Mat Dryhurst: you can simply instruct a reasoning model to be bad at spelling
matdryhurst · x · 2026-11-04
AI researcher Mat Dryhurst notes it's entirely possible to instruct a reasoning model to henceforth be bad at spelling. He found the first viral videos funny, but once one hits, an incentive forms to repeat it — "cui bono".
A reminder that seemingly spontaneous model quirks in viral clips may be deliberately induced and mass-reproduced for engagement.
Related event: Mat Dryhurst Warns Profitable Misleading AI Content Is the Real Risk(3 posts)→
More from AGI Musings
- Security researcher: "no AI needed" breaches indict an industry stagnant for 20 years — chrisrohlf · 2026-11-04
- a16z partner: proof of agent is the 2020s version of GPU mining — adamamcbride · 2026-11-04
- Tuned harness beats the latest model: older GPT gains ~10 points untuned — VictorSanh · 2026-11-04
- What's missing from AI utopias: what the future actually feels like day to day — ryan_t_lowe · 2026-11-04
- Hot take: new AI methods are BS until someone runs them in prod three weeks later — generativist · 2026-11-04
- AI ethicist claims AGI lab leaders all privately believe the risk but stay silent — danfaggella · 2026-11-04