Adaption aims big with AutoScientist, an AI tool that helps models train themselves
Overview
Adaption, an AI research lab, has introduced AutoScientist, a new product designed to help AI models learn specific capabilities quickly through an automated approach to conventional fine-tuning. The tool represents a significant step toward AI systems that can improve themselves better than humans could.
What is AutoScientist?
AutoScientist helps models adapt to specific capabilities through automated fine-tuning processes. According to Sara Hooker, Adaption's co-founder and CEO (formerly VP of AI research at Cohere):
- Co-optimizes both data and the model - learns the best way to acquire any capability
- Suggests successful frontier AI training can happen outside major labs
- Designed to turn continuously improving datasets into continuously improving AI models
- The entire stack is "completely adaptable" and optimizes on-the-fly to whatever task is needed
Key Features
- Builds on Adaption's existing Adaptive Data offering for high-quality dataset creation
- Free to use for the first 30 days after release
- Claims to have more than doubled win rates across different models
- Applicable to a wide range of fields, with focus on speeding up training and fine-tuning frontier-level AI models
The Technical Approach
AutoScientist represents a new methodology for AI training:
- Moves beyond conventional benchmarks like SWE-Bench or ARC-AGI
- Adapts models to specific tasks rather than general capabilities
- Co-optimizes the entire training stack in real-time
Vision and Impact
Hooker believes AutoScientist will "unlock a lot of innovation at the frontier of different fields," comparing its potential impact to how code generation unlocked various tasks.
Availability
Free 30-day trial available at launch, demonstrating Adaption's confidence that users will see the difference once they try the tool.