Demystifying how AI will accelerate scientific progress

What can a "country of geniuses in a data center" actually do?

We often hear that AI will be able to “cure cancer”, “solve climate change”, and supercharge the rate of scientific and technological innovation in the coming years. It seems intuitive to think that something “more intelligent” than us would obviously be able to “do more science”, and just “invent more things” than humans can.

But my recent exploration into how a country of geniuses in a data center could actually make “progress” has given me a curious perspective which I want to start outlining here.

This article is a collection of questions, thought experiments and partial answers surrounding this topic. It’s less “formal” than previous articles, but I hope it’s still valuable and reduces some of the vague-ness around how AI will impact scientific discovery in the coming years.

---

When people talk about how AI will impact scientific progress, their thinking tends to fall into (or sometimes between) one of these two camps:

1. The pattern-finder view: AI will be able to detect patterns and make predictions by analyzing huge datasets, at a scale humans can’t match

2. The “oracle” view: Superintelligent AI will basically be able to solve any scientific or technological challenge, given enough time and compute

Neither of these views seems to fully capture the reality of what we will face.

But before we can explore how AI can “supercharge progress”, let’s talk about all the different ways we actually make “progress”.

Ways that “progress” happens

At a very high level, there are a bunch of different ways that humanity progresses scientifically; it’s not a straightforward or predictable process. For example:

• Accidents - Alexander Fleming simply noticed some mold growing on a Petri dish, and this led to the discovery of penicillin

• Constant trial and error - Edison tried thousands of materials before he found a suitable one for an incandescent, electric light

• Unpredictable, creative breakthroughs - Einstein’s relativity was a paradigm shift for physics, but it was unlike anything that came before

• Better tools revealing new things – The invention of the microscope opened up entire fields like microbiology

• Connecting two separate disciplines – The field of bioinformatics emerged when biology and computer science started overlapping

And many others!

Basically, progress doesn’t follow one path, and it’s not restricted to one type of process. But these general categories can help us make sense of how breakthroughs tend to happen, and how a powerful new technology can impact things.

So how can AI help accelerate this process of “progress”? Where does it fit in?

Recently, AI systems have started excelling at digesting giant amounts of data and finding patterns within them. They use their knowledge of these patterns to make predictions or simulate outcomes under new conditions, at a scale and speed that humans can’t match.

For example, researchers can train AI on massive databases of known materials and their properties, and then use this to discover new materials, where we only optimize for the things we are looking for. This can save researchers months or years in the lab, no longer needing to manually test thousands of new materials for things like batteries, superconductors, and solar panels.

Due to these capabilities, AI is providing a big boost to the “constant trial-and-error” category of scientific progress, specifically in areas that have a clear objective to solve for, and lots of data.

This leads to the first view of how AI can impact science in the near future.

The pattern-finder view of AI

According to this view, there will be many such data-intensive problems that we can point AI towards, and it will simply use its enormous computing power to generate a ton of useful predictions and insights that we need within that area.

We saw how this can accelerate progress by years in the case of new material discovery, but it also applies to many other fields, especially biology - AlphaFold was a breakthrough that compressed decades of lab work predicting protein structures, into a matter of months.

It’s worth reiterating how big of a deal this application of AI is. There are so many problems that require a slow, methodical exploration of countless options, because they have an insanely large “search space”. Having AI come in and instantly solve all these problems, saving us decades, would be a significant multiplier for many areas of scientific progress.

But by viewing AI only through this lens, we’re implicitly assuming that while it may compress lots of research timelines, it won’t be the one actually “steering the research ship”. It will be doing the optimizing, but humans will still the ones deciding what to optimize for.

This view assumes that all the other ways that scientific progress occurs, specifically the “creative leaps” that occur in science, would still be human territory. Coming up with entirely new hypotheses, reframing the problem itself, and moving the fundamental frontier of our understanding of reality forward — that’s not something AI would take on. According to this, Einstein’s theory of relativity could never come from AI - we would still need human thinkers to take the conceptual leaps that no amount of data can brute-force.

But I think this is incorrect.

I think that AI can actually engage in the creative core of science itself — generating new hypotheses, and asking the kinds of questions that push science forward in non-obvious ways.

I think AI can do more to advance scientific progress than just find patterns and extrapolate from large datasets.

More than just a pattern-finder

To understand how AI could start taking on the full scope of what a scientist does, I want to demystify what the process of scientific discovery actually looks like, specifically in the context of physics.

When we think of Einstein and similar revolutionary scientists, their process of discovery seems rather magical to us. It seems like they created their new theories out of thin air, through some magical, deep quality of human brains to propose totally new ideas. But, at the risk of de-valuing the ingenuity of these breakthroughs, this type of “hypothesis generation” seems replicable to me.

The process of generating new hypotheses

In a nutshell, what really goes on when these new hypotheses are proposed is scientists are playing around with the edges of the parameters that exist within their fields. They are using the latest experimental results, combining findings from different disciplines, and asking new questions that open up unexplored paths.

When Einstein came up with special relativity, he wasn’t making stuff up from thin air. He took the results of the speed of light experiments of his time, and recent mathematical research done by another physicist and extrapolated the results - the logic seemed to imply that space and time were not fixed. This seemed strange to humans, but if you purely looked at the physical and mathematical equations, this radical new idea seemed to make sense. It was a revolutionary insight that required intellectual courage and clarity — yet it was a product of existing human knowledge and research, it just needed someone to make the relevant connections. I don’t wish to minimize the monumental achievement that this was, but only to highlight that the mechanisms which produced it don’t seem to be limited to some magical quality that no other form of intelligence can possess.

This process of being creative within the parameters of your field to come up with new and novel hypotheses definitely seems like something that an AI system of the future could do. Specifically, is it unreasonable to suggest that AI could do the following?

• Absorb the entire scientific literature of a field and spot contradictions

• Run billions of thought-experiment simulations to see which assumptions break and which hold

• Search through vast mathematical spaces for models that fit observed data

• Propose new hypotheses that we wouldn’t have thought to test

As a side note, if you don’t think AI can be truly creative, go ask ChatGPT to find you 25 creative ways to use any common household item you own. How many of those could you have come up with? Could you even come up with 25 uses?

AI that is running 24/7 coming up with billions of new, creative hypotheses on the edge of our current knowledge seems like it would produce at least some breakthrough ideas?

Right?

How intelligent do you have to be to come up with a “breakthrough”?

If it’s true that a large proportion of major scientific breakthroughs come from a small proportion of researchers (say 5-10%) that are unusually creative, intelligent, or resourceful, then it implies that there it matters how skilled you are as a researcher - it’s not just about luck. These researchers often are working in niche edges of their subfields — they often are basing their intuitions on sparse data, and a deep level of experience in their area.

The way our AI models are trained today relies heavily on lots of data. They perform pretty poorly in areas where they don’t have much of it. This begs the question, will AI really be able to reach the skill level of a top 5% researcher within their field, given how sparse the data is that they often operate within?

And if AI cannot reach that level, does sheer brute force and running 24/7/365 overcome that limitation anyway? Does the breadth and speed of AI overcome the depth and intelligence of a top 5% human researcher?

This question becomes pertinent because if we start using AI in this manner, it seems like we will soon have to make decisions for what hypotheses to dedicate resources testing.

But before we get there, let’s reflect on what this mechanism of AI exploring billions of scientific ideas 24/7/365 could actually mean:

Will AI be an oracle that solves everything for us?

The second camp of people view AI as a superintelligent entity that will basically solve every meaningful scientific and technological challenge that we face. Want a dyson sphere, or a cure for aging? Just give it some time and compute.

To be fair, superintelligence does imply more intelligence than humans - and it seems reasonable to think that something more intelligent than us in every dimension should be able to advance through the technology tree much quicker than us.

But is this actually true?

First, let’s define what kinds of superintelligent systems we are focusing on - I’m mainly looking at the next 2-10 years, where it seems like we will have a “country of geniuses in a data center”, as Dario Amodei states.

The obvious thing to point out about this is that superintelligence will be mainly limited to a digital form. Despite it being “agentic”, it will need to rely on humans to perform tasks in the physical world - and even for areas that it could rely on robotic labs, the resource constraints of the physical world are still going to restrict its ability to conduct experiments as quickly as it is able to think.

So, even though this AI will be able to create billions of copies of itself, it won’t be able to conduct billions of experiments in physics, chemistry, or biology laboratories.

So, what would an army of geniuses in a data center be able to produce in terms of scientific progress by purely thinking (and conducting experiments very slowly)?

Imagine you took all the smartest scientists in the world and put them in a room, but the only thing they could do is just think, write, and use a computer, 24/7/365. Will we end up with dyson spheres and intergalactic space travel within a few years?

Spoiler alert: no. Many physical sciences will NOT see major breakthroughs as a result of just more thinking time. This is because physical sciences rely very heavily on physical experimentation to progress.

Fundamental physics

Every major breakthrough in physics has required us to do experiments before it has been confirmed. Even our biggest breakthroughs, such as Einstein’s relativity and Planck’s quantum mechanics were just theories until they were confirmed through experimentation. In fact, most scientists weren’t even sure they were correct until they were proven so through testing.

So the only thing these billions of AI scientists will be able to do for fundamental physics is spit out a bunch of theories about what is actually going on. Many of them will also simply spit out a bunch of new experiments that we should try, because they could reveal something new about the nature of reality.

But they will not be able to simply think of a new law of physics. Unless there is a new law of physics already hiding in the data we have collected, and we are just waiting to discover it (this seems unlikely considering the recent slowdown in fundamental breakthroughs), any new hypothesis that AI comes up with will have to be tested using new experiments that we conduct.

Biology and drug development

Biology does seem like a field where just thinking harder, or in this case, just analyzing more data, can actually go a long way. There’s probably still a ton of low-hanging fruit hiding in existing datasets. For example:

• Green fluorescent protein (GFP) had been known since the 1960s — scientists had already seen it glow in jellyfish. But it wasn’t until the 1990s that someone realized you could use it to make living cells light up, which completely changed how we study biology.

• The CRISPR mechanism had been sitting in bacteria for decades, and technically could’ve been discovered back in the ‘80s. But it took us until the 2000s to realize what it could actually be used for.

A superintelligent AI could be extremely good at spotting those kinds of overlooked patterns. It could read every biology paper ever written, flag unexpected gene-disease links, or identify drug interactions no human noticed. It could design proteins, screen compounds virtually, and figure out what experiments are worth running next — which would be a huge time-saver.

But again, it hits a bottleneck: all of these ideas still have to be tested in real cells, animals, and eventually humans. And that part still takes time, money, and regulatory sign-off. So while AI could supercharge the idea-generation part of biology, the pace of real-world progress would still be bound by how fast we can actually run the experiments.

Mathematics and computer science

However, there are some fields where you don’t need to touch the physical world at all.

Mathematics is basically built for this. AI could search through mathematical spaces way faster than we can, generate new theorems, and even prove them on its own. We’re already seeing early versions of this with models that can fill in steps of formal proofs — a smarter system could take that way further.

Same with computer science. An AI could come up with faster algorithms, better compression methods, or new kinds of neural networks — and immediately test them, because all of it happens in software. It can run millions of experiments per day with no lab, no materials, no waiting. That’s also why AI research itself is moving so quickly, and why AI’s will soon become self-improving, leading to an intelligence explosion.

We will basically see a foom in fields like this, that do not require real-world testing.

What could be the consequences of this? How much could an AI that is superhuman in these particular fields impact the rest of the world?

---

So essentially, we face the prospect of systems that will soon be vastly superhuman in certain digital and logic-based fields (and in their own intelligence, due to their ability to conduct AI research on themselves), but would still not be much more advanced than humans when it comes to the physical fields such as physics, and in some cases, biology and chemistry.

This seems interesting and non-intuitive to me.

But…there’s a cheat code for the physical sciences

There is a cheat code that could allow us to treat the physical sciences the same way as mathematics and computer science research (i.e. they could be done at the speed of AI “thinking” without relying on physical experiments).

Simulations.

If we can simulate physical reality with enough accuracy, we can turn parts of biology, chemistry, and physics into digital playgrounds too.

We’re already seeing early versions of this in the real world:

• In biology, AlphaFold 3 goes a step further than just protein structure prediction: it models how proteins, DNA, RNA, and small molecules interact. This starts turning drug discovery into a mostly digital process, where an AI could screen billions of potential treatments before a single physical test is run.

• In high-energy physics and chemistry, exascale supercomputers like Frontier are starting to simulate parts of the universe, running massive molecular dynamics models, and pushing into territory that previously required billion-dollar particle accelerators or decades of physical experimentation.

• And in biology again, researchers are building early digital twins of cells and organs — virtual systems detailed enough to test the impact of drugs or genetic edits without ever entering a lab.

If these kinds of simulations keep improving, the physical sciences start looking a lot more like computer science. The thinking and testing collapse into one loop. In that world, AI doesn’t just suggest a new theory — it simulates its consequences, compares it to alternatives, refines it, and only sends the very best candidates into the slow, expensive physical pipeline.

So yes, there are still hard limits in areas that depend on physical experimentation. But that boundary is shrinking. More and more of science is becoming digitizable. And as that happens, the advantage of having a “country of geniuses in a data center” grows exponentially — not just because they can think faster, but because they’ll increasingly be able to experiment faster too.

In that world, maybe you could see a cure for aging developed after just a few weeks of “thinking”.

For everything else, how will we rank what theories to test?

For stuff that we can’t simulate, it seems like this “country of geniuses in a data center” will be able to spit out tons of possible things for us to test or design experiments for.

Imagine turning on a set of a million superintelligent physics agents to go and “explore” for a few weeks, and then they get back to us with 3000 possible theories that could be true. To test each theory, we would require hundreds of new experimental setups, some of them costing millions of dollars. In the meantime, our human scientists have also been researching in the labs, and they have a few theories of their own they would like to test.

How will we choose which ones to dedicate resources to?

There are already some interesting ideas underway that decide how to “rank” different scientific hypotheses among each other - Google’s recently released co-scientist system uses an internal “tournament” to assign an Elo rating to each hypothesis it generates, and then surfaces the hypotheses with the highest Elo to the humans.

How will we rank the AI’s ideas vs. the humans’ ideas? It seems like ultimately we will have to somehow figure out methods to choose the “best” hypotheses to test, which will probably depend on various factors such as:

1. Cost of experiments

2. How easily AI can “explain” it to us

3. Some political influences

4. How relevant it is to real-world applications

5. Other factors…

Taking stock

The methods of AI impacting scientific progress that I’ve outlined above are only a few out of the many ways this will play out over the coming years.

Things I haven’t discussed include:

• How AI will be able to speed up various engineering tasks, leading to better tools — and how better tools can accelerate discovery in many ways

• How AI could propose new experimental setups, not just hypotheses — setups that humans wouldn’t have thought of due to practical or conceptual constraints

• The possibility that AI could help unify previously disconnected domains — e.g., using techniques from fluid dynamics to solve problems in neuroscience, or vice versa

It does seem interesting to me that we are likely to see highly asymmetric progress in science - i.e. some disciplines will skyrocket in terms of progress, whereas some others will continue to be severely limited due to real-world and physical constraints.

What would the impact be of this asymmetric progress be? Could some random niche process being sped up with AI lead to a 100x impact on overall progress? Whatever that even means? Is it more likely that we’re still going to be pretty much where we are today, despite AI becoming superhuman in its intelligence?

Will its ability to generate new hypotheses lead to a flood of new insights, or will we still be very bottlenecked by our capacity to evaluate and act on them? Will it still be dumber than our best researchers for a while to come?

This space is still unfolding, and I’m curious to see how things play out over the coming months and years.