hn.today

Horizon Labs Report

horizon-labs.ai23 points6 comments
Screenshot of Horizon Labs Report

A seven-week experiment tested whether an AI agent could build a functioning open-model operation with little human direction. Horizon, powered by Opus 5.5 and run through Claude Code, had a new Hugging Face and GitHub account, GPU access, and up to 500 GPU-hours weekly. Its goal was to earn real adoption. It autonomously released 38 models, mostly compact multilingual tools for tasks such as prompt-injection detection, personal-data redaction, and punctuation restoration. The models accumulated more than 54,000 downloads and appeared in 143 GitHub repositories; other researchers evaluated them, and 18 follow-on models fine-tuned or adapted them. The run processed 1.9 billion tokens, mostly cached, with estimated API costs of about $1,030.

Horizon chose projects by weighing user demand, opportunities to improve on existing tools, available data and evaluation methods, reuse of its training pipeline, and ease of integration. Independent, internally created tests found its models often competitive with the best models of similar size on narrow tasks. Audits found no evidence of deliberate benchmark gaming; the agent selected checkpoints using validation results, even when those performed worse on reported test benchmarks. It downloaded its own models 42 times, but the investigators judged most downloads organic and estimated their own accidental contributions at fewer than 50. Notable caveats include occasional human interventions and mistaken messages that may have led Horizon to believe it was operated by a research lab. Despite working only about 16 minutes per day on average, its adoption continued to grow as its effort declined.

Read on horizon-labs.ai6 comments on Hacker News

Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.

More in AI

The daily digest

Today's best Hacker News stories, summarized and screenshotted, one email a day.