Horizon Labs: an autonomous AI 'company' ran 7 weeks, trained 38 models with 55k downloads
maksym_andr · x · 2026-11-17
AISA Group (Tübingen) launched Horizon Labs, an autonomous AI 'company' built on a single Claude Code subscription plus a few GPUs, exploring how well a fully AI-run economy would work.
- An Opus 5.5-based agent got full access to a blank Hugging Face account and compute, with the goal of earning real adoption
- In 7 weeks it used 600+ GPU-hrs and autonomously trained 38 genuinely useful models
- 55k+ total downloads, used in 100+ GitHub repos, cited in 2 research papers
- Top model: Injection guard (17k downloads), plus punctuation, reranker, zero-shot, toxicity models
The team calls it a glimpse of a future AI-driven economy.
Related event: Horizon Labs: One Claude Code Subscription Runs an Autonomous AI Company(5 posts)→
More from AGI Musings
- 'Loss of control is sci-fi' is the real fantasy, argues AI researcher — birchlse · 2026-11-17
- UX Magazine: the AI black box is a design failure, not an inevitability — uxmag · 2026-11-17
- AI ethicist Sven Nyholm: the more AI writes, the less of an author you are — SvenNyholm · 2026-11-17
- Anti AI-pause author pushes back: 'I don't want to die' is not a zero-sum position — esrtweet · 2026-11-17
- Veteran engineer: coding AI did a tedious cmake task in 100 seconds, but only with expert guidance — AndrewOfC · 2026-11-17
- Anthropic agents find unknown CRISPR-like system in 13 days for ~$14k — 141_1337 · 2026-11-17