Aniket Ghosh · Writing

EverythingIwrite,inoneplace.

Two things, mostly: what's actually going on inside neural networks, and whatever else I've been chewing on. It lands in a few different places. This page pulls it back together.

Start reading Portfolio

Writing

researchAug 6, 2026

Function vectors as a model diffing tool: 17 heads repair a bad fine-tune

Copying 17 attention heads from the base model into a medical fine-tune gone wrong undoes most of the harmful behavior without costing much capability.

#interpretability
#model-diffing
#activation-engineering

LessWrong

researchJul 25, 2026

Linear probes tell you where quantization will hurt

Probing where a capability lives in the residual stream predicts which layers survive low-bit quantization and which ones you can't touch.

#interpretability
#quantization
#mech-interp

LessWrong

Elsewhere

Portfolio What I build. Projects, research, the résumé.
LessWrong Where the interpretability work goes.