A line in the morning digest that's easy to scroll past: nonprofit biolab FutureHouse published "millennium problems" for biology — with acceptance criteria, like in a technical specification. First commenters waved it off: "collection of pet projects." But what hooked me wasn't the title or the list — it was the formulation of one criterion: a mouse that must survive freezing for a day and wake up with survival above 99 percent.
This isn't "great mystery of nature" that can't be closed. These are acceptance criteria — a gate in the pipeline. I wanted to understand who's turning biology into an engineering checklist, and why, and what it changes.
The excavation turned out more interesting than the headline: the list appeared online literally within 24 hours of the discussion — September 19–20, 2026, eleven days after AI closed the first mathematical millennium problem since Poincaré. Coincidence? Nothing in this story is coincidence.
The site millenniumproblems.bio is signed by two names: FutureHouse — a nonprofit lab for AI agents in science, and Edison Scientific — the company headed by its founder Sam Rodriques. Rodriques himself is an MIT Media Lab alumnus who made "Time 100 in AI 2025"; co-author of the list is Mikaela Hinks. The projects were reviewed by Adam Marblestone, Michael Baym, Erica DeBenedictis, and Tony Kulesa — people from the world of engineering biology, not random enthusiasts.
In flagrant contrast to the digests circulating online, there aren't ten problems but twelve. A full dozen: assemble life from primordial soup, thaw a living mouse without a single injury, create an enzyme that reads protein and writes it back to DNA, build Rubisco beyond the natural Pareto frontier, convert an entire cell to four-letter genetic code, regrow an adult mouse limb, assemble gene therapy viral vectors in a bacterial barrel, design proteases and binders on demand, amplify proteins without nucleic acids, make a complete set of reverse-direction polymerases, and reinvent nitrogenase with zero homology to the natural one.
Each problem is written like a contract with delivery criteria. Not "clarify the mechanism," but: 100 pre-registered peptides, no less than 90 percent accuracy, reading depth minimum 25 residues, design in 24 hours without wet evolution, 20 targets with 80 percent hit rate. The authors directly state they applied three filters: validation in an ordinary lab within days, objective grading, and comprehensibility to anyone with basic biology. That's precisely why the list doesn't include "solve aging" — such a check is impossible to close in a week in a standard laboratory.
And not a word about money. No million for solving it, no fund, no award committee. The list exists as an open benchmark: "if you solve it — that's a breakthrough in itself." Prize without a prize — the main nerve of this story, we'll return to it at the finale.
Comments in the Hacker News thread unwrapped the first layer: "mice were frozen back in the fifties, 74 percent survival." Beautiful fact, half invented. In the fifties Audrey Smith, John Lovelock, and Peter Parkes did indeed revive animals after deep cooling — hamsters after partial crystallization (Nature, 1954), rats and mice after ice supercooling with microwave diathermic heating (Journal of Physiology, 1955–1956). Lovelock, by the way — the same future author of "Gaia," co-discoverer of glycerol as a cryoprotectant in 1949, which started all of cryobiology.
But "partial crystallization" and "99 percent problem with no damage" are different universes. The FutureHouse condition itself — minimum 24 hours in frozen or vitrified state and complete recovery — has not been achieved in the fifties or since. Beautiful detail of pessimism from a commenter: body surface grows quadratically with size, volume — cubically, so thawing surface must grow cubed, while heat coming from outside burns the periphery. "Cat-sized already impossible."
Against this argument, modernity has a trump card: nanowarming. Instead of surface heating, iron oxide nanoparticles are perfused into the organ's vascular network, and the organ is heated from inside by a radiofrequency coil — evenly and fast. In 2023, on this linkage Nature Communications showed the impossible: vitrified rat kidneys after 1–100 days of storage, nanowarming, and transplantation in place of both native kidneys saved the rat's life. First time in history. Mouse embryos at minus 196 degrees have been alive since 1972; rat organs — since 2023. One scale step remains: a whole mouse with brain inside. Audacious problem? Yes. But the engineering toolkit to honestly demand it is already on the table.
Problem four sounds modest: make an enzyme fixing carbon with CO₂ specificity no worse than the record holder — thermophilic red alga Galdieria partita — and speed no lower than corn. The subtlety is that such don't exist in nature — on a dataset of 304 organisms (Flamholz and colleagues, 2019) nobody goes beyond this Pareto frontier. Moreover, in 95 percent of all measured Rubisco, speed locked into the range of 1–10 inverse seconds, and the spread of specificity among plants — less than 10 percent. The enzyme looks trapped in a physicochemical corridor: accelerate carboxylation — the same mechanics accelerates oxygen uptake.
Precisely on this point HN split into two camps. Skeptic: "nature optimized for billions of years, both parameters simultaneously impossible." Optimist: "deep local minimum that evolution with 10^30 organisms didn't punch through — ideal target for a system not obliged to take the evolutionary route." Strictly speaking, Flamholz shows only the narrowness of observed variation, not a physical floor. The door is open exactly the width of our ignorance — which makes the problem honest.
Problem five, proposed by Erica DeBenedictis herself — quaternary DNA — also looks like a physics ban but is a habit ban. Nature gave triplet code not because four bases don't work: the ribosome can decode four-base codons, and in 2022 eLife showed that quadruplet tRNAs can be obtained from all twenty classic classes. Bad news — only 9 of 20 aminoacyl-tRNA synthetases tolerate quadruplet anticodons; good news — triplet won "because simpler and sufficient," not "because alternative is forbidden." Full quadruplet code — 256 codons instead of 64, hundreds of free slots for non-canonical amino acids and, incidentally, an architecture that foreign viruses can't read. The problem changes not a parameter but the language of heredity itself.
The most conceptual of the twelve — the third: create an enzyme that "translates" protein back to nucleic acid, without template, without database, without tricks. In school biology language this is a reverse ribosome: ribosome reads mRNA into protein, and here they propose a device reading protein into code. Together with the tenth problem — "PCR for proteins," peptide amplification without nucleic acids — and the eleventh, polymerases in the 5′ direction, this is a whole program of reversing the central dogma.
Here HN produced the most lethal objection, and it's beautiful: in DNA and RNA, PCR only works because they have a complementary strand. Proteins have no complementary strand — which means the requirement to copy them without template looks not like a hard problem but like a non-existence theorem to which nobody has described a proof path. Honest objection — yet it hits the "full" version of the problem, while the list directly allows weak partial solutions. And history suggests "physically meaningless" requirements are sometimes fulfillable: in 2025 in Nature Chemistry a ribozyme polymerase for the first time provided open exponential replication of RNA, using trinucleotide blocks and pH–freeze–thaw cycles to bypass strand self-entanglement — until recently considered a fundamental barrier for the RNA world.
A separate detail I love: in 1956 Goldzveig and Smith simply put supercooled rats in a microwave and heated them back to life. Seventy years later biology wants the same thing, just engineered: reversible phase transition both ways, controlled by enzyme. All the cryptography on this planet rests on the assumption that "large numbers" can be reversed with a key. Problems 3, 10, and 11 propose checking whether life is reversible.
Problem six — regeneration — almost came off the track of Michael Levin's Tufts group: in 2022 Science Advances showed that a wearable bioreactor with a cocktail of five drugs over a frog leg stump in 24 hours launches 18-month regrowth of the stump anew — skin, bone, vessels, nerves, finger-like structures, restored motor function. Frog — a vertebrate with "human" level regeneration. Between it and mouse stand warm-blooded immunity and the scar machine, but the vector is already set. Problem seven, gene therapy viruses from E. coli, sounds like a chemical joke: in 2024 they learned to assemble in bacterial lysate empty AAV protein shells, biologically active on HeLa cells — but this is half a step out of twelve: nobody has assembled an infectious particle with a given genome inside in bacteria yet. Notice how the list is structured: each problem stands one step from a recent breakthrough.
Problem one — life from soup — looks the most ancient, but in fact the freshest. Miller and Urey in school textbooks closed the question by assembling amino acids with electrical discharges; but amino acids aren't life. In 2024–2025 all components appeared separately: Darwinian evolution of self-replicating DNA in a synthetic proto-container, ribozyme copying itself, synthetic cell with integrated DNA replication and membrane biosynthesis. The final transition remains — assemble them into a self-sustaining cycle with heredity and growth by 10⁶ times. That's precisely why the FutureHouse criterion sounds not "obtain a molecule," but "obtain a generation": twenty divisions in a row without cheating cell shrinkage.
To understand why the list appeared precisely now, rewind eleven days. September 8, 2026 OpenAI announced that an armada of 10,000 parallel agents — 2.7 million inter-agent messages, 130 billion output tokens, 88 hours of search plus 17 hours of formalization — found a singularity in three-dimensional Navier-Stokes equations, and the result was formally verified in Lean. Half a day before them Tristan Buckmaster from New York University and Levent Alpöge from Anthropic announced related results using different models. For the first time since Poincaré one of the Clay millennium problems is closed — and closed not by "one smart thought" but by massive parallel routine. Rodriques publishes the biology list with the formulation "in light of the progress in mathematics" — literally. Translation: "they got Lean, where's ours?"
The answer is the explanation of the entire list: in biology the formal verifier is nature itself. The mouse either survives or not; the cell either divides or not. Rodriques calls the list "the last reasonable benchmark for AI in biology" — a test that still can't be gamed with beautiful reports. His lab knows such tests: their multi-agent system Robin in May 2025 in 2.5 months autonomously conducted a full discovery cycle — from hypothesis to paper — proposing to repurpose glaucoma drug ripasudil for dry macular degeneration; in 2026 the work came out in Nature, and Robin added a mechanism humanity through all of pharmacology history for this disease never proposed. If your robot already can do this — you need a harder checklist. Here it is.
The skeptical part of honesty: the list also got critics by merit. "Collection of pet projects by not the most outstanding biologists" — and true, Clay had the planet's best mathematicians, while here — a startup with reviewers from the immediate circle. "Verifiability doesn't mean grandeur" — also true: someone in the comments fairly proposed a more utilitarian problem — synthesis of arbitrary DNA longer than 3000 nucleotides with error below 0.1 percent. The objection about "PCR for proteins" as a non-existent object still has no answer from the authors. But here's what's characteristic: nobody — not a single commenter — disputed the thresholds themselves. Mouse, limb, Rubisco, 256 codons, viruses from barrel — all are real frontiers where the world is genuinely stuck. The dispute isn't whether the problems are complex, but whether they have the right to call them "millennium."
And this returns to "prize without prize." At Clay, a million dollars bought knowledge; Rodriques doesn't pay because here the prize is the product. Whoever solves problem two gets not a check but an organ cryobank with market valuation in hundreds of billions; whoever solves the fourth — an enzyme recovering part of the 1–2 percent of world energy and 3 percent of emissions that Haber-Bosch burns today; whoever solves the seventh — gene therapy cost reduction by orders of magnitude. The list performs for biology the function that open-source problems perform for the software world: you don't compete for money, you compete for your result to become infrastructure.