Can AI Rediscover Relativity? Early Tests Reveal Model Limits

In 1915, Albert Einstein introduced the general theory of relativity, redefining gravity as the curvature of space-time and establishing a cornerstone of modern physics. The framework now guides everything from GPS navigation to gravitational wave detection. Today, however, it has become a benchmark for artificial intelligence leaders seeking to determine whether machines could independently achieve such paradigm-shifting insights.

During the India AI Summit in February, Demis Hassabis, co-founder of Google DeepMind, proposed a rigorous challenge: train a large language model (LLM) exclusively on data available before 1911 and see if it could reconstruct general relativity. “That would be a good test for AGI,” Hassabis stated, referencing the industry’s pursuit of artificial general intelligence.

This concept builds on earlier work by Owain Evans of Truthful AI, who in late 2024 explored what “vintage” or historically constrained LLMs might be capable of. Since then, several teams have attempted to build these temporal models, yet their initial results highlight significant limitations rather than breakthrough capabilities.

Ido Kaminer, a quantum optics specialist at the Technion – Israel Institute of Technology, co-authored a July preprint suggesting that while a relativity-like discovery is not impossible for AI, current architectures require fundamental redesign. Tom Zahavy, another Google DeepMind researcher, argued in a January position paper that modern LLMs lack the ability to make the necessary “jumps” in reasoning. Zahavy noted that historic scientific advances rely on abductive reasoning—creative leaps that invent causes for singular phenomena—rather than the inductive pattern recognition that characterizes current AI.

Sendhil Mullainathan, a computer scientist at MIT, provided empirical evidence for these limitations. In a July conference paper, he tested an orbital mechanics foundation model against synthetic planetary data governed by Newtonian laws. The model failed to derive the true law of gravitation, instead producing unique but incorrect laws for each system.

Despite these struggles with physical laws, LLMs have shown promise in pure mathematics. In May, an OpenAI chatbot disproved an 80-year-old conjecture by mathematician Paul Erdős. Mullainathan described this as a genuine conceptual advance because the solution required novel abstractions rather than brute-force searching. However, he noted that the achievement still relied on existing mathematical literature, differing from the isolated intuition associated with Einstein.

Independent researcher Michael Hla attempted to replicate Hassabis’s test using a model named Machina Mirabilis, trained on pre-1900 data. While Hla observed glimpses of intuition—such as when prompted with the photoelectric effect, the model suggested light breaks into distinct impulses—he concluded the system ultimately lacked a robust internal representation of physics, often parroting plausible-sounding phrases without true understanding. Additionally, training these models proves difficult due to artifacts from digitized texts and broken English in historical records.

Nick Levine, an independent computer scientist, faced similar hurdles creating a model limited to knowledge up to 1930. He found that filtering data was nearly impossible due to “leaky” training sets; the model frequently answered questions about post-1930 events correctly by accident. Despite these issues, Levine remains optimistic that prediction-type AI could eventually forecast scientific discoveries, predicting that a model trained on early 2026 data could reproduce findings made later in the year.

In Switzerland, a team led by economist Daniel Göttlich created the Ranke-4B family of historical models with cut-offs ranging from 1913 to 1946. Acknowledging that these models lack the strength of modern counterparts due to data scarcity, Göttlich stated they are testing for “sparks of genius” rather than full-blown innovation.

Jacob Andreas, an MIT computer scientist, argued that the core issue is not the generation of theories but the evaluation of them. He noted that an LLM could probabilistically output the correct theory alongside countless erroneous ones, and unlike in mathematics, there is no immediate way to verify which physical hypotheses are correct. This distinction underscores a fundamental difference between human and machine theory-building: humans rarely generate vast probabilistic ranges of theories and then filter for validity.