Core Mechanism: Joint Data-Model Optimization

AutoScientist advances beyond standard fine-tuning by co-optimizing both training data and model parameters simultaneously for targeted capabilities. This automated process builds on Adaption's Adaptive Data product, which generates evolving high-quality datasets, turning them into self-improving models tailored to specific tasks. CEO Sara Hooker explains it 'learns the best way to basically learn any capability,' adapting the full stack on the fly to user needs. Result: models acquire skills faster without manual dataset curation, applicable across fields but aimed at frontier AI.

Proven Gains and Accessibility

Adaption reports AutoScientist more than doubles win-rates across various models for task-specific adaptation—though standard benchmarks like SWE-Bench or ARC-AGI don't apply directly due to its customization focus. To prove value, it's free for the first 30 days post-launch, letting builders test real-world impact on their workflows. Hooker positions it as unlocking 'innovation at the frontier of different fields,' akin to how code generation expanded capabilities.

Broader Implications for AI Labs

This tool democratizes elite training: Hooker claims it enables 'successful frontier AI trainings outside of these labs,' countering the scaling race dominated by resource-heavy players. From her Cohere background, it shifts from human-led processes to AI-driven ones, aligning with investor bets on research labs pursuing self-improving systems.