Multiple teams train LLMs on historical data to test whether AI can reproduce general relativity

In 1915, Albert Einstein proposed the general theory of relativity, which changed humanity's understanding of space-time and gravity. More than a century later, a thought experiment about artificial intelligence (AI) has once again brought people back to that era.

In February of this year, at the India AI Summit held in New Delhi, India, Demis Hassabis, co-founder of Google DeepMind, put forward an idea—setting 1911 as the knowledge cutoff, training a large language model (LLM) on the knowledge humanity had accumulated before that point, and seeing whether it could rediscover general relativity. Hassabis believes this could test the capabilities of general AI.

According to a recent report in the British journal Nature, since the beginning of this year, multiple teams have begun trying to build "historical AI" or "retro LLMs," capping training data at a certain point in history to see whether the models can rediscover scientific ideas that emerged later. The results show that AI is still a considerable distance from "Einstein-style" scientific breakthroughs.

Independent AI researcher Michael Hla trained a language model called "Machina Mirabilis" using only pre-1900 data, hoping it would produce quantum mechanics, special relativity, and general relativity. He fed the model physical phenomena that could already be observed at the time, such as the photoelectric effect, and also provided the "elevator thought experiment," which later became an important source of ideas for general relativity. The model did occasionally come up with some rather interesting answers. Faced with the photoelectric effect, it once proposed that light would "decompose into a large number of mutually independent pulses." This statement was already close to the edge of the concept of the light quantum.

But such flashes of insight went no further. In most cases, the model failed. It could produce plausible-sounding physics language, yet "did not know why." More often, it was merely imitating statements in historical texts, rather than truly understanding the relationships among these physical phenomena.

Another team set the cutoff at 1930. Independent computer scientist Nick Levine and colleagues trained a model with knowledge only up to 1930, hoping to use it to predict scientific results that emerged later. But doing this in practice was not easy. The dates of historical materials are not always accurate, and errors and mixing can occur during digitization. They found that this so-called "1930 model" could sometimes correctly answer questions from the 1950s.

In January of this year, Google DeepMind researcher Tom Zahavy pointed out in the paper "LLMs Do Not Jump" that current LLMs are still unable to make the kind of inferential leap that Einstein made. Major scientific advances often do not come from inducing general laws from large amounts of data, but from starting with a specific phenomenon and proposing an explanation that did not exist before; this kind of reasoning is called abductive reasoning.

Leaps of imagination, or what philosopher of science Thomas Kuhn called "paradigm shifts," often occur when data are extremely scarce, or arise from anomalous phenomena that do not fit existing theories. When scientists face problems that are difficult to explain, they usually try to build a "world model," that is, infer the underlying basic laws from a small amount of data.

For example, in the 17th century, astronomer Johannes Kepler summarized mathematical laws through observations of planetary motion. On this basis, Newton further established the law of universal gravitation and the system of classical mechanics. The key is that Newton was not solving a problem of deriving an answer from data. He had to judge which phenomena were worth explaining, and then propose a theoretical framework that had not existed before. Planck was the same. He found that his blackbody radiation formula, although it matched experiments fairly well, was never perfect enough. He did not treat this as an error, but continued to probe further, ultimately proposing energy quantization.

So can AI models likewise reason out universal world models from sparse data? Sendhil Mullainathan, a computer scientist at the Massachusetts Institute of Technology in the United States, believes they cannot yet. In July of this year, he and colleagues gave an "orbital mechanics" AI model multiple virtual planetary systems that follow Newtonian mechanics, hoping it would derive the law of universal gravitation. The result was that the model did learn something, but it did not find the true law. It generated different laws of gravity for different planetary systems, each with its own errors.

In May of this year, AI overturned a conjecture proposed by Hungarian mathematician Paul Erdős that had stood for 80 years.

Mullainathan considers this a "genuine conceptual advance." He said that AI did not find the answer through simple brute-force search, but introduced some new mathematical concepts in the process of solving the problem. However, AI still integrated ideas that already existed in the mathematical literature, and did not, like Einstein, suddenly produce a breakthrough discovery from seemingly unrelated places.

Therefore, some scientists believe that future AI scientists cannot simply be a larger language model. Mullainathan proposes that models need to simultaneously understand humanity's existing knowledge, real-world data, and "mathematical space," that is, how to manipulate mathematical objects. The evaluation criteria for AI systems should also not be limited to prediction accuracy, but should also examine the simplicity of theories, their scope of explanation, and so on.

In an interview with Life magazine editor William Miller during his lifetime, Einstein once said that the most important thing is not to stop questioning; curiosity has its own reason for existing.

4; New model ignites AI circles: it does not generate text, but makes rapid decisions specifically for machines

1; Advanced Oncology publishes its first cover article: AI helps innovate Helicobacter pylori detection technology

2; ECE paper interpretation ▎How can a P&ID drawing be transformed into computable engineering knowledge with AI?

4; Paper topic inspiration: "Sustainable Food" research direction | MDPI Sustainability

5; FDE From lesson preparation to grading, AI is reshaping the classroom, but there is one thing it cannot do

6; FDE Under the frenzy of technology, who is bearing the cost of AI education?

7; Paper topic inspiration: "Sustainable Urban and Rural Development" research direction | MDPI Sustainability

8; Paper topic inspiration: Cerebrovascular diseases (I) | MDPI Brain Sciences

1; Western cryosphere is accelerating its "degradation": glacier melting sounds the alarm for alpine safety

2; The 6th National Agricultural Science and Technology Development Strategy Think Tank Alliance Conference was held in Lanzhou

3; The 2026 Nobel Prizes are about to be announced, and authoritative experts will provide you with a comprehensive interpretation!

8; Without a paper, he still obtained a doctorate and became the first person in the national medical field

Famous Doctoral Dissertations 163: Nobel Laureate Dr. Peter Doherty's 1970 doctoral dissertation at the University of Edinburgh

My elite university complex and elite university experience

More>>