Do induction heads and attention sinks count? Debate over interpretability's missed milestone

aryaman2020 · x · 2026-09-03

Replying to stuhlmueller's pessimism about interpretability progress, aryaman2020 questions the premise: induction heads, function vectors, and attention sinks should all count as novel algorithmic insights extracted from LLMs, so how has the field not met the 2026 goal?

Related event: Researchers Clash Over Whether Interpretability Is Delivering Real Insights(3 posts)→

Original post →

More from Research

Research channel →