Mitchell Hashimoto explains why models get locked onto your examples instead of the task
mmitchell_ai · x · 2026-10-24
Chris Albon complained that communicating by examples makes AI models "lockjaw" onto them and stop thinking. Mitchell Hashimoto explained the mechanism: when your input supplies examples of ingredients rather than the recipe, the "recipe" token has no matching pattern, so the model tries to complete the output along the lines of the examples instead of addressing your actual ask. Hashimoto followed up with the workaround: restructure examples as a list the model must complete (one-shot / few-shot prompting).
More from Apps
- Morphic launches Recast Pro to swap people in any video while keeping motion, cuts and audio — Scobleizer · 2026-10-25
- CapCut plugin launches in early access on ChatGPT desktop, mobile and web due mid-November — Aiden_Tech_Ai · 2026-10-25
- Hospital hands him a Windows XP disc, Claude Code reads his X-ray instead — korymath · 2026-10-25
- Founder: apps should be agents, not apps with agents — the edge is proprietary data plus great UX — xkonjin · 2026-10-25
- Latest Pixel phones caught faking zoom details with AI image generation — NathanpmYoung · 2026-10-25
- Anthropic x Adaptyv protein design contest challenges participants to design Ozempic-like drugs — AllThingsApx · 2026-10-25