Researchers showed an AI a video of a robot arm. Two joints, waving around at random. No label was there, It was just the raw footage.

The AI watched it and figured out, on its own, that exactly two things were moving. It pulled the structure of the arm straight out of the pixels. Nobody told it the arm had two joints. It just knew.

Then they changed one thing. Same arm, same camera. But instead of letting it wave around randomly, they let it try to reach for a target.

This time the AI failed. Badly.

Everything was same. The only difference was that the arm now had a goal.

I'm bringing this up because of where it comes from. The paper went up last week. It's called "When Does LeJEPA Learn a World Model?" and one of the three authors is Yann LeCun.