Adaption's AutoScientist trains models on its own

Less than a thousand people on Earth actually know how to train and shape a frontier AI model. Most of them sit inside a handful of elite labs, leaning on unwritten tricks and tribal knowledge while everyone else is left contorting requests through prompt engineering.

AI lab Adaption launched AutoScientist, a tool designed to tear down that walled garden by automating the entire research and fine-tuning loop end-to-end.

Automating the black art of fine-tuning

Outside frontier labs, model fine-tuning usually fails for the same familiar reasons. Models suffer catastrophic forgetting, overfit on tiny datasets, or choke on mixed training signals.

AutoScientist tackles this by co-optimizing both the training data and the training recipes at the same time. Building on the company's existing Adaptive Data product, the system continuously adjusts data selection and hyperparameters side by side until the model converges on the target behavior.

According to Adaption, this turns a weeks-long ordeal of babysitting parameter sweeps into an afternoon project for developers.

Outperforming human AI researchers

In head-to-head internal benchmarks, Adaption pitted AutoScientist against configurations set by its own in-house research staff. Both had access to the same model types, domain information, and dataset sizes.

AutoScientist won 64% of head-to-head runs, up from 48% when human researchers set the configs. Across all runs, it delivered an average performance gain of 35% over human-configured setups.

The benchmark spanned 8 different verticals with dataset sizes ranging from 5,000 to 100,000 samples across models hosted on Together AI. CEO Sara Hooker, formerly VP of AI research at Cohere, told TechCrunch that the system is designed to unlock frontier-level training outside proprietary labs. The tool is free for the first 30 days.

Why it matters

For developers and enterprises, this could be the off-ramp from massive proprietary API bills. Many teams know a fine-tuned smaller model would cut latency and costs, but they cannot afford a specialized research team to babysit sweeps. A 30-day free window gives teams a real shot at testing whether they can turn proprietary data into owned models.

Across the industry, this shifts the battleground from prompt engineering to automated capability adaptation. The closely guarded secret of frontier labs has never just been compute; it is the human recipe book for training runs. Turning that craft into software chips away at that moat.

We think Adaption's decision to optimize data and recipes in tandem is clever, avoiding the trap of treating messy training data as fixed. However, there is a catch: conventional benchmarks like SWE-Bench or ARC-AGI do not apply here, meaning these win rates rely entirely on Adaption's internal domain evaluations. Whether that 35% edge holds up on messy real-world corporate data remains to be seen.