Most engineers use the tools they depend on every day without knowing what they are. These guides fix that. Each one starts at the surface — a simple mental model anyone can grasp — and descends through five depths to the abyss, the gory details the framework authors actually live in. Stop at any depth; each is complete on its own.
Into the machine: predictor → tokens → attention → training → inference.
What's actually inside the thing you talk to all day: tokens, the generation loop, how a text predictor learns to chat, and why it fails the way it fails. Watch it pick its next token.
Why the model can't count the r's in "strawberry", why some languages cost triple, and how meaning becomes geometry. Train a real tokenizer in your browser.
Every token asks a question, and softmax decides who gets heard. Queries, keys and values, the residual stream, and where a "70B" actually sits — watch attention resolve a pronoun live.
One move — predict, score, nudge every dial a hair — under four scoreboards: the internet, demonstrations, human taste, checkable answers. Train a real model in your browser.
Prefill reads your prompt in parallel; decode writes the reply one pass at a time; the KV cache bridges them. Why input is cheaper than output, and the bandwidth wall under it all.
One storey up: the engineering built on top of a model you now understand.
A search engine whose only user is a language model. Chunking, embeddings, hybrid search, reranking, lost-in-the-middle — and why the answer's fate is decided before the model speaks.
A loop with hands — and the hands are yours. The model only ever writes text; your code executes it. Tool calls, compounding odds, and why agent safety is architecture, not a prompt.
The last lever: change what the model is. LoRA's skinny matrices, catastrophic forgetting demonstrated live, and the prompt → retrieve → tune ladder — the honest trial of the last rung. The series finale.
Two ways to know a model: measure the behaviour from the outside, open the mechanism from the inside.
Every leaderboard number is a proxy, and every proxy can be gamed. The anatomy of a benchmark, metrics that disagree live in your browser, the statistics of small differences, and Goodhart's law — the honest epistemics of "which model is better?"
Nobody wrote the model — it grew. The logit lens, features as directions, superposition unpacked by a toy model training live in your browser, activation patching, and the first circuits ever read out of a transformer.
A ledger nobody owns, and the two ways the network pays for the right to write it.
From a village notebook down to hashes, Merkle trees, forks, and why a ledger nobody owns can still be trusted. Tamper with a live chain and watch it break.
Mining isn't "solving math problems" — it's a lottery where tickets cost electricity. Down through difficulty retargets, pools, 51% economics, and the energy bill that pushed Ethereum out. Mine a block in your browser.
Security from capital at risk instead of energy burned — validators, attestations, slashing, finality — and the 2022 Merge: swapping a live chain's engine mid-flight.
A junior needs a working mental model. A senior needs the failure modes. The framework author needs the allocation cost. Most explanations pick one audience and lose the rest. These guides serve all three by making depth a scrollable axis — you descend exactly as far as you need, and the metaphor (surface light fading to the pitch-black abyss) keeps you oriented the whole way down.
A single mental model, no jargon. Enough to use the thing correctly.
The real terms and control flow, with an interactive widget to watch it move.
The machinery underneath — compiler transforms, data structures, protocols.
Scheduling, context, the famous bugs, and how to reason about them.
Cost, allocation, edge cases. Where reciting ends and understanding begins.