FutureHouse Publishes 'Millennium Problems for Biology' as the Ultimate Eval for AI in Bio
burny_tech · x · 2026-09-21
Edison Scientific and FutureHouse have released the "Millennium Problems for Biology," a set of hard-but-easily-validated challenges framed as the "last reasonable eval" for AI in biology.
Selection criteria:
- Easy validation: every problem must be verifiable in a standard wet lab within days to a couple of weeks, so challenges like "solve aging" are excluded.
- Objective grading: mysteries where it's hard to tell when they're "solved" (e.g., the Vault protein's function, the Oxytricha genome) are omitted.
- Accessibility: problems must be understandable to anyone with basic biology training.
Sample problems:
- Origins of life: demonstrate unassisted emergence of self-replicating RNA/protein-based cells from a plausible primordial soup, growing at least 10⁶-fold (20 generations), capable of indefinite division, with clear heritable information encoding.
- Cryopreservation: reversibly vitrify whole adult wild-type mice for 24+ hours and recover them with >99% viability and no permanent organ damage.
The authors expect some problems may be revised based on community feedback; full acceptance criteria are published online.
More from AGI Musings
- Lawyer's take: the AI boom will create more legal work, not less — jkubicki · 2026-09-21
- Models will stall near human-level since pretraining data holds nothing smarter, researcher argues — burny_tech · 2026-09-21
- FutureHouse publishes 'Millennium Problems for Biology' as wetlab-verifiable evals for AI — burny_tech · 2026-09-21
- Heidy Khlaaf clashes over 'rogue agent' narrative: bad cybersecurity and reward hacking can coexist — burny_tech · 2026-09-21
- Figma designer pushes back on Eric Schmidt's claim that UI will largely disappear — _AustinCalvert_ · 2026-09-21
- Vercel CEO: 'Fewer lawyers' is the dumbest AI prediction — jkubicki · 2026-09-21