Mini Blog Post 7: Problems Are for Fixing
Source: https://www.neelnanda.io/blog/mini-blog-post-7-problems-are-for-fixing Author: Neel Nanda Date: 2025-12-08
Summary
Short but sharp post from Neel Nanda (Google DeepMind, mechanistic interpretability) on research culture. Core thesis: the most important skill in research is not cleverness but the disposition to actually fix problems when you find them — most researchers document and move on.
Key Claims
- The failure mode: researchers find a problem (data quality issue, eval bug, unexpected result), note it, and continue. The problem persists.
- “Problems are for fixing”: the right response to finding a problem is to fix it, not to caveat it in a paper.
- This applies to interpretability research specifically: finding a feature or circuit is only useful if you use it to fix something about the model.
- Corollary: papers that just describe phenomena without interventions are less valuable than they appear.
- Meta-point: this disposition is learned, not innate. Most researchers weren’t taught to fix — they were taught to describe.
Entities
- Neel Nanda — Google DeepMind; leading mechanistic interpretability researcher
Concepts
- Mechanistic Interpretability — Nanda’s domain; critique of describe-only research
- Autonomous Research — relevant to how AI research agents should behave