Can AI Reproduce Einstein's Relativity? Researchers Test 'Vintage' Language Models

Key facts

• Demis Hassabis proposed at the India AI Summit in New Delhi in February that training an LLM on pre-1911 data to reproduce general relativity would be a good test for AGI.
• Owain Evans of Truthful AI in Berkeley, California, gave a December 2024 talk about 'vintage' or 'historical' LLMs trained only on historical data up to a certain date.
• Independent researcher Michael Hla trained an LLM called Machina Mirabilis on pre-1900 data; it showed 'glimpses of intuition' but failed in most cases and lacked true understanding.
• Nick Levine and colleagues attempted a vintage AI model using only pre-1930 knowledge but found training data 'maddeningly leaky,' with the model often correctly answering questions about the 1950s.
• The University of Zurich's Ranke-4B project created historical models with cut-offs of 1913, 1929, 1933, 1939 and 1946; Daniel Göttlich, now at ETH Zurich, says they test for 'sparks of genius.'
• A July preprint by Ido Kaminer of Technion—Israel Institute of Technology in Haifa argues a relativity-like breakthrough is not inherently out of reach but requires rethinking current model principles.
• Google DeepMind researcher Tom Zahavy's January position paper 'LLMs can't jump' highlights that transformative advances require abductive reasoning, not inductive correlation.
• MIT's Sendhil Mullainathan found in a July conference paper that an orbital mechanics model never inferred the true law of gravitation, instead producing a different wrong law for each planetary system.
• In May, an OpenAI chatbot disproved an 80-year-old conjecture by Hungarian mathematician Paul Erdős, which Mullainathan calls a 'genuine conceptual advance.'
• The article was published in Nature 657, 338-340 (2026), doi: 10.1038/d41586-026-02804-x.