科学智能体最难的不是写代码，而是知道什么算真正的进步

Turning a scientific question into a scorable task makes automated exploration possible, but the score must still represent the question the researcher actually wants to answer.

ERA combines a language model's prior knowledge with a structured search over executable notebooks, so proposed changes are informed guesses rather than random mutations.

A model that predicts held-out examples accurately has not automatically discovered an explanation that remains valid outside the conditions represented by its data.

More powerful optimization increases the need for protected holdouts and careful validation, because a rising score can reflect a weakness in the evaluation rather than scientific progress.

An autonomous inner loop can reduce routine coding work while leaving humans responsible for reframing the problem, checking assumptions, and deciding which experiments deserve further resources.

In the contrail project, the useful advance was a simpler counterfactual model with a better combination of confounders, not merely a more complicated predictor.

Weather prediction and climate projection share physical foundations, but changing long-term conditions make climate modeling a different problem from fitting abundant short-term observations.

A simple baseline establishes what the data already makes easy, so a sophisticated model must demonstrate useful improvement rather than merely additional complexity.

Useful scientific AI must connect its predictions to measurements, interventions, and real-world constraints, because faster computation cannot remove every physical bottleneck.

Keeping time for difficult hands-on learning and open-ended exploration can preserve the domain judgment needed to supervise increasingly capable AI tools.