Adaption launches AutoScientist to automate frontier AI model fine-tuning
Co-founder Sara Hooker says the automated system represents a shift toward fully adaptable AI stacks, with performance claims that defy conventional benchmarking.
Adaption has introduced AutoScientist, an artificial intelligence tool designed to accelerate the training and fine-tuning of frontier-level models through an automated approach. The system co-optimises data and model parameters to help AI systems adapt to specific capabilities quickly, marking a departure from conventional fine-tuning methods.
The launch positions Adaption as a key player in a sector where investors are increasingly funding research-driven labs with the goal of creating self-improving AI systems. Co-founder and chief executive Sara Hooker, who previously served as vice-president of AI research at Cohere, stated that the technology aims to enable successful AI training outside of major research laboratories.
AutoScientist builds upon the company’s existing Adaptive Data offering, which focuses on constructing high-quality datasets over time. According to Hooker, the new tool is designed to convert those continuously improving datasets into continuously improving AI models. “Our view at Adaption is that the whole stack should be completely adaptable, and should basically optimise on the fly to whatever task you have,” she said.
In its launch materials, Adaption reported that AutoScientist has more than doubled win-rates across different models. The company noted that because the system is built to adapt models to specific tasks rather than general benchmarks, conventional evaluation metrics such as SWE-Bench or ARC-AGI are not applicable.
To allow users to test the technology, Adaption is offering AutoScientist free of charge for the first 30 days following its release. Hooker compared the potential impact of the tool to the unlocking of code generation, suggesting it will facilitate significant innovation at the frontier of various fields.
The specific methodology behind the reported performance improvements is not detailed in the source material, making it difficult to verify the magnitude of the gains. Furthermore, the tool’s effectiveness in real-world scenarios outside of internal testing remains unproven, as the lack of standardised benchmarks complicates external validation.
The vision for enabling frontier AI training outside major labs remains a stated goal rather than an established outcome. As the industry moves toward more automated and adaptable systems, the long-term viability of this approach will depend on how well the tool performs in diverse, uncontrolled environments.
