AIANALYSISPODCAST
University of Amsterdam professor Max Welling, co founder of the materials AI lab CuspAI, argues across a 2026 paper, a book, and a keynote at the major machine learning conference ICLR that physics belongs in AI's architecture, not just its
Welling published a May 2026 arXiv paper, a Cambridge University Press book in pre-order, and an ICLR 2026 invited keynote in the same week in August 2026. The argument runs across all three: the next architectural lever in AI is not larger models but physics, and the specific mechanism is a particle-physics concept called a Goldstone mode.
When a system that looks the same under some transformation suddenly stops looking the same, the residue is a Goldstone mode: a low-energy wave that carries information about how the symmetry was broken. Welling and collaborators (Iqbal, Keller, Song, Miyato) argue in arXiv 2605.14685, titled "Spontaneous symmetry breaking and Goldstone modes for deep information propagation," that deep neural networks cannot route useful information through many layers without breaking a symmetry on purpose. The transformer, the architecture behind today's large language models, has its own way of pushing signals through depth, but the authors argue it does so implicitly and partially. A purpose-built design, they claim, would let Goldstone modes do the work directly.
Welling has a documented stake in the outcome. CuspAI, the lab he co-founded, applies generative AI to design new materials for semiconductors, batteries, carbon capture, and clean energy. The pipeline uses foundation models for chemistry, agentic workflows, simulation, and automated experimentation to search material spaces faster than lab work alone. CuspAI is a working example of physics-informed AI in production rather than a thought experiment, which is also why his argument deserves to be read as a stakeholder's bet rather than a settled result. The same symmetry-breaking and Goldstone-mode primitives the arXiv paper floats as architectural levers are, in principle, a natural fit for a foundation model purpose-built for chemistry rather than adapted from a general-purpose language model.
The argument is wider than symmetry breaking. Episode 774 of the TWIML AI podcast, released August 25, frames four physics-to-AI bridges: connections between machine learning and thermodynamics, waves as a possible new computational primitive, symmetry breaking as architectural inspiration, and statistical-physics ideas for AI architectures beyond the current scaling paradigm. The book, "Generative AI and Stochastic Thermodynamics," co-authored with Lu and Holdijk, sits at the thermodynamics end of that list. Welling's own post directs the proceeds to AIMS, the African Institute for Mathematical Sciences, a small signal that the book is meant to seed a community, not just a citation.
Scaling has not stopped working. The 2026 cohort of frontier models continues to show that bigger training runs and longer contexts move capability forward, and most labs are not abandoning the scaling paradigm Welling calls incomplete. His bet, instead, is that physics offers a complementary lever. The next test of that bet is the arXiv paper itself: if the symmetry-breaking argument survives peer review at ICLR, the next round of architecture work can build on a defined primitive rather than a metaphor. If it does not hold, the paper, book, and keynote still mark the moment a serious research program was given a public shape outside the dominant scaling story. The marker to watch is whether the paper draws replication or critique at the same venue Welling is scheduled to speak.