- Joined
- Feb 2026
- How they joined
- Directly
- Before Cognition
- ImbueMember of Technical Staff
- Status
- Current
- Factored RSA-260 (a 260-digit, 862-bit RSA Challenge number), announced 2026-09-03 on X with a 130-digit factor and the words "divides RSA-260"; the Cognition write-up describes Devin building a GPU lattice siever (modified CADO-NFS) run on spare cluster compute, ~4,900 GPU-days.Sep 9, 2026
- Cognition blog, in his own words: "Over the past few weeks, the Cognition research team and I have been optimizing our job scheduler to better use disaggregated compute. As a proof of concept, and because I've enjoyed factoring numbers as a hobby for the past ten years or so, I drove a bevy of Devins to obtain a factorization of RSA-260.".Sep 9, 2026
- Cognition blog: "my Devins built the world's highest-performance GPU lattice siever, which enables factoring numbers at 10x lower cost than the previous public state of the art."Sep 9, 2026
- Listed first author of 'Introducing FrontierCode' (Jun 2026), Cognition's benchmark of mergeable, high-quality code built with 20+ open-source maintainers; the best model scored 13.4% on the Diamond subset.Jun 8, 2026
- Joined Cognition as Member of Technical Staff (research) in Feb 2026, directly from Imbue rather than via Cognition's July 2025 acquisition of Windsurf.Feb 2026
- Factoring is a long-running hobby: he ran P-1 factoring for the Great Internet Mersenne Prime Search and held 4th place on the GPU to 72 largest-factor leaderboard as of Dec 2024, and published penlu/factoring for cracking FactorDB numbers.Dec 2024
- Said the factoring was squeezed in around other work: "this was the high priority workload that I was fitting the factoring around. more to come!" alongside a post that visualizations were made with SWE-2.Sep 10, 2026this was the high priority workload that I was fitting the factoring around. more to come!
- Co-author on 'SWE-1.7: Frontier Intelligence at a Fraction of the Cost' (Jul 2026), covering RL entropy preservation, multi-cluster fault-tolerant training and self-compaction; on X he wrote "scaling this has been giga fun. more to come."Jul 8, 2026
- Co-author on 'Introducing SWE 1.6: Improving Model UX' (Apr 2026), a model shipped in Windsurf post-trained to cut overthinking, sequential tool calls and looping, at up to 950 tok/s.Apr 7, 2026
- Co-author on 'An Early Preview of SWE-1.6 and Research Update' (Mar 2026), which reports a 6x training speedup and SWE-1.6 scoring 11% higher than SWE-1.5 on SWE-Bench Pro.Mar 1, 2026
- Before Cognition, MTS at Imbue (Feb 2025-Jan 2026): product engineering for Sculptor, a parallel background agent product, and automated rubric optimization for LLM bug-finding and instruction adherence.Feb 2025
- Posted that Devin rebuilt in Lean in under 15 minutes a formalization that took him much of a semester nine years earlier: "cost of formalization has fallen off a cliff."; links to his undergraduate Coq formalization of special relativity cited by HepLean.Oct 29, 2026what nine years ago took me much of a(n admittedly inefficient) semester to partially formalize took devin... less than 15 minutes to rebuild in lean and finish off. cost of formalization has fallen off a cliff. it might even be a net speedup now. times are changing!
research
- Product engineering for Sculptor, a parallel background agent product - Automated rubric optimization for LLM-based bug-finding and instruction adherence
PhD student under the supervision of Prof. Eddie Kohler.
Conducted applied tensor decomposition research: - Applied tensor decompositions towards replicating results from Alzheimer’s gene expression study - Clustered drugs based on Broad Institute RNA expression microarray data - Extended C library with alternating Poisson regression tensor decomposition algorithm - Contributed technical documentation and project result summaries
Developed core services and infrastructure in Scala: - Added support for future timeseries correlation matrices - Improved cone plot asymptotic performance (producing 100-1000x speedup) - Added specs2 test suites for key data structure - Debugged MemSQL database queries during transition from MongoDB
Advisor: Eddie Kohler Dissertation: "Efficient Symbolic Execution and Reasoning for Low-Level Code"
joint concentration in computer science and physics Thesis: "A Formalization of Elements of Special Relativity in Coq"
what nine years ago took me much of a(n admittedly inefficient) semester to partially formalize took devin... less than 15 minutes to rebuild in lean and finish off. cost of formalization has fallen off a cliff. it might even be a net speedup now. times are changing!
this was the high priority workload that I was fitting the factoring around. more to come!
these visualizations were made with SWE-2, which is now available! https://t.co/mM72XWaMEX
this effort was a testament not only to the prowesses and conveniences of Devin Cloud but also to the sophistication of CADO-NFS and the absurd strength of modern GPUs!
@makmmas @cognition both (block Wiedemann in this case). discussed in the post; check out the perf numbers in the appendices!
here's the writeup of the factorization! https://t.co/z25MUI9hjI
among the LLM behaviors, waytoomany LoC output hurts me most I am not aware of SWE benchmarks other than FrontierCode wherein writing more code will eventually reduce your score
@cognition scaling this has been giga fun. more to come.
@vh1ne @cognition https://t.co/kkAwFcAgpl
(the vague question: why is the past different from the future? yeah sure second law; I mean why should the direction in which cognition seems to proceed be the same as the direction of increasing entropy? is that universal among observers or incidental to humans?)
ok the volition part seems well in hand. but it is still always either incoherent or interpolated
we continue to try to capture the true distribution of usage. there is still a long way to go. more to come!
