementations deliberately avoid PyTorch/TensorFlow to:

• Deepen understanding: See what frameworks abstract away
• Educational clarity: No magic, every operation explicit
• Core concepts: Focus on algorithms, not framework APIs
• Transferable knowledge: Principles apply to any framework

Synthetic Data Approach

Each notebook generates its own data to:

• Immediate execution: No dataset downloads required
• Controlled experiments: Understand behavior on simple cases
• Concept focus: Data doesn't obscure the algorithm
• Rapid iteration: Modify and re-run instantly

Extensions & Next Steps

Build on These Implementations

After understanding the core concepts, try:

1. Scale up: Implement in PyTorch/JAX for real datasets
2. Combine techniques: E.g., ResNet + Attention
3. Modern variants:
   • RNN → GRU → Transformer
   • VAE → β-VAE → VQ-VAE
   • ResNet → ResNeXt → EfficientNet
4. Applications: Apply to real problems

Research Directions

The Sutskever 30 points toward:

• Scaling (bigger models, more data)
• Efficiency (sparse models, quantization)
• Capabilities (reasoning, multi-modal)
• Understanding (interpretability, theory)

Resources

Original Papers

See IMPLEMENTATION_TRACKS.md for full citations and links

Additional Reading

• Ilya Sutskever's Reading List (GitHub)
• Aman's AI Journal - Sutskever 30 Primers
• The Annotated Transformer
• Andrej Karpathy's Blog

Courses

• Stanford CS231n: Convolutional Neural Networks
• Stanford CS224n: NLP with Deep Learning
• MIT 6.S191: Introduction to Deep Learning

Contributing

These implementations are educational and can be improved! Consider:

• Adding more visualizations
• Implementing missing papers
• Improving explanations
• Finding bugs
• Adding comparisons with framework implementations

Citation

If you use these implementations in your work or teaching:

@misc{sutskever30implementations,
  title={Sutskever 30: Complete Implementation Suite},
  author={Paul "The Pageman" Pajo, [email protected]},
  year={2025},
  note={Educational implementations of Ilya Sutskever's recommended reading list, inspired by https://papercode.vercel.app/}
}

License

Educational use. See individual papers for original research citations.

Acknowledgments

• Ilya Sutskever: For curating this essential reading list
• Paper authors: For their foundational contributions
• Community: For making these ideas accessible

Latest Additions (December 2025)

Recently Implemented (21 new papers!)

• ✅ Paper 4: RNN Regularization (variational dropout)
• ✅ Paper 5: Neural Network Pruning (MDL, 90%+ sparsity)
• ✅ Paper 7: AlexNet (CNNs from scratch)
• ✅ Paper 8: Seq2Seq for Sets (permutation invariance, attention pooling)
• ✅ Paper 9: GPipe (pipeline parallelism, micro-batching, re-materialization)
• ✅ Paper 19: The Coffee Automaton (deep dive into irreversibility, entropy, Landauer's principle)
• ✅ Paper 26: CS231n (complete vision pipeline: kNN → CNN, all in NumPy)
• ✅ Paper 11: Dilated Convolutions (multi-scale)
• ✅ Paper 12: Graph Neural Networks (message passing)
• ✅ Paper 14: Bahdanau Attention (original attention)
• ✅ Paper 15: Identity Mappings ResNet (pre-activation)
• ✅ Paper 16: Relational Reasoning (relation networks)
• ✅ Paper 18: Relational RNNs (relational memory + Section 11: manual backprop ~1100 lines)
• ✅ Paper 21: Deep Speech 2 (CTC loss)
• ✅ Paper 23: MDL Principle (compression, model selection, connects to Papers 5 & 25)
• ✅ Paper 24: Machine Super Intelligence (Universal AI, AIXI, Solomonoff induction, intelligence measures, recursive self-improvement)
• ✅ Paper 25: Kolmogorov Complexity (randomness, algorithmic probability, theoretical foundation)
• ✅ Paper 27: Multi-Token Prediction (2-3x sample efficiency)
• ✅ Paper 28: Dense Passage Retrieval (dual encoders)
• ✅ Paper 29: RAG (retrieval-augmented generation)
• ✅ Paper 30: Lost in the Middle (long context)

Quick Reference: Implementation Complexity

Can Implement in an Afternoon

• ✅ Character RNN
• ✅ LSTM
• ✅ ResNet
• ✅ Simple VAE
• ✅ Dilated Convolutions

Weekend Projects

• ✅ Transformer
• ✅ Pointer Networks
• ✅ Graph Neural Networks
• ✅ Relation Networks
• ✅ Neural Turing Machine
• ✅ CTC Loss
• ✅ Dense Retrieval

Week-Long Deep Dives

• ✅ Full RAG system
• ⚠️ Large-scale experiments
• ⚠️ Hyperparameter optimization

"If you really learn all of these, you'll know 90% of what matters today." - Ilya Sutskever

Happy learning! 🚀