Physics of AI Requires Mindset Shifts

Physics of AI Requires Mindset Shifts, is an article by Ziming Liu.

We usually think of AI as a bunch of abstract math or black-box code that mysteriously works, but this fascinating new post argues that we should actually be looking at it through the lens of physics. The author makes a compelling case that training an AI model is strikingly similar to physical processes we see in nature, like a hot gas cooling down into a perfect crystal or a drop of ink dispersing in water. It suggests that intelligence isn't just software magic; it is a fundamental physical phenomenon that follows the same laws of nature that govern atoms and energy.

The core technical insight here revolves around the concept of energy landscapes. Imagine a rugged mountain range where the high, spiky peaks represent chaos, errors, and noise, while the deep, comfortable valleys represent order and correct answers. When an AI learns, it is essentially gravity pulling a ball down into the deepest valley to find stability. This connects deep learning to thermodynamics and the way systems naturally try to minimize energy. So, when a model generates an image or writes a sentence, it isn't just guessing; it is mathematically freezing a chaotic cloud of random noise into a stable, structured reality.

Why this matters for the water cooler chat is that it signals a major maturity moment for the industry. For a long time, building AI has been compared to alchemy—we mixed ingredients in a pot and hoped for gold without really understanding the reaction. If this physics theory holds up, we are finally moving from alchemy to chemistry. It implies we might stop wasting billions on trial-and-error training runs and start predicting exactly how much energy or computing power is needed to create intelligence. It also opens up a wild philosophical debate: if thinking is just energy minimization, are our own brains just biological GPUs running the same physics equations?

Link: https://kindxiaoming.github.io/blog/2025/physics-of-ai/