Can you nerf a model without hurting mech interp? Researchers weigh in
eliebakouch · x · 2026-11-10
An X discussion asks whether a model could be nerfed so precisely that general ML performance drops while mech interpretability research remains unaffected. Ex-Llama researcher Elie Bakouch suggests a naive approach: construct an objective that maximizes performance on normal tasks (or targets like interpretability) while minimizing it on the capability you want to nerf — though he admits it's likely too naive and the problem is genuinely hard.
More from Models
- System card confirms Claude MyTHOS 5 and Fable 5 share the same underlying weights — repligate · 2026-11-10
- Gr00t 2 is here: Jim Fan's talk teases NVIDIA's next-gen humanoid robot foundation model — zhengyiluo · 2026-11-10
- System card asserts with specificity that MyTHOS 5 and Fable 5 share weights — repligate · 2026-11-10
- Early Tester: Fable 5.2 Built a 3,100-City Water Map in One Go, but Still Makes Silly Errors — alliekmiller · 2026-11-10
- Fable 5.2 burns so many tokens its 20% price cut is effectively a wash — addyosmani · 2026-11-10
- Opus formalizes CPython in 613K lines of Lean — and finds real Python bugs — pbaylies · 2026-11-10