GPT creates stunning figures but completely fails when asked to fix a small detail
anshulkundaje · x · 2026-08-02
Stanford Professor Anshul Kundaje shared his experience using GPT: initially, the model generated a phenomenal figure based on a complex descriptive paragraph. However, when he asked it to fix a technically incorrect detail (making the DNA a correct right-handed double helix), the model kept doing the stupidest things and got worse with each prompt.
He noted that this amazement-to-frustration rollercoaster happens daily, highlighting the brittle nature of current frontier AI models.
Related event: Frontier AI Models Excel at Image Generation but Fail at Precise Edits(2 posts)→
More from Fun
- Robot long jump hits 7+ meters, just 1 meter shy of human world record — TinfoilTricorn · 2026-08-24
- 'Robots could do that decades ago, they're called cars': humanoid sprint debate — teortaxesTex · 2026-08-24
- OpenAI patches goodwill glitch; t3gg's Theo quietly deletes posts to dodge blame — ns123abc · 2026-08-24
- Developer haider makes his own LLM tier list after disagreeing with theo's rankings — haider1 · 2026-08-24
- AI video model fails to understand Jenga physics, creating absurd collapse — Sad_Coach_1433 · 2026-08-24
- AI Buzzwords Depletion: Era of 'Neo-Harness' Incoming? — cocktailpeanut · 2026-08-24