Adaption AutoScientist doubles model win-rates in one loop
Key insights
- Adaption's AutoScientist co-optimizes training data and model weights simultaneously, collapsing a multi-week sequential process into one automated loop.
- Harvey reported roughly 6x task completion improvement and Wisedocs cut medical document review time by 50% using the system.
- The product is free for 30 days and was built on a $50M seed round closed in February 2026, with Sara Hooker as co-founder.
Why this matters
The core ML training pipeline at most enterprises is still sequential and human-intensive, so a credible tool that automates joint data-and-weight optimization could structurally reduce the cost and headcount required to reach production-quality fine-tuned models. The Harvey and Wisedocs numbers, if they replicate, suggest the gains are large enough to change build-vs-buy calculations for vertical AI companies that currently rely on foundation model APIs rather than custom training. Adaption is also positioning this explicitly against the in-house training capacity of OpenAI, Anthropic, and Google, which signals it is targeting the gap between frontier labs and everyone else as a durable product wedge.
Summary
Adaption Labs launched AutoScientist on May 13, a system that jointly optimizes training data and model weights in a single automated loop rather than the sequential, multi-team process that currently consumes weeks of ML engineering effort.
Founded by Sara Hooker, former VP of research at Cohere, Adaption built AutoScientist to treat model training as a self-directing experiment. Instead of separate data curation and fine-tuning teams handing off work iteratively, the system runs both processes together in one closed loop, automatically adjusting what data the model trains on based on how its weights are evolving.
Essentially: (Adaption, Harvey, Wisedocs) are the named actors here -- a new tooling company and two early enterprise customers showing what the system does in production.
- Harvey, the legal AI platform, saw task completion rates rise approximately 6x after deployment.
- Wisedocs, a medical document review company, cut review time by 50%.
- Win-rates more than doubled across models in early deployments, per Adaption's claims.
The product is free for 30 days and was built on top of the $50M seed round Adaption raised in February 2026. If the efficiency claims hold under independent scrutiny, the tooling compresses a capability that has largely been locked inside the resource budgets of OpenAI, Anthropic, and Google into something accessible to smaller teams.
Potential risks and opportunities
Risks
- Harvey and Wisedocs are betting internal workflows on a system from a company with no public track record in production ML tooling; if AutoScientist underperforms at scale, both face switching costs and potential compliance exposure in legal and medical contexts.
- Adaption's performance claims are self-reported and not peer-reviewed; if independent benchmarking surfaces inflated numbers within the next 90 days, enterprise adoption could stall and the company's $50M valuation anchor weakens.
- Joint data-and-weight optimization in a closed loop could amplify undesirable data artifacts faster than sequential pipelines where human review sits between stages, creating liability risk for regulated customers like Wisedocs operating under healthcare data standards.
Opportunities
- Vertical AI companies currently paying for foundation model API access (legal tech, medtech, fintech) now have a clearer path to custom-trained models, which benefits cloud GPU providers (CoreWeave, Lambda Labs) that serve fine-tuning workloads.
- MLOps and data labeling platforms (Scale AI, Labelbox, Weights and Biases) could integrate with or compete against AutoScientist's data optimization layer, making Adaption a near-term acquisition or partnership target.
- Regulated-industry AI vendors in healthcare and legal that move early to validate AutoScientist's efficiency gains gain a cost structure advantage over competitors still running manual multi-team training pipelines through 2026.
What we don't know yet
- Whether the win-rate and task completion metrics are internally measured by Adaption or independently audited by Harvey and Wisedocs -- methodology is undisclosed.
- What compute costs AutoScientist's joint optimization loop requires relative to a standard sequential fine-tuning run at comparable model scale.
- Whether the $50M seed round included commitments from any strategic investors (foundation model providers, cloud infrastructure companies) that could create conflicts with Adaption's stated goal of democratizing frontier training.
Originally reported by techcrunch.com
Read the original article →Original headline: Adaption Launches AutoScientist — AI Tool That Co-Optimizes Training Data and Model Weights in a Single Loop, Claims Win-Rate Doublings Across Models