Transformer Circuits 2: The Transformer, Redrawn
Reading the Framework paper, part 2. Redrawing the residual stream as a communication channel and attention heads as independent information movers, through virtual weights and the QK/OV split.
Transformer Circuits 1: Summary of Results
Reading the Framework paper, part 1. Treating tiny attention-only models as model organisms for fully reverse engineering transformers, and a map of the key results.
Mechanistic Interpretability: Free Materials to Start
Reverse-engineering neural networks, circuit by circuit. Anthropic’s Transformer Circuits thread and a few companion resources are all free. Where to start, to open this series.
Steering GPT-2's Emotions with Sparse Autoencoders
Finding emotion-related features in GPT-2 using OpenAI’s pretrained SAE, then training one from scratch. Feature patching turns ‘good person’ into ‘shit’.
K-means, GMM, EM: Three Nested Russian Dolls of Clustering
K-means is actually an extreme case of GMM, and GMM is the canonical application of the EM algorithm. How these three connect within a single framework, and how information geometry explains the relationship.
Reinforcement Learning Lecture 10: RL in Games
David Silver’s Lecture 10. A case study. Best response and Nash equilibrium seen through game theory, minimax search with a binary-linear value function, self-play reinforcement learning, combining minimax with RL, and imperfect-information games like poker. One recipe running through Chinook, Deep Blue, Logistello, TD-Gammon, and Maven.
Reinforcement Learning Lecture 9: Exploration and Exploitation
David Silver’s Lecture 9. Use what you know now (exploitation), or find out more (exploration)? This note ties that dilemma into five principles, then follows the multi-armed bandit through regret and lower bounds, UCB, Thompson sampling, and information-state search, and on to contextual bandits and the extension to MDPs.
Reinforcement Learning Lecture 8: Integrating Learning and Planning
David Silver’s Lecture 8. Learn a model of the environment from experience, then plan with that model. Model learning as supervised learning, the AB example, Dyna and Dyna-Q+, Monte-Carlo Tree Search and Go, TD search and Dyna-2.
Reinforcement Learning Lecture 7: Policy Gradient
David Silver’s Lecture 7. Optimize the policy directly, without going through value. The likelihood-ratio trick and the score function, REINFORCE, baselines and advantage, actor-critic, natural policy gradient, and a summary that ties six faces into one.
Reinforcement Learning Lecture 6: Value Function Approximation
David Silver’s Lecture 6. When states are too many to write down in a table, approximate the value with a handful of parameters. The incremental methods that learn features by gradient descent, mountain car and the bootstrapping debate, Baird’s counterexample where off-policy diverges, and the batch methods and DQN that pool experience.