An AI system just ran its own research program and came out the other side with over a hundred new neural network designs that beat the state of the art.
Researchers built ASI-Arch, a closed-loop system that cycles through research, experimentation, analysis, and updating its own approach, without a human choosing the next step. They pointed it at linear attention, a technique that lets models handle long sequences without the usual heavy compute cost, and let it run. Over 1,773 iterative experiments, ASI-Arch discovered 105 architectures that outperformed existing baselines. Its best design improved on DeltaNet, a recent linear-attention architecture, by nearly three times the gain that Mamba2 managed over the same baseline.
That ratio matters more than the headline count. Architecture search is not new, but most of it has involved humans steering the machine through a search space. Here the system set its own research agenda and reportedly beat a result from a well-known hand-designed architecture. If that holds up, it is a real data point for AI labs offloading early-stage architecture hunting to AI itself, freeing researchers to focus on which results deserve deeper scrutiny.
Worth remembering: this is a preprint, already revised once, and the 105 "state-of-the-art" architectures were judged by the same closed loop that produced them. Self-graded homework still needs a second opinion before anyone retires the researchers.