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.

10 pages, newest marked. Hover a card for the one-line version; the list view has it written out, with reading times.

How models are builtthe blocks

From GPT-2 to Kimi K3Jul 2026 How multimodal models workJul 2026 How reasoning models workJul 2026

How models are traineda pipeline

Pre-training vs post-training01 · May 2026Pre-training, step by step02 · Sep 2026 · newPost-training, step by step03 · coming

How models become agentsa loop

The agentic loopAug 2026The agent harnessMay 2026

Four protocols, one conversationMay 2026

Elsewhereoff the map

The GPU stackJul 2026 Dark matterJun 2026

How models are built

The architecture from the inside, one mechanism at a time.

From GPT-2 to Kimi K3
22,580 GPT-2s fit inside Kimi K3. The lineage: the KV cache, linear attention, DeltaNet, gating, MoE, AttnRes. Every code block explained line by line.
updated Jul 30, 2026 · ≈ 38 min
How multimodal models work
The same next-token machinery, reading image, audio and PDF. What the tokens look like.
updated Jul 23, 2026 · ≈ 21 min
How reasoning models work
Not a new kind of brain: the same predictor trained to spend its tokens differently, and served on hardware that feels the difference.
updated Jul 15, 2026 · ≈ 16 min

How models are trained

The overview first, then each phase taken apart. Reads best in order.

01Pre-training vs post-training
One enormous run buys raw capability. A cheap stack of stages after it buys everything you actually talk to. Phase by phase, with the numbers.
updated May 30, 2026 · ≈ 15 min
02Pre-training, step by stepnew
Eight steps from crawl to base model, each opening into the idea, the code and the scaffolding. Real data records, code read from the real repos.
updated Sep 17, 2026 · ≈ 65 min
03Post-training, step by stepcoming
Chat format, SFT, reward signals, DPO, RL with GRPO, and the systems that run it. Same three layers.

How models become agents

From a token predictor to something that reads files and runs commands.

The agentic loop
Special tokens, a parser, and a small Python loop. Four interactives, from sampling to the while loop.
updated Aug 4, 2026 · ≈ 8 min
The agent harness
Two ways to wrap a model. Claude Code, reverse-engineered. OpenClaw, wide open. Mechanic by mechanic.
updated May 29, 2026 · ≈ 9 min
Four protocols, one conversation
API, MCP and the two ACPs, and how they talk to each other over DNS, TLS and HTTP.
updated May 31, 2026 · ≈ 13 min

Elsewhere

Because not everything is a language model.

The GPU stack
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 · ≈ 22 min
Dark matter
Most of the universe is something we have never seen and cannot make. We are still sure it is there. Here is why.
updated Jun 7, 2026 · ≈ 9 min