Learn by taking
it apart.
Notes, experiments and small tools for understanding how things actually work. Start with the explanation. Change something. See where it breaks.
October 2026
Neural Networks from First Principles (Raw WASM & JS)
Train and inspect neural networks directly in your browser with zero external ML libraries—powered by an in-memory compiled wasm32 f32 GEMM kernel and a pure Float32Array autograd engine. Walk through Perceptrons & Backpropagation, Convolutional Neural Networks (CNNs), Transformers & Self-Attention, Image DDPM & Discrete Masked Text Diffusion, Decision Pointer Heads, and a composable Block Builder.
Switch live between Raw WebAssembly and Raw JavaScript kernels, benchmark GFLOP/s, draw 8×8 glyphs, inspect multi-head attention matrices, and register new architecture blocks (Mamba SSM, Sparse MoE) at runtime.
October 2026
celld & Durable Object Fleet Architecture
How single-writer SQLite actors coordinate ownership via S3 Conditional Writes (If-None-Match / If-Match) without Raft or ZooKeeper, fence split-brain zombies with monotonic epochs, balance load across peer tunnels, and account for usage across 10,000+ Durable Objects.
Includes interactive simulators for S3 CAS epoch fencing, fleet vs bucket durability, hot-cell 503 shedding, and Credit Leases + Write-Behind Alarm rollups.
September 2026
Jev, Kev & the decision layer
What changes when a model picks from your options instead of writing an answer? Typed decisions, calibrated probabilities, unfamiliar labels and the software around them.
Real on-device inference via Laya (LiteRT.js) and Kev (ONNX Runtime Web), hosted Jev, and multimodal visual deciders from JevImageBench v0.1.5 tested interactively on images in your browser.
September 2026
Chronicles of OPT-175B: Training at Scale
What training a 175B parameter foundation model on 1024 A100 GPUs actually looks like in practice. Interactive graphical timeline of Meta AI’s 148-page logbook, 90+ hardware restarts, the accidental deletion of the entire cluster (“The Omicron SEV”), loss curve stabilization, and automated self-healing.
Derived from public training logs and notes published by Meta AI's Metaseq team (Zhang, Roller, Goyal, Shleifer, Ott).