What are world models — and are they Europe’s winning AI bet?

0:02Hello and welcome to the Sifted podcast. I'm Freya Praty, Senior Reporter and Investigation lead at Sifted, filling in for your usual host, Amy Lewin. AI has long been dominated by large language models or LLMs like ChatGPT, Claude and Gemini. They're trained on text and work by predicting the next word in a sentence. But buzz is building around different kinds of models, world models. Instead of working on text, world models are designed to model the physical world and plan actions. Excitement around this new form of model ramped up towards the end of last year after Jan Lecun, who was formerly Meta's chief AI scientist, announced he was leaving the company to launch his own startup focused on building world models.

0:46The company is reported to be raising 500 million euros straight off the bat, and luckily for Europe watchers like us, it's part-based in Paris. Joining me to talk us through world models, and if Europe has any chance of producing winning companies in the category, are my colleagues Daphne Le Prince-Renge, based in Paris, and Anne Schraders, based in Berlin. Thank you for joining us. Thank you for having us. Thanks for having us on the pod. Daphne, start us off. How do these models actually work? So great questions. As you said, LLMs are based, LLMs, large language models, which are the form of, which are the AI models that we've largely seen so far that power AI assistants like ChatGPT or Claude, which are largely behind all the buzz around AI so far.

1:35those models they're based on what we know as next token prediction which sounds quite technical but essentially on the basis of huge amounts of texts that they have been trained on they can predict what comes next in a sentence but fundamentally or maybe conceptually you could say they don't really understand why they're making the predictions that they're making um world models don't work like this they also make predictions but the idea is that these predictions are based on their understanding of notions of cause and effect, of maybe you could say the laws that govern the world, like gravity, for example.

2:10That's why we often hear that they are reasoning models. So for example, if you push a cup, a world model can predict that it will fall, not because it's mathematically the most likely outcome that it has seen in its own training data, but because it understands that pushing is going to lead to falling. So that means that world models can understand sequences of actions they can plan accordingly um and even in situations they haven't seen before which is seen as the big big difference with llms um they're capable of adapting again because they understand the world around them um so yeah that's that's what's caused a lot of excitement um for world models compared to llms um because i think we've recently come to a point where uh we're understanding that llms can be limited particularly when it comes to tasks that are occurring in the real world um predicting tokens can only get you so far if you're having an interaction with the real world it's too complex um and yann lequin the the the scientist who's behind uh the the latest buzzy company in in europe amy labs which is building world models i think he put it in a very uh in very good words obviously he says if we take the example of autonomous driving we have millions of hours of cars that are being driven by people and despite that training huge amounts of training data, we still don't have full driving automation for cars.

3:35Contrary to that, you get a 17-year-old who's able to learn to drive with just 20 hours of practice. I think this kind of embodies exactly what world models are. It's in a way, a technology that functions more like a human brain in terms of how they learn about the world and how they use that learning to make decisions. You've touched on it a bit there, but can you go into more of the examples of the training data these models are absorbing? So it's still quite dependent on what kind of world model you are talking about. So world models are quite an emerging technology. It's kind of a term that's used still to refer to different types of architectures within world models.

4:16So all of these different architectures will have different data training requirements. Overall, though, I think you could say world models still need to be trained on a lot of data to get that understanding. Typically, they'll be trained on things like sequences over time, like video frames or actions taken in the real world. And that's still very capital and data intensive. But the idea of how these models are trained is very different. So you're not going to be labeling training data for the model to learn on and then use as a basis to make predictions. As I said, it's a different concept. Yeah, so that's kind of the kind of data you'll be looking at for world models.

4:58And Anne, what applications could these models unlock? Yeah, so Daphne touched on a couple of them in her previous answer, but I think, you know, clearly autonomous driving is a big one. I also hear a lot about robotics and industrial manufacturing, something that's a huge one that in particular Europe is looking into as kind of a potential competitive edge. And I think you're also seeing a lot of applications within gaming, for example, or wearables. We've already mentioned his name a few times, but Daphne, Jan Lecun and his lab. Tell us more about why he's such a big deal, the stage the lab is at.

5:30And you and Anne also got hold of a pitch deck looking at what the company will do. So tell us what that showed. Yeah, so a lot of excitement around Jan Lecun because he was Meta's chief AI scientist for over a decade. And around the end of last year, he announced he would be leaving Meta to launch his own company to build world models after having, yeah, Nukhans kind of made a name for himself for stating that he thought LLMs were not going to be the key to unlocking the full potential of AI. So that obviously caused a lot of buzz, a lot of hype. and then he went ahead and announced the new company at the start of this year.

6:14He announced he would be appointing Alexandre Lebrun, who is the founder of Paris-based AI startup Nabla as CEO. And as we said, at Sifted, we got our hands on the pitch deck, which was dated from December. It describes plans to develop the first foundational world model for business, targeting industries that Anne has already mentioned. So wearables, manufacturing, robotics. And I think this is going to be quite key, actually, is showing that it's not just a research project. It can be turned into a product in the short or medium term. So that will be one to watch. What else to say about AMI?

6:59It's made quite key hires from pretty high profile labs like Meta, Google DeepMind. um the team includes engineers from open ai from xai and what's interesting about ami is that it's going to be hq'd in paris as you said which has caused a lot of excitement but the company really wants to be global from day one so it's already spread across four regions uh paris singapore montreal and new york and as you said it's raising quite a lot of money which is reflecting quite high investor interest for the tech and for Yen Le Kien as well. Do you think those investors are interested because of Le Kien and the rest of the team he's pulling in or is it because of wild models or potentially a mixture of the two?

7:45I would probably say it's a mixture of the two. Yen Le Kien is one of the biggest names in AI. So definitely him going and launching a company based on a technology that is different from the one everyone has been excited about for several years and saying that he thinks this is the right way to unlock AI's potential as opposed to LLMs, obviously is bringing a lot of investor interest onto that technology. But I think it's also coinciding quite nicely with a time where LLMs are, there's perhaps a bit more skepticism around how much LLMs can do exactly. um and so this idea of a almost brain-inspired AI is certainly very resonant with investors looking for the next big thing yeah I guess to add on to that point I think you know some VCs have said that you know they expect for world model startups building in this space that they're really going to need top tier talent to win so I think you know that's something we definitely saw you know with more traditional LLMs although now I think you know in AI in general everyone's kind of doing an AI startup.

8:52But I think with World Models, because of the complexity and because of how much money they're going to have to raise, I think that people are expecting, you know, really the end of one talent, essentially, to be the ones that are the winners here. And that's certainly what Yannick has in his startup. That's a really good point. I think the talent point, as Anne's saying, Yannick is really seen as someone who is able to attract that talent because people are obviously very excited to work with him. So I think that's also what's causing a lot of excitement about AMI at the moment. And aside from AMI, who are the other key European companies listeners should know about?

9:23Yeah, so I think sort of as a more of an overall point, I think, you know, the investors we've spoken with say that, you know, a lot of the more interesting companies and teams are still in the US. But I think of those in Europe, there are a couple that are coming out. We've seen Swiss company General Intuition, which is also building world models and recently raised 134 million in seed funding, which is obviously quite a large seed round. We're also seeing German startup Cereax, which develops kind of a different type of world models than AmiLabs to deploy in robots. There's UK-based Stanhope AI, which launched about two years ago to build brain-inspired work models that are currently deployed on robots in unmanned aerial vehicles.

10:03And I think what's also interesting and what I've heard a lot about from investors in Europe in particular is that there's a number of startups across the continent trying to build world models for really specific use cases. For example, industrial production, and robots. And some examples VCs pointed out include Norway's OneX, which you might know from their humanoid robots, very dystopian robots, and also Germany's NeuroRobotics, which also does humanoids, but also focuses on kind of like robotic arms. And so I think that's where a lot of investors are seeing kind of their angle in, in particular in Europe, is kind of that robotics set of companies.

10:39That's interesting that you're hearing that, Anne, about deploying or developing world models for specific use cases? Because I've had various investors tell me that actually we're going to see few players given the amount of capital and talent that's needed. So clearly it's still very much emerging and these dynamics are still being implemented and we'll see how it actually turns out. I think that's true. I think the other point is I hear a lot that companies are, their main product maybe is a robot, but they need some AI to power it. And now they're kind of developing what maybe they're loosely calling world models focused on that particular product.

11:17So I think that there is kind of that, maybe for some companies, that hardware plus software element where maybe it would traditionally have been considered more of a hardware company, but now there's this AI element in there. So I think, yeah, we're going to see a lot of different types of companies addressing different angles, but clearly Amulabs is sort of trying to do the whole, the big foundational model aspect of it. To zoom in on this comparison between Europe and America, because I know our listeners love to hear about that relationship. Daphne, what are your thoughts on where Europe's edge can be and if we can have an edge in world models?

11:47So the sector is still very much early stage. So there are a few players who are building the tech. Anne mentioned a couple in Europe. There's obviously teams in the US. Notably, there is World Labs, which actually just raised$1 million in funding. So a significant competitor. But there is, I think across the board, there's agreement that there isn't yet a state of the art in world models. And it's unlikely that the tech giants that have now poured billions of dollars into LLMs are suddenly going to pivot to world models. So we're not necessarily expecting the biggest competition from these players.

12:30That's a way to say it. And so in that sense, I think a lot of investors and founders are excited about the fact that Europe could actually be ahead of the curve for this technology. If you look at how much AmiLabs is currently raising, it was reportedly raising a 500 million euro round. Since then, Anne and I have reported that the round is potentially going to be bigger than that. So that's a seed round. You compare it to one of the leading companies in the space, World Labs, that has raised a$1 billion round. It's not that far back or potentially it's even on par with it. So I think Europe is very hopeful that it's going to be able to build finally a leading giant in that field because the competition isn't there yet.

13:19And often there is a comparison that is made to Mistral, for example. Whether that comparison is fair or not is maybe up for discussion. But Mistral raised its seed round in 2023 to build LLMs. And that was already many years after OpenAI had started working on these technologies, as well as other tech giants in the US. So the idea is that perhaps Mistral was always playing catch up, Whereas AmiLabs is actually one of the first companies, one of the first handful of companies rather to be developing that technology. I think another area where European investors are kind of expecting Europe to have a bit of an edge is kind of with its industrial heritage and its prowess and things like robotics.

14:04So I think, you know, clearly that's only one particular use case for world models. But I think that's something that people are quite bullish about for Europe in particular. But I do think and I guess I would caution from what I've heard from some investors that some folks still do believe that U.S. is, you know, a bit ahead of some of these companies in Europe. I think we're seeing, you know, models like Genie from Google and Sora from OpenAI that are working on these types of technologies. And, you know, one investor kind of told me that most of the exciting teams she's seeing are actually in the States.

14:35And so I think, you know, it's certainly not clear that Europe has like much of an advantage here, but I think it is definitely in the ring, I suppose. for this fight for world models. The other sort of wild card I think that's been flagged to me by some folks I've spoken with is that, you know, China likely has good world models as well. There's less transparency there, so we don't necessarily know what's going on as much. But I think that's something to watch out for. There's another, actually, there's another reason that people are pitching Europe as an interesting place that you could be developing world models in.

15:06And that's Yann Lecun has spoken about this as well. He has said before that Silicon Valley is currently hypnotized by LLMs and it's very hard to be building a technology outside of that because of this obsession that there is for LLMs. So if you're going to build a new paradigm, if you're going to approach AI in a different way, I think Lecun was saying you need to come out of Silicon Valley and Europe is therefore the natural place to build this technology because there's a lot of talent. I mean, obviously, Yenouk is French, so there's obviously that link as well. But there's this idea that Europe is perhaps, and maybe this is idealistic, but perhaps this place where you can have more creative approaches to AI than in Silicon Valley.

15:53And when can we expect to see real world outcomes from these models? What timelines are companies coming out and saying they're working to? I think it's pretty unclear. Some people who spoke to us believe the first use cases could emerge as soon as in the next, you know, 18 months or so. And as I mentioned, you know, companies like OpenAI have already been working on their video model SOAR for a while, and it looks quite realistic and quite good. So I think, you know, it's still, I guess, as unsatisfying an answer as that is, it's still quite unclear when exactly we'll see a lot of mainstream use cases of world models, but certainly a lot of people are pushing into those use cases we talked about with, you know, robotics and video and things like that.

16:32So I think, you know, certainly people are pushing full steam ahead. And do you think they will live up to expectations, perhaps in comparison to the impact we're seeing from LLMs? I think, you know, of course, time will tell. I think based on how people are talking about world models in the future down the line, they're going to be massive. I think a lot of people I've spoken to have also thought, you know, we're going to kind of see this intersection between LLMs and world models in the future where maybe they work together or maybe LLMs are even kind of a subset of world models. And I think, you know, that's all seems to be still quite down the line from where we are right now.

17:07But I think this certainly is a huge thing that, you know, it's not going to kind of go away. Whether or not that lives up to, you know, precise expectations people have right now, I don't know. But I think it seems to be a trend that's not going to slow down. Yeah. And to continue your point, and I think the interesting thing about world models is that it's really opened people's minds about what the future of AI could look like when LLMs, I think, were largely considered the main way to reach AGI, if that's what you want to be your objective. And certainly in the public's imagination, chatbots like ChatGPT and Claude are really seen as this kind of frontier technology.

17:47And world models, I think, have opened this up and shown that the future of AI actually might be more plural. LLMs are not going anywhere. Obviously, They're a very good technology and they're very good at what they do. But probably what's going to happen is that we're going to have an ecosystem where world models will be very good at some use cases. LLMs will be very good at other use cases. It's not going to be just one technology that's going to unlock the potential of AI. From a personal perspective, I find the idea of a brain inspired AI model very exciting and a bit scary. But yeah, we'll see if it delivers.

18:23And that's all we've got time for. If this has piqued your interest in all things world models, sign up to our AI and deep tech newsletter, which lands in inboxes every Monday morning. You can find a link in the episode description. As usual, please rate and review the podcast. It really helps us reach new ears. And this podcast was produced by Maya Durampo-Hornby.