Sara Hooker is an Adaption cofounder studying efficient AI systems that learn as their environment changes. She previously led research at Cohere and worked as a research scientist at Google DeepMind.[1][6]
Start with The Hardware Lottery →
X · sarahookr Career
At Cohere, Hooker was Vice President of Research and led Cohere Labs. Before that she worked at Google DeepMind. Her owned biography reports a computer-science PhD at Mila supervised by Hugo Larochelle and Aaron Courville. Her current work at Adaption explores how data, models and interaction can adapt together rather than depending only on larger static training runs.[1]
Contributions
AutoScientist searches training recipes across data preparation, model configuration, evaluation and alignment. Adaption’s launch describes a loop that adapts these choices to the task instead of asking a researcher to hand-design every configuration. Hooker’s talk connects that automation with efficient post-training and continual learning. Internal comparisons with staff-designed configurations are the lab’s reported evaluations, not evidence that the system independently performs general scientific discovery.[4][6]
Ideas
The Hardware Lottery argues that a research direction’s success depends partly on whether available hardware and software fit it. An idea can be difficult to explore because the systems needed to test it do not yet exist. Hooker uses this history to argue for closer work across algorithms, software and hardware rather than treating the current compute stack as a neutral measure of an idea’s potential.[5]
Tools, writing & talks
Writing, research & talks
Start here
The Hardware Lotteryarxiv.orgAn essay about how hardware and software availability influence which research ideas can succeed.
Start with the paper’s historical argument about hardware and research direction. General machine-learning familiarity is enough; use the conference talk to connect it to Adaption’s current training experiments.
Author[5]Sara Hooker estimates fewer than five thousand people know how to train a frontier model at scale, knowledge passed on like an apprenticeship rather than a literature, and argues the route to the frontier has narrowed into a single funnel.
Catalog workOriginal sourceA talk about access to training tools, automated model recipes and adaptive intelligence.
Speaker and cofounder[6]
Tools
- AutoScientistadaptionlabs.ai
Adaption’s automated model-training system.
Cofounder’s company product; launch authored by Adaption Research Staff[4]