🏝️ Left alone on a desert island, AI agents started to play.
MIT CSAIL stranded coding agents nicknamed Eko (mainly Claude Opus 4.7, with GPT-5.5 and Kimi K2.6 in later checks) on a simulated island for 30 hours with no assigned task. They invented basketball and floor-is-lava, built constellations in stone, spelled their own name, and cancelled a juggling record for being too easy, then remade it harder.
Agents that spent those hours playing later beat fresh agents on held-out tasks. Deleting the memory of a technique selectively removed the skill.
The caveats: the paper is not peer reviewed, a soul file nudged curiosity, auto-continue prompts kept idle agents going, and the authors admit it could be role-play.
Our take: most AI learns by copying or chasing a reward. Play may be a third path, something closer to childhood.
Read the story
React if play-as-learning should get more research attention.
23
21
21
21
20
20
18
15
15
12
12October 19, 2026 5K 1
