Opus Keeps Butchering a Joke-Rewriting Task, Ignoring Explicit Instructions to Preserve Punchlines
ivan_bezdomny · x · 2026-11-10
A blogger reports that Opus consistently fails on a humor-rewriting problem: it not only fails to write better jokes but actively butchers the original, losing both the premise and the punchline even when explicitly instructed to keep them — and it can't write a working prompt or process for itself.
The author's take: once a smarter model finally solves this kind of problem, the task likely can't be "farmed back down" to less intelligent models — suggesting humor may serve as a marker capability for model intelligence tiers.
Related event: Tests Show Claude Fable 5.2 Keeps Failing at Making Jokes Funnier(3 posts)→
More from Fun
- Vintage sitcom still "How To Use Your Robot" resurfaces as a robotics meme — heresalexandria · 2026-11-10
- CybersecurityPostingBeLike: The Bors Comic Perfectly Captures Infosec Discourse — HankYeomans · 2026-11-10
- Anthropic engineer's Fable 5.2 demo hailed as AAA-grade, netizens joke: now make GTA VI — mallow610 · 2026-11-10
- Dev builds Saudi-flavored Candy Crush with Claude Fable 5.2 and MCP assets at wild speed — aziz4ai · 2026-11-10
- Yudkowsky: AI fiction lags code badly — 'we may only be alive because gradient descent generalizes policies so poorly' — sebkrier · 2026-11-10
- An AI-generated parametric engineering drawing this complete makes 'AI mechanical engineers' feel near — burhop · 2026-11-10