Architecture study sparks monitoring concerns over a multi-hop planning spike
ChowdhuryNeil · x · 2026-11-12
Commenting on part 4/10 of jeffreygwang's architecture research thread: comparing three architectures at shared base-model perplexity, perplexity predicts performance across architectures for most no-CoT evaluations.
Neil Chowdhury's read on the subtext: looping doesn't reduce monitorability more than other perplexity-reducing architectural changes — except "Architecture 2" shows a notable spike in multi-hop planning and serial computation relative to perplexity, a detail worth watching.
More from AGI Musings
- In 2026 an obviously Claude-written cover letter makes recruiters doubt your whole CV — Afinetheorem · 2026-11-12
- Dev complains security features are gated behind expensive enterprise plans — AccBalanced · 2026-11-12
- The AI trap: doing unneeded work faster—prioritize before you scale — jiayuan_jy · 2026-11-12
- Rutger Bregman: The AI 'Panic- mongers' Have a Decade of Correct Predictions — birchlse · 2026-11-12
- Toby Ord argues predeployment testing misses the real risk: internal frontier models — ben_j_todd · 2026-11-12
- "It's Probably Fine": AI doom skepticism draws support in risk debate — NathanpmYoung · 2026-11-12