Hugging Face has released ML Intern, an agent that carries a machine learning experiment from a chat message to a finished artifact on the Hub without anyone opening a terminal. In the run the company demonstrated, it worked for roughly six hours and cost under $0.50. The tool arrives while Hugging Face is itself being acquired by Nvidia, whose chief executive Jensen Huang has promised to keep the platform open and not tied to any particular hardware.
Once started, ML Intern handles the whole sequence on its own: creating datasets, training models, monitoring running jobs, uploading results to the Hub, writing reports and assembling demos. Each training run gets its own dashboard for following progress.
Source: the-decoder.com
The price is the number worth sitting with. Under $0.50 across about six hours works out to less than nine cents an hour, which is far below the going rate for on-demand accelerators. That tells you what the demo actually was: a small model on a small dataset, with most of those six hours spent waiting rather than computing. It does not tell you what a real experiment costs, and Hugging Face did not say.
The stated effect is to lower the barrier for starting new machine learning projects on the platform, and it plainly does that. It also changes what Hugging Face is. The company already hosts the datasets, the models and the training runtime; an agent that decides on its own to spin up a job turns the Hub from a place where people put artifacts into a place that generates compute demand.
That is the part I would watch through the acquisition. Huang's commitment to hardware neutrality is about what the platform will run on. It says nothing about how much the platform will run. An agent that makes it trivial to launch training jobs is, for the company buying the storefront, the most valuable feature Hugging Face could have shipped this year, and it shipped it mid-deal.
Missing from the announcement is any account of what ML Intern produces when it is not being demonstrated. There is no benchmark, no comparison against a human-configured run, and no description of what the agent does when a job fails or a result is bad — whether it stops, retries, or writes a confident report about a model that does not work. Automating the writing of the report is the easy half. Knowing whether the report should be believed is the half that has not been shown.