Adaption, co-founded by ex-Cohere VP of AI research Sara Hooker, unveils AutoScientist, which can automate the research loop behind model training and alignment

For years, AI researchers have anticipated the moment when AI systems will be able to improve themselves better than humans could.

TechCrunch 2026-05-13 Russell Brandom

Context & Ripple Effects

Adaption previously raised a $50 million seed around continuously learning AI systems that it said could be cheaper to run than leading models. AutoScientist extends that premise from model operation into the research process itself, specifically training and alignment.

The announcement arrives alongside public efforts by OpenAI and Core Automation to automate AI research workflows, while Reflection has applied agentic systems to software-engineering knowledge work. The common arc is moving AI from assisting discrete tasks toward running larger portions of technical iteration.

First-order effects

Adaption now has a named system positioned to automate parts of the experimental loop behind model training and alignment, making research-process automation central to its product and company narrative.
Researchers working on training and alignment are the immediate intended users or counterparts: the system is designed to take on work that otherwise requires repeated human-led research cycles.

Second-order effects

Labs pursuing autonomous research agents, including OpenAI and Core Automation, face a clearer competitive benchmark around automating not just coding or analysis but the training-and-alignment loop itself.
If such systems reduce the human effort required per experiment, competition among AI developers can shift toward the quality of automated iteration, evaluation, and alignment workflows rather than model training alone.

Third-order effects

If automated research loops prove reliable, AI development could become more continuous and more concentrated in organizations able to operate and validate automated experimentation at scale.
Alignment becomes a more consequential control point: automating research may increase iteration speed, but it also raises the importance of determining whether automated changes remain aligned with intended objectives.

The trend: AutoScientist is part of the emerging push to turn AI R&D from a human-directed workflow into an increasingly automated, multi-agent research system.

Discussion

@sethbannon Seth Bannon on x
Very cool release. AutoScientist self-improves and automates the full research loop behind model training and alignment. Model training + RL are powerful but have been hard to get right outside frontier labs. This is a step towards letting anyone build models for their task.
@sarahookr Sara Hooker on x
We have been hard at work on this for the last few months. I think a few things are very interesting, self-improvement for only model training alone is difficult unless you co-optimize with data. Our work on adaptive data paved the way for incredibly exciting results here.
@sarahookr Sara Hooker on x
Without co-optimizing both data and model learning, AutoScientist has much less predictable gains because data quality is too high variance. Makes me more excited about end-to-end optimization and long task horizon AI R&D.
@sarahookr Sara Hooker on x
We found that AutoScientist beats hand engineered training configs from our research staff that took into account domain, dataset size, model type. I attribute this in part to the fact frontier AI staff are very used to working with a single family of models. In contrast, here [i…
@adaption_ai @adaption_ai on x
Introducing AutoScientist. Most model training fails outside of frontier labs. AutoScientist automates the full research loop so it doesn't have to. [video]
@sudip_r0y Sudip Roy on x
Less than 1,000 people know how to shape a frontier model. AutoScientist is our attempt to change that. Describe the outcome. It automates the rest.
@sarahookr Sara Hooker on x
What was remarkable was we found consistent results across verticals, model types and dataset sizes. AutoScientist unlocks on average much more predictable performance. I actually think this is underestimate, as we configured some of our compute budget to stop search after [image…

Chronicles