Summary: Adaption launched AutoScientist on May 13, 2026 to automate the training and alignment loop around a user-defined goal. Reported win rates improved from 48% to 64%. Those results used internal domain benchmarks rather than public leaderboards.
On May 13, 2026, Adaption, the AI lab started by former Cohere VP of AI research Sara Hooker, took the wraps off AutoScientist — a system that hands the entire research loop behind model training and alignment over to automation. Training data and training recipes are tuned side by side, run after run, until the model settles on the objective the user has spelled out. By the company’s telling, an idea turns into a fine-tuned model you fully own “in an afternoon.”
Where it sits in the market is what makes this notable. Fine-tuning tooling today generally takes on a single step — scrubbing the data, or sweeping hyperparameters. AutoScientist is selling automation of the whole training-research pipeline, and it is free for the first 30 days after launch. Hooker’s remark to TechCrunch reached past the product: it “suggests we can finally allow for successful frontier AI trainings outside of these labs.”
How the AutoScientist Loop Runs
Adaption bills itself as a research-driven “neolab.” Before founding it, Hooker ran AI research at Cohere, and in an October 2025 interview with TechCrunch she made clear her bet sits outside the scaling race. AutoScientist is that bet turned into shipping software — no one-shot fine-tuning utility, but a self-improving system that walks the training research loop end to end on its own.
The mechanics are easy to state: you name the capability you want, and the system reworks dataset composition and the training recipe at the same time until the model converges on that goal. Two groups are in its sights — ML engineers who would rather skip manual tuning, and non-technical builders whose only tool so far has been the prompt. The tagline on the company’s site says it without ornament: “Intelligence should not arrive preconfigured, and building AI shouldn’t require a PhD.”
What the Figures Show: Win Rates 48% to 64%
Three batches of numbers came out with the launch. Batch one: set against training configurations proposed by human or AI researchers, the training AutoScientist produces averages 35% better results. Batch two: win rates rose from 48% under researcher-recommended configurations to 64%. Batch three: the gains were said to hold across 8 verticals, datasets running from 5k to 100k examples, and the model architectures open to fine-tuning through Together AI.
The catch is where the evidence comes from: evaluation ran on Adaption’s own in-house, domain-specific benchmarks rather than public leaderboards. TechCrunch’s note stands — without external comparison points, win-rate claims like these are hard to place in context.
From Adaptive Data to a Stack That Adapts Itself
AutoScientist didn’t come out of nowhere. Adaptive Data, the earlier Adaption product, is about piling up high-quality datasets over time; AutoScientist turns that stream of ever-better data into ever-better models. What Hooker wants is a stack that is “completely adaptable,” optimized on the spot for any task it faces.
One line of the roadmap goes further still: adaptation techniques that skip training entirely. If they materialize, the cost and latency of adapting a model drop by yet another order of magnitude.
Why Off-the-Shelf Benchmarks Fall Short
The pipeline builds a model fitted to one task at a time, so general benchmarks like SWE-Bench and ARC-AGI can’t score it directly — aim the same automation at a different objective and a different model comes out. That cuts both ways: users weigh “more accurate on my task” above leaderboard placement, while outside observers have nothing but the vendor’s own benchmarks and a hands-on trial to inspect — which is why the tool is free for 30 days in the first place.
What Developers and Researchers Should Take Away
Three grounded calls. First, fine-tuning’s entry bar is moving from “can write a training script” down to “can describe a goal,” and that visibly shortens non-technical teams’ route to a proprietary model. Second, the claim that frontier training can come unhooked from the big labs finally has a usable product behind it, matching the accelerating model cadence that has run through 2026 so far. Third, the verification job has moved to the user: with no public leaderboard to lean on, testing against your own eval set before rollout beats taking a vendor’s win-rate chart on faith.