Deep dive: World Models: Simulators That Learned by Watching

Why today

In the last two weeks, two launches used the same phrase. On October 17, Odyssey posted a video of a simulated fruit fly brain flying through a world its model generated on the fly. Two days later, Skyfall AI released SKY-001, which it calls an engineering world model for computer-aided design and building plans. Behind Waymo's new driving simulator sits Google DeepMind's Genie 3. Waymo has built its driving simulator on top of it. The money is serious too. In June, Odyssey raised a $310 million Series B at a $1.45 billion valuation. Still, ask ten investors what a world model is and you will hear ten answers. Part of the reason is that the term now covers at least three different things. So today we take the idea apart. What does a world model learn? How is it different from a language model? And which of the claims around it have actually been shown?

Plainly

Start with the difference. A language model predicts the next word. A world model predicts the next moment, given an action. You tell it you turn left, and it shows what you would see after turning left. Then it does that again, many times a second. Here is the analogy worth keeping. Pilots train in flight simulators because crashing one costs nothing. A traditional simulator is hand-built. Engineers write the physics, model the runway and code every gust of wind. A world model is a simulator that taught itself by watching. Odyssey says its model was trained on internet video, gameplay recordings and simulated physics. Nobody wrote down that cups fall or that doors swing. The model picked up those patterns from footage, the way a toddler learns gravity long before reading Newton. The catch is that this simulator works like a dream. Each moment looks believable next to the one before. But nothing guarantees the rules hold. Turn around and the street behind you may have changed. Leave the room and come back, and the furniture may have moved. Under one label sit three families. The first draws pixels. Genie 3 generates each new frame in real time, at 24 frames per second, so a person or an agent like DeepMind's SIMA can navigate inside the world it creates. The second predicts in shorthand rather than pictures. Meta's V-JEPA 2 does not draw the future. It predicts a compressed summary of what comes next, and that summary is enough for a robot to plan. Yann LeCun at Meta has long argued that this is the route that matters. The third family stretches the word. SKY-001 does not render a world at all. Its world is the state of a design file, and it plans the sequence of modeling steps needed to build a part or a floor plan. All three share one idea: learn how a state changes when you act on it. That is what lets you rehearse before you commit. A robot can try a grasp a thousand times in imagination, and a car can meet a flood it has never seen. With that map in hand, let's look at what has actually been shown.

What the evidence shows

Start with what is solid. DeepMind says Genie 3 produces interactive worlds at 720p and 24 frames per second. It says they stay largely consistent for several minutes, with visual memory going back about a minute. That is DeepMind describing its own work. But since January, paying subscribers on Google's $249 per month AI Ultra plan have been able to try it through Project Genie, so outsiders can now test the basic behaviour. DeepMind also lists the limits itself. In its words, the model supports a few minutes of interaction, rather than extended hours. An analysis on TechTalks adds physics errors, such as people appearing to walk backward. When DeepMind put its game-playing agent inside Genie 3, the agent could only move around. It did not pick things up. Meta reports that V-JEPA 2 needed only 62 hours of robot video to learn to plan. It says zero-shot pick-and-place success in new settings ran from 65% to 80%. That is self-reported, on Meta's own lab tasks. Waymo says its Genie-based model produces both camera and lidar data, and can stage tornadoes, floods and even elephants. As The Decoder noted, Waymo shared no benchmark results or independent evaluations. Now the claims that go beyond the evidence. The fly demo is a video on X. There is no paper. Odyssey has not said which fly brain map it used. A public map called FlyWire has 139,255 neurons and roughly 50 million connections. We don't know how the generated frames are turned into signals for the fly's eyes. We also don't know whether the virtual fly behaves anything like a real one. For now, it is a striking demo from a company that has just raised a large round. SKY-001 has a benchmark, but Skyfall built it. Every figure is self-reported, with no weights, no API and no outside replication. That is a tie. One independent write-up reads the preview as producing sketches and visual states, with editable design histories still future work.

How people read it

The same facts support quite different readings. The AGI-forward reading goes like this. Robotics is short of data, and every real-world trial is slow and costly. A model that has absorbed physics from video can generate endless practice worlds for agents and robots. On this view, Genie 3's memory, which DeepMind says it did not program explicitly, hints at understanding emerging from scale. Waymo, which runs a commercial robotaxi fleet, putting this into its pipeline counts as a serious vote. The AI-realist reading starts from the dream problem. A few minutes of consistency is fine for a demo. It is not enough for training. Teach a robot inside a simulator that gets physics slightly wrong and it learns the wrong physics, with confidence. Robotics has wrestled with this gap between simulation and reality for decades. None of the sources we checked shows a robot or car trained mainly in generated worlds outperforming one trained on real data. On that view, today's value is in testing and rare-scenario coverage, not in replacing real data. The market reading looks at who is paying. Odyssey's backers include Amazon, AMD Ventures and GV, and AWS is its preferred cloud. Chip and cloud companies are funding a customer that consumes a great deal of compute. That is a fair bet, but it means the valuation says as much about compute strategy as about demand. Odyssey has not disclosed revenue.

What it means for builders and investors

Our read is that world models are real infrastructure for simulation and testing. As a source of training data they are still early, and in some launches they are just a label. Each reading would change with one piece of evidence, which we come to next. For builders, the general-purpose pixel model is a capital game, and Google and well-funded labs are already playing it. Serving cost matters as well. In 2025, Odyssey's earlier model streamed from clusters of H100 GPUs at $1 to $2 per user-hour. A video game runs on a console the player already owns. A consumer product built on a hosted world model has to cover that GPU bill every hour, for every player. Seed and Series A companies seem better placed where the world is narrow and the output can be checked. A CAD part can be measured against a geometry engine. Owning interaction data, such as robot logs or sensor feeds, gives a company something to fine-tune with that others lack. Waymo's lidar step is the template here. Tooling that measures whether simulation results hold up in reality looks undersupplied. The exposed companies are thin wrappers on a model someone else hosts. One question worth exploring in diligence might be this. How often have the company's simulated results been checked against the real world, and what happened when they disagreed?

What to watch

Three things would settle a lot of this. First, a published result, from Waymo, DeepMind or a robotics lab, where training mainly in a generated world beats training on real data, with numbers. That would move us towards the AGI-forward reading. Its continued absence would support the realist one. Second, outside testing of SKY-001 once the waitlist opens. We would want to know whether the cost gap holds on someone else's parts, and whether engineers get editable design files or just pictures. Third, a paper from Odyssey showing how the fly brain is wired into the model, and whether the fly behaves like a fly. None of these has a date. Over the next year, falling serving costs per user-hour will show which consumer uses can pay their way.