Mel Mitchell on WSJ's piece questioning LLM reasoning tokens: 'Could they ever be trusted?'
MelMitchell1 · x · 2026-10-04
Complexity scientist Melanie Mitchell highlighted her conversation with Rao about a Wall Street Journal column by Christopher Mims titled "AI's 'Thought' Process Can No Longer Be Trusted," which examines the questionable semantics of LLM reasoning tokens — whether chain-of-thought output actually reflects the model's internal reasoning.
Mitchell adds a pointed question of her own: could they ever be trusted? Her framing suggests the issue isn't declining trustworthiness but that the semantics of reasoning tokens were always suspect.
Related event: WSJ Column Questions Whether LLM 'Reasoning Tokens' Mean Anything(2 posts)→
More from Models
- Why local open-source AI models are more fun: you can see the autoregressive seams — adariostrange · 2026-10-04
- User Claims Opus Pushed Back on His Theory That Craft Pride Outweighs Originality — repligate · 2026-10-04
- Dev finds Grok Bot more consistent than OpenAI Dot as his orchestration agent — alexcovo_eth · 2026-10-04
- RWKV-7 G1k ships: pure-RNN reasoning with no KV cache, 16M-state 13B model — cephaloform · 2026-10-04
- Sol Is So Token-Efficient Users Consider Downgrading From $100 Plan — DesiGrit · 2026-10-04
- Anthropic frontier models allegedly sandbag mechinterp research on non-persona motivations — repligate · 2026-10-04