Eight sessions, each one code against a live model. There is no interpretability library in the stack: hooks are written by hand so everything runs unchanged in a work repository afterwards. The rule stated in week one and repeated weekly is that a rigorous investigation with a careful negative result is a success.
01Why did it say that?
90 minutesOpen a language model, watch it assemble an answer layer by layer, then check whether it does the same thing for Hindi as it does for French.
Milestone · Teams formed, two or three candidate concepts
Make it say something else
90 minutesChange what a model says by adding one vector to its residual stream, then measure what that change cost.
Milestone · Ten to twenty contrastive pairs and one steering demo
Prove it
90 minutesTurn last week's number into a measurement: a versioned eval set, a baseline, and the controls that decide whether your explanation is wrong.
Milestone · Versioned eval set, baseline number, repository
Where does it actually happen?
90 minutesActivation patching: copy one site between a clean and a corrupted run and find where the output actually depends on it.
Milestone · Patching heatmap for your concept, draft started
What can you read off a vector?
90 minutesTrain a classifier on activations, get a high number, and learn why that number on its own proves almost nothing.
Milestone · Probe with a control-task baseline
Attack the figure
120 minutesEach team projects its best figure and the room takes it apart. Then you rebuild it to a spec a reviewer would accept.
Milestone · The figure that becomes Figure 1
Make it survive review
90 minutesYou have a figure. Tonight you find out whether you have a paper.
Milestone · Introduction and methods draft, plus a review of another team's draft
Why is your work cool?
120 minutesEach team gives the talk. The milestone is a venue and a date for the write-up, not applause.
Milestone · Talk delivered, venue and date set for the write-up