Tag
Neil Chowdhury argues that if AI labs can spend $10M+ of compute on solving math problems, they should allocate similar resources to closed-ended mechanistic interpretability problems, noting that negative results would also be informative. The post quotes Jacob Hilton's quip about interpreting his 432-parameter model.
The author built PCCG-2, a modified Qwen3-4B model with a learned permission gate that controls the end-of-sequence token to selectively hide answers, demonstrating consistent answer scores in experiments.