Building a truly novel game exposes model limits: Opus cuts scope, DeepSeek melts down
teortaxesTex · x · 2026-11-10
A developer asked frontier models to build FLATROOM V3, a genuinely novel offline HTML three.js game with dual realities and folding rooms. Opus prudently proposed cutting scope, admitting in its reasoning that the task is 'essentially impossible to fully implement' while planning a strong playable demo; DeepSeek melted down in confusion; Astra assumed games are slop and didn't try — though it's the only one fit to write the relevant math. Novel tasks expose each model's character.
Related event: Asking frontier models to build a novel game exposes their weaknesses(3 posts)→
More from Fun
- CVPR deadline collides with ICLR rebuttal window, and researchers are dying — mariyaivasileva · 2026-11-10
- 'Marketing cofounder is the new technical cofounder': AI startup circle meme takes off — xkonjin · 2026-11-10
- OpenAI launches Supply Co. merch store with $200 ModRetro Chromatic Codex handheld — btibor91 · 2026-11-10
- The first mass consumer robot will be a hairdresser, not a laundry folder — katieruthmishra · 2026-11-10
- OpenAI merch shipping to Europe costs €82.94 on a €181.95 handheld — Angaisb_ · 2026-11-10
- One Piece creator Eiichiro Oda: 3-hour sleep nights for 15 years, 600M copies sold — floguo · 2026-11-10