In The Matrix, it becomes obvious it’s a simulation when the black cat appears twice. World models fail similarly: visually the generated simulation looks realistic, but the underlying distribution is wrong. Introducing CaliBench, a measurement of the randomness of reality!

Introducing Agora-1, a world model that's learned to simulate multi-agent experiences. It's so fun. Today we're launching a playable research preview, where you can relive your childhood and enjoy a multiplayer simulation of GoldenEye. So excited about this new capability!

Introducing Starchild-1 from @odysseyml, the first ever real-time multimodal world model. This a model that can generate interactive simulations of the world that you can—for the first time ever—hear. Starchild-1 represents a big step towards a general-purpose world simulator.

TOMORROW, 9:30AM PT We go live with an @odysseyml announcement (or two) I've been looking forward to for some time. I think these releases will make where world models are headed, and how powerful they can become, much more obvious. So hyped. GPUs scaling as we speak.

New blackboard lecture w @ericjang11 He walks through how to build AlphaGo from scratch, but with modern AI tools. Sometimes you understand the future better by stepping backward. AlphaGo is still the cleanest worked example of the primitives of intelligence: search, learning from experience, and self-play. You have to go back to 2017 to get insight into how the more general AIs of the future might learn. Once he explained how AlphaGo works, it gave us the context to have a discussion about how RL works in LLMs and how it could work better – naive policy gradient RL has to figure out which of the 100k+ tokens in your trajectory actually got you the right answer, while AlphaGo’s MCTS suggests a strictly better action every single move, giving you a training target that sidesteps the credit assignment problem. The way humans learn is surely closer to the second. Eric also kickstarted an Autoresearch loop on his project. And it was very interesting to discuss which parts of AI research LLMs can already automate pretty well (implementing and running experiments, optimizing hyperparameters) and which they still struggle with (choosing the right question to investigate next, escaping research dead ends). Informative to all the recent discussion about when we should expect an intelligence explosion, and what it would look like from the inside. Timestamps: 0:00:00 – Basics of Go 0:08:06 – Monte Carlo Tree Search 0:31:53 – What the neural network does 1:00:22 – Self-play 1:25:27 – Alternative RL approaches 1:45:36 – Why doesn’t MCTS work for LLMs 2:00:58 – Off-policy training 2:11:51 – RL is even more information inefficient than you thought 2:22:05 – Automated AI researchers

In Minecraft, PROWL dramatically improved world model performance across physics, visuals, action following, and long-horizon consistency. A really promising lever of scale for world models.

What if world models could learn by discovery? Today we’re sharing PROWL: RL agents that explore game environments, simulators, and eventually robots to discover failures in a world model. This loop of learning is fully automated!

Introducing PROWL! We’ve built RL agents that explore game environments, tasked with discovering failures in world models across physics, visuals, and actions. Those failures then become training data in an automated loop that advances world model performance.