The lab journal — dated milestones from building a neural-network framework from zero, newest first. Each entry links to the evidence it produced.
Lyceumai.dev redesigned as an editorial research journal — single warm light theme, serif statements, and every number on every page rendered from the same abort()-on-fail gates that verify the code. Custom 404, error boundary, robots and sitemap ship alongside.
See the researchThe complete 30k-epoch grokking reproduction lands: MLP 28→128→14 on all-but-diagonal (a+b) mod 14 pairs, both weight decays run to completion. Held-out accuracy reaches 1.00 at epoch 9,000 (wd=1.0) and 25,000 (wd=0.3) — stronger decay crosses first, matching the reference ordering.
Read the curvesStage C closes: Embedding, LayerNorm, causal SDPA, multi-head attention, TransformerBlock — each with its numeric gradient check. MiniGPT (Embedding + learned positions + 2 blocks + seq head) fits a period-8 char grammar at train 97.7%, held-out 94.5%.
Model resultsArena autograd and optimizers land in the same window: a dynamic Graph of Value nodes with backward dispatch, then SGD, momentum, Adam, and AdamW with decoupled weight decay. The generic trainer follows — fit/predict/evaluate over any model meeting the NNBase trait.
How an epoch runsThe ml/ package takes shape: extended Tensor ops on row-major Lists of Float64, cross_entropy / mse / binary_cross_entropy losses, and the synthetic-task dataset layer with deterministic LCG shuffling. Every op ships with its central-difference gradient check from day one.
Module by module