An interesting and worthy read. As some commenters note (and the author admits), some of these parallels inspired by the thermodynamics metaphor shouldn't be pushed too literally for all ML systems.

During my PhD work I was immersed in these parallels, having worked with Hinton's Boltzmann Machine (and its "Mean Field Theory" variants), and having applied ML to the protein folding problem. The "spin glass" model from Physics seemed to provide insights into both protein structures and neural networks — though, again, the differences between these different complex systems had to be noted as important caveats.

I miss that intellectually stimulating work, and thank the author for the fun reminders.