The first cohort ran the full curriculum remotely: eight sessions, each one code against a live model, with teams formed in week one and a talk delivered in week eight. Participants wrote their own hooks, built a versioned eval set with controls, produced a patching map and a probe with a control-task baseline for a concept of their choosing, and rebuilt their main figure after the room attacked it.
Week by week
- Why did it say that? Teams formed, two or three candidate concepts.
- Make it say something else. Ten to twenty contrastive pairs and one steering demo.
- Prove it. Versioned eval set, baseline, repository.
- Where does it actually happen? Patching heatmap, draft started.
- What can you read off a vector? Probe with a control-task baseline.
- Attack the figure. The figure that becomes Figure 1.
- Make it survive review. Introduction and methods draft.
- Why is your work cool? Talk delivered, venue and date set.
What came out of it
Two team projects continued past week eight and are now under review at NeurIPS 2026 workshops, with the participants as authors.
- Doubt Is a Decision: A Sparse Circuit Commits Language Models to “I Don’t Know”. Sheryl Mehta, Shagun Chadha, and Kanika Singhal, co-first authors, with Manan Wadhwa and Shivam Dubey.
- Steered Affect Moves a Partner Agent’s Activations, and the Logistic Probe That Reads It Returns a Tie-Break, Not an Estimate. Debojeet Ghosh, Shreyansh Kumar, and Ayushman Garg, with Manan Wadhwa and Shivam Dubey, co-first authors.