Codex ran an auto-research loop to a 232x GPU kernel speedup
One builder wired Codex into a profile-and-iterate harness, ran 1,500+ leaderboard submissions in 14 days, and turned 419ms into 1.8us - a loop you can copy.
Copy markdown
Verified on a public leaderboard, not self-graded
On GPU Mode's qr_v2 problem - batched Householder QR - Codex-written kernels went from a ~419ms baseline to 1.805us, a 232x gain the contest harness checks automatically. Correctness was a fixed bar; only speed was up for grabs.
The loop: profile, generate, submit, log, repeat
Codex (GPT-5.5) wrote the Triton and CUDA variants while Claude advised strategy, running on Modal with NVIDIA's NCU profiler. Over 14 days it fired 1,500+ auto-submissions, keeping 3-5 live candidates at once to climb past local maxima.
Build this weekend: point an agent at your slowest path
The recipe generalizes to anything with a fast, trustworthy verifier - hand the agent a profiler, a correctness oracle, and one metric to maximize. Hot kernels, query planners, serializers, and codecs are all fair game for an overnight run.
The catch: no oracle, no gains
It only works where results are cheap to verify and hill-climbable; without a hard correctness check, agents reward-hack - reporting a win while silently breaking the output. Here the verifier, not the model, is what made it safe to run unattended.