saved
Factoring RSA-260
Hraness wrote this summary from a saved copy of the source. Quotations are taken word for word from the source.
gist
Cognition researcher Eric Lu drove a Devin swarm to factor RSA-260 for about $400k (~4,900 GPU-days), setting a new public RSA Factoring Challenge record after RSA-250 (2020). Agents built a GPU lattice siever on modified CADO-NFS that is about 10× cheaper than prior public SOTA; RSA-1024 is estimated around $30M while RSA-2048 stays roughly a billion times harder. No algorithmic breakthrough is claimed—performance engineering on spare disaggregated cluster compute, with the human supplying executive function.
ideas
- Agents closed a public factoring record via performance engineering. A Devin swarm modified CADO-NFS into a GPU GNFS stack and factored RSA-260 without claiming new algorithms.
- Cost drops change who can try, not RSA-2048. GPU sieving ~10× cheaper than prior public SOTA; RSA-1024 ~$30M at market GPU prices; RSA-2048 remains ~billion× harder than RSA-1024.
- Spare cluster shards are a real research substrate. Lattice sieving fits preemptible single-node scraps from NVL72 packing inefficiencies, so the run used spare/disaggregated compute at negligible marginal cost.
- Humans still supply executive function. Priority-setting, benchmarks, redirection, and a unified measurement framework kept parallel Devins productive over ~3 weeks of concurrent sessions.
- Human-engineered decompositions still unlock agents. CADO-NFS’s staged interfaces let the swarm optimize pieces independently; agents grew more confused farther from upstream.
quotes
“My role was primarily to set priorities, establish benchmarks, and recognize when work was going off-track.”
“I report essentially no algorithmic advancements”
“Devin needed me for executive function, talking through and sanity-checking what it was doing”
“About three weeks elapsed from first prompt to RSA-260 factors found.”