CuspAI's Max Welling: ‘We are building molecules to remove forever chemicals from water’

{"text":[[{"start":7.92,"text":"Max Welling is a computer science professor at the University of Amsterdam and the co-founder of CuspAI, a UK-based start-up aiming to revolutionise the discovery of new materials. The company’s pitch is that it will use AI to achieve breakthroughs in “months, not millennia”."}],[{"start":23.52,"text":"He has flitted between the worlds of research and commerce for three decades. His academic life includes posts at University College London, University of Toronto, University of California, Irvine and California Institute of Technology, while his business career ranges from a stint as a start-up founder as well as roles at Qualcomm and Microsoft."}],[{"start":42.48,"text":"Welling co-founded CuspAI with Chad Edwards, a former executive at Quantinuum, in 2024, to create a “search engine” for materials science. The company has assembled a heavyweight board of advisers, including the AI researchers Geoff Hinton and Yann LeCun, Lord John Browne, the former chief executive of BP, and Martin van den Brink, the former president of ASML."}],[{"start":64.76,"text":"This month, CuspAI announced a $450mn second stage venture capital fundraising that valued the company at $2.6bn, led by Kleiner Perkins and NEA. Amazon’s Jeff Bezos also participated in the round. At the same time, CuspAI launched its AI Materials Foundry, bringing together more than 50 partners from around the world to develop new materials, including Nvidia, Meta, Samsung, Hyundai, Henkel and Tokyo Electron."}],[{"start":92.96,"text":"In this conversation with the FT’s innovation editor, John Thornhill, Welling talks about his plans to harness AI to create new materials that can help address climate change, manage the energy transition and solve the continued miniaturisation challenge facing the semiconductor industry."}],[{"start":93.46,"text":"John Thornhill: How did you first become interested in AI?"}],[{"start":111.68,"text":"Max Welling: I started off in theoretical physics in Utrecht, in the Netherlands, working with a Nobel laureate on quantum gravity. But I wanted to make more impact in the real world. And doing quantum gravity in two dimensions is a path to making very little impact. And so I decided to go into the very small, budding field of AI and machine learning. We were talking about models with 100 parameters back then, which is now a trillion. So you can see it has followed its own exponential path."}],[{"start":140.72,"text":"I was at Caltech as a postdoc for a few years. Then I moved to London to work with Geoff Hinton for one year and then to the University of Toronto with Hinton, who moved there."}],[{"start":150.88,"text":"I was definitely intrigued by machine learning. You could teach machines to learn from experience rather than hard coding things in. This was a whole new way to build AI. What we call deep learning and transformers and LLMs [large language models] and all that are based on that particular view of the world."}],[{"start":169.52,"text":"JT: Why did you decide to launch CuspAI?"}],[{"start":173.08,"text":"MW: There were a couple of things which were converging. The first was that I had developed a number of models in my academic lab. One was equivariant models in which the symmetries of the world are encoded in the neural nets. People were starting to use these models and I said this is the perfect time to make the jump."}],[{"start":189.56,"text":"But possibly a more important reason was that I was very worried about climate change and the energy transition. I saw a huge opportunity that these technological advances could bring. The beauty is that we are really working on these problems now. We are building molecules that remove PFAS [forever chemicals that do not break down in nature] from water. We have a whole programme to remove carbon dioxide from the atmosphere."}],[{"start":211.98,"text":"JT: You’ve pulled together quite an impressive board, including both Geoff Hinton and Yann LeCun. How do you stop them arguing about existential risk all the time?"}],[{"start":221.36,"text":"MW: It’s totally an interesting question. So, I think Geoff Hinton sees there are also good applications of AI, right? So not all applications of AI are bad. In particular, he supported us because he thought this was one of the better applications of AI that doesn’t do a lot of harm."}],[{"start":237.92,"text":"And of course, Yann LeCun has a very different view. So he’s not aligned with Geoff Hinton on those [existential risk] concerns. But he is aligned with the good things that can come out of AI in this case. There are also other very strong candidates on our board who have different views on the world. But those diversities of views are also very powerful."}],[{"start":null,"text":""}],[{"start":null,"text":""}],[{"start":260.2,"text":"JT: CuspAI says it is creating a search engine for materials science. Can you explain what that means to FT readers?"}],[{"start":267.96,"text":"MW: The first thing is that when you search for information on the web, you search for existing materials, existing articles and existing information. Nowadays when you type something into a search engine, you have the AI version of it, which has also created new information. And it can reason about and combine information."}],[{"start":286.52,"text":"So that analogy is actually very good with what we build. We also have a prompt or a sentence you can put in a platform and you express the type of materials and the properties that you want out of that material. So you could say something like “I want a semiconductor that has a high dielectric constant”. You could put that into the search box. It would start searching for existing materials because we have a proprietary database with a huge number of scientific articles that talk about materials and their properties."}],[{"start":null,"text":""}],[{"start":314.92,"text":"But then the next step is it will start to create entirely new materials. So these are materials that have never seen the light of day in this universe, which I find extremely fascinating, because the number of possible molecules the universe could have made is enormous, like unimaginably large."}],[{"start":331.08,"text":"Of course, only a tiny fraction has materialised in this universe. But now we can imagine these new materials and actually make these new materials in machines. Then these new materials will be tested using a digital twin. We do all sorts of simulations on these materials. We also send them to an experimental set-up where they will be made. And then in the end, a small list of those will be returned to the user."}],[{"start":353.8,"text":"JT: As I understand it, other materials science companies like Schrödinger are running forward simulations to try to find new materials, but you’re using inverse generative design. Could you explain the difference between those two approaches?"}],[{"start":364.12,"text":"MW: I don’t know precisely what they do at Schrödinger, of course. But I know there are companies that basically build simulators based on the laws of physics. So they take, let’s say, a quantum mechanical calculation and they simulate the evolution of atoms. [At CuspAI we use machine learning] to accelerate those simulations. And that is important because if you want to do inverse design, which is what you just mentioned, you try to improve the material by giving it better properties in these simulators."}],[{"start":393.2,"text":"If the simulator is very slow, it’s very hard to assess whether the material is doing well or not. But if you can accelerate this a lot, you can then simulate them and figure out if this is improving the material for the properties that you want."}],[{"start":405.88,"text":"JT: Can we take a real example? You have launched a project with Kemira, the Finnish chemicals company, to help remove forever chemicals. And I understand it took you six months to explore 300tn possible material structures and 5,000 novel material designs with 20 priority candidates. Could you walk us through that process?"}],[{"start":425.88,"text":"MW: So this involves a deep partnership with the domain experts at Kemira. They tell us very precisely what the molecules are that we want to capture in what’s called a metal-organic framework. Our task was for three of these molecules to find the metal-organic framework that would hold on to those PFAS molecules but also would be able to then release it once it’s full. You know, of course, the materials cannot be toxic because they go in drinking water. And there’s all these other things that the experts at Kemira listed that we would need to deliver."}],[{"start":null,"text":""}],[{"start":455.8,"text":"That 300tn is the total number of possible metal-organic frameworks you could potentially consider in this space. The first thing is, as I described, we just go through the literature and find the things which are potentially interesting. But we also have a dedicated generative model that would generate all of these possible metal-organic frameworks that would have these properties."}],[{"start":475.68,"text":"What we do then is we generate the molecules with our generative models and we search. We don’t literally instantiate 300tn of them. It’s more like searching in the space in the direction of better and better molecules. Once we have generated the right ones, we use a pretty expensive simulator. This figures out whether the molecule is stable. So we wiggle the molecule around and if it collapses on to itself it is not a stable molecule. But if it stays strong, it’s stable."}],[{"start":506.64,"text":"Then we go through this funnel and do all these tests in a computer and most of them are dismissed. Then there’s a list of a few that remain that are created in a lab."}],[{"start":517.08,"text":"JT: Where does this project stand today? You’ve identified priority candidates. What are the next steps?"}],[{"start":523.66,"text":"MW: There are still a few steps between the identification of these materials and putting them in an actual process. We need to do more experimental tests on these molecules to ensure they have all the right properties. But, of course, Kemira is the expert in that part of it. Then there is a whole certification process that follows. And this could be many years."}],[{"start":524.16,"text":"JT: What’s CuspAI’s business model?"}],[{"start":547.08,"text":"MW: We have milestone payments for the delivery of the sort of things that we agreed on. Then there is, you could say, a performance bonus. It can come in the form of a royalty on the actual material after commercialisation. Or it can come as another bonus or payment when the actual material makes it into a real device."}],[{"start":562.88,"text":"JT: Everyone is saying that we’re moving into a world of physical AI. Is that right? What does that mean for the world?"}],[{"start":null,"text":""}],[{"start":570.7,"text":"MW: This is very interesting. I do believe the next wave in AI is physical AI. What does it mean? In physical AI we want to instantiate things in the physical world. We want to do as much of that in the digital world. And this is called a digital twin. So, for instance, if you zoom out a bit you could imagine an LLM would connect to an AutoCAD tool that designs a bridge. The request will be: design for me a bridge with these properties. It needs to be sturdy. It needs to be able to do all these things. It needs to hold so many cars. You put all this in this AutoCAD system."}],[{"start":600.76,"text":"And it will start to design this bridge and optimise this bridge. And then, of course, in the end you have to print it or you have to make it. But this step where you build this digital twin and you run it, using an LLM, is what we can now think of as the next step in physical AI. That’s where you can see a lot of companies are getting interested in building these kinds of intelligences that directly interact with the physical devices."}],[{"start":624.2,"text":"JT: The LLMs that we have at the moment have ingested all of the world’s known knowledge from the internet. But the next revolution is going to depend on us generating new data, isn’t it? And that’s going to involve robots going out to amass it all, so there’s going to be a very big land grab for new data sets, right?"}],[{"start":642.24,"text":"MW: It doesn’t have to be robots running around in the real world. It could be RL [reinforcement learning] environments, which I believe is the more realistic thing. You build simulators of the real world, which really model the physics. Then you optimise models to operate well in the simulator. So this is more like a digital twin version of what you just described, where the data gets collected by simulators."}],[{"start":665,"text":"Maybe another way to think of it is that there are game consoles. If you run games, they are actually very good physical simulators and people control these games. That creates a lot of really interesting data. And that data gets collected. That is very good data for training robotics platforms. So I think that’s probably more realistic as a data collection effort."}],[{"start":684.76,"text":"JT: So people who play video games are helping your research to develop new materials?"}],[{"start":689.88,"text":"MW: That may be a bit of a big step. I think you can collect data in a sort of simulated physical world. And then you can also have cars and robots in the real world. Or for us this would be having experimental devices, which would collect real data from the real world. So these are two modalities in which we can collect more data. But what we will see is that there is a lot of intelligence in what we call world models, which is our three-dimensional geometrical understanding of the world."}],[{"start":714.68,"text":"JT: There’s a massive debate at the moment about how far LLMs can scale. And you have people like Yann LeCun talking about world models being the way forward. Where do you stand in that debate? Have we still got a lot further to go in scaling the basic LLM architecture or are we going to need new hybrid approaches to take us to the next level of AI?"}],[{"start":736.64,"text":"MW: I think all of the above will happen. The scaling will continue as far as it can be pushed until the economic model stops making sense. So if it becomes so expensive to put GPUs [graphics processing unit] in space and it doesn’t pay for itself, then people will stop doing it. But otherwise, this is the most predictive way to get performance gains. So it’s the easiest, least risky way for a company to get performance gains. And so they will keep doing it until the economic model doesn’t make sense anymore."}],[{"start":766.4,"text":"At the same time, I think there will be other ways in which we can improve these models. World models are a clear way forward but I think there will also be alternative hardware developed."}],[{"start":null,"text":""}],[{"start":776.12,"text":"There’s a really ridiculous difference between the efficiency of a human brain and the GPUs that we are using. This is like 1mn to 1bn times more efficient. And at some point we’ll figure out if we can create intelligence with brain-like structures. We want to get hardware that sort of starts to imitate that. And people have done this with spiking neural nets, not with great success, I would say. But as the pressure mounts and we need more intelligence out of every kilowatt hour, then I think people will go back to coming up with new hardware designs, which are more analogue."}],[{"start":806.96,"text":"JT: One of the big claims made by quantum computing companies is that they will be able to develop new materials. But there’s been some pushback on that as an idea as the boundaries of classical computing expand. Do you see these two as being complementary? Are there things that quantum computers can do that classical computers can’t in the field of materials science?"}],[{"start":826.96,"text":"MW: It’s definitely true that a quantum computer can do different things than a classical computer. The question is how big is the piece of cake for quantum? And how big is the piece of cake for classical? What has happened is that classical models have been very successful, like tensor networks, for instance. They’ve been eating away a little bit at the sort of domain where quantum would shine. But I think there are definitely material classes, which are extremely hard to simulate classically, where quantum could at least generate the relevant data for neural networks to try to approximate them. And there are certain phenomena, which are deeply quantum mechanics, where quantum computers could play an important role in the future."}],[{"start":868.36,"text":"JT: At CuspAI, what are the most promising industries for you?"}],[{"start":873.04,"text":"MW: Our most important application domain is semiconductors. I think over the next few years there’s going to be 1,000x performance gains needed. And Moore’s Law alone will not cut it. The other issue is that in two dimensions we are reaching the limits of lithography. We can go maybe a bit smaller but when we hit the level of atoms it becomes very hard to push any more out of it. We are now moving into three-dimensional stacking where we are putting materials on top of each other. It’s called packaging. But that comes with thermal issues. It gets very hard to radiate out the heat from these chips. This is a very clear material problem. And we are very happy to step into that space because we can actually create these new materials for the semiconductor."}],[{"start":null,"text":""}],[{"start":916.04,"text":"JT: How will the launch of Anthropic’s Claude Science affect CuspAI? How much of a competitive threat is this to you?"}],[{"start":923.36,"text":"MW: The arrival of Claude Science validates what we’ve known for years: science is the ultimate frontier of AI. But there’s a risk of these companies losing focus on their core mission as they push harder into vertical domains."}],[{"start":935.64,"text":"It’s easy to underestimate the friction when software solutions meet the physical world of labs, devices and supplier networks. That’s why CuspAI’s focus is connecting models with proprietary data, compute infrastructure, lab validation and real-world industrial deployment."}],[{"start":951.2,"text":"I personally don’t believe in the “superintelligence will solve everything” narrative. I believe in solving some of humanity’s hardest materials science problems through deep partnership."}],[{"start":962.4,"text":"JT: How ambitious are you for CuspAI?"}],[{"start":965.72,"text":"MW: Well, I want us to be the company in the world that develops breakthrough materials. We have seen already an enormous amount of acceleration using our platform. Our chief scientist types in things he wants to develop and this sort of motor spins up and creates all sorts of new insights, which he says in 20 minutes can create the equivalent of a three-year PhD."}],[{"start":988.12,"text":"Right now, I think we’re on a trajectory to be the global leader in developing new breakthrough materials. And I honestly believe that this could completely change the world if we can truly — in a very hot, fast way — develop all sorts of new materials with all sorts of surprising properties. For instance, the energy transition is just starting and it depends on creating these new materials. And that will create an enormous market as well."}],[{"start":1013.2,"text":"JT: How will the AI Materials Foundry work?"}],[{"start":1016.2,"text":"MW: The Foundry is a global network of data, labs, compute and scientific expertise for designing new materials, all orchestrated through CuspAI’s agentic platform. We have more than 50 founding partners now, including Nvidia, Meta, Hyundai Motor Group, Merck and Henkel."}],[{"start":1034.08,"text":"There are lots of ways for partners to take part: some will work with us on specific discovery programmes, while others are data partners, labs, compute or model providers, or will be part of sharing industrial expertise and best practice. The aim is to create a beneficial flywheel, where every collaboration and experiment strengthens the wider network and helps us move more quickly from materials design and simulation to synthesis and real-world validation."}],[{"start":1059.52,"text":"JT: How will CuspAI use the proceeds of your latest fundraise?"}],[{"start":1063.68,"text":"MW: The funding gives us the ability to move faster, growing our teams across Europe, the US and Apac, where we’ve just opened an office in Singapore. It means we can invest further in Cusp’s agentic platform, obtain more relevant data, and build the infrastructure needed to take new materials from prediction to the real world."}],[{"start":1083,"text":"The transcript has been edited for brevity and clarity"}],[{"start":1090.96,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1785567197_7859.mp3"}