A New Startup Wants AI Models to Train Themselves

Adaption Labs, the AI startup co-founded by former Cohere VP of research Sara Hooker and former Cohere director Sudip Roy, came out of stealth this week with a product that aims to take a lot of the human grunt work out of fine-tuning. The tool, called AutoScientist, treats data curation and model training as a single closed-loop problem rather than two sequential steps. According to a TechCrunch piece on May 13, the company is offering it free for the first thirty days.

The pitch is concrete. A developer or research team uploads a base model and a description of what they want it to do, and AutoScientist iteratively generates training data, runs evaluations, identifies weaknesses and adjusts the dataset, all without a human pulling the levers between rounds. Hooker, in a launch post on the Adaption Labs blog, claims the system more than doubles win rates against baseline fine-tuned models across a range of internal benchmarks, and gets developers from idea to an owned, adapted model in an afternoon.

The strategic bet

That last clause is the strategic bet. The conventional view in 2024 and 2025 was that frontier capability was a function of model scale: more parameters, more compute, more data. Adaption's thesis is that the next frontier is smarter training rather than bigger models, and that small teams with focused datasets and a tight feedback loop can outperform much larger generalist systems on the tasks that actually matter to them. It lines up with what several researchers, including Hooker's own published work at Cohere, have been arguing for two years.

How novel is it?

How well it holds up in practice will take longer to judge. The benchmarks Adaption is citing are internal, and the company has not yet released the underlying methodology in a peer-reviewed paper, which is the more demanding test. Outside reviewers quoted in AIChief's writeup have noted that the closed-loop framing is not entirely novel, and that variants of self-distillation and active-learning fine-tuning have been around for years. The novelty, if there is one, is in the productisation: tying together what was previously a researcher's bespoke pipeline into something a small team can use without a PhD.

For the open-source community, the launch is good news. AutoScientist works on top of any base model, including the most recent open-weight releases from Mistral, Meta and the Chinese labs, and it lowers the barrier for non-frontier teams to specialise those models for their own data. Whether the same dynamic helps or hurts incumbents like OpenAI and Anthropic is less clear. A world in which fine-tuning gets dramatically easier is one in which a lot of customers stop renting frontier capability and start owning a smaller, sharper version of it. Adaption is betting that world is closer than the major labs would prefer.