Zhengyao Jiang describes Aiden, Weco's experimental autonomous research system, through its participation in the Parameter Golf competition. He argues that useful research should be judged not just by benchmark scores but by whether a community reviews, adopts and builds on the work. The talk's performance and compute figures are attributed to his account of that experiment.
Zhengyao Jiang traces a reported improvement to the combination of human research ideas, an implementation under a model-size limit and a tokenizer change from another contributor. His interpretation emphasizes idea discovery, execution and combination rather than a claim that agents independently originate all the advances they produce.
Zhengyao Jiang compares research-system design to model training: evaluations set the optimization target, while code abstractions constrain the search. A fraud-detection example illustrates how a permissive API can leak test information and how a stricter boundary can improve the result. He argues that better evaluations, interfaces and research judgment become more important as execution is automated.
Watch on YouTube
