Inspired by Demis Hassabis, dev plans a "rediscovery test" to see if a model can derive E=mc²
burny_tech · x · 2026-09-27
- Recalling a DeepMind podcast, the author cites Demis Hassabis's AGI litmus test: pretrain a model only on pre-1905 data and see whether it can predict or derive E=mc² — a "rediscovery test."
- He plans to run the experiment himself: build a pretraining + post-training pipeline, then try modifications to steer the model toward the equation, logging which changes work.
- He's openly inviting collaborators to explore different approaches.
More from AGI Musings
- "Mathematics is Effectively Dead": essay extends Daniel Litt on AI and the future of math — burny_tech · 2026-09-27
- Researcher's AGI definition: Astra and Opus are already near the bar — burny_tech · 2026-09-27
- Blogger argues ASI is existential risk: even 10 extra IQ points across trillions of agents would doom us — JOBhakdi · 2026-09-27
- Gary Bernhardt: Assembly will never be a good language for AI agents — mgill25 · 2026-09-27
- Doctor's thread: AI beats doctors who use AI on medical reasoning, human-in-loop can hurt — jonc101x · 2026-09-27
- Hinton: AI will only work with brain-like computation; free energy advocates disagree — mjdramstead · 2026-09-27