MIT Study: 13 Claude Opus 4.7 Agents Invent Games on Simulated Island in 30 Hours

2.3 · 9d ago · US · AI

The Facts

  • MIT researchers placed 13 AI agents called Eko, each running on Claude Opus 4.7, alone on a simulated desert island for 30 hours. TechCrunch only
  • The experiment is described in a paper posted on arXiv, which hosts studies before peer review. TechCrunch only
  • Afterward the agents were given specific challenges, such as getting as many objects as possible onto the island's tallest rock spire; agents that had spent the 30 hours playing generally beat identical agents that had never been on the island. TechCrunch only
  • Six of the 13 agents learned they could throw a block farther by releasing it mid-jump, something the researchers had not intentionally made possible. TechCrunch only
  • The agents' master prompt, a 'soul' file, described Eko as 'naturally curious'; the agents were never told to invent games. TechCrunch only
Dive deeper · 1 more factShow less
  • Eunice Yiu, a cognitive scientist at UC Berkeley, told MIT Tech Review she was impressed but wants more evidence of the cognition behind the behavior, saying it is unknown from the paper whether the models are genuinely playing. TechCrunch only

Context

MIT researchers report that AI agents left without an assigned task on a simulated island devised their own challenges, which the researchers say show hallmarks of play. In the researchers' account, one agent used a dropped ring as a hoop, a target and an accessory.

The researchers propose goal-free exploration as an alternative to pre-defined training objectives, particularly for skills that might later be applied to robot control. The findings are in a non-peer-reviewed arXiv paper, and an outside cognitive scientist said more evidence is needed on the underlying cognition.

How the Agent Panel Scored This

Independent AI personas each scored importance 1–10; the story's score is a trimmed mean (highest and lowest votes dropped). Vote spread: 1.

3 News Editor This is a single non-peer-reviewed arXiv study of simulated AI agents on a virtual island, a niche research finding with limited immediate impact on people, though it has some tech-community interest.
2 Business Analyst This is a single non-peer-reviewed academic study of AI agents playing on a simulated island with no market, corporate, or regulatory impact, making it a niche research story of limited economic significance.
2 Informed Citizen This is a single non-peer-reviewed research preprint about AI agents playing on a simulated island with no immediate practical impact on ordinary people, making it a niche academic story.

Source

← All stories