LLM from pre-1930 derives quantum mechanics and relativity
An LLM trained exclusively on pre-1900 text was tested on its ability to derive modern physics from experimental data, showing limited but notable intuition. Despite failing most physics tasks, the model independently suggested concepts resembling quanta and the equivalence of gravity and acceleration. The experiment raises questions about the nature of intelligence and whether AI can perform out-of-distribution reasoning akin to scientific breakthroughs.
- ▪The LLM was trained solely on text from before 1900 and prompted with experimental observations from physics history.
- ▪The model generated statements resembling early quantum theory and the equivalence principle in general relativity.
- ▪It failed most standard physics tasks but showed glimmers of intuitive reasoning when confronted with anomalous data.
- ▪The experiment was inspired by Demis Hassabis as a test of whether LLMs can perform true out-of-distribution scientific reasoning.
- ▪The results suggest that AI may simulate aspects of human-like scientific discovery under constrained conditions.
Opening excerpt (first ~120 words) tap to expand
Machina Mirabilis Michael Hla March 2026 An experiment to see if an LLM trained from scratch on text prior to 1900 can come up with quantum mechanics and relativity. While it fails at most physics related tasks, the model shows glimpses of intuition. When given experimental observations the model can declare that “light is made up of definite quantities of energy” and can even suggest that gravity and acceleration are locally equivalent. (For results, click here) (For a mini essay on what this taught me about intelligence, click here) Annus Mirabilis "The result is one of the greatest achievements of human thought." — J. J. Thomson, on the confirmation of Einstein's theory in 1919 The late 19th century was an interesting time to be a physicist.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at Michaelhla.
Discussion
0 commentsMore from Michaelhla
