Feb 25, 2026 · 34m · latent-space
🔬Max Welling: Materials Underlie Everything
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Renowned AI researcher and physicist Max Welling explores the emerging frontier of 'AI for Science,' explaining how foundational physics principles and machine learning algorithms unite to revolutionize material discovery. Through his startup CuspAI, Welling demonstrates how computational search over chemical space can address existential global challenges like climate change and environmental contamination.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The hosts as informed peer 3.4 Guest teaching 3.0 Guest disagreement 0.5 The hosts pushing back 0.6
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 23:15 Pushback against autonomous dark lab fantasies
Welling dismisses the unrealistic industry hype of dark labs where AI completely automates discovery without human domain experts, calling for humility in the face of deep chemical complexity.
Hardest push from the hosts ▶ 29:55 Host pushes devil's advocate argument on data augmentationThe co-host directly challenges Welling's emphasis on equivariance by asking why engineers shouldn't simply use brute-force data augmentation across orientations instead.
Biggest teaching moment ▶ 11:29 Welling explains how materials underpin the entire tech stackWelling educates the hosts on why software LLMs are ultimately physical problems, walking down the chain from software to GPUs, semiconductor wafers, EUV lithography, and energy transition bottlenecks.
The host holds their own ▶ 30:59 Host raises the Bitter Lesson critique of hand-crafted inductive biasesThe host demonstrates deep ML literature familiarity by questioning whether hard-coded physical inductive biases will ultimately lose to raw compute and scale, directly invoking the Bitter Lesson.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Career Evolution: From Pure Physics to Real-World Impact | 4 | 2 | 0 | 0 | The host opens with an informed summary of Max Welling's foundational work across variational autoencoders, GNNs, equivariance, and quantum gravity. Welling provides an open, reflective explanation of his career arc transitioning from curiosity-driven theoretical physics to impact-oriented climate tech. | |
| Physics as the Unifying Thread in Machine Learning | 3 | 3 | 0 | 0 | The co-host asks whether physics is the connective thread across Welling's body of work. Welling expounds on symmetry groups, gauge equivariance, and non-equilibrium thermodynamics bridging to diffusion models. | |
| The Rise and Momentum of AI for Science | 3 | 2 | 0 | 0 | The hosts ask about the sudden surge and investment boom in AI for Science, and how non-domain AI engineers can transition into the space. Welling outlines the success of protein folding and ML force fields along with educational pathways. | |
| Materials Underlie Everything: Transforming Discovery into Search | 3 | 4 | 0 | 0 | The host probes what AI for science offers beyond pure software/bits. Welling delivers an educational monologue explaining how all compute and energy technologies fundamentally bottleneck on material discovery, which AI converts from slow empirical trial to search. | |
| CuspAI’s Mission: Accelerating Carbon Capture and Materials | 2 | 3 | 1 | 0 | The co-host asks for an introduction to CuspAI's mission and scope. Welling describes the carbon capture imperative and introduces his conceptual model of experimental labs acting as nature's 'Physics Processing Units' (PPUs). | |
| Architecture and Team Behind the CuspAI Platform | 3 | 3 | 0 | 0 | The host inquires about the platform's architectural design and development process. Welling outlines the multi-scale digital twin ladder, agentic orchestration, and the scientific team executing it. | |
| Tool Building and the Human-in-the-Loop Philosophy | 5 | 4 | 2 | 2 | The host asks whether CuspAI started fully autonomous and added humans or vice versa, then attempts to summarize Welling's philosophy as tool-building rather than automation. Welling gently corrects the framing, explaining automation is a progressive retreat of the expert, while pushing back against unrealistic 'dark lab' fully automated hype. | |
| Moonshots, Partnerships, and Breakthrough Materials | 3 | 2 | 0 | 0 | The co-host asks whether CuspAI chases single transformative moonshots or incremental commercial wins. Welling explains their balanced strategy, highlighting specific domain partnerships like PFAS water filtration with Kemira. | |
| Demystifying Equivariance and Inductive Biases in AI | 6 | 4 | 2 | 4 | The co-host plays devil's advocate by questioning why mathematical equivariance is needed when data augmentation exists, and the host prompts with Rich Sutton's 'Bitter Lesson.' Welling gives a nuanced technical answer on optimization surface constraints versus scale. | |
| Generative AI, Stochastic Thermodynamics, and Upcoming Book | 2 | 3 | 0 | 0 | The co-host asks about Welling's forthcoming book. Welling details how the core mathematics of generative diffusion models unifies with non-equilibrium statistical mechanics, citing historical links from Hinton and Neal to Friston. |
Statements from this episode (19)
Max Welling: 2D quantum gravity research yields zero real-world impact
“Working in two-dimensional quantum gravity, you pretty much guarantee there's going to be no impact on what you do relative, you know, maybe a few papers, but not in this world, at this energy scale.”
Max Welling: CuspAI was founded because politics is failing on climate change
“I got pretty worried about climate change. I think we, and I think we should, you know, and politics seems to have a hard time solving it, especially these days, and so I thought better work on it from the technology side, and that's why we started CASP AI.”
Welling: Diffusion Models Share Exact Mathematics With Non-Equilibrium Stochastic Thermodynamics
“It turns out that the mathematics that we use for diffusion models, but even for reinforcement learning, for Schrodinger bridges, for MCMC sampling, has the same mathematics as this theory, this physical theory of non-equilibrium Systems.”
Welling: Billion-Dollar Funding Rounds Signal Exploding AI for Science Bubble
“It's not just emerging, it's exploding, I would say. That's the better term, because I know you go from investments into like in the hundreds of millions, now in the billions. So there's now actually a startup by Jeff Bezos that, you know, is that 6.2 billion …”
Welling: Protein Folding and ML Force Fields Drove AI for Science Boom
“I think there's two big examples, you know, protein folding is a big one, and the other one is machine learning force fields, or something called machine learning inter-atomic potentials. Both of them have been actually very successful.”
Welling: Semiconductor Scaling Limits Require Materials Innovation for Hardware Advances
“More or less we've reached the limits of You know, scaling things down, and now we are trying to improve further by new materials, so that's the fundamental materials problem.”
Welling: Perovskite-Silicon Solar Panels Can Theoretically Reach 50% Light Capture
“They can now make solar panels with new perovskite layers on top of the silicon layers that can capture, you know, Theoretically up to 50% of the light, where now we're at, I don't know, maybe 22 or something, right?”
Welling: AI Enables Searching the Space of All Possible Molecules
“Now we can treat this as a search engine. Like we search the internet, we now search the space of all possible molecules, not just the ones that people have made, or that they're in the universe, but all of them.”
Welling: Keeping warming under 2°C requires century-long atmospheric carbon removal
“In order to get, you know, to stay within two degrees, let's say, we would not only have to reduce our emissions to zero by 2050, but then, you know, another half century or even a century, Of removing carbon dioxide from the atmosphere, not by reducing your e…”
Welling: CuspAI has raised $130 million and expanded to 40 people
“We've grown to about 40 people. We've kind of collected a hundred thirty million investment in the, into the company, which is for a European company is quite a lot.”
Welling: Physical Lab Experiments Act as Nature-Powered 'Physics Processing Units'
“I want to think of it as what I would call a sort of a physics processing unit, like a PPU, right? Which is you have digital processing units, and then you have physics processing units. So it's basically nature doing computations for you. It's the fastest com…”
Welling: Materials Discovery Moat Lies in Data and Platform Engineering
“Where the moat is in the data that you can get your hands on, and the, and actually building the platform”
Welling: CuspAI Only Pursues Materials Discoveries with Industrial Partners
“So we always, we only start to invest in the direction if we find a good industrial partner to go on that journey with us.”
Welling: CuspAI is partnering with Kemira to remove PFAS from water
“We want to remove PFAS from water, right? So we do this with a company, Camira. So they are a deep partner for us, right?”
Welling: AI materials discovery breakthroughs will require humans in the loop first
“I think that the breakthrough will happen with a lot of human in the loop, because there is the chemists who have a whole lot more knowledge of their field, and it's us who will, you know, help them with AI, training AI and new methods, and in that kind of int…”
Welling: Lab Automation Fails to Generalize Across Material Science Problem Classes
“And also, it is very vertical specific. So it's like completely automating something for problem A, You know, you can probably achieve it, but then you'll sort of have to start over again for problem B because, you know, your experimental setup looks very diff…”
Welling: Equivariant neural networks need far less training data
“Where if you build equivariance in, basically, once you've trained it in one orientation, it will understand it in any other orientation. So that means you need a lot less data to train these models.”
Welling: Simple Data Augmentation Can Outperform Hard-Coded Neural Network Equivariance
“Sometimes actually data augmentation works even better than hard coding the equivalence in, and this is something to do with the fact that if you constrain the optimization, the weights, before the optimization starts, the optimization surface or objective bec…”