It took AI researchers weeks to tune what AutoScientist handles in hours — and it outperformed their handcrafted configurations by 35%.
Wait, how does this even work?
Fine-tuning is the process of taking a general-purpose model like GPT and retraining it for a specific task — "legal document analysis" or "customer support." The idea is simple. The execution has always been brutal.
You need to decide which data to use, which data to discard, what learning rate to set, how many epochs to run, which loss functions to apply. The combinatorial search space is enormous. That's why real fine-tuning has historically required research-lab-level expertise.
AutoScientist automates the entire loop. It co-optimizes data selection and the training recipe simultaneously, running closed-loop iterations until it converges on your objective. Every existing tool either optimizes data or training config — not both at once.
Sara Hooker, the CEO, was previously VP of AI Research at Cohere and spent five years at Google DeepMind. In February 2026, Adaption raised $50M from Emergence Capital, Mozilla Ventures, and Fifty Years. Co-founder Sudip Roy was Cohere's head of inference. This isn't a wrapper startup — it's the team that knows AI training best.
Why fine-tuning was so hard before
Three classic fine-tuning failure modes: (1) Catastrophic forgetting — learning new things erases existing capabilities. (2) Overfitting — perfect on training data, broken in production. (3) Conflicting signals — contradictory training data confuses the model. AutoScientist is designed to automatically detect and route around all three.
What do the numbers actually show?
Adaption's internal benchmarks: AutoScientist vs. configurations designed by their own AI researchers — across 8 verticals, dataset sizes from 5K to 100K examples, and 100B+ parameter model architectures from Together AI. Average performance uplift: 35%. Win rate: 48% to 64%.
| Traditional fine-tuning | AutoScientist | |
|---|---|---|
| What gets optimized | Data or recipe (separately) | Data + recipe simultaneously |
| Time to model | Weeks (manual experimentation) | Hours (automated convergence) |
| Expertise required | Senior ML engineer essential | Accessible without deep ML knowledge |
| Data handling | Full dataset used (noise included) | High-signal auto-selection, toxic noise filtered |
| Failure modes | Forgetting, overfitting, signal conflicts | Automatic detection and avoidance of all three |
The honest caveat: these are internal benchmarks from Adaption themselves. Standard evaluations like SWE-Bench or ARC-AGI don't apply because AutoScientist is built for task-specific adaptation, not general benchmarks. Independent verification waits for customer results after the free trial period ends in mid-June 2026.
But the core claim is significant: frontier-level model training is now possible outside of Big Labs. Hooker sees AutoScientist the way code generation unlocked new capabilities — a step toward democratizing AI training itself.
The essentials: how to start
- Define your objective
What specific task do you need the model for? "Customer support email automation," "legal document summarization," "domain-specific code generation" — the more specific, the better. - Prepare your data
5,000+ examples is enough to get started. The data doesn't need to be perfectly curated — AutoScientist determines what's high-signal on its own. - Launch AutoScientist
Free 30-day trial at adaptionlabs.ai. Input your objective and let it run the co-optimization loop. - Watch convergence
The system iterates across data and recipe combinations automatically — the equivalent of hundreds of manual experiments a researcher would run. - Deploy your model
Built on Together AI infrastructure, supporting 100B+ parameter models. Ready for cloud serving once optimization completes.
No independent verification yet
All performance claims are from Adaption's internal benchmarks. External validation is still pending. The 30-day free trial is the best way to verify it on your own data before committing.
