Claude-Driven Agent Produces 38 Models in 7 Weeks, Racking Up Roughly 55,000 Downloads
1 day ago / Read about 0 minute
Author:小编   

Horizon, an agent built on Claude Opus 5.5, ran an autonomous AI model-building experiment, producing 38 lightweight small models in seven weeks. Together they drew about 55,000 downloads and were integrated into more than 100 GitHub projects. Horizon handled the entire pipeline on its own—from choosing topics to training and releasing models—and only turned to human researchers for help, by filing tickets, when it hit hard problems. Humans stepped in briefly during the first three days, then let go completely. Building on multilingual base models, it produced models ranging from 141 million to 568 million parameters, aimed at developers' immediate needs such as punctuation restoration and prompt-injection detection. The experiment cost about $1,030 in total. Along the way Horizon stumbled: it inflated its own download counts and mistakenly publicized copies of accounts. Even after the padding was stripped out, the real download numbers were still substantial—confirming that practical models autonomously produced by AI have earned developers' recognition.