Evidence · Science
Decades of science at once
AI's deepest benefit does not show up in anyone's daily routine: it shows up in laboratories, years before it becomes a drug, a battery or a vaccine. It is where a new tool changes the speed at which humanity discovers.
The map of proteins
For 60 years, solving the structure of a single protein could cost years of work and hundreds of thousands of dollars. At that pace, humanity accumulated about 200,000 lab-solved structures. In 2022, an AI model predicted the structure of nearly every protein known to science: 200 million, published for free, for any researcher to use.
More than 2 million researchers, from about 190 countries, have used that catalog to design drugs, vaccines and enzymes. In 2024, the work received the Nobel Prize in Chemistry. It was the first time the prize recognized a discovery made, in part, by an AI model.
Figure 1
200M
protein structures predicted and published for free
1,000×
the catalog built by 60 years of experiments
2M+
researchers using it, in about 190 countries
Source: AlphaFold Protein Structure Database; Nobel Prize in Chemistry, 2024.
Materials that did not yet exist
After centuries of experimental chemistry, humanity knew about 48,000 stable crystalline materials, the catalog from which batteries, chips and superconductors come. In 2023, an AI model proposed 2.2 million new candidates and flagged 381,000 as stable, multiplying the catalog by nearly ten.
It did not stop at prediction: dozens of these materials have already been synthesized independently in the lab, some by a robotic system that plans and runs its own experiments. The cycle that took years between hypothesis and physical material now fits in days.
Figure 2
Stable crystalline materials known to humanity
9×
Before, with centuries of experiments
After AI predictions, in 2023
2.2 million candidates proposed; 381,000 flagged as stable by the model.
Source: Merchant et al., Nature, 2023. GNoME project.
Why this matters to everyone
Scientific discovery is the only benefit that compounds on its own: every mapped protein becomes the input for a drug, every new material the input for a battery, and each of them funds the next question. When AI speeds up that cycle, it does not help one user; it helps everyone who comes after.
These results are not ours, and that is why they are here: they are the public ruler of what frontier models already do for humanity when pointed at the right problem. It is the ruler we use to measure our own work.
