我为什么还在做 AI4SWhy I'm Still Doing AI4S

These days, doing research with AI, I often feel there isn’t enough time. I used to write code by first designing the structure, then implementing, compiling, and testing it piece by piece. Now the agent and I iterate on a plan back and forth a few times, and it quickly finishes a whole round on its own: it reads a few papers, changes a few thousand lines of code, produces a batch of results, and even writes up a respectable-looking summary along the way. The execution log keeps scrolling down the screen. Sometimes a step strikes me as somehow interesting without my being able to say why, and all I can do is make it stop and re-explain what it just did; and when it explains too fast, I have to ask it to slow down again. Gradually I’ve come to feel that I’m the one holding the agent back from working more efficiently, and lately this has left me increasingly anxious — and a little worried: when the speed of human understanding itself becomes the slowest link in the whole chain, is there still any value in a person staying in the loop?

I didn’t start out thinking about questions like this. In fact, my education wandered a long way, and at first I wasn’t even sure where I was headed.

I did my undergraduate degree at Tiangong University. For the first two years, I was a headless fly. When I enrolled in 2017, the news of AlphaGo beating Ke Jie was still fresh; artificial intelligence blew into the School of Computer Science like a sudden gust of wind. Nobody was quite sure what it was, but everyone knew they ought to at least look in that direction — as if the word itself carried a faint glow of the future. The school already had people doing network security, people writing front-end and back-end, people playing with embedded systems; now, all at once, more of them began talking about neural networks. I tried a bit of everything too. I sat in on a few network-security classes, but there were so many concepts that I was lost in the fog and soon stopped going; I dabbled a little in embedded systems and IoT without building anything worth showing. For neural networks, I’d go every evening to an upperclassman’s dorm and listen to him talk. He built his own models for competitions and had reached the top of the leaderboard on quite a few Kaggle datasets; the cost was just as plain — he’d failed a good number of courses, and whether he’d graduate seemed the more pressing worry. After about half a year, the thing I actually stuck with longest turned out to be the English club, because now and then we could use the school’s budget to go out and eat somewhere — the most concrete benefit of all.

Back then I also emceed events, took part in Model UN, pursued a second degree in public administration, and tried a children’s coding project with some classmates. We translated a set of visual programming tools into Chinese, hooked them up to dev boards and little robots, and actually ended up running after-school clubs at two schools; for a while we even thought about renting a storefront and registering a company. But everyone’s coursework gradually piled up, and we found that a team in Beijing had already turned something similar into a company, so the small bit of courage we’d just gathered slowly dispersed. My head was full of the liberal-arts ideal back then — or, to be more honest, I had no idea what I wanted and just wanted to look at everything. After class I often went to the library and read a great deal of philosophy, history, and social science. I was especially drawn to the more speculative parts of philosophy and jurisprudence, though it’s hard to say now how much of it I truly understood. To dress this period up as “interdisciplinary exploration” would be to gild the past; at the time it was mostly just seeing a door and wanting to stick my head in for a look, usually with no plan for what I’d do once inside.

What really began to gather up those scattered interests was a frontiers-of-computer-science course I took as an undergraduate. It covered the intersections of artificial intelligence with bioinformatics, drug discovery, speech, and vision. The material looks like only a very basic introduction in hindsight, but at the time it happened to be my awakening. Computers, it turned out, weren’t only for writing software; they could reach into other disciplines and touch problems that had nothing to do with me before. Afterward I got in touch with Professor Ding Liu, who taught the course, and did research under him together with my roommate. The first project I picked up had to do with RNA, proteins, and CRISPR gene editing — put simply, we wanted to use machine learning to predict where a certain protein would cut a sequence.

I had almost no background in biology at the time, and had forgotten much of the genetics I’d learned in high school. The first time I read a review paper, every sentence seemed to bristle with thorns — Latin terms and abbreviations crowded together, and looking up one word would grow three more I didn’t know inside its explanation; a single paragraph could take half an hour to get through. But I seemed to have a lot of patience in that period. Each concept I cleared up added one small dot in my mind. At first these dots had no relation to one another, just scattered there; as they slowly accumulated, they could suddenly be joined into lines, and later even yielded a faint sense of the whole. Why a protein can find a particular sequence, why its cutting is selective, how exactly the inputs and outputs in the data should correspond — these things were like a crumpled ball of paper being slowly unfolded and pressed flat, until at last you could even draw arrows on it and turn it into a concrete model. Transformers had just been proposed then, and we tried applying them to sequence tasks. I spent a long time reading papers, changing code, and cleaning data; at first the model did worse than random guessing, and then it climbed bit by bit from below fifty percent to sixty, seventy, and around eighty. The project never became a paper because of issues with the collaboration, and judged purely by a table of results it probably doesn’t count as a successful piece of research; but it nearly changed the way I see learning. Learning a new field brings a very direct kind of joy, like walking into a room that had been locked: when the door first opens it’s pitch dark inside, and you can only grope your way to the corner of a table and the wall; stay long enough and the furniture, doors, and windows slowly take shape, until at last you can move around in it. The room hasn’t grown any smaller — I’m just no longer shut out entirely by the dark. Ever since, I’ve been reluctant to first ask whether a problem is “my direction,” and more inclined to first ask whether it’s interesting. As long as it’s interesting enough, I always want to find a way in and take a look.

After that project, Professor Liu didn’t rush me on to running new data; instead he had me go study machine-learning theory, statistical learning, and information theory. He believed that if the theoretical foundations weren’t clear, whatever you built would easily amount to nothing more than a layer of surface tricks.

During that time, I began to encounter the theory behind neural networks. To make sense of the relationships among different models, I read many papers and tried to rewrite them into a common mathematical form. One day, the early neural energy models, associative memory networks, and modern attention suddenly revealed a similar skeleton on the page. A neural network no longer seemed to be just a black box; behind it there was a dimly visible structure that could be understood. So nearly all my enthusiasm at the time converged on one thing: “opening the black box.” That excitement seems a little naive to me now. Every time I’d just strung a few things together, I’d get the illusion of having understood a great deal, only to look again a few months later and find I’d merely walked out of one black box into a larger one. Still, research sometimes needs this kind of brief self-confidence to keep it alive.

Later we wrote a paper on information measures in neural networks, trying to define a new kind of entropy and mutual information in the high-dimensional parameter space. It came out in a not-very-prominent journal, but a few years on I found it cited by some work related to large-model training, and a study from LeCun’s group had used one of its measures to construct a training objective.

In the second half of my senior year, I settled on Tianjin University for graduate school, and came into contact with a collaborative project using AI to accelerate electronic-structure and quantum-transport calculations. It involved non-equilibrium Green’s functions, electronic structure, and semiconductor devices. At first I assumed it would be roughly like the biology project: look up the concepts I didn’t know, and once I’d looked up enough, I’d understand. But I soon found it wasn’t like that at all. In biology the difficulty is usually the sheer number of terms; once you’ve clarified a term, you can set it aside for the time being. In physics, a single term often drags an entire house behind it — push open the door and there’s a staircase, and beneath the staircase a basement. To understand quantum transport, you first have to know Green’s functions; to know Green’s functions, you have to go back and shore up quantum mechanics, differential equations, linear algebra, and complex analysis; chase it further down and you always end up at some of the most basic mathematics and mechanics. After failing several times to understand the frontier papers, I had to admit that looking things up on the fly wasn’t going to work. I had to relearn the foundations of physics from the ground up. And so my graduation trip and my whole summer were swallowed together by Green’s functions.

From that summer through my first year of graduate school, I spent a long time catching up on math and physics. After the semester began, I found I had already taught myself much of the engineering mathematics, the numerical computation course, and a good deal of the computer-vision material in my own program, so I happily skipped those classes and went to sit in on physics courses at the School of Science. Tianjin University’s science building is built like the Roman Colosseum, and I got lost every time I went in. There, alongside physics undergraduates, I sat through theoretical mechanics and quantum mechanics. I still remember the engineering-math professor saying in the first class that if you felt you already knew the material you didn’t have to come, and that the participation grade would in the end be computed from the exam score. I thought this arrangement was excellent at the time, only to discover later that knowing the material and knowing how to take the exam are two different things. Because I hadn’t done the problem sets, many calculations were unfamiliar; I derived slowly during the exam, and while others had already handed in their papers I still had a portion undone. The result was that what I did do was almost all correct, and what I didn’t do was very reliably all wrong, so my grades naturally didn’t look good. I didn’t mind at the time — I felt that laying a foundation in physics mattered far more than a dozen extra points on a transcript. But now, when applying for funding requires submitting transcripts, I occasionally feel a little regret over that old nonchalance — though not much; if I had to choose again, I’d probably still go sit in on those courses at the School of Science, and just do a couple more problem sets before the exams.

Mathematical analysis, statistical mechanics, and quantum mechanics each struck me differently. Formulas I’d previously just picked up and used in engineering courses suddenly had origins and boundaries; a few basic assumptions and a handful of equations could, astonishingly, connect the chaotic motion of microscopic particles to the macroscopic world. Learning physics often felt like learning the history of physics: a clean single line of formula in a book is really what many people left behind after taking a long, roundabout path — as if a long and muddy journey had been folded, in the end, into one straight line on the page. To force myself to genuinely clear up the vague spots, I would regularly explain what I’d learned that week to my labmates. But the more I learned, the more I could feel the depth of mathematics and physics. To make something genuinely valuable in this field, you can’t rely only on CS or AI intuition to take a new model and run it once over a batch of scientific data. Without understanding the field itself, it’s hard to know which problems matter and which approximations are reliable — whether a model is solving the core difficulty or just wobbling around the edges, producing something that looks good but is useless.

It was in the quantum-transport collaboration that I first met Professor Hong Guo.

At the time we were more familiar with algorithms and machine learning, and our judgment about semiconductor devices, materials, and real applications was quite limited. Professor Guo, though, could often start from a computational method and quickly point out what device problem it might correspond to, which physical quantities carried experimental meaning and could be used as references and for design. This habit of not being confined to the method itself, of always pressing on to ask how a result connects to reality, is precisely the deepest and most crucial part of this discipline. It made a strong impression on me. Coming from an AI and CS background, I’d seen plenty of research start from a method and then look for an application that could contain it, assembling a story. It’s not that such work has no value, but the one thing AI is never short of is methods — nor the ability to tell new stories about them. Where a method actually belongs often takes more domain experience than pushing it up another few percentage points; it comes from long contact with materials, devices, and industrial problems, from knowing which are problems in theory and which are the ones that will actually get things stuck in experiment.

At that time we tried bringing automatic differentiation into non-equilibrium Green’s function calculations. Traditional simulation is mostly forward: you give a structure and parameters, then compute current or conductance; if the whole computation can be differentiated, you have a chance to optimize materials and devices backward from a target property. This work was later published in PRB, and it was my first paper to appear formally in a physics journal.

What truly renewed my resolve to treat AI4S as a long-term path was reading “The New Dawn of Applied Mathematics.”

That day I’d just finished a bowl of ramen near Yingkou Road in Tianjin. Outside it was overcast, it had just rained, the ground was a bit muddy, and the pedestrians and traffic were noisy. I stood at the roadside and slowly read the article on my phone. It talked about high-dimensional problems, scientific computing, machine learning, and applied mathematics. As it happened, the past year’s work had led me to step through many of these problems myself; I had already seen that AI could indeed change certain computations, but I didn’t yet know how to make sense of these scattered changes. Every case in the article seemed to press exactly on some problem I had just touched; and yet it stood higher, describing from a broader vantage the future that might emerge once AI was combined with theoretical computation. The mathematics, physics, algorithms, and numerical computation that had been scattered separately in my hands were suddenly placed on the same map. It was then that I suddenly grasped what it really means for a computation to go from ten hours down to one minute.

A simulation that takes ten hours can usually only be run occasionally, to explain and to extrapolate; if it can be done in a minute, it can be put into a loop, tuning parameters and recomputing over and over, then optimizing further based on the results. Once quantitative change accumulates to this point, the nature of the tool changes. Simulation is no longer only about observing and explaining the world; it may begin to take part in designing it. Looking back now, this idea was fairly crude — I imagined the road between theory, experiment, and manufacturing as far too smooth, and the defects, fluctuations, and contingencies of real processes were almost never drawn onto that map. But the broad direction, I still believe to this day, was not wrong.

During graduate school I later got the chance to intern at the AI for Science Institute, staying from my second year through around graduation — more than two years in all. The AISI back then had a strong magnetic field: people from completely different fields gathered together, each speaking the dialect of their own discipline, yet all genuinely believing that AI and scientific computing could make the previously impossible feasible. I watched machine-learning potentials go from a research tool to infrastructure for large-scale simulation, and watched machine-learning-driven high-throughput screening predict molecules and materials that hadn’t been found before, entering materials screening, energy devices, and real manufacturing tests. The real process seemed to happen quite naturally; in practice, what we discussed with one another were all the trivial dailies we ran into — data formats, failed jobs, models to be retrained, and programs that ran perfectly well on one system and inexplicably broke on another. These trivial dailies broke the final breakthroughs into a little bit of forward motion each day, so that when the results finally appeared, they weren’t all that surprising. Only looking back did I find that something quite remarkable had, in fact, grown out of that unromantic work.

During that time I also put a large part of my energy into something very important to me: developing DeePTB together with Qiangqiang Gu. We wanted to use neural networks to learn the relationship between a material’s local environment and its tight-binding Hamiltonian. First-principles calculations are accurate but expensive; traditional empirical models are fast but often struggle to adapt to complex materials. So the field has roughly two paths: one keeps accelerating first principles, the other improves the accuracy and applicability of empirical models. From our earlier experience, we gradually realized that in the semiconductor and device industry, what R&D people need isn’t just a model that is as accurate as possible from first principles — it also has to be fast enough, able to handle large systems, and able to be continuously calibrated against experimental data and specific processes. If a model lacks the freedom to be tuned toward particular experimental and manufacturing conditions, it’s hard to keep aligning theory and reality back and forth. So we hoped to build a new kind of model in between, one that learns from high-accuracy data while retaining a degree of physical structure and calibration ability — and that is DeePTB. Later the model gradually became able to handle more complex systems, and the computational scale extended to the order of a million atoms and ten million orbitals. When we first reached this point, I once felt that the problem of large-scale electronic structure had largely been solved, as if the bridge were almost across to the far bank. Soon I found it wasn’t — the bridge had probably only reached the middle of the river. Computing the Hamiltonian doesn’t mean the properties an experimentalist cares about are already in hand. Electrical, optical, and transport properties still have to be computed; temperature, defects, interfaces, and the device environment still have to be considered. Between the ideal material in theory and the thing actually made on the process line lies an entire set of different problems and languages. For someone doing electronic structure, a band diagram is often already the result; for a device engineer, what he cares about might be current, noise, stability, and the process window. I gradually realized that the truly difficult part of this path is connecting the languages of these different worlds — model, property, experiment, device, and process — so they can calibrate one another and then enter real design and optimization.

It was with this problem that I came to McGill and became Professor Hong Guo’s student. I chose to come here because I felt I was still missing two things.

The first is the judgment that comes only from going deep into a concrete field. What experiment and process care about, which quantities can be measured, where devices fail — these all require staying inside the problem for a long time, corrected again and again by real systems. Professor Guo’s experience in semiconductor transport, quantum devices, and real applications is exactly what I had most lacked before. In the past I was more inclined to start from “what this algorithm can do” and look for applications; only after coming here did I slowly begin to learn to design computational methods backward from “what the real problem needs.”

The second is the ability to keep turning large-scale electronic-structure models into observable properties. Machine learning can already build ever-larger material models, but if you can’t extract transport, optical, and other experimentally visible properties from them, large-scale models will still struggle to genuinely enter device design and process problems.

So in my doctoral stage, on one hand I keep studying strong correlation, many-body physics, quantum devices, and condensed-matter materials, trying to advance my understanding of the field’s details through more concrete research; on the other hand I develop the software and numerical methods needed for large-scale property calculations. This kind of work spans physical theory, numerical algorithms, high-performance computing, and software engineering. Some days I spend the whole day checking the applicability conditions of a formula, the next it becomes debugging MPI communication, and the third the program is several times slower simply because the memory layout is off. After doing it long enough, the lines between these areas aren’t so clear anymore; they just jointly decide, from different positions, whether something can actually be computed. Fortunately this work happens to connect with my past: a computer-science background means I’m not too afraid of software and performance problems, physics training lets me understand where formulas come from, and experience with machine-learning electronic structure lets me know how these computations should ultimately connect to AI models. Recently, in a study on moiré materials, an AI agent genuinely took part in my process of scientific discovery. From the complex structural changes, it suggested trying a projection onto a few phonon modes. Following this thread, we found that a reconstruction of tens of thousands of atoms, which had looked extremely complex, could in fact be described mostly by just a few modes. AI was involved almost throughout — helping write code, analyze results, and search a large body of literature for possible mechanistic explanations; some of the key clues even came mainly from its combing through material far beyond what one person could read. This made me more certain that AI will take part in scientific discovery ever more deeply, helping researchers propose and test hypotheses and finding connections in material far beyond any individual’s reading capacity. But after actually working with an agent for a long time, I find I don’t in fact look forward to a fully human-free automatic science system. A more reliable, and to me more appealing, vision is to let humans and machines continually cross-check each other in research, letting AI accelerate every link of R&D while ultimately producing discoveries and theories that can still be understood by people. Because a new phenomenon being found doesn’t mean the research is over; we still have to know why it appears, how it relates to existing theory, what might be seen in experiment, and what further questions can be asked from it. Only after these processes does a discovery become knowledge, folded into humanity’s existing understanding, and then become the starting point for the next discovery.

Looking back, the things I’ve done seem quite scattered. Biological sequences, information theory, neural networks, quantum transport, electronic structure, high-performance software, materials, and devices — put on a résumé, they’d need several separate columns and look rather untidy. Every time I apply for funding or an award and have to explain my direction in a very short space, I find it a headache. Even this essay took this many words to barely trace the path once. But in my heart, what recurs across these experiences is really the same impulse: wanting to know. It’s a very concrete, even somewhat annoying, curiosity. What I concretely do now is build the computational infrastructure connecting AI electronic-structure models, large-scale physical-property calculations, and real device problems. It may take the form of a model, a set of algorithms, or a piece of software, and in a few years it may take some other form. But what I want to do over the long run is broader than any single one of these projects: to let machines carry the ever-growing complexity of science, so that people still have room to understand and to judge, and can still enjoy the pleasure of exploration itself. This is, of course, putting it a bit grandly. The actual daily work is often just changing code, hunting bugs, or arguing back and forth with the agent over some seemingly tiny problem. If a grand goal can’t come down into these trivial things, it’s nothing but pretty words.

Back to the original question: will people be pushed out of the work? For much of the day-to-day execution, the answer is probably yes. Nor do I want to cram people back into a process machines could complete, just to prove that people still matter. If a machine can do something faster and better, there’s no reason to refuse it merely to preserve an old way of working. But at least in the sense I’m willing to call science, I don’t think people will disappear. Science isn’t only about producing correct answers; it’s the process by which people turn the anomalies in phenomena into questions, turn results into explanations, and put those explanations back into a shared body of knowledge, using it to keep discovering. A result that isn’t understood, can’t be folded into existing theory, and offers no idea of how it should be verified is at most a temporarily useful answer, and can hardly become knowledge that keeps growing. Even if every link of R&D can one day be automated, the appraisal, understanding, and adoption of the results will still require human judgment and practice. Because value doesn’t float within the answer itself; it always has to land on some subject — changing a person’s understanding, choices, or life — to truly become value.