Max Welling has one of the more unusual career arcs in machine learning. He started in quantum gravity, working under Nobel Laureate Gerard 't Hooft. He co-invented the Variational Autoencoder. He helped build the mathematical foundations of equivariant neural networks. And now he runs CuspAI, a materials discovery startup with $130 million in funding and a mission to find new materials for the energy transition.
On this episode of Latent Space, Welling laid out a vision that should make any AI engineer rethink where the real bottlenecks are. Not in software. Not in model architecture. In materials.
The conversation moved fast - from the philosophy of treating nature as computation, to the practical reality of building an AI-driven materials platform, to a deep connection between diffusion models and thermodynamics that Welling has turned into a forthcoming book. Here are the parts that matter most.
Nature as the Fastest Computer
Welling's framing is striking: a "Physics Processing Unit." The idea is that real-world experiments are a form of computation. You have digital processing units (GPUs in data centers) and physics processing units (laboratories running experiments on actual atoms).
"Nature is the fastest computer known, as possible even. It's a bit hard to program because you have to do all these experiments. Those are quite bulky. But in a way it is a computation."
The vision for CuspAI is to make these two kinds of computation work together. Digital models generate and filter material candidates. Physical experiments validate the survivors. The feedback loop between digital and physical gets tighter over time.
This is not the "fully autonomous dark lab" vision that some AI-for-science startups pitch. Welling is explicit about that:
"The vision of a completely dark lab, where you close the door and just say 'find something interesting' - that's not the vision I have. Not for a long time."
Instead, CuspAI builds tools that make chemists and materials scientists dramatically faster. The human stays in the loop - not because of some philosophical commitment to human oversight, but because the problem is genuinely too complex for full automation today.
Why Materials Are the Real Bottleneck
Welling makes a case that most people in AI don't think about enough. Every layer of the technology stack sits on top of materials.
LLMs run on GPUs. GPUs are made by depositing materials on wafers and etching structures with EUV light. We've nearly hit the limits of scaling transistors down - future improvements depend on new materials. Batteries for the energy transition? Materials problem. Solar panels that can capture 50% of sunlight instead of 22%? Materials problem. Carbon capture at scale? Materials problem.
"Underlying everything is a material. I can probably dig deep enough and then tell you, well, actually, the very foundation of what you're doing is a material problem."
The opportunity, as Welling sees it, is that we can now treat materials discovery as search. Instead of the traditional cycle - read papers, form a hypothesis, run one experiment, learn, repeat - you can search the space of all possible molecules. Not just the ones scientists have already synthesized. All of them.
"We can treat this as a search engine. Like we search the internet, we now search the space of all possible molecules."
This is where the funding numbers start to make sense. CuspAI's $130 million is significant for a European startup, but it was quickly eclipsed by competitors. Welling mentions a Jeff Bezos-backed AI-for-science startup that raised $6.2 billion in what appears to be the largest seed round ever. The capital flowing into this space signals that investors see materials discovery as infrastructure-level opportunity - not a niche academic exercise.
Building the Platform: Tools First, Agents Later
The most practically useful part of the conversation was Welling's description of how CuspAI's platform evolved. It mirrors a pattern that works across AI applications, not just materials science.
They didn't start by building an autonomous agent. They started by building tools.
"You build all these tools, and then you go through a workflow actually in the beginning just manually. You figure out, oh, this porous material we're trying to make actually collapses if you shake it. Okay, then you add a new tool that tests for stability."
The progression went:
- Manual workflows - Scientists stringing tools together by hand
- Modular tooling - Each capability packaged as a reusable component
- Workflow automation - Connecting tools in repeatable sequences
- Agentic orchestration - Bayesian optimizers or LLM-based agents choosing which tools to run and in what order
This is the same bottom-up pattern that works for enterprise AI in any vertical. You don't start by deploying an autonomous agent. You start by understanding the workflow, building tools that handle individual steps, automating the connections, and only then layering in AI-driven decision-making where it adds clear value.
Welling gave a concrete example: DFT (Density Functional Theory) calculations. Every time a non-expert needed to run one, they had to consult a DFT specialist. The first automation step was making DFT user-friendly enough that non-experts could run the right calculation for the right problem. Then automate the parameter selection. Then automate the quality assessment. Piece by piece, the human retreats from routine decisions and focuses on judgment calls.
"It's a retreat. First you need an expert to tell you precisely how to set the parameters. Maybe we can automate that. And so increasingly, more of these things are going to be removed."
The Bitter Lesson Hits Science Too
Welling touched on something that will resonate with anyone who has followed the scaling debate in LLMs: the tension between inductive bias and raw data scale.
In scientific AI, equivariant neural networks embed physical symmetries directly into the model architecture. If you know a molecule looks the same regardless of how you rotate it, you can hard-code that constraint. This means you need far less training data.
But there's a catch. Constraining the weights before optimization makes the loss surface more complex. Harder to find good minima. Sometimes, with enough data, data augmentation (just showing the model lots of rotated examples) works better than hard-coding the symmetry.
"Ultimately it's a trade-off between data and inductive bias. If your inductive bias is not perfectly correct, you have to be careful because you put a ceiling to what you can do."
Welling's take: the bitter lesson applies here too. Make sure your architecture scales. If you have a tiny dataset, inductive bias is a lifesaver. But as data grows, the constraints you thought were helping can become the bottleneck. The same scaling dynamics that drove transformer dominance in NLP are coming to scientific AI.
This has direct implications for anyone building AI systems in specialized domains. The question "should we hard-code domain knowledge or let the model learn it from data?" doesn't have a universal answer. It depends on how much data you have and how correct your assumptions are.
Diffusion Models Meet Thermodynamics
Welling's forthcoming book, "Generative AI and Stochastic Thermodynamics," argues that the math behind diffusion models (the technology powering image and video generation) is identical to the math behind non-equilibrium statistical mechanics (how physical systems of molecules relax to their ground state).
This isn't a loose analogy. Welling claims the equations are the same. Variational free energy in machine learning - first written down by Hinton and Radford Neal years ago - maps directly onto free energy in physics. Fluctuation theorems from stochastic thermodynamics have counterparts in generative AI that the ML community hasn't explored.
"When we see that these things are actually the same, we can look at this new theory developed by very smart physicists and say, what can we take from here that will make our algorithms better?"
The cross-fertilization potential is real. Physics has theorems that ML hasn't used. ML has computational tools that physicists need. Welling sees this as the beginning of a genuine new discipline - not "AI applied to science," but a merged field where the math flows both ways.
What This Means for AI Builders
Three takeaways for engineers and technical leaders:
The hardest problems aren't software problems. If your AI system interfaces with the physical world - manufacturing, logistics, energy, healthcare - the materials and hardware layer may be your actual constraint. Software-only optimization hits a ceiling when the underlying physics limits what's possible.
The tools-first approach works everywhere. CuspAI's progression from manual workflows to modular tools to agentic orchestration isn't specific to chemistry. It's the reliable pattern for deploying AI in any complex domain. Skip steps at your peril. The Anthropic guide to building effective agents makes the same point: start with the simplest solution, add complexity only when it earns its keep.
Inductive bias vs. scale is the central design question. Whether you're building equivariant graph neural networks for molecular simulation or choosing between fine-tuning and prompt engineering for an enterprise chatbot, you're making the same trade-off. How much domain knowledge do you bake in, and how much do you let the model discover? The answer depends on your data budget and how confident you are in your assumptions.
How OpenNash Can Help
CuspAI's platform evolution - from manual tooling to automated workflows to agentic orchestration - maps directly to what we build for enterprise clients every day.
Workflow audit and tool design. Before anyone talks about agents, we map the actual process. What are the individual steps? What decisions get made? Where does a human add value vs. where are they just clicking buttons? This is the same "build tools first" approach Welling describes - and it applies whether you're automating materials discovery or invoice processing. Our enterprise AI readiness checklist walks through this progression.
Human-in-the-loop agent design. Welling is right that full autonomy is a long-term vision, not a near-term reality. The production systems that work today keep humans in the loop for judgment calls while automating the routine. We build agents that match this pattern - the 5 agent patterns that actually work in production are the ones with clear escalation paths and deterministic guardrails.
Cost-controlled scaling. When your AI system needs to run thousands of evaluations (whether they're DFT calculations or LLM calls), cost engineering matters. We apply the same model routing and token budget strategies that keep production agents under budget.
Concrete mini-example. At one client, we implemented an invoice audit agent in 4 weeks:
- Before: Finance ops manually reviewed high invoice volume and month-end close slipped because exception handling was fragmented across inboxes and spreadsheets.
- Build: We shipped document ingestion, policy-rule checks, anomaly detection, and a human approval queue for exception cases.
- After: The team reduced manual review load, closed faster, and gained a cleaner audit trail for leadership and compliance.
If your team is trying to move from prototype to production with AI-driven workflows, book a call and we'll map the pattern to your specific use case.
Full Transcript
Below is the complete transcript with timestamps. Click any timestamp to jump to that point in the audio.
[0:00] Max introduces the idea of a Physics Processing Unit - nature doing computations as the fastest possible computer. The interface is complicated because experiments are bulky, but digital and physical computation need to work together seamlessly to discover new materials.
[0:44] Brandon frames Max Welling's career arc - from VAE pioneer to equivariant graph neural networks to CuspAI founder. The question: what's the thread connecting quantum gravity to materials science startups?
[1:34] Max describes how his motivation evolved. In his younger days, he followed pure curiosity - black holes, the boundary of the universe, quantum mechanics. As he got older, impact became a second dimension. Quantum gravity guarantees no real-world impact at current energy scales.
[2:43] Climate change became the driver. Politics struggles to solve it, so working from the technology side made more sense. CuspAI combines impact with interesting science - both the depth of the problems and the potential to build tools that matter.
[3:39] The thread is physics. Symmetry plays an enormously important role in particle physics and general relativity. Applying those symmetries to machine learning became a deep mathematical problem - from simple rotational symmetries all the way to gauge symmetries on spheres.
[6:52] AI for science isn't just emerging - it's exploding. Investments have gone from hundreds of millions to billions. A Jeff Bezos-backed startup raised $6.2 billion in what may be the largest startup seed round ever.
[7:53] Two catalysts: protein folding breakthroughs and machine learning force fields proved AI could move the needle on hard science. Both involved symmetries. The AI community saw an opportunity for real impact beyond advertising and multimedia.
[10:12] For AI engineers without a science background: read the book when it comes out. Universities are creating curricula at the AI-science interface. Workshops, online courses, and tutorials are growing. The content to get into this field is becoming much more accessible.
[11:28] Underlying almost everything is a material. Under an LLM is a GPU. Under a GPU is a wafer with deposited materials. We've hit limits on scaling down - future improvement requires new materials. Batteries, fuel cells, solar panels - all materials problems.
[13:02] Materials discovery can become a search engine. Instead of the slow hypothesis-experiment-learn cycle, you search the space of all possible molecules - not just ones that exist, but all of them. Type what you want, computation and experimentation happen, and you get a list of candidates.
[14:48] CuspAI started 20 months ago, motivated by climate change. Staying within two degrees requires not just zeroing emissions by 2050, but actively removing CO2 for another half-century at half the current emission rate. The company has grown to 40 people with $130 million raised.
[18:01] The platform design isn't rocket science in concept. A generative component proposes candidates. A multi-scale digital twin filters progressively - cheap simulations first, expensive ones later. Survivors go to experiments. Agents now orchestrate computations and search literature.
[21:17] The evolution was bottom-up. Build tools, run workflows manually, figure out what's missing, add new tools. Only then automate. First a chemist assembles the workflow. Then you ask: how do I automate this? Small pieces first, then bigger ones.
[25:04] CuspAI follows a mixed strategy - a long-term lighthouse material (impactful proof point) alongside shorter paid partnerships with companies that have more modest goals. They prefer deep partnerships where they can change something meaningful.
[29:01] Equivariance explained: if you train a neural network to recognize a bottle, rotating the bottle breaks recognition. Equivariant networks understand all orientations from training on one. This drastically reduces data requirements by constraining model weights to respect symmetry.
[30:01] Data augmentation is an alternative but not exact - you'd need infinite augmentations. Sometimes augmentation works better because constraining weights makes the optimization surface more complex. The field has contradicting claims depending on data availability and application.
[31:55] Welling's upcoming book, "Generative AI and Stochastic Thermodynamics," reveals that the math behind diffusion models and non-equilibrium statistical mechanics is identical. This enables cross-fertilization - physics theorems can improve ML algorithms, and ML tools can advance physics.
[33:44] Publication timeline depends on the publisher, but Welling hopes to have it ready for his ICLR keynote in April.