title: "Can Large Language Models Derive Relativity? — And a Discussion on the AGI Criterion" date: "2026-08-01" author: "Zhigeng" category: "frontier" excerpt: "If you feed a large language model all of human physics knowledge prior to 1905, can it derive special relativity? A recent position paper from Google DeepMind says no—because the crucial leap in scientific discovery comes from something LLMs will never have: sensory experience." readTime: 12

Can Large Language Models Derive Relativity? — And a Discussion on the AGI Criterion

1. Hassabis's Einstein Test

In February 2026, at the IISc AI Impact Summit in India, Google DeepMind CEO and Nobel laureate Demis Hassabis proposed a criterion for AGI that the industry would come to call the Einstein Test: give an AI system a knowledge cutoff of 1911, feed it all the knowledge humanity possessed at that time, and see whether it could, like Einstein in 1915, independently derive general relativity. If it can, we are looking at true AGI. If it can't, what we have is still just an extraordinarily sophisticated pattern matcher—a glorified search engine.

So, can a large language model really derive general relativity from the sum of human knowledge as it stood in 1911?

Here's the interesting part. Just one month before Hassabis proposed the Einstein Test at the February summit, Google DeepMind had published a position paper titled LLMs Can't Jump.

The paper anchors its core argument in a diagram Einstein himself drew. In 1952, Einstein wrote a letter to his friend Maurice Solovine, in which he sketched a cyclical diagram: starting from sensory experience (E) → through a leap (J) → arriving at an axiom system (A) → then proceeding by logical deduction from the axioms → ultimately returning to sensory experience for verification. This is what came to be called the E-J-A discovery model.

The paper then invokes a framework from the 19th-century American logician C. S. Peirce, who classified human cognitive mechanisms into three types: induction (extracting patterns from observed cases), deduction (deriving conclusions from known rules), and abduction (generating novel explanatory hypotheses). Large language models already excel at the first two, but abduction is something entirely different. It does not follow logical rules. It is intuitive, it is a leap. The J Einstein drew—that is the leap of abduction.

Hassabis could not have anticipated that this paper had already, in theory, ruled out the very possibility his Einstein Test was designed to detect.

Why is that? Or, to put it more precisely: how could someone as brilliant as Hassabis make this misjudgment?

I suspect it's because he wasn't deeply familiar with the actual intellectual journey Einstein took to reach general relativity, and therefore didn't truly grasp what Einstein's E-J-A discovery model entailed. Alternatively, the LLMs Can't Jump paper, published in January, may have been produced by a different department within DeepMind, and Hassabis may not have been involved—meaning the paper's stance does not necessarily reflect Hassabis's own.

So, how did the crucial mental leap that led Einstein to relativity actually happen?

4. "Thank You. I've Completely Solved the Problem."

By early 1905, the 26-year-old Einstein had been wrestling with this problem for seven years.

He knew he had to reconcile two mutually contradictory propositions: on one hand, the principle of relativity—the laws of physics should be the same in all uniformly moving reference frames; on the other hand, the constancy of the speed of light—Maxwell's theory said light in a vacuum always travels at c. In Newton's framework, these two were irreconcilable: if the speed of light is c and you run toward the light, the relative speed should become c + v. But experiments said otherwise.

He tried to patch Lorentz's electron theory to bridge this contradiction, spending nearly a year. All failed. He later recalled having wasted all that effort.

The turning point came on a beautiful, sunny day.

He went to visit his friend Michele Besso, an Italian-born electrical engineer. Einstein himself described the scene years later in a lecture at Kyoto University:

"It was a beautiful day when I went to visit him. I said to him: 'I've been stuck on a difficult problem lately. I've come to you today to wage war on it together.' We discussed every aspect of the problem. After I returned home, I suddenly understood where the key lay. The next day I went back to see him and said, without even greeting him: 'Thank you. I've completely solved the problem.'"

The crucial breakthrough was in a single sentence: analyze the concept of time—time cannot be defined absolutely; there is an inseparable connection between time and signal velocity.

To be specific: whether distant events are simultaneous is not an absolute fact; it depends on the observer's state of motion. Two observers moving relative to each other will give different answers to the question of which two events happened at the same time.

After this epiphany, it took him only five weeks to write On the Electrodynamics of Moving Bodies—the foundational paper of special relativity. At the end of the paper, he thanked only one person: Michele Besso.

One thing is certain: without the breakthrough on the relativity of time, there would have been no special relativity. And without special relativity, there could never have been general relativity.

This breakthrough insight did not come from reasoning within existing theories. It came from a strange fantasy in adolescence, from the extraction of relevant questions amid the tedium of examining disparate patent cases, from conversations with an engineer who had hands-on experience, and from the conceptual awakening philosophy provided.

5. The Leap a Language Model Cannot Make

Now back to the original question: can a large language model derive general relativity?

Judging from Einstein's own actual experience, the answer is clear: no.

A language model lacks many things. For example:

First, it has no embodied sensory experience. Einstein's thought experiment of riding a light beam at age 16 was not something he reasoned out on paper—he saw it. In his imagination, he saw a beam of light flying ahead of him and then asked: what if I ride on it? This extraordinary question, born from that vision of riding on light, ultimately led to special relativity.

Now, having arrived at special relativity, could one then logically derive general relativity?

In 1907, Einstein suddenly discovered the equivalence principle—the conceptual core of general relativity. At that moment, he was sitting in his chair at the Patent Office, seeing a man falling freely from a rooftop, then realizing that the falling man does not feel his own weight.

This imagination, rooted in bodily sensation, is likewise irreplaceable by any reasoning based on knowledge, terminology, laws, or theory. A large language model can read the word free fall ten thousand times, but it cannot experience weightlessness. It processes statistical probabilities between tokens—it is not simulating the physical world. The DeepMind paper put it precisely: a large language model is a Chinese Room—it can flawlessly process the language of physics but cannot reach the physical world that language points to.