Mike Winer wants math, not guesswork, to keep AI in check Statistical physicist Mike Winer, a Marvin L. Goldberger Member at the Institute for Advanced Study and an affiliate of the Alignment Research Center, is developing mathematical methods to predict how neural networks behave rather than relying on Monte Carlo sampling. Winer's approach uses cumulants and Hermite expansions to approximate a multilayer perceptron's expected outputs and the distribution of its internal activations at initialization, targeting rare extreme outputs that sampling struggles to find. Winer presented the work at the Yale FDS Colloquium on October 22, 2025 and is scheduled to speak at UCLA's IPAM Foundations of Interpretability Workshop on September 2, 2026, though he acknowledges the research so far covers only MLPs at initialization, not fully trained models. Mike Winer wants math, not guesswork, to keep AI in check A statistical physicist at the Institute for Advanced Study is building tools to predict how neural networks behave before anything goes wrong Most efforts to make AI safe come down to testing. You poke the model, watch what it does, and hope you poked it in the right places. Mike Winer, a statistical physicist, wants something sturdier: mathematical guarantees about how a neural network will behave, worked out with equations rather than trial runs. Who Winer is and what he is building Winer holds the title of Marvin L. Goldberger Member at the Institute for Advanced Study IAS . His specialty is the statistical physics of disordered systems, along with how that field applies to deep learning. He is also affiliated with the Alignment Research Center ARC in Berkeley. ARC focuses on keeping advanced AI systems aligned with human intentions. His current research targets multilayer perceptrons, or MLPs. Think of an MLP as a stack of filters. Data enters at the top, gets transformed layer by layer, and an answer comes out at the bottom. MLPs are among the most basic building blocks in deep learning. That makes them a sensible place to start if you want to understand neural networks from the ground up. Winer’s work centers on two questions. First, what output should you expect from a network on average? Second, how likely are rare, extreme outputs, the kind that rarely surface but could cause real damage? Swapping dice rolls for equations The standard way to answer those questions is Monte Carlo sampling. You feed the network huge numbers of random inputs and tally what comes out. Run enough samples and a reliable picture of its behavior eventually emerges. The catch is that rare events are, by definition, rare. To catch a one-in-a-billion failure by sampling, you need an enormous number of tries. AI, tech, and the markets they move—in one daily briefing. Daily. Free. Join 34,000+ readers across crypto, finance, and policy. Winer’s approach skips much of that. He mathematically approximates a network’s expected outputs and the distribution of its internal activations at initialization, the point when a model is first set up before training. To build those approximations, Winer relies on cumulants and Hermite expansions. Both are classical mathematical tools for describing the shape of probability distributions, including their tails. The tails are where the safety question lives. A model that behaves well 99.9% of the time can still produce dangerous outputs in the remaining sliver, and sampling is notoriously bad at finding that sliver. A growing push for verifiable AI Winer’s work fits into a wider research movement often called mechanistic interpretability. The idea is to understand neural networks the way engineers understand bridges: by knowing why they hold up, not just observing that they have so far. His research falls under a field ARC calls heuristic estimation. That means producing reasoned, mathematically grounded estimates of a network’s behavior without running every possible scenario. On October 22, 2025, he spoke about heuristic estimation of neural networks at the Yale FDS Colloquium. He is also scheduled to present at UCLA’s IPAM Foundations of Interpretability Workshop on September 2, 2026. That talk is expected to cover his methods built on cumulants and Hermite expansions. What this means for AI safety The caveats are real, and Winer acknowledges them. His work so far focuses on MLPs and on networks at initialization, a long way from the massive, fully trained models that dominate the AI industry. Trained models develop intricate internal structure that is much messier to describe than a freshly initialized network. Winer has said that much work remains in this area. Still, his background is an interesting asset. Statistical physicists have spent decades describing systems with countless interacting parts, from magnets to glasses, where tracking every component individually is hopeless. Neural networks share that character. Millions or billions of parameters interact in ways no one can follow one by one, which is exactly the kind of problem disordered-systems physics was built to tackle. For researchers, the key milestone to watch is whether these techniques scale beyond simple architectures. The September 2026 IPAM workshop should offer a fresh look at how far the methods have come. Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy https://cryptobriefing.com/editorial-policy/ .