AI for Science: Three Walls, Eight Hundred Million People, and a Copernican Revolution

A few days ago, I attended a conference where the topic shifted to how AI is changing scientific research. It was supposed to be a closing session, but the conversation wouldn't stop. Several interesting points didn't get fully explored, so I've been thinking them over for the past two days and decided to jot them down.

1. AI Writing Papers Is Already the Norm; the Real Question Is the Paper Itself

The fact that AI writes papers has become so commonplace it barely needs discussion. In a 2025 Nature survey of 5,000 researchers, 57% admitted to using AI to help write over the past two years, and 72% plan to use it in the next two. These are just the numbers researchers were willing to admit.

Someone at the conference argued this isn't actually good for early-career researchers. Writing papers used to be a form of mental training; with models doing the writing, that step gets skipped.

I didn't immediately react when I heard this. It took some thinking to see clearly. The concern is valid, but the perspective is a bit small. The real problem sits one level higher: Is the paper itself, as a medium, still appropriate?

The vast majority of papers today are still PDFs. PDF is a format designed for humans to read: layout, font size, double-column formatting, following the publisher paradigm from decades ago. But what if the first consumer of knowledge in the future isn't a human but an agent? PDF becomes a translation layer—it has to be parsed back into structured text before being handed to a model.

Why not just use Markdown?

Markdown is far more agent-friendly. Clear structure, machine-readable, easily rewritten, embeddable into larger workflows. If humans want to read it, just render the layout—no harm done. The default form should be flipped: serve the agent first, then the human.

No one is pushing this yet, but it will inevitably change within two years.

2. The Real Bottleneck for AI Doing Research Is Three Walls

Producing papers—the output part—is easy. But the core of research has never been the writing step.

The core of research is connecting broad swaths of knowledge and letting new things grow at the intersections. Single-point intelligence is useful but far from sufficient. A friend at the conference put it very clearly: what is blocking AI from doing research isn't model capability; it's three walls.

The First Wall: The Paywall

The vast majority of the scientific knowledge system sits behind paywalls.

Computer science has arXiv, which is an exception; physics and math are mostly on arXiv too. But outside these circles, the picture changes. In pharmacology, inorganic chemistry, chemical engineering, and similar fields, over 90% of papers are behind paywalls—agents simply can't read them. The open-access portion is maintained by a minority of people in a minority of disciplines.

My friend used an apt metaphor: fog of war. Anyone who's played RTS games knows the map is mostly black; you can only see the small patch you've explored. That's the knowledge map AI faces right now. You think it's doing global reasoning, but the territory it can actually see is pathetically small.

What's worse, the obscured areas are precisely where knowledge connections are densest. A new discovery often requires linking several disciplines. If one or two of those disciplines are blacked out, the connection is severed. No matter how smart the model is, it can't compensate for input that doesn't exist.

The Second Wall: The Wall from the Digital World to the Physical World

Two or three years ago, when AI for science was at its peak, materials science was the focal point. I seriously looked into the field during that period and found a classic bottleneck.

The theoretical phase involves combinatorial arrangements of large numbers of molecules and crystal structures, from which promising candidates must be selected. AI can accelerate this step by several orders of magnitude. In 2023, Google DeepMind's GNoME system identified 2.2 million new stable inorganic crystal structures in just 17 days—equivalent to nearly 800 years of accumulated human discovery—of which 380,000 were predicted to be stable enough for engineering.

It sounds almost too good to be true. But getting from theoretical structures to usable products requires passing through wet labs, process development, and production engineering. If the front end accelerates 100× while the back-end pipeline stays put, the entire chain can only move at the speed of its slowest link.

A promising trend in the last year or two is the self-driving lab. The idea is to fully automate and digitize wet labs, letting AI directly schedule experimental equipment and run closed loops. A 2025 proposal from North Carolina State University used dynamic flow experiments to collect data every half second, more than 10× faster than batch experiments.

But this path has only just begun. In most disciplines, wet labs are still manual, with each run taking anywhere from days to weeks. Without a closed physical feedback loop, no matter how fast the AI side is, it can only wait.

The Third Wall: The Perception Wall

This wall is the most easily overlooked, and possibly the deepest.

AI currently performs best with text. Vision is catching up, but how much of the true information density contained in vision—lighting, texture, spatial relationships—can be digitized remains an open question.

Even harder are the modalities beyond text and vision. Touch, smell, and taste are critical data for many disciplines. For food research, smell is a core signal. In biological research, subtle odor changes in chemicals are often used to judge reaction direction. These signals still lack good digital channels today.

Then there's a fuzzier category called intuition. When a human sees an experimental phenomenon, hears a set of data, or smells a particular odor, the first reaction doesn't come from linear reasoning; it emerges directly from some corner of the brain. This intuition is pattern recognition accumulated over decades of training, yet the pattern itself can't be articulated.

As long as this layer exists, AI cannot independently complete scientific research; it must collaborate with humans.

3. Terence Tao Says Intelligence Is Facing a Copernican Revolution

Speaking of collaboration, I was reminded of something Terence Tao has been saying recently. I jotted down a line from his appearance on the Dwarkesh podcast:

"We are experiencing a Copernican revolution at the level of cognition. We used to think human intelligence was the center of the universe; now we see there are very different kinds of intelligence out there, each with its own strengths and weaknesses."

What Copernicus did was remove humanity from the center of the universe. Tao says the same must happen with intelligence: humanity must be removed from the center of intelligence.

We used to think of intelligence as one-dimensional and linear. Humans in the middle; anything above is superhuman, anything below is useless. That's why AGI makes people nervous: if it surpasses me one day, do I lose all value?

Tao says this framing is fundamentally wrong.

Intelligence is multidimensional. Different dimensions can coexist and collaborate rather than substitute for one another. AI is extraordinarily strong in breadth—rapidly scanning massive amounts of information and making cross-disciplinary associations—capabilities humans simply cannot train themselves to achieve. Humans hold irreplaceable depth: intuition, problem selection, judgment of meaning. Put the two together, and they outperform either alone.

This aligns perfectly with the three walls mentioned earlier. The walls AI cannot climb are precisely where humans have the most leverage. Physical embodiment in the world, cross-sensory intuition, the judgment of whether a problem is "worth doing"—these are things AI cannot replace in the short term, and they are exactly what makes humans most valuable in this collaboration.

When Copernicus removed humanity from the center of the universe, he didn't say humans were worthless. He said the universe operates differently from how people imagined. Tao's point is the same. Stepping down from the cognitive center is not a demotion; it is seeing clearly where one actually stands within a larger system.

4. The "Participation Population" of Research May Leap by Two Orders of Magnitude

The most exciting part of the entire discussion was a shift in perspective.

People worry about AI displacing researchers, but the interesting development is exactly the opposite. AI is democratizing research from the hands of a tiny minority to hundreds of millions of people.

According to UNESCO statistics, in 2018 there were roughly 8.8 million full-time researchers worldwide. That sounds like a lot, but against a global population of 8 billion, it's less than one-tenth of one percent. In other words, the absolute frontier of human knowledge has historically been advanced by only one in a thousand people, while the remaining 99.9% occupy more basic, everyday roles.

This wasn't because people didn't want to participate; they couldn't.

A glance at the history of science makes this clear. Many early scientists were amateurs. Newton dabbled in optics on the side; Darwin hitched a ride on a ship to conduct his surveys. The frontier of human knowledge wasn't yet so deep; a smart, curious person with time on their hands could, after a few years of serious effort, genuinely make contributions.

Later, that stopped being possible. Disciplines fragmented into finer and finer subdivisions; the frontier grew ever more distant. A serious paper now requires reading hundreds of prior works; an experiment requires mastering an entire suite of techniques. The barrier grew higher and higher, shutting most people out. The social division of labor funneled them into other positions—basic work, organizational operations, repetitive labor. It couldn't be helped.

The arrival of AI has flattened that barrier significantly.

This doesn't mean everyone can solve Millennium Prize Problems. It means that when you have genuine curiosity about a field and are willing to dig in persistently, the "prerequisite capabilities" that used to take ten years to build might now take only a few months. What remains truly scarce is judgment about which problems matter, curiosity, and perseverance—qualities that never had much to do with formal training in the first place.

I wrote a few days ago about how AI has turned ignorance into an advantage, describing how 23-year-old Liam Price used GPT-5.4 Pro to crack a mathematical conjecture that had stumped Erdős for 60 years. That's an extreme example, but the direction it points to is real. The number of people capable of standing at the frontier of human knowledge will likely jump from the millions to tens of millions, or even hundreds of millions, in the next 5 to 10 years.

A tenfold, hundredfold increase in scale.

My own feeling is that I never dared to think about touching frontier research before. I work in business; I think about how companies bring innovation to society. At that level, every decision of "should we research X" had to be made with extreme caution. Resources are limited; every path requires pouring in large amounts of people and money, then waiting a long time. That was the strategic puzzle to wrestle with every day: how to pour limited resources into the most critical places.

But once AI slams the research barrier down to ground level, directions that were originally too deep, too complex, or too long-term—such as optimizing a particular inference engine, improving the energy efficiency of a certain device, or addressing some neglected niche problem—become accessible. An ordinary person, as long as they genuinely think about it and are willing to dig in, has a chance to make nontrivial contributions.

The next decade will likely see a knowledge explosion. Not the kind of tiny incremental progress, but a supply-side eruption that inevitably follows when participation jumps by several orders of magnitude.

5. Beyond the Bright Side, the Shadow

Of course, we can't look only at the positive side.

Technological progress never directly equals happiness. How many new possibilities productivity gains create, and how much actual well-being they deliver, are separated by a layer called distribution. Distribution is another topic; I won't expand on it today, but we can't pretend it doesn't exist.

The most certain direct impact: repetitive labor will continue to be eliminated. There's no suspense here. When a form of supply becomes extremely cheap and abundant, it ceases to be scarce. Without scarcity, there is no value; without value, one can no longer make a living from it. This is the most basic lesson in economics.

In elevators, you hear people talking about layoffs: weekly reports, PowerPoints, process-running roles are the first to go. News of this is already happening daily.

But on a longer time scale, I'm still more inclined to look at the positive side. The population participating in research growing from one-tenth of one percent to a few percentage points is a shift of such magnitude that any short-term negative shock will be overshadowed by its long-term impact.

Humanity has rarely wielded leverage of this magnitude in history.

Closing Thoughts

Returning to that discussion at the conference.

Some people worry that AI writing papers will deprive junior researchers of mental training. I now think this concern is somewhat small in scale. AI writing papers is merely the outermost layer of this transformation. One layer deeper: in the past, only 8.8 million people were pushing the frontier of human knowledge; in the future, it may be 800 million. What was once the privilege of a small scientific elite will be open to anyone with genuine curiosity.

Inside the door, the three walls still stand: the paywall, the wet-lab wall, and the perception wall. We can see them today, and tomorrow we will still have to climb them one by one.

Terence Tao says we need to step down from the center of intelligence. It sounds humble, but the crucial subtext is: people must re-examine what their most valuable part is within the new intelligence system, and pour their time and attention into exactly that.

The last time tickets like this were issued was centuries ago.