FutureHouse and Edison Scientific publish twelve ‘Millennium Problems for Biology’, pitched as the ‘last reasonable eval’ for AI in biology
Event confirmedAwaiting review
The takeaway
Rodriques called them “the ‘last reasonable eval’ for AI in biology”; the launch post had ~508k views on X, and cash prizes and a judging panel were promised.
Status
- Claim
Event confirmedAwaiting review
- Our reporting
- High confidence
- Verification
- Unverified claim
- Importance
- 3 of 5
- Last verified
- 3 October 2026
Your AI and this story
- GPT-6.2 Astra111 days after its cutoff
- Claude Opus 5.581 days after its cutoff
- Gemini 4 Hydrogen172 days after its cutoff
- Grok 4.7111 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 81 days before it.
Key facts
- The 12 problems: origins of life; whole-mouse cryopreservation; reverse translatase (protein → nucleic acid); improved Rubisco; quadruplet-codon cell; somatic limb regeneration in mammals; bacterial production of AAV/lentivirus; programmable proteases; cell-penetrating intracellular protein binders; protein amplification chain reaction; 5′ polymerases; new nitrogenases
- Design rule: each has acceptance criteria checkable in a simple laboratory environment
- Sept 21 update (Rodriques): the specs will be revised after community feedback; FutureHouse will run a community process for a longer list of grand challenges, convene a panel to adjudicate success and announce cash prizes
- Sept 28: Andrew White (FutureHouse co-founder) joked that ‘Millennium problems for bio #4 lasted 9 days?’, pointing to a UC Davis/LBNL bioRxiv preprint (posted Sept 9) in which protein-language-model-designed de novo Rubiscos were active and one showed an apparent CO2/O2 specificity beyond the natural Rubiscos assayed; no adjudication has been announced
- Hacker News discussion reached 164 points (Sept 20)
What happened
Rodriques framed the list as biology’s answer to the run of AI results on open math problems: concrete tasks that, if an AI scientist (or anyone) solved them, would be unambiguous to check at the bench.
Why it matters
Benchmarks for AI in science are saturating. A small set of hard, lab-checkable targets from an AI-scientist lab gives a public yardstick for claims that AI systems are doing transformative biology, though how wins are judged is still being worked out.
What is disputed or not yet verified
| Verification | unverified claim |
|---|
Sources
7 sources from 5 sites. Numbers match the chips in the text.
7 sources: 5 primary, 2 reactions
Primary
- Millennium Problems for Biology (site)millenniumproblems.bio, official
- FutureHouse: Millennium Problems for Biologyfuturehouse.org, official
- Sam Rodriques on X: launch post (~508k views)x.com, official
- Sam Rodriques on X: Sept 21 update (panel, prizes, expanded list)x.com, official
- bioRxiv: De novo Rubisco design with protein language models (Kehl, ..., Siegel)biorxiv.org, paper
Reactions
- Andrew White on X: ‘#4 lasted 9 days?’x.com, discussion
- Hacker News discussionnews.ycombinator.com, discussion
Changes
- Filed (Sept 30)