Adaption launches AutoScientist for automated AI fine-tuning

On Wednesday, AI research company Adaption introduced AutoScientist, a new tool designed to automate the conventional fine-tuning process. Fine-tuning is the process of taking a pre-trained AI model and training it further on a specific dataset to help it learn capabilities more quickly. According to Sara Hooker, co-founder and CEO of Adaption and former VP of AI research at Cohere, the tool represents a shift in how models acquire skills. Hooker noted that the tool is exciting because it co-optimizes both the data and the model to learn the best way to acquire any capability, allowing the system to adapt to specific tasks on the fly.

The new tool builds on Adaption’s existing data product, Adaptive Data, which aims to make it easier to build datasets over time. While Adaptive Data focuses on dataset development, AutoScientist is designed to turn those continuously improving datasets into continuously improving AI models. Hooker explained that Adaption’s view is that the entire stack should be completely adaptable and optimize on the fly for any given task. Because Adaption is confident that users will see the difference once they try AutoScientist, the lab is making the tool free to use for the first 30 days after its release.

Adaption claims that AutoScientist has more than doubled win rates across different models. However, evaluating these improvements using standard industry metrics is difficult. Because the system is designed to adapt models to specific tasks, conventional benchmarks are not applicable. These include SWE-Bench, which is an industry-standard benchmark for evaluating AI models on software engineering tasks, and ARC-AGI, which evaluates models on general intelligence tasks.

Despite the lack of traditional benchmark compatibility, Adaption expects the tool to lower the barriers to frontier-level AI training. Hooker compared the tool’s potential to the way code generation unlocked various tasks, noting that it will unlock innovation at the frontier of different fields. “It suggests we can finally allow for successful frontier AI trainings outside of these labs,” said Hooker, co-founder and CEO of Adaption. By automating the optimization process, the company believes the tool could allow successful frontier-level AI training—referring to advanced, state-of-the-art models—to occur outside of major AI labs.

Why it matters

The tool represents a potential shift toward making frontier-level AI training accessible beyond the largest, best-funded labs by automating the co-optimization of data and models.