IBM's open-model AI agents reportedly found a 52-fold speedup over a baseline chip design in synthesis estimates for $26 in inference

An article citing IBM Research says an all-open AI-agent setup searched for improvements to a quantum-error-correction decoder under a 100-hypothesis budget. Its best design showed a 52-fold speedup over the baseline in high-level-synthesis estimates, at about $26 in model-inference costs. The article compares the approach to earlier silicon compilers; it notes that the $26 excludes synthesis and verification costs and that each configuration was tested in one search run.