Can reinforcement learning assign more rides than greedy?
When accepting a ride now can block two later: what DQN, AlphaZero-style search, and three state representations taught us about learning to plan.
Blog
Notes on algorithms, optimization, and machine learning.
Understanding PyTorch: From Tensors to a Model That Learns
Follow one example from a prediction to a parameter update. Tensors, gradients, training loops, neural networks, and inference—with visible shapes and a complete runnable script.
From Language Models to AI Assistants: Understanding Transformers
How we got here, what happens inside one prediction, and how models become useful systems. A worked guide to attention, training, KV caching, reasoning, retrieval, and tools—with the names kept straight.
Policies, Values, and Planning: Understanding Reinforcement Learning
From Q-learning to PPO and AlphaZero: what each method learns, how it chooses an action, and why you might pick one over another. One running example and a glossary to keep the names straight.
From Hexagons to Driving Times
Precompute a small picture of the road network, then estimate the drive. H3 landmarks, a conservative route bound, and three ML models—with measured accuracy and storage tradeoffs in three cities.
Navigating the maze: a dive into Dijkstra’s algorithm for finding shortest paths
From a city full of intersections to a few lines of Python: how a greedy choice finds the shortest path, why it works, and where other algorithms fit.