What was already known
The comparison is against the tested library routines on the paper’s hardware and inputs, rather than the asymptotic comparison-sorting lower bound.
What changed and why it matters
A deployed improvement to a widely used algorithmic primitive has substantive reach despite being a constant-factor optimization.
The AI contribution
Reinforcement learning searched low-level programs; humans translated and integrated the selected routines into the library.
Limitations
Speedups depend on implementation, processor and data. This does not beat the comparison-sorting asymptotic lower bound. Later human routines can outperform particular AlphaDev kernels. AIProofWatch has not rerun the benchmarks.
Verification evidence
Evidence labels show what supports the claim. Importance rates its impact if correct.
Author claim
Reported measurements and library integration.
Accessible artifact / proof
Released routines, not benchmarked by AIProofWatch.
Peer review
Published research; this is not independent reproduction of the speedups.
Correction and publication history
Public-discussion audit found this substantive scoped contribution missing. Source statement, baseline, attribution and importance reviewed; no full proof audit. Historical addition without subscriber alerts.
Read this published revision