The Physics Professor Who Won’t Touch AI, and the Student Who Can’t Stop

Top physics researchers aren’t blind to what AI is doing in the application layer. They’re locked by their own evaluation system and can’t afford to look.

Author: Koutian Wu; GitHub: ktwu01

I was talking with Gemini about this and an analogy came out that I can’t stop thinking about.

Wang Ming and Bo Gu. Early 1930s Chinese Communist Party. These were the theorists, Moscow-trained, with every line of Marxist-Leninist doctrine memorized and internalized. They weren’t stupid. They weren’t lazy. The thing that destroyed them wasn’t ignorance — it was that their entire identity, their authority, their legitimacy, was built on being correct according to the text. When the reality on the ground diverged from the doctrine, they couldn’t adapt, because adapting would have meant admitting the book was wrong. And that was a cost they couldn’t bear. So they voted with their feet, kept citing the correct passages, and led things into the ground.

Then there’s 教员. Mao. What made him different wasn’t that he’d read more books. It was that he fundamentally rejected the idea that theory should come before practice. 实事求是 — seek truth from facts. Figure out what’s actually happening, then act, then adjust. He didn’t have a perfect ideological system. He had ground-awareness and an iteration speed that nobody around him could match.

I watch this split happening right now in academia.

The Wang Ming and Bo Gu types today, I think, are the senior physics researchers at major universities. And I want to be careful here, because this is easy to misread. I’m not saying they’re dumb, and I’m not saying they can’t see what AI is doing. A lot of them have watched AlphaFold happen. They know Neural Operators exist. They can read the papers.

But their book — their sacred text — is the idea that students must understand physics first. That understanding means being able to derive things from first principles, trace the equations, build intuition the old way. Their whole evaluation system for what counts as legitimate knowledge runs through this framework. And their authority, the thing that makes them worth listening to after thirty years of research, rests on being the people who did that hard work and came out the other side with genuine physical intuition.

So when AI shows up and a student starts producing results without that traditional foundation, the defensive move is not to question AI’s capabilities. The defensive move is to question whether understanding has happened at all. They insist AI cannot truly understand physics. They insist a student who uses AI to produce results hasn’t really learned. Not because they’ve thought carefully through the epistemology, but because the alternative — acknowledging that the old path is becoming optional — would be to announce that their entire moat is dissolving.

Wang Ming and Bo Gu didn’t fail because they didn’t know dogmatism was dangerous. They failed because their knowledge system voted them toward dogmatism anyway. That’s the trap.

The UT professor who knows their students are using AI to get 100% on Python assignments and teaches it anyway — I get it. There’s even a real argument in there somewhere about logic training and foundational thinking. But Python as a language is probably going to be replaced. Programming as a required skill might not exist in ten years. And meanwhile Stanford CS students are using AI to build things that were simply impossible to start three years ago.

That delta is not closing. It’s widening.

I’ll be transparent about where I am in this. I’m a physics PhD student and by the traditional measure I don’t know enough physics. I use Claude Code, Gemini Deep Research, Cursor. A lot of what I produce, if you asked me to justify it from first principles, I couldn’t fully do it. My professors aren’t wrong that this carries risk. I can produce results that look right but encode bad physics. This has happened. It’s not hypothetical.

But here’s the turn I’ve made that I think they haven’t caught up with: AI is starting to understand physics too. Not reliably. Not yet completely. But the trajectory is clear. Neural Operators learn PDE solutions from data. PINNs encode conservation laws directly into the loss function. AlphaFold stunned structural biology not because someone explained every folding rule to it, but because it internalized the patterns well enough to beat approaches that took decades to develop. The question of whether AI can understand physics is a moving target, and the target is moving fast.

So my question isn’t “do I understand physics well enough to use AI.” My question is “can I build a system where the physical constraints get respected — whether that happens in my head, in the model, or in the validation loop.” That’s a different frame entirely.

Nobody’s going to talk me out of this frame. I know that. The people who think I’m doing it wrong aren’t entirely wrong about the risks. We’re just operating on different time horizons with different definitions of what “understanding” means.

That’s why I need results. Not the argument, not the seminar room, not a conference panel where everyone agrees to disagree. Real papers. Real money. Working systems that do something people care about. In Texas especially, that’s the only currency that actually settles anything. This is the Yan’an Rectification moment — you can’t argue about correct theory, you demonstrate with outcomes.

The strategy I keep landing on is 80/20. Eighty percent on things that hold regardless of how the technology shakes out — genuine domain knowledge, physical intuition, real understanding of what the systems I study are actually doing. Twenty percent on aggressive AI experimentation, Agent workflows, zero-to-one projects that were impossible before.

The 80 means I don’t drown if the wave breaks. The 20 means I have a ticket if it doesn’t.

The real difference between Wang Ming and 教员 wasn’t intelligence or even total knowledge. It was ownership. Wang Ming’s knowledge belonged to the books and to the authority structure that endorsed those books. The 教员’s knowledge belonged to the problem he was trying to solve and the ground he was standing on.

In the AI era, I think this is still the right frame. It’s not about how much AI you know or how much AI you use. It’s about whether you own the problem. If you’re always solving the hardest, most concrete, most real version of a problem that actually exists in the world, AI will always be a weapon in your hands and never your replacement.

If not, at some point you’ll have to ask yourself honestly: are you studying the map, or walking the terrain.