AI in the lab13 minForget chatbots. In one week, AI learned to simulate the whole planet's storms at 5 km, plan chemistry experiments, steer real robots after training only in simulation, and exploit 948 Linux bugs. Here is what actually shipped, what is hype, and why it matters.Read now →

Could an AI Trained Only on Pre-1911 Physics Rediscover Einstein? A Cosmologist Tried, and the Machine Fudged It
AI is crushing hard math. So a cosmologist asked a sharper question - if you wiped a language model's memory of everything after 1911, could it find general relativity on its own? It matched the data. It never found the idea.
Give a modern AI the orbit of Mercury. Thousands of years of positions. Every wobble.
Then erase everything it knows about the 20th century. No Einstein. No curved space-time. No Interstellar, no wormhole memes, no pop-science documentaries.
Now ask it to explain why Mercury's orbit drifts the way it does.
That's the experiment UC San Diego cosmologist Brian Keating has been running and describing in public for the past year. He calls it the Einstein test. The result so far: the machine matched the planet beautifully. It just never had the idea.
Instead of discovering that gravity is the curvature of space and time, he says, it built an extremely dense three-dimensional mesh of flat space and bolted on fudge factors until the numbers worked.
That gap, between fitting the data and finding the concept, may be the most important open question in AI and science right now.
Why physics, why now
AI systems are already doing remarkable things in mathematics. This site has covered models helping formalize Fermat's Last Theorem and cracking long-standing problems about prime numbers. Mathematicians who were skeptics two years ago are now co-authoring with machines.
Keating's argument is that this success might not transfer to physics, for a simple reason.
Mathematics is a language. Physics is described by a language, but it isn't one.
Large language models are, at heart, machines that learn the structure of language. A proof is made of symbols, and symbols are their native territory. But physics is about the world. The equations are a description of something that exists whether or not anyone writes it down. Keating likes to make the point with translation: the word for bread differs from language to language, but the bread does not. The word is not the thing.
So the question becomes: can a system that lives entirely inside descriptions ever make the jump to a new description of reality?
The happiest thought of Einstein's life
To see what the test is really asking, go back to 1907.
Two years after special relativity, Einstein was sitting in the patent office in Bern when he had what he later called the happiest thought of his life. A person falling freely, he realized, would not feel their own weight. Picture an elevator whose cable snaps. Everyone inside floats. In free fall, gravity effectively vanishes.
That small, almost bodily insight did enormous work. If you can make gravity disappear by choosing the right frame of reference, then gravity isn't quite a force in the usual sense. Over the next eight years, through some of the hardest mathematics he ever did, that thought became general relativity, unveiled in 1915: mass and energy curve space-time, and objects follow the straightest possible paths through that curved geometry.
Keating's point is about what kind of thought it was. It wasn't derived from a dataset. It was grounded in a physical, almost visceral sensation, the drop in your stomach on a roller coaster or a bumpy flight. He asks whether a language model can have anything like that. Can it experience something like happiness? Can it viscerally imagine weightlessness? Could an embodied robot running a language model?
Those are partly philosophical questions. But they lead to a very concrete experiment.
The experiment: a lobotomized model
The idea is to build what Keating bluntly calls a lobotomized model: a language model whose knowledge stops at a cutoff date. He has described cutoffs of 1900 and 1911 in different tellings. The goal is to give it the intellectual ingredients Einstein had, and nothing he didn't.
Then you see whether, with all the compute we now have, it can get to general relativity on its own.
Step 1: Remove the answer
Obviously you have to strip out relativity and quantum mechanics. That part is easy.
Step 2: Remove the culture
This is the part that surprised the team. Physics leaks into everything. Films like Interstellar are full of relativity, much of it grounded in the work of physicist Kip Thorne. Popular science books, TV, jokes and memes all carry fragments of 20th-century physics.
Any of that contaminates the test. If the model has absorbed even a pop-culture shadow of curved space-time, rediscovering it proves nothing. And scrubbing popular culture, not just physics papers, turns out to be extremely hard.
Step 3: Give it the data
The team fed the model the long orbital record of Mercury, drawn from NASA's Jet Propulsion Laboratory. Mercury is the classic test because its orbit precesses (its closest point to the Sun slowly rotates) by an amount Newtonian gravity can't fully explain. General relativity accounts for the leftover drift exactly. That anomaly was one of the theory's first triumphs.
Step 4: Watch what it does
The model kept forcing a solution. It could get arbitrarily close numerically. But in Keating's account, it couldn't derive the mathematical machinery of curvature, things like the Riemann curvature tensor and the Ricci tensor. It matched the planet by mimicking, not understanding.
His summary is brutal: it can pattern-match better than a thousand grad students, but it can't come up with the insight that gravity is not a force but a manifestation of curved space-time.
He's careful to say this doesn't mean it will never happen. The team simply hasn't had good luck yet. But his hunch is no.
The wider research: "LLMs can't jump"
Keating isn't alone. The Einstein test has become a small research movement in 2026.
In September, Nature ran a feature by science writer Philip Ball on exactly this question: whether language models trained on historical data can reproduce creative breakthroughs. The feature describes "vintage" models trained on pre-1911 data.
The most cited intellectual case is a position paper titled "LLMs can't jump" by Tom Zahavy, presented at ICML 2026. Its argument maps neatly onto the Mercury result.
- Induction (spotting patterns in data): current AI is excellent at it.
- Deduction (logically deriving consequences, as in proofs): also increasingly excellent.
- Abduction (the intuitive leap to a brand-new explanatory hypothesis): this, the paper argues, is what's missing.
Using Einstein's route to general relativity as the case study, the paper argues current models are structurally unable to create new foundational axioms, especially when data is scarce. Its proposed fix is integrating physically consistent, multimodal world models.
Notably, Zahavy, who contributed to DeepMind's AlphaProof, stresses that the paper is narrower than headlines suggested. It doesn't claim AI can't make discoveries. He says he's seen first-hand that AI systems already are. The claim is about one specific kind of leap.
The pushback
Not everyone buys it. One response, titled "Hot Take: LLM can 'jump'," argued that general relativity could also have been reached by a more deductive route, and that models able to spot gaps between fields may not need an abductive step at all.
There's a deeper theoretical counterargument too, one AI safety researchers often raise. In principle, the data up to 1911 contained enough information to find general relativity. A system trained only to predict observations, such as where Mercury will be or how starlight bends during an eclipse, would, if it predicted perfectly, have to internalize the underlying physics. Prediction can, in theory, push a model past what any human understood.
The honest gap is between what the data permits a model to learn and what today's architectures actually manage to learn. The Einstein test measures the second.
The cosmology problem AI can't touch
Keating's second argument is even more interesting, because it isn't about intelligence at all. It's about data.
Take one of the biggest open crises in cosmology, the Hubble tension. We know the universe is expanding. We don't agree on how fast.
There are two main ways to measure it:
- The nearby universe. Measure distances to galaxies (still billions of light-years away) using Cepheid stars and supernovae, and see how fast they recede.
- The early universe. Use the cosmic microwave background, light released about 380,000 years after the Big Bang, roughly 13.8 billion years ago, and project its expansion rate forward with the standard cosmological model.
Keating uses a pediatrician analogy. Measure a two-year-old's height, and decades of growth data let you predict their adult height with good accuracy. Measure the early universe's expansion, and the standard model should predict today's.
It doesn't. The two methods each claim roughly 1% precision and disagree with each other. Published analyses put the discrepancy above 5 sigma, and Keating describes it as even larger, odds of a fluke in the one-in-a-trillion range. James Webb Space Telescope observations of more than 1,000 Cepheids rejected one popular explanation, measurement crowding, at 8.2 sigma, a higher confidence than the tension itself.
Now hand that problem to the most powerful AI you can imagine. Ask it which number is right.
How could it know? It can't collect its own data. It can't build a new microwave background observatory or a new telescope. It can't measure how much dust sits in a distant host galaxy. And it can't do an experiment, because in cosmology there's only one universe. You can't order a supernova to go off a billion years after the Big Bang.
Why biology is different
Contrast that with biology. Labs can run robotic experiments all day: folding proteins, mixing reagents, generating fresh data and feeding it straight back into models. That closed loop is a big reason AI has been so successful there.
Keating also notes, pointedly, that biology is where AI labs are most protective. Try to use a frontier model for certain biology research and it may refuse on safety grounds, while the labs pursue those discoveries themselves.
In cosmology, the training data comes from human observation, and the universe doesn't take requests.
Theory, experiment and why both matter
Keating is an experimentalist. He co-founded the BICEP experiment, which in 2014 announced what looked like primordial gravitational waves from the Big Bang. The signal turned out to be largely cosmic dust, a story he told in his book Losing the Nobel Prize.
That experience shapes his view of what AI can and can't do.
The two personalities of physics
In quantum mechanics, he notes, there are two camps. One cares deeply about interpretation: Copenhagen, many-worlds and the rest. The other says "shut up and calculate." He sees validity in both, but warns that the ultimate "why" question never ends. Like a child asking why over and over, eventually you have to stipulate something.
Why the best physicists do both
To do an experiment, you have to understand the theory behind it. To do theory, strictly speaking, you don't need to understand experiment. But the best theorists do. He points to this year's physics Nobel: Francis Halzen won the 2026 prize for decisive contributions to the IceCube Neutrino Observatory, a cubic kilometer of instrumented Antarctic ice, and the discovery of high-energy neutrinos from beyond our solar system. Halzen is more theorist than builder, but he understood the experiment as well as any experimentalist. Likewise the LIGO builders understood Einstein's gravity deeply.
And Einstein himself, the theorist's theorist, did his greatest work through thought experiments. What would I see riding alongside a beam of light? Would two observers on different trains agree that two flashes were simultaneous? Physics thought experiments are thought experiments for a reason: many can't be done in reality.
Hardware is slow, software is fast
Keating makes an interesting contrast with AI itself. New model versions now arrive every couple of months, with a dizzying taxonomy of names. But data centers and new chips arrive slowly, because hardware is genuinely hard. He's skeptical of runaway thought experiments like the famous paperclip maximizer, noting that even a superintelligence can't conjure more iron than the Earth contains. A thought experiment, he argues, is often a thought experiment because it can't be done.
Is string theory a hallucination?
One provocative line in this debate is the claim that string theory is a kind of hallucination produced by physicists, a theory that grew because there wasn't enough data to constrain it.
Keating pushes back on the reasoning, if not entirely the conclusion. String theory dates to the late 1960s, he notes, the golden age of particle physics, when data was plentiful. It was created for a real reason: to unify gravity and quantum mechanics, and it can do things in that domain nothing else can.
But it also doesn't make useful, testable predictions. That, he says, makes it effectively metaphysics. Not necessarily wrong. Just not yet physics you can check.
History offers a warning in both directions. Johannes Kepler gave us his three laws of planetary motion, which astronomers still use to discover every exoplanet. Kepler also believed planetary orbits were set by nested geometric solids tied to musical harmony, which was nonsense. Ptolemy's epicycles once matched the data better than early Newtonian attempts.
The lesson: be wary of theorists, and be wary of experimentalists too. All theories are wrong, Keating likes to say, but all are useful at some level.
That lesson cuts straight back to AI. A model that adds fudge factors until it fits Mercury is doing something uncomfortably like epicycles.
Falsifiability isn't the whole answer
Popper's falsifiability criterion is the usual line between science and non-science. Keating thinks it isn't enough on its own. His joke: if an astrologer predicts you'll meet someone at a conference, and you don't, the astrology was falsified. Does that make it science?
He borrows a principle from organizational psychologist Adam Grant, in paraphrase: don't believe everything you can prove, and don't automatically dismiss everything you can't falsify. Mathematics has Gödel's incompleteness theorems marking limits on what can be proven. Physics has no equivalent rule for what counts as physics.
Why does math work at all?
The deepest part of the conversation is a mystery physicists still can't explain.
Galileo said the universe is written in the language of mathematics, specifically triangles and geometry. But you can't actually draw a triangle. Any line you draw has thickness. Ink is three-dimensional. Perfect triangles exist only in minds. Even a computer can't represent a true zero-dimensional point. Its smallest floating-point number is still nowhere near zero.
And yet math works, as Eugene Wigner famously marveled. Poll margins of error use the Gaussian distribution, which contains pi. What does a circle's ratio have to do with an election?
Keating's favorite example: the classical equations that let you know a particle's position and momentum precisely, and the quantum equations that forbid it, differ in a striking way. The quantum version carries the square root of negative one, the imaginary number i. Higher-dimensional analogues exist (2x2 and 4x4 matrix versions), then they seem to stop. Nobody knows why.
If physicists can't explain why mathematics maps onto reality, it's fair to ask whether a system made purely of mathematical language can find the next mapping.
The rest of the big questions
Keating's broader worldview runs through all of this:
- Time is not like space. Physicists talk about space-time as one fabric, and he calls that misleading. You can fly to St. Louis. You can't fly to yesterday. He increasingly thinks of time as the perception of change, a dual to motion. He admits he has no idea if he's right.
- The beginning of time is metaphysics. Whatever happened "the Tuesday before the Big Bang," his favorite way of posing it, can't be grounded in physical law, because the origin of time can't sit inside time.
- Instruments change minds. He traces a chain from the printed Bible to mass demand for spectacles, to Dutch glassmaking, to Hans Lipperhey's telescope, to Galileo seeing Jupiter's moons, the phases of Venus and the Moon's craters. Experiments can crush metaphysical certainty.
- On God, he's a "practicing agnostic." He keeps religious practice because, as a behaviorist, he judges that it improves his life, while agreeing he doesn't believe in "a guy in a white beard."
- Alfred Nobel's lesson. Nobel read a premature obituary calling him a merchant of death and changed how history would remember him. Keating's takeaway: live so you could die at any time and be at peace with what you made.
Why this matters
It's a real benchmark for "AI scientist" claims. Labs increasingly market models as research partners. The Einstein test asks the hardest version of that question: not "can it help," but "can it originate."
It separates two bottlenecks. Even if models learn to make conceptual leaps, some sciences will remain data-limited. In cosmology, more intelligence doesn't create more universe.
It changes where AI pays off. Fields with cheap, fast experiments (biology, materials, chemistry, math) are where AI discovery will compound first. Fields with one sky and slow instruments will move at the speed of telescopes.
Counterpoints
- Small sample. The Mercury run is one team's experiment, described publicly rather than in a peer-reviewed result.
- Contamination cuts both ways. If scrubbing culture is so hard, a "clean" model may also be crippled, missing ordinary 1911 knowledge.
- Architectures are changing fast. Reasoning models, tool use and world models are new. Today's failure may not be tomorrow's.
- Einstein had help. He spent years and leaned on mathematician friends like Marcel Grossmann. Expecting a one-shot leap may be an unfair test.
FAQ
What is the Einstein test? An experiment asking whether an AI given only pre-relativity knowledge (cut off around 1900-1911) can independently derive general relativity.
What happened when it was tried? In Keating's telling, the model fit Mercury's orbit closely but used a dense flat-space mesh and fudge factors rather than discovering curved space-time.
Does this mean AI can't do science? No. Even the most-cited skeptical paper says AI systems are already making real discoveries. The narrower claim is about foundational conceptual leaps.
Why can't AI just solve the Hubble tension? Because it's a data problem. Two precise measurements disagree, and resolving them needs new observations no model can generate.
Who won the 2026 Nobel Prize in Physics? Francis Halzen of the University of Wisconsin-Madison, for the IceCube Neutrino Observatory and the discovery of high-energy astrophysical neutrinos.
The bottom line
The Einstein test reveals something uncomfortable about AI and the heroes of science.
A machine can now read Mercury's orbit and reproduce it to any precision you like. What it can't yet do, as far as anyone has shown, is feel what Einstein felt in 1907, the strange lightness of falling, and turn it into a new picture of the universe.
Maybe that capacity will arrive with bigger models, embodied robots or world models. Maybe it needs something we don't have a name for. Either way, the next great leap in physics may still need two things AI can't fake: new data from a universe that doesn't take orders, and a mind willing to imagine falling.
Sources
- The Einstein test: what happens when AI tries to rediscover relativity? (Nature)
- LLMs can't jump, position paper and responses (Tom Zahavy)
- JWST observations reject Cepheid crowding as an explanation for the Hubble tension (arXiv)
- Nobel Prize in Physics 2026 press release (NobelPrize.org)
- IceCube: Francis Halzen wins 2026 Physics Nobel Prize





