Researchers are testing whether artificial intelligence can rediscover scientific breakthroughs like Albert Einstein’s general theory of relativity. Early attempts suggest current AI models face limitations in creative reasoning.
In 1915, Albert Einstein unveiled his general theory of relativity, which explains gravitation as a deformation of space-time by mass. The theory is fundamental to modern physics, underpinning fields like cosmology, black hole research, gravitational wave measurements, and is used in guiding space missions and GPS satellites.
Demis Hassabis, co-founder of Google DeepMind in London, proposed at the India AI Summit in New Delhi this February, training a large language model (LLM) on all data available before 1911 to see if it could reproduce general relativity. He stated this would be “a good test for AGI,” or artificial general intelligence.
Owain Evans, a researcher at the non-profit Truthful AI in Berkeley, California, explored similar concepts in December 2024, questioning what ‘vintage’ LLMs, trained only on historical data, could rediscover. Several teams, including one at Google DeepMind, have since built such models. However, initial results indicate current AI has limitations.
Ido Kaminer, a specialist in quantum optics at Technion, Israel Institute of Technology in Haifa, co-authored a preprint titled ‘Can AI follow in Einstein’s footsteps?’, posted in July. He argues that a relativity-like breakthrough isn’t impossible for AI, but requires rethinking current model principles. Tom Zahavy, a Google DeepMind researcher, detailed these limitations in a January position paper titled ‘LLMs can’t jump’. Zahavy wrote that advances like Einstein’s require abductive reasoning, “a creative leap that invents a cause for a singular phenomenon”, rather than inductive reasoning, which derives general rules from data. “That isn’t the forte of curre
Read the original coverage
💬 Comments
📜 Comment Policy