Verifiers are slow and expensive: why AI probably won't cure most diseases in a decade

ziv_ravid · x · 2026-08-27

The author (computational biology background, PhD in computational neuroscience) pushes back on Dario Amodei's claim that AI could help cure most diseases within 5-10 years, arguing progress in coding will not transfer directly to biology.

The verifier problem

AI has advanced fastest where answers can be checked cheaply: code has compilers and unit tests, formal math has proof checkers, games have scores. In biology, checking an answer may take months and cost hundreds of thousands of dollars, experiments are noisy, and results may not generalize to humans.

Where drug development time actually goes

The right mechanism: high HDL correlates with lower cardiovascular risk, yet drugs that successfully raised HDL failed to improve outcomes
The right molecule: liraglutide and semaglutide target the same receptor and are structurally similar, yet one lasts 13 hours and the other a week. Binding is only part of the problem — a drug must reach the right tissue, stay stable, and avoid unwanted interactions
Surviving development: what works in cells may fail in animals; what works in animals may fail in humans. Safety in humans can't be established without giving it to humans, and rare side effects need large, long studies

Conclusion

A clinical trial is not a slow unit test — it's a noisy experiment on real people, and much of it takes biological time that can't be compressed. Anthropic's own de novo protein-binder campaigns did produce successful binders, but binder design happens to be the step most compatible with current AI (generate → rank → standardized assay). The mistake is assuming that accelerating this step accelerates every later step at the same rate.

Original post →

Verifiers are slow and expensive: why AI probably won't cure most diseases in a decade

More from AGI Musings