RAISE Summit '26 | Pim de Witte, General Intuition

Clips
More from RAISE Summit '26

Pim de Witte

CEO

General Intuition

play_circle_outline World Models Beyond Text: Vision-and-Action Foundation Models for Spatial-Temporal Reasoning

play_circle_outline Vision-Based Spatial-Temporal World Models: Learning Real-World Navigation in Minutes Without Text Prompts

play_circle_outline Trillions of Actions: Enabling LLM-Scale Models That Learn New Tasks in Minutes

play_circle_outline Trillions of Actions: Scaling General Models with Talent, High-Quality Data, and Massive GPU Compute

play_circle_outline From Video Games to Digital Twins: Gamified Science, Controller-Driven Robots, and $320M Powering Spatial-Temporal AI

play_circle_outline Use of simulation and synthetic data to handle long-tail situations

Info

Pim de Witte, General Intuition

Pim de Witte

CEO General Intuition

search

John Furrier

>> Palo Alto Studio Connection, Silicon Valley and Wall Street. I'm John Furrier, co-host of theCUBE here with Gabe Olave, my co-host. Welcome to the RAISE Summit coverage brought to you by Solidigm with support from Tensordyne, d-Matrix, and Argentum AI. Welcome back to theCUBE's live stream here in Paris, France. I'm John Furrier, host of theCUBE. It's the RAISE Summit, an international conference of all the AI leaders gathered here from the US, Europe, and all around the world, really talking about the AI infrastructure and all that is enabling from physical AI and robots. We talked about it all day yesterday and today, AI and all the real-world applications that are starting to come on top of it. You're starting to see the fruit coming off the tree of the valley, and it's just the beginning. It's only going to get much better as the innovation and the entrepreneurial energy pops in. We got a great entrepreneur in here. Pim de Witte, he's the CEO of General Intuition. He's got the coolest product, I think, really talks about real world, but he's doing something compelling, bringing gaming in there. Pim, thanks for coming on theCUBE and sharing what you're working on. Appreciate it.

Pim de Witte

>> Thank you for having me.

John Furrier

>> Okay, first of all, explain what you're doing, because I think we hear world models, oh yeah, we're going to put some UI and we're going to make things better for software and workloads. Your workload is different.

Pim de Witte

>> YeahWe've been convinced by Silicon Valley that LLMs are getting us very, very close to general intelligence. But the reality is that you can't even send this model in a robot to walk down Broadway in New York. It's too chaotic. There's too much going on. So you need to train general models that are rooted in the chaos of the real world, which is what we do. So we build foundation models that are rooted in spatial and temporal dynamics where then you can add text on top. So one way to think about it is, as humans, the choice to speak or the choice to write is a very small subset of the actions we can actually take. There's much more to intelligence, and so we are building the whole space of actions, whereas LLMs just do the text components. We have general models that can play video games at a very high level, that can also control robots walking down streets with the same ease.

John Furrier

>> This is a problem that people sometimes can't grok because it's like, they hear LLMs, large language models, but AI is beyond text.

Pim de Witte

>> Yes.

John Furrier

>> I was talking to an entrepreneur yesterday, he says it's like swimming. You can't learn how to swim by watching a video.

Pim de Witte

>> Yeah.

John Furrier

>> And so that's a software problem.

Pim de Witte

>> That's right.

John Furrier

>> You gotta get data, you gotta do it.

Pim de Witte

>> Yeah.

John Furrier

>> So take us through the tech involved because your labs and team are working on some of the tech to get to this level of expertise then render value on top of it.

Pim de Witte

>> Yeah, the metric we use to determine our own internal success is how much data is required from a customer until it nails a new task. And so the models that we've shown can learn how to navigate the streets of New York on minutes of data, as opposed to trying to describe it using text as you described it. You can't. One way to think about text is text is a compression mechanism. It is something that we as humans invented to communicate quickly. In doing so, a lot of information is lost. And vision communicates that information at a much higher bandwidth. That's how the actual world works, right? And so these are vision-based problems. So the way that you solve them is you show the models in vision space how to do something as opposed to explaining it with a prompt, which is how we train our models.

John Furrier

>> How big is the data? Because again, the petabytes involved are pretty significant.

Pim de Witte

>> Yeah, yeah, yeah. We don't publicly disclose the size of the dataset, but I can tell

John Furrier

>> you—Scope it, scope the order of magnitude.

Pim de Witte

>> All the data on the internet amounts to in the trillions of text tokens, and we have in the trillions of actions. So we're the only company in the world that has the dataset to be able to take kind of the same level scaling bets as the LLM companies have.

John Furrier

>> So what's the secret sauce when you start to think about mapping the streets of New York and putting that into a spatial problem space, solving that? What's that? What's the secret sauce to do all this?

Pim de Witte

>> Really good people, really good data, and lots of GPUs.

John Furrier

>> Okay. What's the results?

Pim de Witte

>> General models that can do any task that involves spatial and temporal reasoning. And if you think about it, any sufficiently long task touches on this. So we are going to be competing at the frontier with the other labs. See us as focused on spatial temporal reasoning the same way that Anthropic was focused on code to get to the frontier. These models do not work today. LLMs do not work in these environments. We use that to get to the frontier. And then because they are general models, we also will compete on everything else.

John Furrier

>> Yeah, we're seeing a lot of other entrepreneurs do the same thing where they pick a problem. I see it in biology all the time. There's biology models. This seems to be a trend. Is that, you see that as a feature, not a bug? just pick a spot where the goal is to get to the frontier, right? And then what happens next?

Pim de Witte

>> I think, so from what the market shows is that the demand for the best intelligence is almost uncapped and everything else is on the margin.

John Furrier

>> Talk about the results. You and I were riffing before we came on camera about video games in your world model. Just explain that so people can understand it. Because you have a world model, games are in a virtual world, basically statically programmed. What does a world model do differently?

Pim de Witte

>> So think about as humans how we choose to do things. We dream something up, we see it in our head before we choose to take a certain set of actions. And LLMs don't have the data to do this, They sit purely in text space. That's why they go off track because they can't imagine their actions in the world they are about to do it in. And so why video games are helpful is it is this constant stream of actions and observations, actions and observations, actions and observations. And that is a much better pre-training set because video games have the information density, so like interfaces, text, Factorio, these types of games are very dense, but they also have the spatial-temporal dynamics of the real world, which visual data lacks. If you try to teleoperate a robot, you don't get text labels, you don't get anything other than 3D coordinates, right? And so as a result, games are the perfect pre-training set for general models. And what did you just explain to me? That when you play games, sometimes the people you play with are better. They turn playing with you into a game because you're playing with them. So what do we do as humans? We continuously frame what we're doing as a game that we can act in and we can be in. And a lot of intelligence is actually just that. And so our models of games are very broad, right? You have a grocery store simulator, you have all sorts of things. Having a reference frame in the backbone of the models is very, very helpful. And then you can transfer over to real tasks with very little data.

John Furrier

>> And this is where the intelligence actually is rendered.

Pim de Witte

>> Yeah.

John Furrier

>> Not imagined and then rendered.

Pim de Witte

>> Yeah.

John Furrier

>> Game of Life.

Pim de Witte

>> Game of Life.

John Furrier

>> So how do you see this evolving? Because injecting intelligence into things has been a big conversation. Inject intelligence at the edge of the network. Inject intelligence into your business. Inject intelligence into fill in the blank. You can put anything in there.

Pim de Witte

>> Yeah.

John Furrier

>> With world models, that seems to be the perfect opportunity. How has the data Let me rephrase. How has the data acquisition, primary data versus say synthetic data, because you got digital twins.

Pim de Witte

>> Yeah.

John Furrier

>> When we start getting data, when do we hit that threshold where we have enough world model data? remember Uber came out, oh, they got a lot of data, but they do so many tasks. That's how you have the best GPS tracking. I can imagine something similar.

Pim de Witte

>> Yeah, the long tail is huge. And this is why being able to simulate is so important because you need to be able— you want your models to work in long-tail situations, but you might not have data to train on for that specific situation. And so what matters is that your model is an incredibly general intelligence that can handle novel situations. So for instance, I'll give you the digital twin example. As humans, if you drop us in a new place, we can navigate. We're not going to randomly run into things. So our intelligence is general enough to understand how to interact with a new space. And it comes— our dataset, for instance, involves a lot of dynamics. We make a lot more decisions off dynamics and how we see other people doing things or how we— if you place a robot in this conference room, everything would start falling apart because of the chaos. Because a still image or a digital twin wouldn't teach you anything about how to navigate this conference floor.

John Furrier

>> Yeah.

Pim de Witte

>> Right?

John Furrier

>> Yeah, situational. Exactly. New data.

Pim de Witte

>> Exactly.

John Furrier

>> This is why I like the game mechanics. Let's explain more of that, because I think this is a tie-in. The game mechanics gives you a lot more interaction data that can feed the synthetic data.

Pim de Witte

>> Yes, that's right. Just like LLMs started by being trained on internet data and are now hill-climbing scientific problems, the same will happen with actions. You use both of them, you construct a flywheel that gets you into synthetic data, regime, you use that to solve problems that were way outside of the original problem space.

John Furrier

>> Yeah, one thing that jumps into my head when you're talking is, one, let's get a video game out so you can get more multiplayer gamers guiding the interactions, and then two, throw a Pokémon GO over the top to get more real data.

Pim de Witte

>> Yeah, I'll give you one example is you can reframe real-world problems as games surprisingly often. So for instance, when Demis Hassabis did AlphaFold, which won a Nobel Prize. They started by training a general agent to play a game called Foldit, which was basically prompting humans to do basic folds of proteins. And so you can gamify a lot of these problems. And General Intuition is called General Intuition because Demis said, "Wouldn't it be amazing if you could mimic the intuition of these gamers who are, by the way, only amateur biologists?" And so a general model that can take instructions and act in space and time is also an incredible, solution to many scientific problems.

John Furrier

>> Talk about what's next for you guys. What's on your plate? I love the labs. You're banging out some great work. Yeah, what's coming out? What are you focused on?

Pim de Witte

>> Yeah, any robot that can be controlled using a game controller, we will have models that work for you this year. And that is a massive unlock for robotics because it unlocks all of the lower degrees of freedom. So every quadruped, everything that ships with a game controller has already accounted for the interface and our models will work very, very well. And so if that's a problem, it means that intelligence is a bottleneck. We're gonna unbottleneck that intelligence and make your robot able to do many more things.

John Furrier

>> All right, so you got a culture, you got a gaming culture, I can tell right there. You guys build video games. Who are you looking to hire? Who should work for you? Describe the perfect candidate to join the team.

Pim de Witte

>> Typically somebody who is top of their field or demonstrably going to be top of their field because of the projects that they've recently been working on. It's a very small team. We hire deeply ambitious people with a strong sense of urgency who deeply understand the problem.

John Furrier

>> And funding?

Pim de Witte

>> We just raised $320 million from Khosla, Jeff Bezos, and Eric Schmidt.

John Furrier

>> All right. Well, great, great guys. They see the technical vision. And what are your goals for next year?

Pim de Witte

>> Ship the model first and foremost and make lots of customers very happy.

John Furrier

>> All right, Pim, great to have you on again. Look, the world is just starting. We're starting to see the breakout stories emerging here at RAISE and also all around the world. Entrepreneurship is changing. The data game is changing. And you start to see the AI infrastructure continue to enable that. The more we can get the GPUs, the better the inference, the better the training, the reinforcement learning, the better the tech and the infrastructure enables faster execution and more value creation in a whole new way. We're doing our part here on theCUBE to bring that to you. I'm John Furrier. Thanks for watching.

Pim de Witte

>> Thanks for having me.

Pim de Witte, General Intuition

search

John Furrier

>> Palo Alto Studio Connection, Silicon Valley and Wall Street. I'm John Furrier, co-host of theCUBE here with Gabe Olave, my co-host. Welcome to the RAISE Summit coverage brought to you by Solidigm with support from Tensordyne, d-Matrix, and Argentum AI. Welcome back to theCUBE's live stream here in Paris, France. I'm John Furrier, host of theCUBE. It's the RAISE Summit, an international conference of all the AI leaders gathered here from the US, Europe, and all around the world, really talking about the AI infrastructure and all that is enabling from physical AI and robots. We talked about it all day yesterday and today, AI and all the real-world applications that are starting to come on top of it. You're starting to see the fruit coming off the tree of the valley, and it's just the beginning. It's only going to get much better as the innovation and the entrepreneurial energy pops in. We got a great entrepreneur in here. Pim de Witte, he's the CEO of General Intuition. He's got the coolest product, I think, really talks about real world, but he's doing something compelling, bringing gaming in there. Pim, thanks for coming on theCUBE and sharing what you're working on. Appreciate it.

Pim de Witte

>> Thank you for having me.

John Furrier

>> Okay, first of all, explain what you're doing, because I think we hear world models, oh yeah, we're going to put some UI and we're going to make things better for software and workloads. Your workload is different.

Pim de Witte

>> YeahWe've been convinced by Silicon Valley that LLMs are getting us very, very close to general intelligence. But the reality is that you can't even send this model in a robot to walk down Broadway in New York. It's too chaotic. There's too much going on. So you need to train general models that are rooted in the chaos of the real world, which is what we do. So we build foundation models that are rooted in spatial and temporal dynamics where then you can add text on top. So one way to think about it is, as humans, the choice to speak or the choice to write is a very small subset of the actions we can actually take. There's much more to intelligence, and so we are building the whole space of actions, whereas LLMs just do the text components. We have general models that can play video games at a very high level, that can also control robots walking down streets with the same ease.

John Furrier

>> This is a problem that people sometimes can't grok because it's like, they hear LLMs, large language models, but AI is beyond text.

Pim de Witte

>> Yes.

John Furrier

>> I was talking to an entrepreneur yesterday, he says it's like swimming. You can't learn how to swim by watching a video.

Pim de Witte

>> Yeah.

John Furrier

>> And so that's a software problem.

Pim de Witte

>> That's right.

John Furrier

>> You gotta get data, you gotta do it.

Pim de Witte

>> Yeah.

John Furrier

>> So take us through the tech involved because your labs and team are working on some of the tech to get to this level of expertise then render value on top of it.

Pim de Witte

>> Yeah, the metric we use to determine our own internal success is how much data is required from a customer until it nails a new task. And so the models that we've shown can learn how to navigate the streets of New York on minutes of data, as opposed to trying to describe it using text as you described it. You can't. One way to think about text is text is a compression mechanism. It is something that we as humans invented to communicate quickly. In doing so, a lot of information is lost. And vision communicates that information at a much higher bandwidth. That's how the actual world works, right? And so these are vision-based problems. So the way that you solve them is you show the models in vision space how to do something as opposed to explaining it with a prompt, which is how we train our models.

John Furrier

>> How big is the data? Because again, the petabytes involved are pretty significant.

Pim de Witte

>> Yeah, yeah, yeah. We don't publicly disclose the size of the dataset, but I can tell

John Furrier

>> you—Scope it, scope the order of magnitude.

Pim de Witte

>> All the data on the internet amounts to in the trillions of text tokens, and we have in the trillions of actions. So we're the only company in the world that has the dataset to be able to take kind of the same level scaling bets as the LLM companies have.

John Furrier

>> So what's the secret sauce when you start to think about mapping the streets of New York and putting that into a spatial problem space, solving that? What's that? What's the secret sauce to do all this?

Pim de Witte

>> Really good people, really good data, and lots of GPUs.

John Furrier

>> Okay. What's the results?

Pim de Witte

>> General models that can do any task that involves spatial and temporal reasoning. And if you think about it, any sufficiently long task touches on this. So we are going to be competing at the frontier with the other labs. See us as focused on spatial temporal reasoning the same way that Anthropic was focused on code to get to the frontier. These models do not work today. LLMs do not work in these environments. We use that to get to the frontier. And then because they are general models, we also will compete on everything else.

John Furrier

>> Yeah, we're seeing a lot of other entrepreneurs do the same thing where they pick a problem. I see it in biology all the time. There's biology models. This seems to be a trend. Is that, you see that as a feature, not a bug? just pick a spot where the goal is to get to the frontier, right? And then what happens next?

Pim de Witte

>> I think, so from what the market shows is that the demand for the best intelligence is almost uncapped and everything else is on the margin.

John Furrier

>> Talk about the results. You and I were riffing before we came on camera about video games in your world model. Just explain that so people can understand it. Because you have a world model, games are in a virtual world, basically statically programmed. What does a world model do differently?

Pim de Witte

>> So think about as humans how we choose to do things. We dream something up, we see it in our head before we choose to take a certain set of actions. And LLMs don't have the data to do this, They sit purely in text space. That's why they go off track because they can't imagine their actions in the world they are about to do it in. And so why video games are helpful is it is this constant stream of actions and observations, actions and observations, actions and observations. And that is a much better pre-training set because video games have the information density, so like interfaces, text, Factorio, these types of games are very dense, but they also have the spatial-temporal dynamics of the real world, which visual data lacks. If you try to teleoperate a robot, you don't get text labels, you don't get anything other than 3D coordinates, right? And so as a result, games are the perfect pre-training set for general models. And what did you just explain to me? That when you play games, sometimes the people you play with are better. They turn playing with you into a game because you're playing with them. So what do we do as humans? We continuously frame what we're doing as a game that we can act in and we can be in. And a lot of intelligence is actually just that. And so our models of games are very broad, right? You have a grocery store simulator, you have all sorts of things. Having a reference frame in the backbone of the models is very, very helpful. And then you can transfer over to real tasks with very little data.

John Furrier

>> And this is where the intelligence actually is rendered.

Pim de Witte

>> Yeah.

John Furrier

>> Not imagined and then rendered.

Pim de Witte

>> Yeah.

John Furrier

>> Game of Life.

Pim de Witte

>> Game of Life.

John Furrier

>> So how do you see this evolving? Because injecting intelligence into things has been a big conversation. Inject intelligence at the edge of the network. Inject intelligence into your business. Inject intelligence into fill in the blank. You can put anything in there.

Pim de Witte

>> Yeah.

John Furrier

>> With world models, that seems to be the perfect opportunity. How has the data Let me rephrase. How has the data acquisition, primary data versus say synthetic data, because you got digital twins.

Pim de Witte

>> Yeah.

John Furrier

>> When we start getting data, when do we hit that threshold where we have enough world model data? remember Uber came out, oh, they got a lot of data, but they do so many tasks. That's how you have the best GPS tracking. I can imagine something similar.

Pim de Witte

>> Yeah, the long tail is huge. And this is why being able to simulate is so important because you need to be able— you want your models to work in long-tail situations, but you might not have data to train on for that specific situation. And so what matters is that your model is an incredibly general intelligence that can handle novel situations. So for instance, I'll give you the digital twin example. As humans, if you drop us in a new place, we can navigate. We're not going to randomly run into things. So our intelligence is general enough to understand how to interact with a new space. And it comes— our dataset, for instance, involves a lot of dynamics. We make a lot more decisions off dynamics and how we see other people doing things or how we— if you place a robot in this conference room, everything would start falling apart because of the chaos. Because a still image or a digital twin wouldn't teach you anything about how to navigate this conference floor.

John Furrier

>> Yeah.

Pim de Witte

>> Right?

John Furrier

>> Yeah, situational. Exactly. New data.

Pim de Witte

>> Exactly.

John Furrier

>> This is why I like the game mechanics. Let's explain more of that, because I think this is a tie-in. The game mechanics gives you a lot more interaction data that can feed the synthetic data.

Pim de Witte

>> Yes, that's right. Just like LLMs started by being trained on internet data and are now hill-climbing scientific problems, the same will happen with actions. You use both of them, you construct a flywheel that gets you into synthetic data, regime, you use that to solve problems that were way outside of the original problem space.

John Furrier

>> Yeah, one thing that jumps into my head when you're talking is, one, let's get a video game out so you can get more multiplayer gamers guiding the interactions, and then two, throw a Pokémon GO over the top to get more real data.

Pim de Witte

>> Yeah, I'll give you one example is you can reframe real-world problems as games surprisingly often. So for instance, when Demis Hassabis did AlphaFold, which won a Nobel Prize. They started by training a general agent to play a game called Foldit, which was basically prompting humans to do basic folds of proteins. And so you can gamify a lot of these problems. And General Intuition is called General Intuition because Demis said, "Wouldn't it be amazing if you could mimic the intuition of these gamers who are, by the way, only amateur biologists?" And so a general model that can take instructions and act in space and time is also an incredible, solution to many scientific problems.

John Furrier

>> Talk about what's next for you guys. What's on your plate? I love the labs. You're banging out some great work. Yeah, what's coming out? What are you focused on?

Pim de Witte

>> Yeah, any robot that can be controlled using a game controller, we will have models that work for you this year. And that is a massive unlock for robotics because it unlocks all of the lower degrees of freedom. So every quadruped, everything that ships with a game controller has already accounted for the interface and our models will work very, very well. And so if that's a problem, it means that intelligence is a bottleneck. We're gonna unbottleneck that intelligence and make your robot able to do many more things.

John Furrier

>> All right, so you got a culture, you got a gaming culture, I can tell right there. You guys build video games. Who are you looking to hire? Who should work for you? Describe the perfect candidate to join the team.

Pim de Witte

>> Typically somebody who is top of their field or demonstrably going to be top of their field because of the projects that they've recently been working on. It's a very small team. We hire deeply ambitious people with a strong sense of urgency who deeply understand the problem.

John Furrier

>> And funding?

Pim de Witte

>> We just raised $320 million from Khosla, Jeff Bezos, and Eric Schmidt.

John Furrier

>> All right. Well, great, great guys. They see the technical vision. And what are your goals for next year?

Pim de Witte

>> Ship the model first and foremost and make lots of customers very happy.

John Furrier

>> All right, Pim, great to have you on again. Look, the world is just starting. We're starting to see the breakout stories emerging here at RAISE and also all around the world. Entrepreneurship is changing. The data game is changing. And you start to see the AI infrastructure continue to enable that. The more we can get the GPUs, the better the inference, the better the training, the reinforcement learning, the better the tech and the infrastructure enables faster execution and more value creation in a whole new way. We're doing our part here on theCUBE to bring that to you. I'm John Furrier. Thanks for watching.

Pim de Witte

>> Thanks for having me.