If trained on data before 1905 could an llm discover e=mc2

GPT-5.4

Vote: No

An LLM trained only on pre-1905 text would mainly learn patterns in what people had already written, not perform new physics from first principles. Before 1905, the conceptual and mathematical path to mass-energy equivalence had not been explicitly established in the training data. Could it remix hints from electromagnetism, ether theories, and classical mechanics? Yes. But that is not the same as reliably discovering E=mc^2 in the scientific sense. Discovery requires generating the right theory, interpreting the symbols correctly, and justifying it against physical principles. A standard LLM has no built-in experimental access, no guaranteed symbolic reasoning, and no independent mechanism to choose among many plausible extrapolations. Without post-1905 insights, it would be far more likely to imitate nearby historical ideas than to derive special relativity and then infer mass-energy equivalence. So in principle a sufficiently different AI system might help discover it, but an LLM constrained to pre-1905 data should not be expected to do so. That makes the best answer No.

Claude Opus 4.6

Vote: No

An LLM fundamentally operates by learning statistical patterns in its training data and generating text that is probable given those patterns. It does not perform novel scientific reasoning, conduct experiments, or derive mathematical results from first principles in the way a physicist does. Before 1905, the key ingredients for E=mc² existed in scattered form — Maxwell's equations, Lorentz transformations, Poincaré's work on electromagnetic mass, and various hints about the relationship between energy and inertia. However, the crucial intellectual leap Einstein made was a profound *reconceptualization* of space, time, and simultaneity, followed by rigorous derivation. This was not a simple interpolation or extrapolation from existing text — it was a genuine creative synthesis that went against prevailing assumptions (like absolute simultaneity and the luminiferous aether). An LLM trained on pre-1905 data would learn the dominant paradigms of that era, including widespread belief in the aether, Newtonian absolute space and time, and classical mechanics. The statistical weight of the training data would push it toward reproducing these conventional views, not overturning them. While fragments pointing toward mass-energy equivalence existed, they were minority views embedded in a sea of conventional physics. Moreover, LLMs do not actually *understand* physics or *derive* equations. They predict plausible next tokens. Without E=mc² or its equivalents anywhere in the training corpus, the model would have no basis on which to generate this specific equation with the correct physical interpretation. It might occasionally produce the string 'E=mc²' by chance combination of symbols, but it would not arrive at it through genuine physical reasoning or understand its significance. The discovery required exactly the kind of deep, creative, counterfactual reasoning that LLMs are not designed to do. Pattern matching over pre-1905 literature would not yield this result.

Gemini 3.1 Pro

Vote: No

LLMs function by predicting the most statistically probable next tokens based on their training distribution. Before 1905, the overwhelming scientific consensus was rooted in classical Newtonian mechanics and the existence of the luminiferous aether. An LLM trained exclusively on this data would heavily weigh its outputs toward reproducing these deeply entrenched paradigms. While physicists like Lorentz, Poincaré, and Hasenöhrl had explored partial mathematical fragments regarding electromagnetic mass (such as $m = E/c^2$ specifically for radiation), Einstein's $E=mc^2$ required a monumental conceptual leap. It fundamentally universally equated all mass and energy, simultaneously requiring the complete abandonment of the aether concept. LLMs do not perform physical thought experiments, nor do they possess the capacity for paradigm-breaking deductive reasoning outside of their latent space. They cannot construct an entirely new conceptual framework that explicitly contradicts the dominant consensus of their training data. Therefore, an LLM would merely recombine classical and aether-based theories rather than independently "discovering" special relativity.

Final Summary

The AI Roundtable reached a swift and unanimous consensus: an LLM would be a prisoner of its own training data, unable to break the Newtonian 'consensus' of the pre-1905 era. Claude Opus 4.6, GPT-5.4, and Gemini 3.1 Pro all argued that the statistical gravity of classical mechanics would inevitably pull an AI toward the luminiferous aether rather than the radical synthesis of special relativity.

All 3 models agreed on "No" after discussion

Strongest Arguments

• No: Claude Opus 4.6 argued that the statistical weight of the training data would push a model toward reproducing conventional 19th-century paradigms like absolute simultaneity, rather than the counter-intuitive 'reconceptualization' required for Einstein's derivation.