May 14, 2026ResearchRLAgents

Sara Hooker Launches AutoScientist - Models Train Themselves Now

Sara Hooker, the ex-Cohere VP of AI research who started Adaption Labs after walking out, just shipped AutoScientist. The pitch in one sentence: an AI that fine-tunes other AIs.

The system co-optimizes data and model together. Builds on Adaption's earlier Adaptive Data offering, the idea being that continuously improving datasets feed continuously improving models, and the whole loop runs without a human in the middle. Hooker says it more than doubled win rates across the models they tested.

What makes this worth watching is the framing. AutoScientist sits next to a cluster that has been building for months: AutoTTS discovering test-time scaling strategies, AlphaEvolve discovering algorithms, Continual Harness alternating between acting and refining its own prompts and skills, SLIM auto-managing skill lifecycles. The pattern is consistent - the meta-layer is no longer hand-designed. Whatever humans used to do above the model, an agent now does instead.

Hooker's quote is the part to underline: this suggests we can finally allow for successful frontier AI training outside of the big labs. That is a thesis statement. If she is right, the moat collapses from compute-plus-data to compute-plus-data-plus-meta-loop, and the meta-loop is shippable software.

https://adaptionlabs.ai
← Previous
Fractile Raises $220M Series B for Agent-Scale Inference Chips
Next β†’
Map-then-Act Drags Frontier Models From Zero to 88% on ARC-AGI-3
← Back to all articles

Comments

Loading...
>_