Explainers
Pages I build to learn one thing by taking it apart. Each one is a self-contained, scroll-driven page with hand-drawn diagrams, and it works in light and dark. The map reads top to bottom: how a model is built, how it is trained, how it becomes an agent.
12 pages, newest marked. Hover a card for the one-line version; the list view has it written out, with reading times.
The architecture from the inside, one mechanism at a time.
22,580 GPT-2s fit inside Kimi K3. Each change that got us there (the KV cache, linear attention, DeltaNet, gating, MoE, AttnRes), with every code block explained line by line.
updated Jul 30, 2026 · ≈ 41 min
How one next-token predictor reads images, audio and PDFs, and what those inputs look like once they become tokens.
updated Jul 23, 2026 · ≈ 22 min
The same next-token predictor, trained to think in writing before it answers. How that training works, and why all that thinking is costly to serve.
updated Jul 15, 2026 · ≈ 17 min
The overview first, then each phase taken apart, then one you run yourself. Reads best in order.
One huge run teaches a model what it knows. A few much cheaper stages afterward teach it to talk to you. Each phase, with the numbers.
updated May 30, 2026 · ≈ 17 min
Eight steps from web crawl to base model, each opening into the idea, the code and the scaffolding. The data and code come from real datasets and repos.
updated Sep 17, 2026 · ≈ 77 min
Eight steps from base model to chat model, each opening into the idea, the code and the scaffolding. Chat templates, loss masks, DPO and GRPO, with examples small enough to work by hand.
updated Sep 18, 2026 · ≈ 64 min
The hands-on one. Fine-tune a 1.5B model on a Mac in two minutes with QLoRA. Every command is copyable and every number comes from a real run, including the test where it struggled.
updated Sep 23, 2026 · ≈ 32 min
From a token predictor to something that reads files and runs commands.
How a model that only predicts tokens becomes an agent: special tokens, a parser and a small Python loop, in four interactive steps.
updated Aug 4, 2026 · ≈ 9 min
Claude Code (reverse-engineered) and OpenClaw (open source) side by side: how each one turns a model into an agent, part by part.
updated May 29, 2026 · ≈ 11 min
API, MCP and the two ACPs, and how they talk to each other over DNS, TLS and HTTP.
updated May 31, 2026 · ≈ 15 min
Because not everything is a language model.
The AI chip supply chain, from a fistful of sand to a datacenter. What each layer does and who owns it.
updated Jul 15, 2026 · ≈ 25 min
Most of the universe is made of something we have never seen and cannot make in a lab. How we know it is there anyway.
updated Jun 7, 2026 · ≈ 9 min