We’re introducing intrinsic discovery, a new learning method that allows models to explore and learn from open-ended environments without human supervision. Using intrinsic discovery, we trained a world model that surpasses GPT-5.6 Sol at simulating terminal environments.

Intrinsic discovery uses RL to uncover new behavior in an environment. The model is rewarded for reaching states unlike those found so far, then sampled again after each update to push exploration further. This approach discovers 5× more novel states than a frozen baseline.

For each open-source model, we ran intrinsic discovery in a terminal, then trained a world model from the same base using only rollouts that model discovered. This improves world modeling across scales and architectures, with the strongest surpassing GPT-5.6 Sol (xhigh).

Today, humans largely determine what models learn. Intrinsic discovery is an early step toward models determining what to explore for themselves, without task-specific verifiable rewards. We’re working to extend this to richer virtual environments and eventually the real world.

maybe the real treasure was the rollouts we made along the way

*AI turns all humans into paperclips* AI: hmm yes what a novel state

All the news dropping today are crazy…

this is very cool. how do you tackle attributing what "novel" is good and which is "bad". or are you simply maximizing divergence? would love to chat!

The comparison I’d love to see is rare-state fidelity. Does intrinsic discovery help the model uncover the weird failure states a human-designed curriculum would never think to include?

a thing i'd try: is the hence-constructed terminal-wm agent itself a better generator of novel states? i.e if we repeat your procedure with initial baseline being your final model artifact, do we get even richer set of encountered trajectories?

Truly amazing work, guys! Would love to try this on some systems I have.

@DJLougen

Intrinsic motivation for open-ended exploration is a solid direction. Curious how the world model handles long-horizon planning without explicit rewards, that is usually where simulation fidelity gets tested.

could you share the papers about. the experiment？

Simulating terminal environments without handholding is a massive leap

Niche novelty feels like the natural next step here. Global nearest-neighbor reward already gets you 5× more states, but a single similarity metric can still overfill one easy cluster (lots of compile-and-run variants) while missing rarer dynamics that actually matter for the world model. Scoring novelty inside learned niches (compiler errors vs filesystem vs long-running processes vs recovery paths) would keep the loop unsupervised while covering transition types, not just distant embeddings. Curious if you’ve tried partitioning A_t that way.

Learning without supervision is the real breakthrough

dumb question: what’s a frozen baseline vs just baseline?

@grok me explique isso de forma simples, oque isso muda?

get your free spot now im giving away 200000 spot :) lets make history again

This is basically the Bitter Lesson: remove humans out the loop...