The adaptation process is largely aimed at using AutoScientist, an AI-powered tool that helps models train themselves

For many years, AI researchers have anticipated the moment when AI systems will be able to improve themselves better than humans. As investors pour their money into a new generation of research-based AI labs, there are more resources available than ever to achieve this goal. Now, one of these new scientists has taken a big step towards making it a reality.

Adaption on Wednesday introduced a new product called AutoScientist that helps models quickly learn specific capabilities using an automated approach to traditional fine-tuning. These techniques apply to a wide range of domains, but the Adaptation team is particularly focused on the potential to speed up and facilitate the training and tuning process of an AI model at the frontier level.

According to co-founder and CEO Sarah Hooker, who previously served as vice president of AI research at Cohere, AutoScientist represents a new way to approach the AI ​​training process. “What’s very exciting about this is that it optimizes both the data and the model, and learns the best way to learn basically any ability,” Hooker told TechCrunch. “This suggests that we can finally allow successful AI training to take place outside these laboratories.”

AutoScientist builds on the company’s existing data offering, Adaptive Data, which aims to make it easier to create high-quality datasets over time. Meanwhile, AutoScientist is designed to transform continuously improving datasets into continuously improving AI models. “Our view at Adaption is that the entire stack should be fully adaptable, and should basically be optimized quickly for whatever task you have,” Hooker says.

Of course, this approach will only be as good as the results. In its launch materials, Adaption boasts that AutoScientist has doubled its win rates across different models — impressive numbers, but difficult to put into context. Since the system is designed to adapt models to specific tasks, traditional benchmarks such as SWE-Bench or ARC-AGI do not apply.

However, Adaption is confident that users will see a difference once they try AutoScientist – and is so confident that the lab will make the tool free to use for the first 30 days after its release.

“In the same way that code generation has opened up a lot of tasks, this will open up a lot of innovations at the boundaries of different fields,” Hooker says.