Why we're watching this: AutoScientist puts automated model fine-tuning in the hands of non-experts for the first time. If its in-house benchmark claims survive real-world use, it redraws the economics of who can shape a frontier AI model and at what cost.
Key Takeaways
- AutoScientist raised win rates from 48% to 64% over human-configured fine-tuning in Adaption’s internal benchmarks, across 8 verticals and multiple model architectures
- The system co-optimises data and training recipes simultaneously, running the full research loop automatically without requiring an ML engineer to babysit the process
- Free for 30 days from launch, with Adaption targeting developers, non-technical builders, and enterprises sitting on proprietary data they cannot currently surface to AI
Adaption has launched AutoScientist, a system that automates the full research loop behind AI model training, aiming to put a capability previously locked inside a handful of frontier labs into the hands of any team with a dataset and a use case.
The product builds on Adaptive Data, Adaption’s existing data optimisation offering, and extends it into the model layer. Where Adaptive Data shapes training inputs, AutoScientist shapes the model itself, co-optimising data and training recipes in lockstep until the model converges on a defined objective.
“What’s super exciting about it is that it co-optimises both the data and the model, and learns the best way to basically learn any capability,” co-founder and CEO Sara Hooker told TechCrunch. “It suggests we can finally allow for successful frontier AI trainings outside of these labs.”
Hooker, who previously served as VP of AI Research at Cohere and spent years at Google Brain, founded Adaption Labs in late 2025 on the thesis that scaling large language models had become an inefficient path to better AI. AutoScientist is the first concrete product to emerge from that bet.
In Adaption’s internal testing, AutoScientist outperformed configurations set by the company’s own AI research staff by an average of 35%, with win rates climbing from 48% to 64% across runs spanning dataset sizes from 5,000 to 100,000 samples, multiple model architectures via Together AI, and all eight verticals tested. The gains were consistent regardless of domain, which Adaption frames as evidence the system is not overfit to any single task type.
“The same way that code generation unlocked a lot of tasks, this is going to unlock a lot of innovation at the frontier of different fields.” — Sara Hooker, CEO, Adaption Labs
The credibility question is the one Adaption cannot yet answer. Every benchmark in the launch materials is in-house. There is no independent replication, no SWE-Bench equivalent, and no third-party validation of the win-rate claims. Adaption’s 30-day free access window is an implicit acknowledgment that the product needs to prove itself in real environments, not just internal ones.
All published benchmarks are Adaption’s own. No independent third-party validation has been released. Teams evaluating AutoScientist should run their own domain-specific evals before committing to production use.
The market context makes the claims worth testing. OpenAI reportedly requires customers to spend upward of $10 million to access its fine-tuning consulting services, a threshold that locks out most of the companies sitting on proprietary data. If AutoScientist delivers consistent gains without that price tag or the need for specialist ML staff, the addressable market is effectively every enterprise that has ever been told fine-tuning was too expensive or too complex.
Adaption says real-time adaptation without any training at all is next on the roadmap. For professionals tracking where AI capability is being democratised and where new cost curves are forming, Relve is the intelligence layer between those shifts and the teams deciding whether to act on them.
