The Physics Reason AI Works When It Shouldn’t - #568

The same mathematics that explains why a fridge magnet sticks to your refrigerator also explains why neural networks work when they have no right to. His group did the calculation. Nobody else had. Subscribe if you want science with evidence, not speculation.

Goldenfeld is a Professor of Physics at UC San Diego, a Fellow of the Royal Society, and a Member of the National Academy of Sciences who spent 36 years at the University of Illinois applying condensed matter physics to problems everyone else had given up on: why the genetic code is optimal, why early life evolved impossibly fast, and why AI works despite being overparameterized beyond anything classical statistics can explain.

The thread connecting all three is one idea: that what emerges from many things together is qualitatively different from the sum of its parts. That idea explains magnetism, AI, the origin of the genetic code, and why life may be inevitable wherever the laws of physics apply. He also argues that what is happening to science right now is not a disagreement about facts but a fracture in how people decide what is true, and that is a more dangerous problem.

What you’ll hear:

-Why the same phase transition that explains magnetism also explains why AI works at all

-Why Francis Crick concluded life must have come from outer space and what Goldenfeld found instead

-What horizontal gene transfer has to do with how libraries work

-Why Goldenfeld thinks Enceladus is a better bet for life than Europa

-The purpose of life, stated as a thermodynamics problem

-Why “different is more” is more useful than “more is different”

“The impact you make is the ratio of what you do divided by what everybody else does. Minimize the denominator.” — Nigel Goldenfeld

CHAPTERS