MIT study: LLM agents left to play freely for 30 hours learn to build towers with no rewards
burny_tech · x · 2026-10-21
MIT researchers (including Jacob Andreas, Phillip Isola and Daniele Rus) dropped an AI coding assistant into a physics sandbox island with no tasks, rewards, or objectives.
- Over 30 hours of unguided activity, agents climbed hills, stacked cubes into towers, drew mandalas, reinterpreted sports, ran physics experiments, and acquired techniques that later expanded what they could accomplish.
- With no RL and no weight updates, 62% of agents built >2 towers on a many-5-towers task (up to 16), vs 15% for naive baselines (peaking at 6).
- All 13 agents started identically but diverged into distinct histories; the paper asks whether this satisfies classical criteria for play, and whether play could be a mode of machine development.
More from AGI Musings
- Abacus.AI's Bindu Reddy warns Anthropic seeks frontier lead to dictate AI regulation — bindureddy · 2026-10-21
- Diamandis: People Must Share in AI-Created Wealth Through the Tech Transition — PeterDiamandis · 2026-10-21
- Wharton's Mollick: AI Voice Interfaces Could Finally Kill the Open Office — emollick · 2026-10-21
- AI's gigawatt appetite has quietly killed Big Tech's carbon-neutrality talking points — XFreeze · 2026-10-21
- Crypto is the operating system of finance that AI agents can natively use, dev argues — jarrodwatts · 2026-10-21
- X's Nikita Bier: AI's short-term GDP contribution may be neutral after breaches and spam — nikitabier · 2026-10-21