ASI-Arch: An Automated Search for Neural Network Architectures

A 2025 paper presents ASI-Arch, a system for proposing, implementing and testing neural-network architectures. I find it exciting as an example of AI helping research on AI itself. The paper invokes an “AlphaGo moment,” but that is an analogy and the authors’ framing, not evidence that general artificial superintelligence has been achieved.

I argue that human research time can limit the rate of new AI ideas and imagine agents proposing hypotheses and running experiments at greater scale. That is a hypothesis about the future, not proof that human-led research grows only linearly or that automated progress will grow exponentially.

AlphaGo’s match with Lee Sedol provides the analogy. Its 37th move surprised commentators, and its value became clearer as the game developed. AlphaGo used training and search, including self-play; the original system also learned from human game records. Its success in Go does not by itself establish the capabilities of an architecture-search system.

The proposed ASI-Arch loop uses language-model agents and tools to hypothesize designs, implement code, train candidate models and analyze results. I question the use of “ASI” in its name because artificial superintelligence is a loaded term. The system works within a human-designed research setup, with a chosen domain, models, evaluation procedures and external knowledge.

The researcher role draws on previous experiments and human literature to propose new designs, using strong candidates as parents and references. The engineer role implements and tests candidates, handling code errors where possible. This automated debugging can prevent some promising experiments from failing only because of implementation mistakes, though it does not guarantee every error is fixed.

The analyst reviews training logs, benchmark results and other evidence, then records insights for later proposals. The researchers describe a shared memory of candidates and lessons. This creates an iterative search loop in which past results influence the next generation of architectures.

The original paper reports 1,773 initial experiments and approximately 20,000 GPU hours across its search and validation stages. It identifies 106 strong linear-attention architectures under its evaluation criteria. These results concern a particular architecture family, training budget and selected baselines; they do not mean 106 models outperform every existing architecture in every task.

The paper describes a two-stage process: initial exploration at small scale, then training selected candidates at 340 million parameters, followed by a more extensive validation of five designs against selected baselines. These details help define what was actually tested. They do not establish that the discoveries will transfer unchanged to frontier-scale models.

I imagine increasing the compute budget dramatically and running experiments in parallel. More resources could allow more trials, but the paper does not demonstrate that compute is the only limit or that innovation grows exponentially. Data, evaluation quality, implementation, available methods and the validity of the chosen benchmarks also constrain such a system.

I also wonder whether similar loops could contribute to biology or medicine. That is a possible direction, not a result established by this neural-architecture experiment. The authors released the framework and discovered architectures, and I see related work in AlphaEvolve, the Darwin Gödel Machine and Sakana AI’s AI Scientist as part of a broader effort to automate research.