The pursuit of artificial intelligence systems capable of autonomous self-improvement has long been a primary objective for researchers. Adaption, a research-focused AI startup, has taken a significant stride toward this goal with the introduction of AutoScientist. This new tool is designed to streamline the fine-tuning process, enabling AI models to acquire specific capabilities more efficiently through automated optimization.

Led by CEO Sara Hooker, formerly the VP of AI research at Cohere, Adaption developed AutoScientist to co-optimize both data and model architecture. By automating the iterative learning process, the tool aims to democratize access to frontier-level AI training, potentially reducing the heavy reliance on massive, centralized labs. The system functions as a natural evolution of the company’s previous offering, Adaptive Data, which focuses on the continuous curation of high-quality datasets.

While traditional industry benchmarks like ARC-AGI or SWE-Bench are not directly compatible with the highly specialized, task-specific nature of AutoScientist, early internal testing suggests significant performance gains. Adaption reports that the tool has more than doubled win-rates across various models. To encourage widespread adoption and validation, the company is offering free access to the platform for the first 30 days following its release.

Ultimately, Adaption envisions a future where the entire AI development stack is dynamic, adjusting on-the-fly to meet specific user objectives. By lowering the barrier to entry for model refinement, the company hopes to accelerate innovation across diverse scientific and technical fields, effectively allowing for more sophisticated AI development outside of traditional, resource-intensive environments.