Nature is the fastest computer we know — it's just hard to program

From quantum gravity to climate: why a theoretical physicist moved into carbon capture

Welling says that as a young researcher he followed curiosity: what lies inside black holes, where the universe ends and what quantum mechanics means. With age he added a second dimension—impact. Two-dimensional quantum gravity was almost guaranteed to have no effect on the physical world; concerned about climate change and doubtful that politics would solve it quickly, he decided to work from the technology side.

CuspAI follows from that decision, but it is not only an altruistic turn: materials science also contains deep scientific problems. He also discusses his adviser Gerard ’t Hooft—the most brilliant and rarely wrong person he knows—who now argues that quantum mechanics is wrong. Welling admires the courage and continues to seek a more mundane interpretation of quantum mechanics himself.

Physics as the thread: from gauge symmetry to diffusion models

Welling reduces his research trajectory to one sentence: physics is the thread. Symmetry in theoretical physics extends through gauge symmetry; with Taco Cohen and others he brought these structures into machine learning, from rotations to gauge symmetries on spheres. His former student Maurice Weiler later wrote an entire book on symmetry in machine learning.

His more recent thread connects diffusion models to stochastic thermodynamics. He argues that diffusion, reinforcement learning, Schrödinger bridges and MCMC share mathematical structure with non-equilibrium statistical mechanics, and developed an AIMS course into Generative AI and Stochastic Thermodynamics. At the interview date the manuscript was with the publisher and he hoped for publication before his April 2026 ICLR keynote; this note preserves that as a dated plan, not a completed event.

AI for Science is exploding—and Welling calls it a bubble

Welling gives two reasons for the field's surge: successes such as protein folding and machine-learning interatomic potentials showed that AI tools could enter science, while AI researchers wanted to apply their work to health, drugs, energy materials and carbon capture rather than advertising and multimedia. He notes that both success stories involve symmetry.

His read on capital is blunt: rounds have moved from hundreds of millions into the billions. He cites, without naming it, a Jeff Bezos-backed startup raising $6.2 billion, and concludes that the sector is creating a new bubble. The name Cusp comes from the same stage judgment: the field sits at the cusp of something new.

Why materials: beneath the LLM lies a materials problem

Welling's argument moves down the stack: LLMs run on GPUs; GPUs depend on deposited materials and EUV lithography; as dimensional scaling approaches its limits, further gains increasingly depend on new materials. The energy transition is similar—batteries, fuel cells and solar cells are fundamentally materials problems.

He cites tandem perovskite-on-silicon solar cells that could theoretically capture up to 50% of incident light versus roughly 22% at the time, and imagines plastics that degrade into fertilizer after weeks. The transcript renders his central line as 'underlying everything is immaterial'; context makes clear that he said 'underlying everything is a material'—the transcription reverses the meaning.

Turning materials discovery into a search-engine problem

The traditional loop is read papers, form a hypothesis, run an experiment, learn and repeat. Welling argues that materials space is becoming searchable for the first time—not only molecules already made or found in nature, but all possible candidates.

His end state resembles an interactive tool: specify desired properties, trigger computation and experiments in remote data centers and labs, receive a list of materials, then refine the query. It is not one-shot generation, but repeated convergence among computation, experiment and human judgment.

CuspAI's platform: the design is not the hard part

The platform combines a generative model, a multi-scale multi-fidelity digital twin and experimental feedback: cheap calculations remove obvious misses, progressively more expensive calculations narrow the set, and a small number reach experiment. Welling says this is not rocket science and that the early architecture remains largely intact.

Multi-scale modeling and agents for literature search, experiment suggestion and orchestration came later; he describes them as being at different stages of maturity. Self-driving labs were also absent from the original design and are now the next layer. The moat is not the architecture diagram, but the obtainable data and the engineering required to keep the platform running.

Bringing a material into the real world requires domain experts and industrial partners, so CuspAI enters a direction only with the right partner. The interview names leads for scientific-platform engineering and MLOps, but their spellings lack independent confirmation and are omitted here; Aron Walsh's appointment as chief scientific officer is verifiable through Imperial College.

The Physics Processing Unit: treating the lab as another processor

Welling does not want experiments to remain only end-of-pipeline validation. He frames them as a Physics Processing Unit alongside digital processors: nature performs computations through experiments, but the interface is bulky and hard to program. Computation in the data center and in nature must work together to reach a target material.

Immediately after the vision, he draws a vocabulary boundary: CuspAI does not say superintelligence because he does not know what it means and does not want to oversell it. The aim is to put a more powerful tool in the hands of chemists and materials scientists, not replace checkable workflows with a grand label.

Humans in the loop: why a dark lab is not the vision

CuspAI starts from expert workflows, not a fully automated lab: build tools for an application, let a chemist assemble them manually, add tools when new problems appear, then let a Bayesian optimizer or a model trained as a good chemist call them in the right order. One concrete automation test is whether a non-DFT expert can rely on the system to choose the right setup, runtime and quality checks without finding a DFT specialist.

Welling explicitly says that closing the door and telling a lab to find something interesting is not his vision and will not be for a long time. Materials engineering contains tacit expertise accumulated over careers, and automation is highly problem-specific: moving from problem A to B often means different experiments, characterization instruments and retrained or fine-tuned models.

He contrasts materials AI with quantum computing and fusion: the field need not wait decades for one all-or-nothing breakthrough; each thing built may be immediately useful. He expects the real breakthrough to arrive with many experts still in the loop.

Equivariance and the bitter lesson: trading inductive bias against scale

Welling explains equivariance as injecting symmetry into a neural network: once an object is learned in one orientation, the model recognizes it after rotation. Weight constraints can substantially reduce data requirements; relevant groups include translations, rotations, permutations and richer physical symmetries.

Data augmentation is another route, but not exact; in principle infinitely many augmentations would be required for complete coverage. Yet he concedes that with enough data augmentation can outperform hard-coded equivariance because constraints make the optimization surface more complex. The field contains contradictory empirical results, so neither side is an unconditional law.

Asked whether physically grounded priors face their own bitter lesson, he answers that it is ultimately a trade-off between data and inductive bias. An imperfect prior can cap model capability, yet a known symmetry is difficult to ignore. His practical rule is to keep architectures scalable unless the dataset is tiny.