AI can fit orbital trajectories while still missing the physics, study argues
jonippolito · x · 2026-07-28
AI can fit orbital data without learning physics
Jon Ippolito argues that generative AI is the wrong tool for modeling the world, because it can produce accurate predictions while missing the underlying laws.
- He cites a Harvard/MIT study presented at ICML showing that transformers trained on planetary orbits could match trajectories with arbitrary predictions.
- But the learned “force laws” changed from sample to sample, which makes them unlike scientific models built from a small set of generalizable principles.
- The essay contrasts this with real scientific modeling: a model should explain the system, not just simulate outputs that happen to fit the data.
- It uses the history of Ptolemy vs. Copernicus to argue that better fit alone is not enough if the theory does not reveal the physics.
More from AGI Musings
- Study Finds Amazon Flooded with AI-Generated Books, Crowding Out Human Authors — TuhinChakr · 2026-07-28
- Stanford Scholar Asks: Why Do Human Brains Learn Math More Efficiently Than AI? — SuryaGanguli · 2026-07-28
- Ambient co-founder says open source will win and calls closed AI a bad infrastructure bet — IridiumEagle · 2026-07-28
- AI that invents new languages reignites the debate over creativity and consciousness — begusgasper · 2026-07-28
- Bill Gates’ 1996 “Content is King” essay resurfaces amid the creator-economy thesis — jackedAJ · 2026-07-28
- Robotics advances could bring more industry back onshore, the post argues — S_OhEigeartaigh · 2026-07-28