Episode Summary
Executive Summary: Max Welling argues that physics is a powerful design toolkit for AI, not just a scientific domain for AI applications. He describes Cusp AI’s platform for discovering new materials and explains how equivariance, diffusion/thermodynamics analogies, and wave-based neural networks could improve memory, reasoning, and long-range information flow in AI.
Main Topics: AI for Science and Cusp AI’s materials platform (Priority: 5/5): Welling explains how Cusp AI uses generative and agentic AI to design novel materials for carbon capture, semiconductors, batteries, solar, catalysis, and PFAS removal, combining database search, generation, simulation, and lab validation. Equivariance as a practical scientific advantage (Priority: 5/5): He traces his earlier work on geometric neural networks and equivariance into chemistry/materials, where symmetry-aware models accelerate force prediction and molecular simulation by roughly three to four orders of magnitude versus expensive quantum methods. End-to-end materials discovery workflow (Priority: 5/5): The discussion details Cusp’s pipeline: user-defined material constraints, literature/database retrieval, generative molecule creation, fast filtering, force-field distillation, molecular dynamics, higher-scale device modeling, and eventual experiments, increasingly integrated with self-driving labs. Physics as a blueprint for AI architectures (Priority: 5/5): Welling argues that concepts from physics—thermodynamics, entropy, wave propagation, symmetry breaking, and edge-of-chaos dynamics—can directly inspire better machine learning systems, not just interpret existing ones. Generative AI and stochastic thermodynamics (Priority: 4/5): He previews his book on the mathematical equivalence between generative AI methods and non-equilibrium thermodynamics, emphasizing shared information-theoretic foundations and practical cross-fertilization between machine learning and physics. Wave-based neural networks and long-range memory (Priority: 5/5): Welling presents work on neural networks that use spontaneous symmetry breaking to create stable traveling waves, improving information propagation, memory tasks, and operation near the edge of chaos. Scaling from molecules to devices and markets (Priority: 4/5): He emphasizes that the ultimate goal is not just discovering candidate molecules, but scaling them into real devices and commercially valuable products through partnerships, labs, and customers.
Key Arguments: Physics provides not only metaphors but concrete mathematical tools that can improve AI architectures, especially for long-range information flow and generative modeling. Equivariance is crucial in chemistry and materials because rotations/translations should not change physical predictions; baking that symmetry into models yields major efficiency gains. Generative AI can search not just existing objects but entirely new materials, enabling custom-designed compounds for targeted functions like selective CO2 capture. A practical AI-for-science platform must combine literature search, generation, simulation, and experiments rather than relying on a single model. Self-driving labs are a key multiplier because they can generate experimental data much faster, making the discovery loop more iterative and scalable. The math of diffusion models and non-equilibrium thermodynamics is deeply related through information theory, entropy, and free energy. Wave phenomena and symmetry breaking can help neural networks avoid vanishing/over-smoothing problems and sustain memory across deep time/layer horizons. The most important validation is not just academic performance but discovering a material that reaches a real device and creates customer value.
Data Points: Initial investment: $30 million - Seed funding for Cusp AI’s launch and team building Team size: ~50 people - Current size of Cusp AI across geographies Company founding year: 2024 - Cusp AI was started in spring 2024 Acceleration from equivariant surrogate models: 3-4 orders of magnitude - Speedup relative to expensive quantum mechanical approximations in materials simulation Atmospheric carbon removal target discussed: 20 gigaton/year - Welling’s estimate of how much CO2 must be removed annually over the next 50-100 years Time horizon for post-net-zero removal: 50-100 years - Period during which large-scale carbon removal may still be necessary Experimental throughput with self-driving labs: ~100 experiments/day - Illustrative scale of faster lab iteration enabled by automation Traditional experiment count: order ten materials at most - Old workflow for lab validation before AI/self-driving lab acceleration Labs/geographies: Amsterdam, Cambridge, London, Berlin, Asia, North America - Cusp AI’s distributed footprint and expansion Publication timing: A few days / a week ago - He says the open-source molecular dynamics framework COPS was released shortly before iClear University time allocation: One day/week - Welling continues academic work and publishing at the University of Amsterdam
Pivotal Quotes: "Can we actually start using this phenomenon, this wave phenomenon that we see in the brain? Can we also start to use it in neural networks?" — Max Welling: Explaining why wave dynamics could improve deep neural networks and memory propagation "The mathematics that describes modern generative AI... turns out to be equivalent to the mathematics that describes modern non-equilibrium statistical mechanics or thermodynamics." — Max Welling: Core thesis of his book connecting generative AI and physics "We can do now in a few days what took a PhD before." — Max Welling: Describing productivity gains from AI-assisted scientific workflows
Implications: The conversation suggests AI’s next leap may come from physics-inspired architectures and AI-for-science systems that can discover, simulate, and validate new materials faster. For industry, this points to more efficient R&D, better memory-heavy models, and faster translation from discovery to deployable products.