a model's inside state for a statement is a point; a straight line separates true statements from false ones, which means the difference is a direction; trained on told-to-lie dialogues, the same test flags deception before the words confess anything

the model scores every word it knows, a fixed recipe turns scores into shares of 100%, temperature reshapes the shares before one word is drawn, and training is the guess-miss-nudge loop that moves the scores

words become points in a space of meaning, attention lets each word gather notes from earlier words under a no-peeking rule, and that loop repeated in layers is most of what a language model is

pictures are grids of brightness numbers, a small window asks one pattern-question at every position, stacked layers turn edges into parts into objects, and shortcuts are what let the stack go deep

the einsum string is a readable proof of a tensor operation, the 0.0 exercises cost 45 minutes and buy fluency for every later chapter, and the part is meant to be verified rather than studied

convolution is a window that walks and pooling is the vote that follows it, a residual stream is memory that tames gradients, and the eleven exercises end with a ResNet34 verified on real images

plain gradient descent overshoots the steep axis and crawls along the flat one, momentum damps the first and accelerates the second, and four exercises make you implement the optimizers you will later trust

the transformer decomposes into ten exercises around a single residual stream, attention is the only new idea inside the block, and ARENA marks this part compulsory for everything chapter 1 later interprets

the part's one move is instrumenting a trained transformer with hooks, attention patterns and ablations, the induction head is the first circuit you can find attribute and ablate end to end, and ARENA marks it compulsory alongside 1.1

an eval starts as a threat model and a specification, not as a quiz, and the exercises build one end to end from claim to scored questions

the part turns a specification into a scored dataset with prompts variance and concurrency, and the scoring function gets built before the data it will grade

the week replaces hand-rolled glue with the inspect_ai pipeline of datasets solvers and scorers, solvers are where the eval's actual logic lives, and the log file is the deliverable that makes results checkable

the week builds an agent twice, once as arithmetic and once as a Wikipedia game, the loop of think act observe is the whole architecture, and tool permissions are where safety enters the picture

AI control evaluates safety measures against an intentionally subversive model, the part plays both roles with backdoor generation on one side and monitors plus protocols on the other, and it completes the toolbox every serious evaluation project builds on

the part runs real API experiments on shutdown resistance and alignment faking, and the first alignment-science results land as numbers from your own pipeline

the last core part audits whether a model's chain of thought reflects its actual computation, the exercises build faithfulness measurements from categorization to counterfactual importance, and the part closes the guide's core route

the guide deliberately leaves one slot unpicked between SAE literacy and intro to RL, both candidates are single-session light parts with no later hard dependency, and the choice is left to the reader's own direction

saving is not a prompt · every reflection must quote you · one Cloud Run service, Firebase, Firestore, Gemini

QUESTION → DATA CONTRACT → DUCKDB + POLARS → 3 AUDITED CHARTS → ESSAY