A Thought‑Experiment Challenge: Can GPT Derive General Relativity from 1911 Data?

1. What is the “Hassabis Problem”?

The Hassabis problem, proposed by DeepMind founder Demis Hassabis, asks whether an AI whose training data is deliberately cut off at a historical point can independently discover a major scientific theory that emerged later, testing the computability of scientific discovery.

2. How does the model “retrace” Einstein’s path?

Alec Radford selects a language model whose knowledge is limited to physics literature published before 1911. While special relativity exists, the field equations of general relativity remain absent. The model must start from classical mechanics, Maxwell’s equations, and early relativistic concepts, and use autoregressive generation tightly coupled with mathematical reasoning to symbolically derive the curvature of spacetime.

The model receives no explicit answer; it must fill the gap between the observed anomaly of Mercury’s perihelion and the formulation of Einstein’s field equations by chaining physical laws, performing self‑correction, and generating plausible intermediate hypotheses.

If a model can derive general relativity, then it essentially possesses one of the greatest intellectual leaps in human scientific history.

3. Technical significance

This experiment probes the “scientific imagination” boundary of AI. Success would show that AI can go beyond pattern matching to hypothesis‑driven reasoning in knowledge‑scarce regions, a core trait envisioned for artificial general intelligence (AGI).

4. Conclusion: The ultimate test of creativity

Radford’s challenge highlights a fundamental tension in AI development: scaling model size and data versus achieving genuine cognitive leaps. Current models act as sophisticated parrots, excelling at statistical extraction but faltering when required to bridge historical knowledge gaps. The Hassabis problem reframes AI evaluation from memorization to discovery, urging the community to prioritize models that can reason and create when facts are missing.