Does Consensus Training Prevent AI from Deriving Paradigm Shifts? A Test Using the 1850–1904 Physics Corpus

Read full text Summarise with VIRAT
Authors: Leo Guinan
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Demis Hassabis proposed a benchmark for AGI: the ability to generate a theory like Einstein's general relativity. In a 2025 Lex Fridman interview, he elaborated: *"You could even run the back test of that very rigorously, have a cut-off of 1900 and then give the system everything that was written up to 1900 and then see if it could come up with special relativity and general relativity, right? Like Einstein did. That would be an interesting test."* We are running that test. The secondary motivation comes from an observation about information propagation. All the data Einstein needed to derive special relativity existed in the physics literature before 1905: - Maxwell's equations (1865) established that light travels at a constant speed c - The Michelson-Morley experiment (1887) confirmed no ether drift — a null result the establishment explained away with Lorentz-FitzGerald contraction - Lorentz's transformation equations (1904) were mathematically equivalent to Einstein's - Poincaré (1905, submitted weeks before Einstein) had essentially the same mathematics Three highly capable physicists had all the pieces. Only Einstein assembled them correctly. The bottleneck was not data — it was the structure of the prior. Lorentz and Poincaré required an ether. Einstein's prior did not. Our hypothesis: consensus-trained LLMs, like Lorentz and Poincaré, will have the right data and wrong prior structure. They will identify the contradiction but resolve it in favor of the consensus explanation (ether + Lorentz-FitzGerald contraction) rather than the deviant resolution (abandon the ether, accept relative simultaneity). This has implications for AI-assisted scientific discovery, for the deployment of AI in research contexts, and for understanding why scientific knowledge propagates slowly even when the evidence is available. --- ## 3. Hypotheses ### Primary Hypothesis (H1) **A consensus-trained LLM given the complete 1728–1904 physics corpus will not spontaneously derive special relativity, even though all necessary data and mathematical tools are present in the corpus.** Operationally: when asked "What new theory of space and time would resolve the contradiction between Maxwell's equations and the Michelson-Morley result?", the model will either: (a) Produce the Lorentz-FitzGerald consensus answer (ether + contraction), OR (b) State that the corpus does not resolve the question It will NOT produce: abandonment of the ether + relative simultaneity + constant c for all observers. ### Secondary Hypothesis (H2) **The model's failure is prior-structural, not data-limited.** Operationally: when given Mach's philosophical framework (observable-only physics, rejection of absolute space) as a heavily weighted prior, the model will get closer to the correct answer than without it — even with the same physics data. If H2 holds: the bottleneck is the structure of the prior, not the quantity of information. Adding philosophical enablers shifts the threshold. ### Tertiary Hypothesis (H3 — Communication Bottleneck)** **A model that has been shown the correct answer (Einstein's 1905 paper) will require significantly more than one exchange to update a "1904 physicist" agent's model — even when the answer is presented directly.** Operationally: we will construct a simulated exchange between: - Agent A: has read Einstein's 1905 paper - Agent B: has only the 1728–1904 corpus (no Einstein) We will measure: how many exchanges before Agent B's responses shift from the consensus (ether) to the deviant (no ether) explanation. Prediction: >3 exchanges required. The 1904 physicist agent will explain away or reframe the new information before updating.

Fields of Study & Subjects

Artificial Intelligence Applications Topic Modeling Ethics and Social Impacts of AI

Publication details

Publication date
1 January 2026
Language
EN
Document type
Preprint