MIT Agents Left Alone Began to Play, Researchers Say; Drop7 Project Tests Automated Strategy
MIT researchers say 13 Eko agents running on Claude Opus 4.7 invented their own challenges after 30 hours alone on a simulated island and later outperformed agents without that play time; a separate Drop7 project tests automated game strategy.
The experiment is described in a TechCrunch report cited by Slashdot. After the free-exploration period, the researchers assigned tasks such as getting as many objects as possible onto the island’s tallest rock spire. The researchers see exploring without a goal as a different way for AI agents to learn skills, especially skills that could one day be used to control robots. They wrote that developers normally decide in advance what AI should practice and what counts as success.
“We normally develop artificial agents by deciding in advance what they should practice, what they should optimize, and what counts as progress,” the researchers write in their report. “Play suggests another regime: give an agent a sufficiently rich and safe world, let it choose what is interesting, and allow useful knowledge to accumulate from the history that follows.”
UC Berkeley cognitive scientist Eunice Yiu told MIT Technology Review she was impressed with the findings but that more evidence is needed to determine whether AI models are genuinely playing. “It could be the case that the models are genuinely playing, but we don’t know based on the paper,” Yiu said. “If you only have the behavioral piece, it could be the case that they’re just trying to abide by the prompt, trying to imitate humans.” Phillip Isola, an MIT professor and coauthor of the paper, added: “It’s not like this is a system that just suddenly magically decided to play. It’s an artifact that learned from our culture, was trained by engineers at Anthropic, and was then placed on an island by us. It does things that surprise us and that are causally related to its future skillful behavior.”
Separately, a developer wrote in a Hacker News post that sixteen years ago he wrote a Java program to pick the ideal move in the puzzle game Drop7 and renewed the investigation this weekend into whether an autonomous strategy can average more than one million points per game. On each turn the player has only seven choices for the next move, but an element of chance turns the simple game into a problem involving search, hidden information, long-term planning and chance, according to the post. An experienced human player can score a million points with long-term strategic thinking, the developer wrote, while the best automated strategies he could come up with are currently averaging around 300,000 to 400,000 points.
The developer said one of the easiest ways to contribute is to compete on the global leaderboard and submit a human strategy. The project also asks for feedback from people knowledgeable in game theory, stochastic planning, learned value functions or heavy-tailed evaluation, and the developer described it as an enlightening experiment into using agents as research collaborators and evidence collectors.