Stop anthropomorphizing CoT: intermediate tokens just augment the prompt
ThirdWaveCat · reddit · 2026-08-19
The author offers a framing for LLM "thinking": intermediate tokens labeled thinking/reasoning are nothing like human step-by-step reasoning. Humans derive conclusions from steps; LLMs use intermediate traces to augment their own prompt, effectively appending material to context. This explains why a model can produce an excellent answer alongside verbose, rambling "reasoning" — answer quality and trace length aren't inherently linked. The post notes that context flooding and fighting compaction are separate issues that shouldn't be conflated with this.
More from AGI Musings
- After automation: the company with the best clock will beat the one with the best model — every · 2026-08-19
- Professor Warns Using AI for Homework Leads to Exam Failure — paulnovosad · 2026-08-19
- AI Homework Helps Grades But Hurts Exam Performance — paulnovosad · 2026-08-19
- AI Mirrors Outsourcing Impact, Eliminating Mid-Tier Creative Jobs — teortaxesTex · 2026-08-19
- Speculation: OpenAI Paused While Anthropic Advances? Safety Concerns Raised — iruletheworldmo · 2026-08-19
- Century-old Carathéodory conjecture likely disproved with AI help — haider1 · 2026-08-19