AlphaFold is usually told as a story about AI. It is better read as one about measurement: it learned from roughly 170,000 protein structures solved one at a time since 1971, and was judged in a blind contest running since 1994.
Every search has two halves: a generator that proposes candidates and an oracle that grades them. AI has made the generator nearly free. In medicine and materials, the oracles that count — patients, production lines — take years and cost a fortune. Faster generation is capped by the quality of the oracle.
Medicine: The Molecule Was Rarely the Problem
From 1950 to 2010, new drugs per billion dollars of R&D halved roughly every nine years, even as screening and chemistry tools improved by orders of magnitude. Scannell and Bosley showed why: a 0.1 gain in a disease model’s correlation with patient outcomes can offset a 10- to 100-fold gain in throughput.
About nine in ten drugs that enter trials fail, mostly because they don’t work or aren’t safe; only 10–15% fail on drug-like properties. Early AI-discovered drugs fit the pattern: 21 of 24 passed Phase I, well above the historical rate, but only 4 of 10 passed Phase II, in line with it. AI is producing well-made molecules, with no evidence yet that it picks better biology.
Matter: Stable Is Not Makeable
GNoME’s 380,000 “stable” crystals were stable in a zero-kelvin calculation, not in a furnace. Chemists who re-examined Berkeley’s autonomous A-Lab judged about two-thirds of its new compounds likely known, disordered versions of the predicted ones; its authors later clarified that “new” meant new to their platform. MatterGen’s showcase material, a 2026 paper found, dates from 1971.
Even a real discovery is the short part: lithium iron phosphate took more than two decades to reach nearly half of the EV battery market. Composition is cheap. Processing is the product.
Where the Value Goes
A loop of models and robots optimizes what it measures; with a poor proxy, a faster loop converges faster on the wrong answer. When one stage of a value chain becomes a commodity, Clayton Christensen argued, profits move to the stages that are still hard: here, human data with outcomes attached, models with proven predictive power, and the slow physical steps of synthesis, scale-up, qualification and trials. The winners will look less like software companies than like drug developers and manufacturers with unusually good models.
Candidates were never the scarce input in matter and medicine. Truth is, and it still has to be measured.