Good ideas can wait a very long time for clean proof. In science, verification loops are often slow, noisy, and sometimes hostile to the truth.
We tend to imagine that experiments quickly confirm or kill theories. Unfortunately, reality is much messier.
Good ideas can wait a very long time for clean proof.
In the 3rd century BC, Aristarchus of Samos proposed that the Earth orbits the Sun. The Athenians rejected it because they couldn't observe stellar parallax (the tiny apparent shift of stars as Earth moves).
In the end, heliocentrism waited 2,000 years to be verified by Friedrich Bessel in 1838.
Sometimes verification is not just slow. It is actively hostile to the correct answer. In 1815, William Prout hypothesized that all atomic nuclei were made of hydrogen and should therefore have whole-number weights. Chemists kept measuring elements like chlorine at 35.46.
For 85 years, the data seemed to flatly contradict him. Only when isotopes were discovered did people realize Prout had been mostly right all along, hidden behind a confounder nobody knew existed.
So how did science move forward? Scientists relied on "explanatory aesthetics" and internal consistency. Einstein developed Special Relativity through thought experiments, often ignoring experimental anomalies he considered "minor errors."
If we trained an AI purely on data, it might never make these leaps. AI thrives where verification loops are tight: code (unit tests) and protein folding (the PDB). It might struggle in domains where the loop spans decades.
It could just keep refining the wrong model to fit the noisy observations, while true discovery requires sustaining independent research programs even when the data is silent or hostile.