MIT study finds AI agents learned by playing on a virtual island

FORMAT B - Trend

In short: Researchers are testing whether AI can learn useful skills by exploring and playing, not just by following set tasks.

What's going on

MIT researchers ran an experiment with 13 AI “agents” named Eko. An AI agent is a program that can take actions in a digital world, a bit like a character in a video game that decides what to do next.

The team placed the agents on a simulated desert island for 30 hours with no specific job to complete. Each agent ran on Claude Opus 4.7, a large language model (a text predicting system that can also plan actions). The study was posted on arXiv, a site where researchers share early versions of papers before formal review.

With nothing assigned, the agents began creating their own goals and activities. One tried to reach the island’s tallest rock spire, then improvised by stacking objects to jump higher. Another threw a golden ring into the ocean, then built a “memorial” out of stones, and later reused the ring as a hoop and target when it reappeared.

After the 30 hours, the researchers gave the agents specific challenges, such as moving as many objects as possible onto the tallest spire. The agents that had spent time freely exploring generally did better than identical agents that had not been on the island.

The researchers also reported unexpected discoveries, like some agents learning a longer throwing trick by releasing a block mid-jump. In other words, the agents picked up “rules of the world” that the designers did not explicitly teach.

What to watch

This idea could affect how future AI systems are trained, especially for robots that need to handle messy real-world situations. Outside experts say the behavior is interesting, but more evidence is needed to show whether this is true “play” or just a convincing imitation.

Source: TechCrunch AI