Adaption's AutoScientist lets AI models train themselves, doubling win-rates
AutoScientist co-optimizes data and model to automate fine-tuning and boost performance.
Adaption, a neolab founded by former Cohere VP Sara Hooker, has unveiled AutoScientist—a system that enables AI models to train themselves via automated fine-tuning. Unlike traditional methods that treat data and model separately, AutoScientist co-optimizes both, learning the best approach to acquire any new capability. Hooker describes it as a breakthrough that could allow successful frontier AI training outside of well-funded labs, potentially transforming how models are specialized for tasks from code generation to scientific research.
AutoScientist builds on Adaption's existing Adaptive Data platform, which helps create continuously improving datasets. The new tool turns those datasets into continuously improving models, claiming more than doubled win-rates across various architectures. While conventional benchmarks like SWE-Bench don't apply—since the system adapts to specific tasks—Adaption is so confident in the results that it's offering AutoScientist free for the first 30 days. Hooker predicts it will unlock innovation at the frontier of multiple fields, much like code generation did.
- AutoScientist automates fine-tuning by co-optimizing both data and model parameters
- More than doubled win-rates across different models in Adaption's tests
- Free for the first 30 days; aims to enable frontier AI training outside major labs
Why It Matters
AutoScientist could democratize frontier AI training, enabling smaller teams to achieve high performance without massive resources.