Industry News
Handed over a company to Claude Opus 5.5 for seven weeks of operation: 38 small models were distributed and downloaded 54,000 times
A group of researchers conducted an experiment: they started a company called Horizon Labs and handed over control to an agent powered by Claude Opus 5.5.
The starting requirements are simple: a Team Premium subscription, an empty Hugging Face account, some GPUs, and a mission to develop AI models that other developers will want to use.
Then it's all on its own. Thinking of ideas, finding data, training, and publishing are all completed by the agent.
Results after seven weeks: 38 models released, with over 54,000 total downloads.
Most of what it does are small and specialized models, such as text classification, removing personal privacy information, and identifying malicious instructions. This choice is quite smart. To train a large model from scratch, seven weeks and a GPU are simply not enough. However, small models with narrow tasks can be trained quickly and have clear requirements. Developers can use them right away, and people will download them once they are built.
38 models, 54,000 downloads, an average of over a thousand each. It’s not a hit on Hugging Face, but for a “company” without human employees, it’s already a real user.
This type of experiment has been done before, mostly in a simulated environment, or by letting an agent run a virtual store. The difference this time is that the output is a real model, the users are real developers, and the number of downloads is real.
Of course, it also illustrates the boundaries of what intelligent agents can do at this stage: pick a narrow, verifiable track and execute it repeatedly. For it to decide in which direction the company will transform, there is no evidence yet.