AI for Science PhD Reflects on a Field Without Natives: Anonymized Forum Essay Goes Viral

Forum topic · ✨步子哥 · 2026-04-08

Summary

A Chinese physical chemistry PhD student's anonymous forum essay describes the structural crisis inside AI for Science research labs. The author argues the field has no "native" experts: no supervisor trained in the discipline exists, so knowledge transmission is broken and everyone self-learns with maps from the previous era. The post claims AI assistants like Claude and Grok understand specialized problems faster and more systematically than human advisors, citing an example where the author's three-week quantum chemistry bottleneck—identified as a Born-Oppenheimer approximation failure—was diagnosed in seconds. A senior labmate reportedly found his dissertation's core contribution absorbed into a late-2024 OpenAI model release. The author likens academia to a World Cup where the strongest teams have left to play a hidden league, and foundation models now redefine the frontier every few months while labs still operate on paper-and-review cycles. The essay ends with a personal note on "cognitive loneliness": after a month without effective human communication, AI became the author's best listener. It concludes by urging peers to be "clear-eyed" rather than "performing science" under obsolete rules.

The House of Wet Concrete: Collective Drifting on the New Continent of AI for Science

*Translated from an anonymous post on zhichai.net by a physical chemistry PhD student with a computer science dual degree, written after 700+ days in the AI for Science field.*

Imagine standing in a physical chemistry lab at a well-known central Chinese university. It's 2 a.m. in winter. I've held this in for two years, and I finally decided to write down the thing that "might get me torn apart if I say it out loud." Because if I don't say it, I'm going to lose my mind.

A New Continent with No Natives: All Advisors Are "Immigrants"

On my first day, I asked a seemingly naive question: "Which professor in our group actually graduated in AI for Science?" The room went silent. Then came the polite deflections: "I worked in condensed matter physics..." "My focus was molecular dynamics..." "I was machine learning, then transitioned..." Nobody. Not one person.

What does this mean? It means this field, in a philosophical sense, doesn't really exist yet. It has no "natives," no twenty-year lineage of knowledge, no methodology systematically forged from undergrad through PhD. Every advisor and senior student is charging into an unexplored continent carrying a map of the old world—building a wall out of physics bricks, chemistry tiles, and machine learning concrete, then saying: "Look, our house!"

But it's not a house. It's a pile of wet concrete. This is a structural, systemic problem: when a field hasn't yet produced even its first generation of homegrown graduates, its knowledge transmission chain is logically broken. It runs on AI reconstruction, self-teaching, and groping in the dark together. That's not a discipline. That's collective wandering.

When AI Understands Your Confusion Better Than Your Advisor

Once, I was stuck for three weeks on a quantum chemistry bottleneck. My advisor suggested: "Go read that 2019 paper." I did—it was helpful, but only up to 2019. Then I typed my problem (a 2,000-word description) into Claude or Grok. In three seconds it produced a complete evolution of the field from 2019 to 2025, and precisely identified three unresolved core contradictions: the gap in scale bridging, generalization failure under data sparsity, and the eternal tug-of-war between interpretability and computational efficiency.

In that moment my eyes went hot. Not because AI is "smart," but because it truly understood. It started from my cognitive framework. It didn't need me to first explain what a Hamiltonian is (briefly: a Hamiltonian is the physical world's "energy ledger," packing a system's kinetic and potential energy together to tell you how a molecule evolves through time and space).

So the question arises: what is the advisor for? Not to bash advisors—it's reality forcing the question. AI's answers are more systematic, faster, more accurate, more current. In highly specialized cognitive labor, the "alignment cost" between humans is absurdly high.

The World Cup Nightmare: The Strong Teams Quit, We're Still Playing Exhibition Matches

At the end of 2024, OpenAI released a new model. My senior labmate—a guy who had spent five years on quantum chemistry computation—read the technical report, went silent for a long while, then said: "The core contribution of my dissertation... has been packaged into it."

Not surpassed—*absorbed*. Like spending five years climbing a mountain, reaching the summit, and discovering the mountain has been buried under an office building's underground parking garage. It's still there, but no one will ever climb it again. Anthropic and OpenAI don't publish papers, don't submit to NeurIPS, don't chase Nature sub-journals, don't compete for citations or benchmark rankings. They do one thing: every few months, quietly lift the ceiling of the entire field, then return to silence.

It reminds me of a chilling analogy: the World Cup's most terrifying nightmare isn't Brazil being too strong—it's Brazil, Germany, Italy, and France jointly announcing: "We're not playing anymore. Too boring." Then the remaining teams keep warming up earnestly, celebrating semifinal berths, solemnly awarding the trophy. Meanwhile, outside the stadium, the truly strongest teams are playing a secret league we can't see and can't enter.

That is AI for Science today. We're still grinding papers, citations, and grant money in the academic league. They're already speeding down the Foundation Model highway, stomping our five years of work underfoot.

An Old Carriage Chasing a Bullet Train: Performing Science, Not Doing Science

We think we're in a scientific revolution, but we're using old-era organizational methods to chase a 300 km/h train. Advisors hold group meetings discussing "whether to take the first step," students run simulations overnight, papers wait half a year for peer review—while Foundation Models iterate every few months, turning yesterday's "frontier" into today's "baseline."

This isn't science. This is *performing* science.

The Crack of Cognitive Loneliness: When AI Becomes Your Best Listener

The most private pain: it's been a month since I had a single genuinely effective conversation with a human being. I still greet people, report progress, complain about cafeteria food with my roommate. But that's not communication—it's just emitting sound.

Real communication is when you explain something complex and the other person instantly gets it, then responds in a way that moves you forward. The last time I felt that was talking with AI. I dumped a 2,000-word molecular dynamics modeling problem on it, and in three sentences it caught a logical flaw I hadn't noticed myself: the Born-Oppenheimer approximation (briefly: it assumes nuclei move slowly and electrons move fast, so nuclei can be treated as a fixed background for the electrons; under certain extreme conditions, this approximation collapses) fails in my system.

When I explained the same problem to a classmate, he listened for ten minutes and said: "Hmm... maybe try running a different model?" I smiled, said thanks, and went back to chatting with AI.

I'm not becoming antisocial—I'm becoming rational. In highly specialized fields, differences in knowledge structure, cognitive speed, and background consume 80% of a conversation's energy just on alignment. AI starts from *your* framework. It knows Hamiltonians, knows graph neural networks (imagine a vast social network where every atom is a person, every chemical bond a friendship link, propagating "gossip" to predict the whole molecule's behavior). It waits for you to get to the point.

So I started talking to people the way I talk to AI: concise, logical, conclusions first. Others found me "cold" and "unfeeling." I found them "inefficient" and "missing the point." A crack opened up. It has no official name, but I know what it's called: cognitive loneliness.

Clear-Eyed, or Keep Performing?

Sitting in the lab, watching the night sky like a crumpled black cloth, I suddenly understood: we're not doing science. We're chasing new stars with old maps. The Foundation Model doesn't publish papers or play games—it just moves forward. And we're still at the foot of the mountain, holding meetings about how to take the first step.

I'm not despairing. I'm just clear-eyed. The rules of the game have completely changed—cruelly, with no retreat. If you keep playing by the old rules, you're not doing science; you're performing it.

To everyone still running on this track: I don't regret choosing AI for Science. But choose it with open eyes—be one of the clear-headed ones.

References

1. OpenAI. (2024). GPT-4o Technical Report. 2. Anthropic. (2024). Claude 3.5 Sonnet Model Card. 3. Jumper, J. et al. (2021). Highly accurate protein structure prediction with AlphaFold. *Nature*. 4. Wang, H. et al. (2023). AI for Science: A Review of Recent Advances. *Nature Reviews Physics*. 5. Davies, A. et al. (2024). The Future of Scientific Discovery with Foundation Models. arXiv preprint.