For years, experts in artificial intelligence have envisioned a future where AI systems can enhance their own capabilities more effectively than humans. With significant investments flowing into innovative AI research labs, the resources available for this pursuit have never been greater. Recently, one such lab has made a remarkable advancement.
On Wednesday, Adaption launched its groundbreaking tool, AutoScientist, which streamlines the model training process through an automated approach to fine-tuning. This innovative technique has broad applications, but the Adaption team is particularly excited about its potential to accelerate and simplify the training of cutting-edge AI models.
Sara Hooker, co-founder and CEO of Adaption and former VP of AI research at Cohere, emphasized that AutoScientist offers a novel perspective on AI training. "What's super exciting about it is that it co-optimizes both the data and the model, learning the most effective way to acquire any capability," Hooker stated. "This suggests we can finally enable successful frontier AI training beyond traditional lab environments."
AutoScientist builds upon Adaption's existing data product, Adaptive Data, which aims to facilitate the creation of high-quality datasets over time. In contrast, AutoScientist is designed to transform these continually improving datasets into evolving AI models. "At Adaption, we believe the entire system should be fully adaptable and optimize in real-time for any given task," Hooker added.
The effectiveness of this approach will ultimately depend on the results it delivers. Adaption claims that AutoScientist has more than doubled win rates across various models, a noteworthy achievement that is challenging to contextualize. Since the tool is crafted to tailor models for specific tasks, traditional benchmarks like SWE-Bench or ARC-AGI may not apply.
Nonetheless, Adaption is optimistic that users will recognize the benefits of AutoScientist, so much so that they are offering the tool free for the first 30 days post-launch.
"Just as code generation has unlocked numerous tasks, this tool is poised to drive significant innovation across various fields," Hooker stated, hinting at the transformative impact of this development.