Devin factored RSA-260 on spare GPUs — for about $400,000
Cognition has finally published how its record factorization happened, and the answer is more interesting than the record: an engineer with no expertise in the math pointed a fleet of coding agents at a spare-compute problem and got a world-first in three weeks.
Cognition engineer Eric Lu posted the write-up we asked for: RSA-260, the largest number from RSA Laboratories' 1991 challenge ever cracked, fell in about three weeks of part-time work at an estimated 4,900 GPU-days — roughly $400,000 at market GPU prices. The method was the general number field sieve, the same algorithm behind every public record, but the implementation was not: Lu had Devin write a GPU lattice siever, the stage that dominates GNFS cost and that every prior record ran on CPUs. The numbers are extraordinary if they hold — 13.8 billion raw relations sieved down to 8.3 billion unique and 3.99 million free, then a 656-million-row matrix with 98 billion nonzeros solved by block Wiedemann over four days at the end of August. Sieving alone took 3,813 GPU-days and the siever got 14–17% faster twice while it ran, because the agents kept improving it mid-computation.
None of it cost Cognition anything at the margin, and that is the point. Lu rigged the job scheduler to fill single-node gaps left over from LLM training on NVL72 racks — lattice sieving is embarrassingly parallel across billions of tiny work units and can be killed instantly, so it is the ideal tenant for compute nobody else can use. The entire run drew a single-digit percentage of the cluster as a side project whose actual purpose was better scheduling. We covered the announcement and the fake "seven months by hand" story that raced ahead of it in RSA-260 falls after 35 years — AI summaries invented the method; the write-up now answers the question that post flagged as the one worth waiting for.
The honest caveats in Lu's own post are the best part of it. He says plainly that he does not understand the underlying mathematics — he compares himself to "a mid-level car hobbyist" who knows what the parts do but not the physics. Across 233 Devin sessions he sent 82,702 words in 3,328 messages; the agents started 101 child sessions and 36 ran with no human intervention at all. What he supplied was executive function: goal hierarchy, catching the agent measuring the wrong thing, and forcing a unified set of benchmarks to exist. And he noticed something anyone building on agents should note: the further the code drifted from upstream CADO-NFS, the more confused the agents became, which hints that pretraining exposure to the reference implementation was doing real work. The efficiency gains, he stresses, barely dent the feasibility of RSA-2048.
Lu also draws the line the rest of the industry keeps blurring. Writing days after OpenAI's Navier–Stokes announcement, he says he shares the concern about a loss of human understanding and admits he learned less about the number field sieve than if he had done it himself — while arguing the tools let more people make real contributions. That is a much more defensible claim than a lab issuing a press release about a Millennium Prize: this is a verified, reproducible, unglamorous engineering result with a cost attached, published with its own limitations. The gap between the two is what "AI did science" should be held to.
What to watch: whether an independent group reproduces the 10x cost claim, and whether any other lab reports factoring work done the same way on idle capacity.
If a hobbyist-level operator plus agents can set a 35-year record on spare capacity, what does expertise mean in computational research now? Tell us in the comments.
