Physics

Can AI Follow In Einstein's Footsteps?

Michael Shalyt, Nathan Regev, Marin Soljačić, Ido Kaminer

Featured August 15, 2026

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

AI is great at guessing answers, but to make big new discoveries like Einstein, it needs to learn how to ask the right questions and invent new rules, not just predict outcomes.

In depth

The paper argues that while AI excels at prediction, its current trajectory in physics discovery is a reverse epistemic order, moving from explicit equations to black-box predictors, unlike human scientific progress. To achieve paradigm-shifting discoveries, AI must develop the ability to pose the right questions, sustain high-risk hypotheses, and construct new theoretical frameworks guided by abstract principles and new mathematical languages, rather than merely optimizing for predictive power.

Key Takeaways

1
AI's current success in physics discovery primarily lies in prediction without understanding, mirroring human scientific history in reverse.
2
Achieving paradigm-shifting discoveries requires AI to move beyond prediction to principle-guided theory building, including inventing new mathematical frameworks.
3
Proposed strategies include training AI to propose research questions, strive for simple interpretable theories, and utilize symbolic computing in theory-building sandboxes.

Conceptual Flow

HIGH LEVEL

Methodology: The Reverse Trajectory of AI Discovery

Humans learned by first guessing patterns, then finding simple rules, and finally inventing big ideas; but AI often learns by first finding simple rules, then just guessing patterns without understanding.

Human Discovery

AI Discovery

Compare Learning Paths

Pattern Guessing

Simple Rules

Big New Ideas

Simple Rules

Pattern Guessing

Results: Guiding AI Towards Principle-Based Theory

Instead of just predicting, the paper suggests teaching AI to ask smart questions, look for simple explanations, and use math tools to test new big ideas, helping it invent new scientific rules.

Current AI: Predicts Outcomes

Add New Skills

Propose Questions

Seek Simple Rules

Test New Ideas

Invent New Theories

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.