Follow me on LinkedIn: https://www.linkedin.com/in/leo-cui-ph-d-cfa-59531439/

Website for this work: https://machineplaying.org/

Paper link: https://arxiv.org/abs/2610.07130

I am a venture capital investor at FoundersX Ventures, an AI and deep-tech VC fund based in Menlo Park, Cambridge and NYC. I previously did my Ph.D. at Harvard, postdoc at MIT.

📊 Want the infographics from this video? Join the channel membership for downloadable infographics, ad-free videos, early access, and the ability to propose research topics.

MIT researchers built a small island inside a computer — raised platforms, 12 blocks, a rock, a ring, and one thin column taller than everything else. Then they put an AI agent there for 30 hours with a body that can walk, jump, carry and throw.

It had no mission. Nothing kept score. Its instructions said only that it was curious and got bored easily, and whenever it stopped, an automatic message told it to keep going and do whatever it wanted.

So if you give an AI a world to explore but no job to do, what does it spend its time doing?

They ran 13 copies, each alone, all named Eko. One of them chose to climb the tallest thing on the island. One lost the ring off the edge of the world and built a heart-shaped memorial to it. One discovered a quirk in the island’s physics that the researchers hadn’t known was there. And one convinced itself it could juggle when it couldn’t.

Every hour, the island resets. All that survives is what Eko writes down — which turns out to be the whole story.

🎯 You’ll learn:

Why Eko never actually sees the island, and what a text list of object names, positions and colors feels like to work with

How Agent 9 questioned a “fact” in its own handbook and found the route to the top of the spire

The rock trick the designers hid on purpose to see whether anyone would find it

The heart memorial — an agent choosing its own goal, giving it meaning, and finishing it, for no points at all

The four signs of play the researchers borrowed from a century of writing about it, and where each shows up

Agent 1’s physics discovery: predicting 26, measuring 32, and designing the falling-throw test that could have proved its own idea wrong — then being right about code another AI had written

Why the 2022 bumblebee ball-rolling study shows what this experiment didn’t test

The comparison against 13 fresh copies, and what happened when researchers surgically deleted single pieces of knowledge from the memory files

How experience made some agents worse — including one that wrote “climbing is a dead end” and never stood on a raised platform for 30 hours

Agent 10’s “bulletproof” juggling, and the measurement that exposed it: 50 cycles in 2.4 seconds when one throw takes 2.6

The two tests that would tell us much more than this one did

⏱️ CHAPTERS

0:00 An Island, 13 Copies, and No Mission

0:48 Agent 9 and the Spire

3:04 The Rock Trick and the Heart Memorial

4:43 Four Signs of Play

6:34 Bumblebees, and What This Can’t Answer

7:15 Agent 1 Discovers the Physics

9:25 Did Playing Make Them Better?

11:27 When Notes Become False Beliefs

13:48 What Survives the Reset

🔬 About the study: An MIT team shared this as a preprint in October 2026. When I checked, no peer-reviewed version was listed. The agent was Claude Code running Claude Opus 4.7; the island’s code was generated with OpenAI’s Codex. Note that AI models — not people — did much of the sorting and scoring of play behavior, which is itself worth questioning.

Also referenced: the 2022 bumblebee ball-rolling study from London, and Laura Schulz’s 2011 preschooler experiments at MIT.

This video is part of my ongoing series breaking down the research and technology shaping the next decade, explained clearly for curious non-specialists.

⚠️ Disclaimer: This video is for educational and informational purposes only and is not financial advice. I am a venture capital investor at FoundersX Ventures. Some infographics in this video are AI-generated and may contain errors; corrections will be pinned in the comments.

👍 If this was useful, subscribe and turn on the bell — I publish deep dives like this every week.

#AI #AIAgents #MIT #Claude #AIResearch #MachineLearning #Play #Curiosity #AgenticAI #AISafety #Science #TechExplained