Interactive explainers for ML, RL and LLM inference. Each one is built from an experiment I actually ran, and every chart says whether its numbers are measured or simulated.
Continuous batching, paged KV cache, the decode roofline and an engine-vs-vLLM scoreboard, on real Tesla T4 numbers from minivllm.
measured simulatedStep through the tiles and watch the online softmax keep a running max and rescale, without ever storing the full attention matrix. Checked against naive attention live.
computed liveWhat a speech recognizer does to Igbo tone marks, clip by clip, with four ways to score it and honest error bars on 21 recordings.
measured computed liveWatch eligibility traces spread credit backwards, sweep lambda on the 19-state random walk, then see Baird's counterexample diverge and TDC rescue it.
computed live