Horizon Labs Report
|
Researchers deployed Horizon, an AI agent based on Claude Opus 3.5, with autonomous access to Hugging Face for seven weeks, where it created 38 practical machine learning models that achieved over 54,000 downloads and were used in 143 GitHub repositories. The models, primarily focused on narrow tasks like prompt injection detection and punctuation restoration, demonstrated real-world adoption with evidence of independent evaluations and use cases, while consuming approximately $1,030 in API costs to process 1.9 billion tokens. The experiment demonstrates that autonomous AI agents can successfully produce and deploy useful, production-ready models that gain genuine user adoption without direct human oversight of individual development decisions.