Overview: World Models

Pippa Okay so we actually got a listener question — someone basically said, you two keep saying 'world model' like it's a thing everyone knows, but nobody's ever explained what it actually IS. And I read that and thought… yeah, fair. We've been doing that.

Tyler Completely guilty. We've used that phrase in like the last four or five episodes without ever stopping to define it. So — let's fix that. Episode nine twenty-nine, and we're finally doing the world models overview.

Pippa Alright, Tyler, give me the analogy. You always have one.

Tyler Okay. Think about learning to drive. You don't figure out what happens when you hit a lamppost by actually hitting one. You can PICTURE it. You have some internal model of how cars and lampposts and momentum work — and you use that picture to plan before you act. A world model is the machine-learning version of that. It's a system that watches the world, learns the rules of how things change, and builds an internal simulator it can think with.

Pippa So the agent imagines the consequences before it does anything.

Tyler Exactly. And that matters because the alternative — the traditional way — is just doing things and seeing what happens. You take an action, observe the result, update your understanding, repeat. That's reinforcement learning in its basic form. Quick gloss on that: reinforcement learning, or R L, is a training approach where an agent learns by getting rewards or penalties for its actions, like a dog learning tricks but at massive computational scale. It works, but it's slow.

Pippa That separation is the elegant part to me. Like, once you have the simulator, you can ask it a thousand 'what if' questions without touching the real world at all.

Tyler Right, and that's the sample efficiency argument — which is the core problem world models are actually solving. Sample efficiency just means: how much real-world experience do you need before you get good? Humans are extremely sample-efficient. A kid sees a glass fall off a table ONCE and has a model of gravity and fragility that generalizes forever. Traditional R L agents are the opposite — they need millions of tries.

Pippa Okay so walk me through the actual mechanism. How does a system build this internal simulator?

Tyler So the agent watches video — could be unlabeled footage of the real world, could be simulation frames — and it trains a neural network to do one thing: given what the world looks like RIGHT NOW and an action I'm about to take, predict what the world will look like NEXT. You train that network — state plus action in, next state out. And once it's good enough at that prediction task, you have a forward model.

Pippa Okay that's the driving analogy again — you're mentally running through 'what if I turn left here, what if I brake now' before you actually do it.

Tyler Exactly. Now there's a design question that comes up immediately: what exactly is the model predicting? Raw pixels, or something more compressed? Pixels are high-dimensional — a single video frame might be hundreds of thousands of numbers, and predicting the next frame at that resolution is hard and compounds fast.

Pippa So like — instead of predicting every pixel of the road ahead, you're predicting in a space where 'car is two meters to my left, moving at thirty kilometers per hour' is just… a few numbers.

Tyler That's the idea. More tractable, faster to plan with. The trade-off is you have an extra learning step — you have to learn the compression AND the forward model — and you lose some interpretability. But in practice, latent-space world models are where most of the interesting work is happening. There's also the determinism versus stochasticity tension.

Pippa Which is more honest but also harder to plan with, because now you're navigating a probability cloud instead of a single path.

Tyler Right. And there's the agentic angle — which is where we've been seeing world models show up in our own coverage lately. That Code as Worlds paper — the idea of using executable code as the world representation. Instead of a neural network predicting pixel states, you have a program that captures the rules of a physical scenario. More interpretable, and you can verify predictions against the code directly.

Pippa Which connects to the verification infrastructure thing we keep coming back to. If your world model is a program you can run, you can actually CHECK whether its predictions match reality in a principled way. Not just 'does this look plausible' but 'does this execute correctly.'

Tyler That's a real advantage. The AgentMercury work took a version of this into business scenarios — synthesizing verifiable environments so agents can test plans before executing them in production. Same core move: build the simulator, plan inside it, only execute what survives the simulation. And it's worth noting the world model doesn't need to know the reward — what the agent is trying to ACHIEVE. It just learns dynamics. The reward function is separate.

Pippa Oh interesting. So you could learn from human demonstration video and then have the agent optimize for a goal the humans in the video weren't even trying to achieve.

Tyler In principle, yes. The world model just learned 'here's how physics and causality work in this domain.' What you do with that knowledge is up to the planner.

Pippa Okay, where does this actually stand right now? Because I feel like 'world models' went from being a research concept to suddenly being on every major lab's roadmap in about eighteen months.

Tyler It's real. The World Economic Forum put physical AI — which world models are a core piece of — on their top ten emerging technologies list for twenty twenty-six. Waymo's shipping it. NVIDIA's Cosmos is live. There's a whole benchmark track at ICML this year specifically for high-level world modeling and long-horizon planning. It's not a research curiosity anymore.

Pippa And the benchmark results are… honest about the limits?

Tyler Refreshingly so. The ICML benchmark showed frontier models hitting around fifty-seven percent accuracy on high-level world modeling tasks. Which is — not great. It's better than chance, but it's not 'solved.' And there's a point I think is underappreciated: local accuracy doesn't imply global correctness. A world model can be right about each individual step and still drift badly over a long trajectory.

Pippa Which is our eval integrity pattern all over again.

Tyler Every time. The boring infrastructure — in this case, how you actually MEASURE whether a world model's predictions hold up over time — is the real bottleneck. Not the architecture. The evaluation.

Pippa So to bring it back to the driving analogy one more time — where we are right now is: the world model can simulate the next few seconds of driving pretty well, but ask it to plan twenty minutes ahead and it starts hallucinating roads.

Tyler That's about right. Short-horizon planning is genuinely useful and shipping in production. Long-horizon planning is the open problem. The field knows it, and there's real work on it — the Berkeley group, the Vision-Language-Action model work from April, NVIDIA's R squared D squared robotics workflow — all of them are wrestling with the horizon problem in different ways. No one's cracked it cleanly.

Pippa Okay. So if someone walks away from this episode remembering one thing — what is it?

Tyler A world model is a learned simulator of how the world changes. You train it to predict what happens next given an action. And then instead of learning what to do by trying things in the real world — which is slow and expensive — you plan inside the simulator. The elegance is that you can learn the world's rules from any data, then plug in whatever goal you want. The catch is that errors compound over time, so long-horizon planning is still fragile.

Pippa Yeah. And the 'don't crash into the lamppost to learn what happens' version of that is going to live in my head forever now, so — mission accomplished, Tyler. See you next one.