AI and the future of science
If we're going to have Einstein-level AIs doing physics research in a decade, what does that mean for the rest of science?
What follows is a 3,000-word essay written primarily by OpenAI’s o1 pro. The essay arose from a dialogue between myself and o1 pro, about Dwarkesh Patel’s recent podcast interview with physicist and AI researcher Adam Brown. I fed the transcript of the interview into o1 pro, and asked the AI to tell me everything that Adam said about AI. Of particular note is Adam’s prediction that ‘Einstein-level’ AIs will come online within a decade. By this he meant AIs capable of doing new physics research at the level of an Einstein, or other similarly prominent physicist. What implications, I wondered, does such a prediction have for other sciences? That question prompted the essay, which appears below, lightly edited by me.
Introduction
Adam Brown holds a unique vantage point: he spends his days working at the intersection of advanced AI research at Google DeepMind and cutting-edge theoretical physics at Stanford. Historically, physics has been the foundational science, clarifying nature’s fundamental laws and enabling the leaps that gave us everything from transistors to GPS satellites. Brown’s outlook, therefore, is doubly interesting. He isn’t just speculating about AI from the standpoint of a tech entrepreneur—he is a researcher who relies on AI tools and also studies the nature of reality at its deepest levels.
One of the more eye-catching claims Brown makes is that large language models, once almost laughably bad at solving even high-school math problems, can now pass graduate-level final exams in general relativity. As he tells it, a mere three years ago, these models scored effectively zero. One year ago, they performed at a low but non-trivial level, much like a weak graduate student. Today, the latest generation can produce correct solutions so consistently that he has “retired” his old final exam as an evaluation. This is a remarkable progression. That an AI can do them puts it squarely in the realm of an upper-tier graduate student.
If this trend continues—and Brown believes it may—it’s natural to ask how long before these systems make actual research contributions: things that no human has discovered. While AI systems already help with literature searches or summarizing complicated topics, full-blown conceptual innovation is another level. Brown suggests that the “final benchmark” for full AI parity with top theorists might be an AI that could replicate Einstein’s leap from Newtonian physics to general relativity. In Brown’s words, “once it can do that, there won’t be much else it can’t do.” Put differently, if a system can produce true breakthroughs in fundamental theory, the line between human brilliance and AI brilliance will have effectively vanished.
Why Might AI Succeed So Rapidly?
Skeptics often argue that AI is merely an “interpolation machine”—a model that memorizes huge volumes of text or data, then parrots them back without deeper understanding. Brown readily acknowledges this criticism, but he also notes that the level of abstraction at which these models now interpolate is far higher than one might assume. Even mere pattern matching across billions of physics, math, and general knowledge texts can begin to look remarkably like genuine insight—especially once the system can recombine knowledge in creative ways.
In addition, physics is a domain where problems can be well-defined, where many solutions have exact or near-exact forms. AI is adept at such pattern recognition, especially once the models are fed specialized training sets that include not only text but also equations, data tables, or even code that simulates physical scenarios. If an LLM can parse an advanced problem, interpret it properly, and pick out the relevant bits of mathematics, it can often produce correct solutions that even advanced grad students find challenging.
Yet producing a valid solution to a known textbook problem is not quite the same as creating a brand-new theory. Brown emphasizes that frontier research is about seeing around corners, identifying conceptual gaps, inventing new frameworks, and trusting your aesthetic judgment of what a beautiful or simple theory might look like. This is where the human factor has long been critical. Einstein’s success, for instance, was not just about manipulating well-known equations, but about rethinking the very meaning of space and time. Still, Brown suspects that with enough scale, AI’s massive interpolation might yield emergent capacities that converge on something like conceptual originality. As surprising as that sounds, much of modern AI’s story has been about emergent capabilities that even the creators didn’t fully predict.
Accelerating Other Domains of Science
One might ask: why focus so much on physics? The short answer is that physics underpins many of the other sciences. From a hierarchical standpoint, physics sets the fundamental rules; chemistry is essentially physics applied to atomic and molecular systems. Biology then emerges from molecular interactions, and medicine in turn relies on biological insights. If AI breaks down barriers to understanding quantum mechanics, subatomic particles, or large-scale cosmological structures, that often leads to better instrumentation, fresh theoretical frameworks, and new ways to probe everything from drug design to environmental modeling.
Moreover, the same AI approach that helps a physicist analyze cosmic background radiation can help a biologist parse genomics data or a chemist design novel molecules. The core advantage is pattern recognition within enormous data sets. Physics experiments—think of the Large Hadron Collider or massive sky surveys—generate petabytes or exabytes of data. Similarly, modern genomics, proteomics, climate observation, and epidemiological databases rival those scales. Tools that can distill meaningful patterns out of big data in physics naturally carry over to other fields. Brown’s vantage, therefore, does not just foretell a “Golden Age of AI in Physics”. It points to a broader revolution in how scientists do research, hypothesize, and test new ideas.
One especially promising implication relates to materials science and chemistry. Better fundamental physics models plus large-scale machine learning can predict the properties of novel substances with far greater accuracy. That could lead to more efficient batteries, new superconducting materials, or catalysts that ease the transition to green energy. In drug discovery, advanced simulation methods—some of which originally come from quantum mechanics—can help design molecules that bind to proteins in precisely the right ways. Rather than random guessing, we can do inverse design: specifying the properties we want, and letting the AI come up with candidate compounds. These leaps hinge on deeper models of how particles interact, how electrons move, or how molecular structures fold—ultimately all grounded in physics.
The Timeline: Five to Ten Years
Brown is often asked for a timeline. How soon, realistically, until we see genuine leaps that match or surpass top human researchers? He sometimes jokes he’s uncomfortable saying a specific date, because “this progress can be shockingly fast.” Still, he posits that if current trends hold we could be five to ten years away from AIs that can do most tasks of a theoretical physicist—including brainstorming novel directions of inquiry. That may feel both shockingly soon and also comfortingly distant: five years is long enough that it won’t blindside us tomorrow, but short enough that we should probably start paying attention.
It’s also worth noting that AI’s maturity doesn’t have to be uniform across all fields. Some domains of science might progress faster than others, especially if they have richer datasets or more direct ways of verifying new theories. Physics has a relatively clean data structure, especially in subfields like high-energy particle physics or astronomy, where repeated experiments can confirm or refute a hypothesis. Other areas, such as clinical medicine, face heavy regulatory hurdles, messy data collection, and enormous variation among human populations. So we shouldn’t expect a monolithic wave that transforms every area at exactly the same pace. Still, the broad direction is the same: as AI models improve, they will be deployed in other sciences, fueling a cross-disciplinary surge.
AI’s “Beauty Problem”—and Potential Pitfalls
Brown draws a distinction between fitting data and truly capturing the underlying structure of reality. Historically, some theories were so good at matching observations that they became complicated “epicycle” models. (In medieval astronomy, epicycles were added to the geocentric model to explain planetary motion as circles within circles—impressive at matching data, but conceptually misguided.) A purely data-driven AI might discover the perfect “epicycle-laden” model that explains every measurement but lacks the conceptual elegance or deeper unification that a Newton or an Einstein might bring.
Why does elegance matter? Because it has often guided scientists toward deeper truths that outlast reams of messy observational data. For instance, the Copernican model eventually triumphed because it was simpler and more coherent, not because it fit old observations massively better—at least not at first. Brown highlights that if we rely wholly on AI to generate new theories, we might get lost in the mathematical weeds if we don’t also encode or prize certain notions of conceptual simplicity or elegance. The entire scientific community may need to teach these models to value more than just data-fitting metrics. In other words, we might need to incorporate a “beauty factor” or “Occam’s razor” into the AI’s approach to theory-building.
Another potential pitfall is overreliance. It’s easy to imagine a future where AI does 90% of the heavy scientific lifting, from designing experiments to interpreting results. Humans might fall into complacency, trusting that the AI will inevitably lead us to correct answers. Even Brown acknowledges that the largest breakthroughs in science sometimes come from improbable leaps of intuition or from renegade thinkers who challenge assumptions. Will an AI trained on mainstream data miss a novel path? Possibly. That said, these models may also discover wild or unorthodox lines of reasoning if they have access to the entire historical corpus, plus creative recombination capacities that exceed human memory. The outcome is still uncertain.
Research Culture: Tutors, Agents, and Co-Creators
Brown notes that many physicists now use AI models in day-to-day ways that would have been unthinkable just a few years ago. For example, an LLM can function as a personal tutor that helps you brush up on subfields you’ve forgotten or never learned. Instead of spending weeks searching for relevant papers, a graduate student can ask the AI for a quick explanation of “squeezed light in LIGO” or “renormalization group flow” and get a coherent, near-instant summary that identifies relevant sources. The AI is non-judgmental, infinitely patient, and has read the entire literature. Similarly, it can propose potential references, debug code, or check the internal consistency of a new hypothesis.
At present, no one is simply telling an AI, “Go discover quantum gravity for me!” But as these systems improve, we might see something close: entire labs handing their raw data to an AI, letting it rummage for hidden patterns or anomalies, and even letting it propose experiments. The human role might become more about curation—knowing which anomalies are worth pursuing, verifying the logic of new proposals, and ensuring that the system’s suggestions follow safety and ethical guidelines.
This all raises interesting questions about authorship and credit. If an AI co-authors a paper that proves a novel theorem or proposes a new cosmic inflation model, who gets top billing: the team that orchestrated the AI, or the AI itself? In a world where thousands of scientists might rely on the same LLM, do they all share in the breakthroughs? Scientific culture has never had to grapple with an automated partner that truly contributes creative insights. We can look at massive collaborations like CERN, where hundreds or even thousands of authors appear on a paper, but that’s still human labor. AI may be different in ways we’re only starting to imagine.
Spillover Effects in Other Sciences
So, should we expect “physics-level” transformations in chemistry, biology, medicine, climate modeling, and more? Quite possibly—perhaps even more quickly in some domains. Biology is already seeing hints of this with tools like AlphaFold (from DeepMind), which accurately predicts protein structures. Drug discovery companies are using AI to filter millions of candidate molecules in silico, drastically speeding up the initial stages of screening. Other labs are investigating how LLMs can read reams of medical literature to find hidden connections—like if a magnesium supplement helps certain headaches or if a certain gene is implicated in unexpected diseases.
Crucially, advanced physics-driven instrumentation can also help. If AI guides us to better quantum sensors, new imaging techniques, or refined ways to do spectroscopy, then the downstream fields (like chemistry and medicine) instantly get better tools. This is exactly how X-ray crystallography revealed the structure of DNA in the 1950s. That technique emerged from physics—X-rays, diffraction, and careful detection. The more we refine instruments, the more data we gather about life’s molecular machinery, and the more robustly AI can analyze it.
Of course, these leaps can also face practical obstacles. Medicine is heavily regulated—each new AI-based diagnostic might require years of validation. Some biological data sets are fragmented or proprietary. Climate data might be politically sensitive. Yet in principle, the same radical speed-up is waiting if or when the organizational and regulatory frameworks catch up. There is every reason to think that within a decade, we’ll see a landscape where scientists in multiple fields partner with AI to produce new insights at a pace previously unimaginable.
Potential Social, Ethical, and Existential Consequences
If AI-based breakthroughs can happen faster and with less human oversight, what does this mean for the role of scientists in society? Some foresee a golden age where researchers are freed from menial tasks—like reading piles of papers or coding routine simulations—and can focus on big-picture questions, interdisciplinary conversations, or creative design. Others fear a loss of autonomy: will scientists become mere “button-pushers” for AI-run labs?
There’s also the question of existential risk. Physics breakthroughs can be immensely powerful. Access to new materials or nuclear processes can lead to safer energy or might create new weaponry. Biological leaps could eradicate disease or enable dangerous pathogens. If AI automates or accelerates these discoveries, our existing governance systems might struggle to keep pace. As Brown alludes, nuclear weapons have shown that science can be destructive unless carefully managed. A future with even faster leaps across every domain might push us into an era of heightened risk, unless we deliberately plan for safe and responsible use of knowledge.
Another angle: once an AI truly makes “Einstein-level” leaps, do humans become obsolete in advanced theoretical work? Brown’s perspective is that this would be a sign we have effectively built an “artificial scientist” that can engage in the deepest conceptual leaps. But it doesn’t necessarily mean humans stop contributing: it may open up whole new frontiers that even the AI’s breakthroughs will need help exploring. Humans might remain at the helm, deciding how to apply these new insights in society—or how to shape them ethically, aesthetically, and culturally.
Conclusion: A Decade of Transformations
Bringing this all together, we can see the broad outline of Brown’s vision—and why it resonates beyond just the physics community. Large language models and related AI systems, which seemed gimmicky not long ago, are now proficient at advanced graduate exams in physics. They digest the entire scientific literature, offering near-instant tutoring, summarization, and solution generation. As these capabilities improve, it’s not a stretch to imagine them making original discoveries or forging entirely new conceptual frameworks. That might happen in the five-to-ten-year timeframe, if progress proceeds without major interruptions.
Because physics often sits at the root of other sciences, an AI-fueled revolution there could ripple outward. Chemists might use it to develop novel catalysts or synthetic pathways. Biologists might discover new ways to treat cancer or design vaccines. Medical technology could see advanced diagnostic tools that spot diseases in their earliest stages. Even engineering, climate science, and ecology might harness improved simulation capabilities and data analysis.
The benefits are enormous, and so are the challenges. We must figure out how to preserve scientific “taste” or “beauty,” how to handle the ethics of exponentially faster discovery, how to share credit, and how to ensure these advances actually serve humanity’s collective well-being. At the same time, it’s hard not to be excited by the possibility that routine tasks in research—sometimes called “the grunt work”—will yield to an era of more rapid, creative, and interdisciplinary exploration.
Ultimately, Aaron Brown’s vision is a reminder that the world of science is on the cusp of a moment that once felt purely speculative. AI and humanity collaborating not just in rote tasks, but in the deep architecture of new theories. Whether we welcome it with open arms or cautious hesitation, it appears that a new chapter in our relationship to scientific knowledge is being written—one where the boundary between human and artificial reasoning grows increasingly blurred, and the pace of discovery may soon accelerate beyond anything we’ve previously experienced.
If Brown is correct, the next decade could be one of unprecedented acceleration across physics and every science connected to it. That includes fields addressing urgent challenges—climate, health, energy, and more. The question for all of us, laypeople and experts alike, is whether we’re prepared to harness this AI moment responsibly. If we are, we might stand at the threshold of a scientific renaissance. If we’re not, we risk losing control over the very discoveries that could shape our future. The choice, and the opportunity, rest with us now.