Establishment of the
Research Centre on
AI for Science (RCAI4S)

I am a computer scientist. Why did I stay in a university, and not go to industry? Because the problems of science interest me far more than the problems of industry — and I am lucky to work in the field that is building the most powerful tool of our time. I have always wanted to use that tool with scientists. Let me show you three of them.

In my PhD I worked with animal scientists, Roland Kays and Meg Crofoot, on animal traces. Meg told me that baboons are democratic.

The collars recorded each baboon's location once every second. That is an enormous amount of data. I built efficient computational tools to find the leader–follower patterns in it — here, one baboon in red, one in blue, and you can see who is following whom.

At Penn State I worked with Sue Brantley and Tao Wen on water-quality data. Hydrofracking had become the main source of natural gas in the US, and we asked whether it was affecting our water.

We had about eleven thousand groundwater samples from one county. I developed a sliding-window technique: move a small window across the map, and in each window test whether methane goes up closer to a geological fault. Red is where it does. Most of the water was fine — but the method found the few places where it might not be.

In recent years I have worked with Zhonghua Zheng on urban climate. Most of us live in cities, yet in a global climate model cities are only about eight percent of the grid cells — and each cell is about a hundred kilometres across. A whole city can be one cell. Urban climate is underrepresented.

Together we built UCformer: we put the physics inside the neural network. It predicts a city's temperature and humidity from the atmosphere above it and its surface — walls, roofs, roads — and the network is built to follow how radiation and heat move between them. The result is better than either side alone: better than the physics model the field uses, and better than a neural network without the physics.

And there are many more — Corina Graif and Stephen Matthews in sociology and criminology, Vikash Gayah in transportation engineering, Thomas Lauvaux in climate science. Many faces, many interesting stories. Each of these collaborations lit one light — nine lights. And all around them, the rest of science stays dark.

Look at what each of these collaborations had in common. They knew what was worth asking — the vast unknown of their science, and which part of it mattered. I knew how to answer it at scale: I could write the program. They brought the judgement; I brought the code.

I wish I had had more time to go deeper with each of them — to make each of these lights brighter. And I wish I could work with many more — to light up the dark ones around them. But people like me are few, and our time is limited. This does not scale.

Let me show you what this looks like in practice. On 22 September, at half past one, Jiong Zhao, his students and I sat around one laptop and put an agent to work on their data. We were laughing half the time. This is the kind of research I enjoy.

Jiong showed me images like this one. For me it was like an ultrasound — the first time I was looking at atoms. The file said "STEM HAADF". I had no idea what that meant. I suspect most of this room doesn't either. Neither, as far as we told it, did the agent.

In my old way of working, a result like this took at least a year. The agent took an hour. At 13:32 I gave it the data and one line of task. Fifty-one minutes later the model was trained, evaluated, and written up. By 14:51 it had run on the real micrographs and published a viewer we could all look at.

One day in heaven, a year on earth. This is not a piece of me any more. It is a much faster Jessie.

And it is not only faster. This is the report it wrote — it is scrolling on the left so you can see how complete it is. The method, an ablation, a paired significance test, the caveats, the code to reproduce every number.

I could not have written a report this good in a year. So it is not just a faster me. It is a better Jessie.

(Jiong) We never told the agent what the data was — and it read the physics out of the data, at the level of a postdoc. Let me show you.

(→) It read "ADF" in a file name and knew it was looking at an electron-microscope image of atom columns.

(→) It counted only about eleven grey levels in each input patch, and concluded: this is low dose — Poisson counting noise. That surprised me. It is exactly the bias my students built into the simulated data.

(→) The data came from three sources, 20, 25 and 30 in the file names. It guessed they are different thicknesses or doses.

(→) And because the data had the atom positions, it scored its own model the way a microscopist would — by whether it finds the atom columns, not just by PSNR. This is the knowledge I would expect from a postdoc in my group. (Jessie) And I still don't know what dose means.

So it is not only a faster and better Jessie. It carries a piece of Jiong too. An agent is not a programming tool. It is already a cross-disciplinary expert.

(Jiong) On real micrographs it is not there yet. Here, on the left — watch the divider sweep across — the raw frame against the model. Look closely and it smears some atom columns into streaks. On other images it paints a lattice into empty vacuum — and it told us so itself: those images are four times coarser than anything it was trained on.

But every one of these shortfalls is the next experiment, and it proposed them: resample to the right scale, simulate new data at the right dose, retrain. That part an agent can already do — it is our next step. The step after that is real data: the same field imaged at low and high dose, on the microscope. That still needs the scientist. Together, that is a loop — and the loop keeps going.

We find students — from computer science, and from the sciences themselves — they learn alongside me, and then they go out to more scientists and do what we did with Jiong last month — sit down together, put an agent to work on their data, and let them see what becomes possible. The first question we ask is always the same: what is your bottleneck?

This is labour. It is outreach, and more outreach, one lab at a time. One of me becomes ten students; ten students reach a hundred scientists.

Doing this by hand does not scale either — not with me, not with ten students. But after a hundred scientists, our own steps become standard. We will have seen which questions to ask, which workflows work, where agents fail. Whatever is standard can become code, and code can become a system: a general workflow that does what Jessie and Jiong did that afternoon, for scientists we will never meet in person. That system is the Centre's technical output.

So where do the first hundred come from?

Right here. This is PolyU at night. Every faculty is a district, every department a block — and every dot is one of us. The dots are real: one dot for each tenure-track professor listed by the department, about nine hundred and fifty in all. (Hover over a block and it opens out: the department's number, then every name by rank.)

Tonight one dot is lit — Jiong, whom you just met. And the yellow rings are you — the colleagues who signed up for this morning. Not lit yet, but next. Why only one? Because he happened to sit next to me. With agents alone, that is what you get: one light, wherever someone like me happens to be. The agent makes it possible. The Centre is the mechanism that makes it happen — for the other nine hundred and forty-nine. What "lit" really means, I will come back to.

Look at where the science is. Not only in the Faculty of Science. PolyU is mostly engineering, but inside engineering there is a great deal of science — the environment in Civil and Environmental, energy in Building Environment and Energy, materials and chemistry in Fashion and Textiles, the body in Health Technology and Rehabilitation, the eye in Optometry. And of course Life Sciences. (Hover the blocks of the departments in the room — say the department aloud: the session is recorded.) These problems are dark not because they don't matter, but because nobody could afford them: dark is an artifact of cost, not a judgment of worth. And dark is not small — Sue Brantley's Shale Network, Roland Kays' Movebank with ten billion animal locations: whole databases with no agents on them.

Our plan is to go block by block — every faculty, every department. The hundred scientists I talked about are about one in ten of the people on this map. So: which of your problems is still dark? Come and find us. And if you are a student or a postdoc who wants to learn to speak to agents — come and work with us; this is how the outreach gets done.

Here is what "lit" means — and what we commit to over three years. (The lights come on.) Electricity was once a privilege of the few; today it is a utility, in every home. With agents, science stops being a privilege too — every scientist can plug in.

A hundred lit: a hundred problems on this campus, each with data, each taken from question to result the way we did with Jiong. Ten of them brighter: the loop closes — the agent runs simulations, the scientist runs the experiment, the data comes back and the model gets better. Making that loop improve itself is frontier research, and it is the Centre's own. And one, brightest: an open, hard question a field has been stuck on, answered.

If tokens are how a problem gets lit, what are they? The new electricity. Every hour an agent works on a problem, it consumes tokens — and the more tokens a problem consumes, the more work is being done on it.

This is PolyU in three years, lit by tokens. And this (the picture turns) is East Asia at night, seen from space — that bright knot is us, the Pearl River Delta. Economists measure how developed a place is by exactly this light — electricity consumed, not power plants built. Capacity is not activity. Tokens let us draw the same map for science — lab by lab, discipline by discipline: first PolyU, then the ten thousand academics in Hong Kong's eight universities.

But having electricity was never the point.