I write and teach Redarc Fellows, the eight-session mechanistic interpretability cohort at Redarc Labs. There is no interpretability library in the stack: hooks are written by hand, because people learn the thing faster when nothing is hidden from them.
One rule is set in session one and repeated every session after it: a rigorous investigation with a careful negative result counts as a success. Two papers came out of the first cohort.
In person, in Delhi. The same curriculum now runs as an offline bootcamp with Secure AI Futures Lab at their Delhi hub, where I am a special projects fellow. Alongside it we run reading groups on the weekends.
Curriculum
90 min
Why did it say that?
Open a language model, trace how it builds an answer layer by layer, and compare Hindi and French.
hooks, logit lens, residual stream
Milestone: Teams formed, with two or three candidate concepts.
90 min
Make it say something else
Change the model’s output by adding a vector to its residual stream, then measure the cost.
steering, contrastive pairs, working range
Milestone: Ten to twenty contrastive pairs and one steering demo.
90 min
Prove it
Turn the previous week’s result into a measurement with a versioned eval set, a baseline, and controls.
evaluation, controls, model-written evals
Milestone: Versioned eval set, baseline number, repository.
90 min
Where does it actually happen?
Use activation patching to locate where the output depends on a given site.
activation patching, causal evidence, attribution patching
Milestone: Patching heatmap for your concept, and a draft started.
90 min
What can you read off a vector?
Train a probe on activations and learn why a high accuracy alone proves little.
probes, control tasks, representation geometry
Milestone: Probe with a control-task baseline.
120 min
Attack the figure
Teams present their best figure for group critique, then rebuild it to a reviewer-acceptable spec.
figures, triangulation, seeds
Milestone: The figure that becomes Figure 1.
90 min
Make it survive review
Turn the figure into a paper.
writing, peer review, methods
Milestone: Introduction and methods draft, plus a review of another team’s draft.
120 min
Why is your work cool?
Teams give talks.
demo day, talks, submission
Milestone: Talk delivered, with a venue and date set for the write-up.
The curriculum at Redarc Labs →
Also
I mentor three first-author projects in Cohort 01, built around one thesis: expressed confidence in LLMs is separable from internal knowledge. Two of those papers are under review at NeurIPS 2026 workshops.