Sara Hooker

Sara Hooker is the co-founder and chief executive of Adaption Labs, which builds AI systems that adapt continuously without depending on ever-larger training runs. Previously a research scientist at Google Brain and vice president of research at Cohere, she investigates how computing infrastructure, language, and geography determine who can build advanced AI.

From research access to frontier AI

Hooker grew up in Mozambique and other African countries and founded Delta Analytics, a nonprofit connecting technical specialists with organizations lacking access to their expertise. She later researched interpretability, efficiency, fairness, and robustness at Google Brain, now part of Google DeepMind, and completed a computer-science doctorate at Mila.

Her research demonstrated that model compression can disproportionately degrade performance on underrepresented examples, even when overall accuracy remains relatively stable. In her 2020 essay on the hardware lottery, she argued that scientific ideas often succeed because they suit prevailing chips and software, not necessarily because they represent the best technical approach. The hardware lottery framework connects technical infrastructure to research diversity and the concentration of scientific opportunity.

In 2022, Hooker became the inaugural head of Cohere For AI, later known as Cohere Labs, and subsequently served as Cohere’s vice president of research. She co-authored the Aya model research, describing an open-access instruction-following model spanning 101 languages, and the Aya 23 technical report, which examined the tradeoff between broader language coverage and greater model capacity per language.

Hooker left Cohere in 2025 and founded Adaption Labs with Sudip Roy. Their work focuses on making specialized AI development more accessible without requiring organizations to assemble frontier-scale infrastructure.

Research themes & current work

• The hardware lottery: Hardware and software shape which algorithms appear practical, influencing both scientific progress and who can afford to participate.
• Efficiency without hidden bias: Compression should be evaluated beyond aggregate accuracy because smaller models can fail disproportionately on underrepresented examples.
• Multilingual AI: Hooker helped develop Aya’s 101-language model and subsequently reported that her team had expanded data automatically across 122 languages, a separate effort at Adaption.
• AutoScientist: Adaption’s automated model-development system jointly optimizes domain-specific data, model configuration, and training recipes. Hooker argues that coordinating these choices can improve customization while reducing the expertise and computing required.
• Tiny AutoScientist: This small-model initiative targets constrained infrastructure, sensitive data environments, and latency-sensitive applications.

Hooker argues that stronger data, post-training, and adaptive inference can increasingly outperform indiscriminate increases in pretraining scale. Her goal is continually adaptive intelligence: models that learn from their environments while giving more organizations practical control over their own AI systems.