This is a special post for quick takes by Shantanu Darveshi. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
I think we are at a point in history where interpretability research needs to expand beyond mechanistic explanations and explore approaches that are practically useful for humanity's AI Safety efforts. The fact that current techniques would require an absurd amount of compute to run on frontier models is strong evidence we’re working at the wrong level of abstraction.
This pattern isn’t new. When mechanistic explanations become too difficult relative to the complexity involved, new fields emerge that adopt higher-level abstractions and practical heuristics.
Neurology ➝ Psychology
Electrical Science ➝ Computer Science
Physics ➝ Statistical Mechanics
The present moment is precious and calls for deliberate effort in the right direction. Incremental improvements are not enough when capabilities are growing rapidly e.g. Gemini 3 performance across all benchmarks and METR's report on long-horizon tasks. A low quality seat belt is better than no seat belt at all.
If you found this interesting I recommend reading this discussion from 2 years ago.