The physics Nobel went to neural networks
Hopfield and Hinton won the physics prize, and Hassabis and Jumper shared chemistry for AlphaFold. A mechanical engineer's reaction to physics and AI converging.
This week the Nobel Prize in Physics went to John Hopfield and Geoffrey Hinton “for foundational discoveries and inventions that enable machine learning with artificial neural networks.” The next day, half the Chemistry prize went to Demis Hassabis and John Jumper for AlphaFold’s protein structure prediction, and the other half to David Baker for computational protein design.
The physics prize started an argument I’ve enjoyed following. Plenty of physicists are asking why a prize for physics went to computer science. I think the Nobel committee has a real case, and it’s more interesting than “AI is popular.”
Hopfield’s 1982 network came directly out of physics. He described a network of binary neurons with symmetric connections that has an energy function, like a physical system, and showed that it settles into energy minima that act as stored memories. Give it a corrupted pattern and it rolls downhill to the nearest stored one. It’s built from the mathematics of spin glasses, disordered magnetic systems where each atom’s spin wants to align or anti-align with its neighbors. Memory as a physical system finding low-energy states.
Hinton’s Boltzmann machine, with Terry Sejnowski, extended that with statistical mechanics. Its units are stochastic, it samples states with probabilities given by the Boltzmann distribution from thermodynamics, and it learns by adjusting weights so the distribution of its states matches the data. The name isn’t a metaphor. The learning rule comes from physics.
So the prize honors the path by which physics ideas became machine learning. What made neural networks work at scale later, backpropagation, GPUs and huge datasets, came from elsewhere, and Hinton has more of a claim to those than the citation mentions.
My own reaction is partly personal. I came into this field from mechanical engineering, and for a while I felt like a visitor. My thesis was geometry and linear systems. My intuitions came from springs, plates, fluids and energy minimization. Over the years I’ve found those intuitions useful again and again in machine learning, from damped springs making faces look alive, to thin-plate splines, to energy-based models in LeCun’s world-model paper. This week’s prizes felt like official recognition that the two fields share deep roots.
The chemistry prize matters in the other direction. In 2020 I wrote about AlphaFold as a physics problem solved mostly by learning from data. Now it’s a Nobel. That’s the part I find most exciting: not that physics produced AI, but that AI is starting to produce physics, chemistry and biology results. I’d be surprised if this is the last Nobel for a machine learning system’s discoveries.