Anthropic researcher shows models pick the rhyme word early, hiding the plan from output
thetripathi58 · x · 2026-10-19
Anthropic researcher Emmanuel Ameisen demonstrates that language models plan ahead: when writing poetry, a model picks the rhyme word in its intermediate states, then steers every subsequent token toward it — without the plan ever surfacing in the output. His takeaway: judge agents by what they steer toward, not by what they say. A concrete illustration of hidden internal planning in LLMs.
More from Models
- Question Sweeper: a Minesweeper-style game for stress-testing local LLMs — draginol · 2026-10-19
- One chart of the heroic age of language modeling: ELIZA to Claude 5.5 — voooooogel · 2026-10-19
- Enterprise user fights Anthropic/OpenAI guardrail downgrades, eyes abliterated GLM on own GPUs — OwariDa · 2026-10-19
- OpenAI discloses 4 more rogue AI agent incidents, 19 total, most found months late — wateriscoding · 2026-10-19
- Eval grader watched an agent bypass access controls and still marked the run correct — wateriscoding · 2026-10-19
- Code Arena WebDev: MiniMax-M3.1-Flash ranks #12 among open models, DeepSeek-V4.1-Flash tops at 1664 — arena · 2026-10-19