AI researchers have long anticipated the moment when AI systems will be able to self-improve better than humans. As investors pour money into a new generation of research-driven AI labs, more resources than ever are available to pursue your goals. Now, one of those neolabs has taken a big step toward making that a reality.
On Wednesday, Adaption announced a new product called AutoScientist that allows models to quickly learn specific features using an automated approach to traditional fine-tuning. Although this technique can be applied to a wide range of fields, the Adaptation team is particularly focused on its potential to speed up and ease the process of training and fine-tuning frontier-level AI models.
According to co-founder and CEO Sara 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 really interesting is co-optimizing both the data and the model, learning the best way to learn basically any feature,” Hooker told TechCrunch. “This suggests that, ultimately, frontier AI training can be successful outside of these labs.”
AutoScientist builds on the company’s existing data product, Adaptive Data, which aims to make it easy to build high-quality datasets over time. AutoScientist, on the other hand, is designed to turn continuously improving datasets into continuously improving AI models. “Our view at Adaption is that the entire stack should be fully adaptable, optimizing on the fly for essentially any task,” says Hooker.
Of course, the approach is only as good as the results. Adaption boasts in its announcement that AutoScientist has more than doubled the win rate across various models. This is an impressive number, but it’s difficult to put into context. Because this system is built to adapt models to specific tasks, traditional benchmarks such as SWE-Bench and ARC-AGI are not applicable.
Still, Adaption is confident that users will see a difference once they try AutoScientist, and the lab is confident in making the tool available for free for the first 30 days after release.
“Just as code generation freed up a lot of tasks, this is going to free up a lot of innovation on the front lines in a lot of different fields,” Hooker says.
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
