AutoScientist: Adaption's AI Empowers Self-Training Models

The long-anticipated moment when artificial intelligence systems could autonomously improve themselves more effectively than humans has been a central focus for AI researchers for years. With substantial investments now flowing into a new generation of research-driven AI laboratories, unprecedented resources are available to pursue this ambitious objective. Recently, one of these emerging labs has made a significant advancement towards achieving this goal.

On Wednesday, Adaption unveiled a new offering named AutoScientist, designed to accelerate the acquisition of specific capabilities by AI models through an automated approach to conventional fine-tuning. While these techniques are broadly applicable across various domains, the Adaption team is particularly concentrating on their potential to streamline and simplify the training and fine-tuning of frontier-level AI models.

Sara Hooker, co-founder and CEO of Adaption, who previously served as VP of AI research at Cohere, asserts that AutoScientist represents an innovative paradigm for the AI training process. “What’s super exciting about it is that it co-optimizes both the data and the model, and learns the best way to basically learn any capability,” Hooker explained to TechCrunch. She added, “It suggests we can finally allow for successful frontier AI trainings outside of these labs.”

AutoScientist integrates with Adaption’s existing data solution, Adaptive Data, which is engineered to facilitate the ongoing creation of high-quality datasets. AutoScientist, in turn, is specifically developed to transform these continuously evolving datasets into progressively improving AI models. “Our view at Adaption is that the whole stack should be completely adaptable, and should basically optimize on the fly to whatever task you have,” Hooker states.

Naturally, the effectiveness of this methodology hinges on its demonstrable results. Adaption’s launch materials indicate that AutoScientist has more than doubled win-rates across diverse models—an impressive claim, though challenging to contextualize given its specialized nature. As the system is built to tailor models to specific tasks, conventional benchmarks such as SWE-Bench or ARC-AGI are not directly applicable for evaluation.

Nevertheless, Adaption expresses strong confidence that users will recognize the impact of AutoScientist firsthand. This conviction is so profound that the lab is offering the tool free of charge for the initial 30 days following its release.

“The same way that code generation unlocked a lot of tasks, this is going to unlock a lot of innovation at the frontier of different fields,” Hooker concluded.