Cognition says Devin-built GPU sieve factored RSA-260 for about $400,000
- Cognition reports that its Devin agents factored the 260-digit RSA-260 challenge number, exceeding the prior public RSA-250 record set in February 2020.
- The work used a heavily modified CADO-NFS implementation of the general number field sieve, with GPU lattice sieving and sparse linear-system solving; Cognition reports no algorithmic advance beyond performance engineering.
- Cognition estimates the run consumed 4,900 GPU-days, or 13.5 GPU-years, and cost about $400,000 at market prices: 643 GPU-days for polynomial selection, 3,813 for sieving, and 467 for linear solving.
- Using standard GNFS scaling, Cognition estimates that factoring RSA-1024 requires 78 times the RSA-260 computation and would cost about $30 million, while RSA-2048 remains roughly a billion times harder than RSA-1024.
- Cognition says Devin autonomously handled measurements, cluster operations, and optimization, while the author set priorities and benchmarks and redirected work that went off track.
Hacker News opinions
I would normally downweight a follow-up, but this article adds enough substance to stand on its own.
This feels like the first "I had an agent swarm do it" post that was actually written by a human. That is a refreshing change.
The interesting part is using stranded compute. The claimed $400k factorization apparently ran on spare or fragmented capacity that could not have been sold for other work.
I am skeptical that modern idle compute is really free. Unlike 1990s machines, current CPUs and GPUs draw substantially more power under load, so using idle capacity still has a marginal electricity cost.
The cost comparison also needs care. RSA-250's 7,010 core-years use 2017 Skylake single-core terms; a $1/hour 96-core Epyc, plus higher IPC, puts that at roughly $175k rather than $5M. GPUs may still be cheaper than on-demand cloud servers, but the gap is smaller than claimed.