V-JEPA: A new architecture for self-supervised learning

So far we've trained our biggest AIs on text, so they understand how we (humans) understand the world.

And we've trained other AIs on images, so they understand what things look like (in a static way).

But the frontier that we couldn't cross before is understanding how things work, independently from our (human) filter, to make better predictions that match the real-world better.

Today, at the same time OpenAI released SORA, Meta (Facebook AI Research) released V-JEPA, which also aims to do exactly this.

So on the same day we've got TWO huge announcements of an epic increase in AI capabilities.

I believe this will ripple into robotics extremely fast, because a robot that is capable of predicting the outcome of a movement becomes much more autonomous. The next few months are going to be a wild ride.