# World models should learn broadly before task-specific tuning Markdown copy of https://realalpha.app/s/world-models-should-learn-broadly-before-task-specific/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy. - Speaker: [Oliver Cameron](https://realalpha.app/people/oliver-cameron/) · Co-founder & CEO, Odyssey - Host: [John Coogan](https://realalpha.app/people/john-coogan/) · Co-host, TBPN · Co-founded Soylent and Lucy; OpenAI bought his show TBPN in April 2026. - Host: [Jordi Hays](https://realalpha.app/people/jordi-hays/) · Co-host, TBPN · Founded Branded Native; co-founded fintech Capital, which Rho bought. - Show: TBPN - Episode: [AMD CEO Lisa Su Live on TBPN](https://realalpha.app/v/8Jqi1yCniHk/), Jul 23, 2026 - Moment: 9:00 to 12:54 of the episode (4 min), playing on https://realalpha.app/s/world-models-should-learn-broadly-before-task-specific/; the episode from there: https://realalpha.app/v/8Jqi1yCniHk/?t=540 ## Quote > “you have **a base that's just very general**, and then you provide driverless car data at the very last step.” > — [Oliver Cameron](https://realalpha.app/people/oliver-cameron/) at [12:24](https://realalpha.app/v/8Jqi1yCniHk/?t=744) · TBPN ## Takeaways 1. Train on broad world data, not just dashcam or robot examples. 2. Fine-tune for driving only at the final stage. 3. A few hours of experience can adapt a model to a driving task. ## About this segment Odyssey argues that world models should learn from broad data, then get tuned on driving examples only at the end. The approach could make autonomy more robust than systems trained on narrow task data. From “AMD CEO Lisa Su Live on TBPN”, Jul 23, 2026. [Watch the full episode](https://realalpha.app/v/8Jqi1yCniHk/) ## Transcript From the transcript TBPN publishes with the episode. [9:05](https://realalpha.app/v/8Jqi1yCniHk/?t=545) Is there still a market in self driving cars? Because it feels like the world models a lot of that research was done at Cruise and Waymo and Tesla, and that was where I was first introduced to them. I was like, oh, they built a video game version of San Francisco, and of course they're training it there. Is there a world where you wind up powering future self driving initiatives and autonomy initiatives? [9:26](https://realalpha.app/v/8Jqi1yCniHk/?t=566) I think so. I think world models can power all sorts of virtual and physical systems. Mhmm. The virtual systems might be gaming for example, Might be education. And then the physical systems will be driverless cars, robotics, drones, all that sort of stuff. Sure. What I believe is that driverless cars today and lots of automated systems look very much like NLP systems in the twenty tens. Mhmm. They are these very complex, very hand tuned systems. They're intelligent, but they're intelligent in sort of isolated ways. Yep. And that those systems hold back what these technologies can do. And so I think you can replace these very hand tuned driverless car systems with a single world model that has this very deep understanding of physics, of cause and effect, of human behaviors, and just lightly tune that world model to the task of driving. [10:11](https://realalpha.app/v/8Jqi1yCniHk/?t=611) Yep. What we're seeing early evidence of is that you can indeed do that. Mhmm. You can take these very big general world models, tune them with just a few hours of experience, and then the car drives itself. Pretty cool. How much What [10:21](https://realalpha.app/v/8Jqi1yCniHk/?t=621) are what are your interactions and relationships, partnerships look like with various types of robotics companies? So think of like, if someone's building a humanoid or we we were talking with Travis Kalanick yesterday doing autonomy and mining. So like, what what does an Odyssey relationship look like with those companies today? [10:37](https://realalpha.app/v/8Jqi1yCniHk/?t=637) Absolutely. So we very much believe that the robot companies are customers of us. We provide that base intelligence. They then have their robot, their embodiment, their task, and they can take that base intelligence and tune their model to make that task amazing. Mhmm. What we see earlier evidence of is that using a world model versus a vision language action model, the number of examples you need to show the world model is dramatically less than a vision language action model. Which means it's just much more adaptable to the world Sure. [11:04](https://realalpha.app/v/8Jqi1yCniHk/?t=664) Than a vision language action model. Makes sense. Talk about data right now. A lot of these world model systems, it feels like they were bootstrapped on Unreal Engine systems or just video game systems generally. I imagine that the amount of data that's getting poured in the pot and mixed together is huge at this point. But what are some of, like, the power law data sources generally that are useful in building world models? [11:29](https://realalpha.app/v/8Jqi1yCniHk/?t=689) Yes. So I I'll get to the point eventually, promise. Yeah. Take your time. So what I observed in driverless cars Yeah. And we've seen in robotics, is that you typically train those systems on very narrow data distributions. Right? So you're a driverless car company. You collect lots of [11:42](https://realalpha.app/v/8Jqi1yCniHk/?t=702) data from the road. Dash cams, basically. Dash cams. But also cameras around the car. Exactly. We've all seen the street view cars. [11:50](https://realalpha.app/v/8Jqi1yCniHk/?t=710) Exactly. And you're really teaching the car just from the perspective of a car. Right? It's it's solely seeing what cars see. Let's get the blinders on. What what those cars don't see in training Yeah. Is video games or is inside a conference or in an office or in a home or just the multitude of scenarios that exist in the world. And what we very much believe is that a world model shouldn't learn from a narrow distribution of the world Yeah. Like a series of driving examples or robot examples. It should be that you learn from every possible thing that could exist in the world. [12:25](https://realalpha.app/v/8Jqi1yCniHk/?t=745) And then you tune the model to the task of driving. So you have a base that's just very general, and then you provide driverless car data at the very last step. And so really, our belief is that that will produce a more robust, more intelligent system than these very narrowly trained models. [12:41](https://realalpha.app/v/8Jqi1yCniHk/?t=761) I love that. Makes sense. Completely tracks with the GPT two, GPT three progression and language models. There was no point when people said, oh, well, we really wanna solve IMO level math. So Yes.