AI News

This study asks whether AI agents that improve hardware designs understand the underlying machine or merely search effectively over parameters. AutoTuring gives the same agent the same 15-dimensional accelerator design space in two forms: one exposes named architectural knobs and simulator counters, while the other presents anonymous variables ranging from 0 to 1. The evaluator, legal design space, and reachable optima remain identical, so the performance gap measures the value of meaning in the problem representation. On a nine-kernel FP16 GEMM workload, the architecture-aware setup outperformed a modeled H200 by 5.4% on average and exceeded the blind setup by 12.3%, while using 70.1% fewer simulator calls. The advantage was not unique to architectural knowledge: adding a critic loop to the blind agent recovered most of the gap, while giving the architect agent the same loop produced no additional benefit. The authors therefore interpret architectural knowledge and structured critique as substitutes in this setting. They describe the results as preliminary, based on five to six runs per condition on one modeled accelerator, and identify the controlled comparison itself rather than the resulting accelerator as the paper's main contribution.