A golden age of science is coming

extract

It’s almost hard to remember, but just a few years ago, a lot of very smart people thought that science was in terminal decline. Respected researchers were writing papers with titles like “Are Ideas Getting Harder to Find?”, “Papers and patents are becoming less disruptive over time”, and “Slowed canonical progress in large fields of science”. In 1996, science writer John Horgan even published a book called The End Of Science: Facing The Limits Of Knowledge In The Twilight Of The Scientific Age. The fundamental idea was that the Universe only has so many insights to discover, and that we’ve picked the low-hanging fruit. It was the idea that Joseph-Louis Lagrange expressed when he said that “we cannot find more than once a system of the world to establish.” As the big ideas and the simple ideas get mined out, we’re forced to throw ever more brains at the mysteries of nature just to keep scientific progress going at a fraction of its old pace: It looked like science was hitting a wall. Then along came AI. It’s far too early to tell how much AI will eventually speed up the growth of total factor productivity. But already we can see signs of a speedup in science. Anthropic Founder Jack Clark assembled some of the hopeful data points in a recent lecture: Anthropic is accelerating its investment into drug discovery, and is already announcing some impressive preliminary results: Claude also appears to be enabling rapid progress toward a treatment for food allergies. OpenAI, meanwhile, announced a possible breakthrough treatment for neurodegenerative disease. Meanwhile, Google Cloud Chief Scientist Pushmeet Kohli has a list of recent scientific advances that have been enabled by its new Gemini model, including breakthroughs in nuclear fusion reactor design, rare disease research, and quantum computing: But as cool as these stories sound, they’re just very minor, preliminary efforts. There are a number of reasons to think that AI is going to propel us into a golden age of scientific discovery. That age might not last forever — AI might still run into diminishing returns eventually. But the next few decades are probably going to be an absolutely amazing time for human science. (Note: This post is written under the assumption that human civilization survives. A golden age of science could also easily turn into a golden age of bioterrorism, in which case all bets are off.) The number of “effective researchers” is no longer limited The most important input into science is scientists themselves. Until this year, that meant human brains — talented, highly trained human brains. The number of trained human brains could be increased by sending more people to grad school, but only to a point — the smartest people tend to be sent into the hardest research fields first, and if you 10x the number of researchers, the average quality will go down. This constraint was already starting to bite, as Ekerdt and Wu argued in 2024: We argue that a substantial portion of the downward historical trend in U.S. research productivity is a composition effect resulting from self-selection in researcher ability and the dramatic expansion of the research sector…[W]e find that this composition effect has lowered average researcher ability by between 37% and 53%[.] And of course, low fertility was making the problem worse, because fewer children meant fewer people who were able to become future scientists. AI just blows up this constraint completely. If an AI agent can function as an autonomous researcher, there’s no constraint on the number of researchers in the world — you can just build more of them, by building more computer chips and data centers that enable the running of more AI agents. Whether or not massive hordes of agents can collaborate to solve individual problems much faster, increasing the number of agents allows them to tackle many different research questions in parallel. What’s more, these AI “researchers” can be trained incredibly quickly compared to human beings. A human scientist takes decades to train, and more years to reach the frontier of their research field by diligent reading and conversation. This is the “burden of knowledge”, and it’s commonly believed to be one reason ideas have been getting harder to find — by the time human scientists can get up to speed, they’re often so old that their ideas are no longer novel. AI just annihilates this constraint — it can assimilate the entire literature and reach the frontier of research extremely quickly, and it can also tell human scientists where the frontier lies without forcing them to discover that for themselves. It can also pivot to new research directions more easily than humans can. And there’s also a self-reinforcing economic cycle here. Because scientific advances enable higher GDP, they enable more resources to be invested into the creation of yet more electronic “scientists”. Now that doesn’t mean the amount of scientific research goes to infinity — there may still be some essential bottleneck task that human researchers are still needed to perform. There are some indications that AI’s progress on open-ended conceptual research tasks is slower than in other dimensions, and it may be that humans are still needed to provide some of this long-term conceptual thinking. I’d bet against this being true for much longer, but it might be true for a long time. Even in this most “pessimistic” case, though, the amount of effective capital that scientists can use to leverage their work just went up by a lot. AI eliminates the theory bottleneck Science proceeds by an iterative process of theory and experiment. The theory part of that cycle is typically short, but non-trivial — it took a couple of years after the blackbody radiation spectrum was observed for Max Planck to come up with the quantum hypothesis to explain it, and decades longer before quantum theory as we now know it was created. It took about a year for Watson and Crick to figure out the shape of DNA after the requisite data was gathered. AI will basically eliminate the theory bottleneck entirely — or at least shorten the turnaround time between experiment and theory by an order of magnitude. AI is an absolute beast at math. Recently, an OpenAI model solved some of the most important open problems in computer science, math, and theoretical physics, and no one even batted an eyebrow because that sort of triumph is becoming increasingly commonplace. One of those solutions was a proof that the Standard Model of Particle Physics could be mathematically well-defined — a problem that human theorists had so far been unable to tackle. Not all research relies equally on theory, but understanding novel experimental results almost always involves some important component of either mathematical or spatial reasoning. That part of the task has now been made much, much faster. And on top of that, science is now open to a greater number of humans who aren’t naturally good at theorizing. AI can read the whole literature Much of the work of human scientists involves reading and understanding endless amounts of papers and going to endless conferences and seminars. The purpose of that work is to understand the existing scientific literature. Understanding the literature allows scientists to do the following things: - Come up with new ideas for hypotheses to test - Avoid doing research that someone else has already done - Synthesize approaches from different areas or fields AI can reduce this effort to almost zero. Consider this excerpt from OpenAI’s blog post about their new research into neurodegenerative disease: The team divided the analysis among thousands of AI agents powered by GPT‑Rosalind, an approach we call distributed biological reasoning. The agents examined a whole-brain CRISPR atlas…a dataset showing how gene disruptions affect different types of cells across the mouse brain. Each agent investigated a gene disruption, analyzed cellular responses, searched scientific literature, and proposed explanations for the results. Their reports were combined into a reasoning atlas linking genes, cellular effects, relevant publications, and possible biological mechanisms. The research team estimated that comparable human analysis and literature review, conducted one perturbation at a time, would take months to a year. The agents reproduced findings from human analyses while uncovering additional leads. Their assessments of novelty and scientific value also aligned with judgments from blinded human experts. These findings help prioritize further experiments needed to test the proposed biological mechanisms. Some of these insights point to new leads for understanding brain disease. [emphasis mine] This is one case in which parallel work by a large number of agents really does make a huge difference: Not only does instantaneous literature review vastly speed up the research process, but it allows AI to make novel conceptual leaps from combining approaches from wildly different fields of research. AI regularly does this in math research, combining results, approaches, and theorems from far-flung places that no human mathematician would have thought to combine. It can do it in science, too. Much of scientific innovation, historically, has been “recombinant” innovation — scientists applying insights from one field to another. With AI, essentially all research is now cross-disciplinary — AI can read not just the literature on a particular subject, but the literature on a million other subjects that might have produced useful techniques or insights for the problem at hand. Superscience So far, all of the factors I’ve listed are ways that AI can speed up science as it’s already done. I’ve basically described how AI can send gigantic hordes of superhumanly fast, superhumanly well-read, superhumanly mathematically gifted researchers to re-accelerate scientific discovery. But it’s also possible that AI will enable types of discovery — and ways of understanding the natural world — that no network of humans, no matter how large or how brilliant it was or how long it had to work, would ever be able to discover. In a post back in July, I wrote about how that might work: What if there are lots of…aspects of the Universe that [are] too complex to understand in terms of simple laws, but not so complex that they just dissolve into unknowable chaos? It’s possible that we can reliably control these complex phenomena with AI, even if we never reduce them to the kind of principles that we can teach a grad student in a textbook… Another way of saying this is that there may be laws of the universe that humans can’t understand but AI can. I call these “cloud laws” — causal regularities that can be exploited by technology, but which are too diffuse and complex for an individual human being to either intuit or communicate… Perhaps social sciences like economics, sociology, and political science obey similarly complex regularities, and AI can help us find them. Perhaps there are physical processes — plasma, or topological materials, or aerial turbulence, etc. — that obey cloud laws instead of chaos?… [H]uman collaboration is bottlenecked — it’s limited by our ability to intuit patterns at an individual level, and to communicate these patterns from one individual to another. AI, being a computer, just doesn’t have this sort of limitation; it can work with vast, diffuse patterns without having to break them into pieces or simplify them in order to compress them into the tiny pipelines of person-to-person explanation… If much of the Universe turns out to be ruled by cloud laws, we could be on the precipice of a scientific renaissance. In order to discover cloud laws, AI will need huge amounts of data about the natural world. AI figured out human speech by assimilating essentially all of the text humans have ever written. If it’s going to figure out hidden patterns in biology, neuroscience, materials science, etc., it’s going to need similarly massive amounts of data from physical instruments. That data will be far, far more than humans — individually or collectively — would be able to parse, label, characterize, or model. But AI was built to do this. If scientists simply build huge arrays of sensors to measure complex phenomena like the brain, the cell, the economy, etc., they may be able to feed this flood of data to AI directly and bypass human cognitive intermediation entirely. AI would then learn about the world directly, and it might learn things that humans never could. This “superscience”, if you want to call it that, would actually go beyond the kind of science humans have done heretofore. Many of the technologies it yielded would be beyond human comprehension — indistinguishable from magic. But to the AIs, they would make perfect scientific sense. In any case, it seems clear that the scientific slowdown that threatened human progress will now be delayed — at least for a while. Unless humans have somehow found most of the actual exploitable regularities in our Universe, we’re about to enter a new golden age of science.