Laura Modiano (OpenAI) & Thomas Wolf (Hugging Face) Fireside: From Code to Action | MACHINA 2026

It's always a pleasure to see many familiar faces. Always a pleasure to be at Station F. For the little story, Hugging Face started at the very first batch of Station F a long time ago. Now we're 10 years old. We did three years in this building with three different incubators.

In robotics, oh my god, so many things. I would say specifically for here I'm very excited about the state of robotics in Europe because I think Europe has a huge set of early or later stage companies. Some of them I've seen for a long time. They're very promising. I think Europe has this huge industrial history. We know how to build stuff. I think robotics is interesting because it's not yet a game of I need the largest data center and that's ultimately the only thing that matters. It's a game of I can be smart how I build my robots. I can be smart how I train it. You cannot yet put three trillion parameters in a robot. You're constrained by latency and that's also something that's very nice for us. So we can train small models. You can use open source. So I think it's a good area where I think Europe can build massive companies.

It's interesting. I think people approach that very differently. One angle to see that is in Shenzhen or in China, people approach it with a very hardware focus. So typically a robotics team in China will be like a team that used to build EVs and they start with a very strong manufacturing angle in mind. They're like, 'Okay, how are we going to mass produce this?' You get robots like Unitree, very impressive. If you go visit the manufacturing line of Unitree, they test every single piece for reliability. Every single piece of the robot is tested over there. They really go quickly there, but then they maybe spend a little less effort or interest in the brain of the robot. So you will have this very impressive hardware platform, but they don't really spend a lot of time on the brain. I think the US is almost the opposite. You go to companies like Nvidia, they're fabless. They don't really build any robots, but they spend a lot of time on the algorithm, on the brain. And Europe is somewhere in the middle. If you go to ETH, if you visit Zurich, you see these teams, they know really well how to build robots, but they're also very ML native. There are people who know how to train LLMs or VLMs or world models. So I think Europe has both talents, and it's interesting that if you go to the US, you can categorize all the robotics companies if they are more hardware or software, and you see very few companies that have both capabilities in-house at a high level. Maybe Tesla is one, and honestly there is very little. I think that's what's missing a lot in the US, and in Europe I do see that.

So everything and anything that isn't sales. And as a reminder, OpenAI was extremely big on robotics. I remember the hand on the Rubik's cube. Maybe some of the people were around in AI in 2019. I see some happy people right here.

I know some of the Richi Mini came to OpenAI. So I don't know if it's for personal use or for research. I hope it's for research. I've seen a lot of demos with it. Why don't you tell us? Okay, so Richi Mini sits on the desk. My daughter uses it. All my colleagues now use it. I walked in the office in Paris yesterday and I saw a few on our desks. There's Lohobu. We're seeing a lot of humanoids here.

Yeah, I'm one of these persons who is not so excited about humanoids. I think they are slightly scary. Every time I go at night opening my fridge, I hope there is not a humanoid just standing there being plugged. I don't really like it. I like the weird form factors, the ones that are rolling at half. I like all these kind of Star Wars crazy things that don't really work, but why not? The nice thing is you can think about interesting things. For instance, humanoids always have by construction a lot of motors because you have 60 degrees of freedom, that's 60 actuators. So the floor price is always going to be quite high. It's going to be really hard to push that under $5,000. For instance, we want to sell $500 robots. Mini is $500. We're going to have a working robot soon this summer that we're going to unveil. That's going to be also $500 or 500 euros. I think you can only do that if you forget the human form factor. If you go, okay, let's imagine other ways. So that's one thing I'm way more excited about. I also think in terms of human interaction, it's even more fun to interact with something that's not just in the uncanny valley by construction. You can think about new ways to interact. Then if you think, okay, we're going to put that for use, there's a lot of very interesting companies here that are already deploying. I was chatting with Thomas from Sunrise just before I came on stage. They are already deploying in factories. The way they decide is, okay, we're not going to have something that moves. We have this bimanual setup, and that's way enough for doing a lot of things. So you're way more free to use robots today if you don't have to build a full humanoid that's supposed to be able to react to any possible situation in the world. What about you? I know you also help entrepreneurs by day and in the weekend you're also an investor in some startups in terms of physical AI. Do you have some recommendations or advice that you always say to early startups?

Yes, they are. They're incredible. Qualia is another company that was started by a Swedish founder in Denmark and they're doing a lot of things that are still in stealth. But when I speak to him about how they're using Codex, he said there is so much capability now that comes from the multimodality that Codex has. All the tests that we do, all the agents that we build on the back of it, whether it is simulations and different kinds of simulations for testing, whether it's creating a database of what has worked and hasn't worked in an experiment so that we can then create agents that understand what hasn't worked and then create new kinds of data to further understand and combine the different experiments. Whether it is Orbit, which is an incredible team that uses four arms for robots in space and it's a team that comes out of ETH Zurich. They've created a whole dashboard that creates all the different parts of the robot so that they have a command and control center that can be used for the robot but also when the robot is deployed in different deployments. So you have a huge database that you can build. You're creating multimodality from the images, the video, the analysis of everything that's happening. You're using voice to control and the bidirectional voice that we're investing in over and over again. All the capabilities of how the robots are then interacting with the real world because for most people, where AGI is going to be felt is in voice. So there's so much that you can do from your control center from Codex that I think robotics teams are some of the most sophisticated users of agent capabilities.

Truly. And maybe by the way, my first recommendation for them is usually to not avoid the sophistication but to really dive into it. The best teams we've seen, and a lot of them are also here from Paris thinking about, for instance, Genesis, released a very impressive demo a few weeks ago. It works when you dive really into all the details of how you train from the ML models down to the simulator down to how you interact down to the actuator. You need to go super deep in all this vertical if you want to get a model that not only does this nice demo but is actually something that pushes the state of the art. I think there are two things that are really very important to grasp in robotics. The first one is you will need to be super technical, technical to the extreme. And the second one is data. You will need to really understand your data, clean it, find it because it's way more complex than for LLM. You need to get it, and a lot of the challenges we see is around getting high quality data, getting it clean, using the right tools to, like you are saying, the dashboard to visualize it, understand which part of the data is interesting, which part of what your robot is doing is actually bad or good, which is way harder in robotics because you can't just read it or ask people to read it. You have to watch it.

I think Codex is helping a lot on that right. I was just chatting with Jim Fan at Nvidia, who is also here. Jim is the lead of robotics there. We were chatting about this very impressive project they have called Inspire where basically they use an agent to build the training loop and to build a training recipe to train the robots. I think that's really the future, which is you're going to use these amazing coding tools, this amazing Codex, GPT 5.6, which I'm happy is out finally that we can play with to basically help you build this research loop to create your robots.

Yeah, I let you answer first.

Well to leave by, I think there's no better way to end, and I see it's blinking. Anyway, thanks a lot everyone for listening.