Sara Hooker

ref · 2026 年 5 月 25 日午前 6:39

AI researcher, VP of Research at Cohere, and co-founder and CEO of Adaption — known for the Hardware Lottery thesis, pioneering work on model compression and bias, building multilingual AI through the Aya initiative, and a sustained contrarian argument that scaling alone cannot sustain the next decade of AI progress.

Profile

Field	Detail
Born	Dublin, Ireland
Nationality	Irish / grew up across sub-Saharan Africa
Current Institution	Adaption (Co-founder & CEO)
Research Areas	Model Efficiency, Continual Learning, Multilingual AI, Model Compression, Interpretability, Algorithmic Fairness, Benchmark Methodology
PhD Advisors	Hugo Larochelle; Aaron Courville
Alma Mater	Carleton College (B.A.); Mila – Quebec AI Institute (Ph.D.)
Personal Website	sarahooker.me
X / Twitter	@sarahookr
GitHub	@sarahooker
Google Scholar	scholar.google.com

Overview

Sara Hooker is an Irish-born computer scientist who has made distinctive contributions across model compression, interpretability, multilingual AI, and the sociology of AI research incentives. She is best known for “The Hardware Lottery” (2020), a conceptual essay published in Communications of the ACM arguing that the success of AI ideas is shaped less by their intrinsic merit than by how well they happen to exploit the hardware available at the time — a thesis that reframed debates about research bias in the field. As VP of Research and head of Cohere Labs from 2022 to 2025, she led the Aya initiative, a crowdsourced multilingual model project that mobilized over 3,000 researchers across 119 countries and more than doubled the number of languages covered by open generative AI. In October 2025 she co-founded Adaption with Sudip Roy, raising a $50 million seed round in February 2026 to build AI systems that learn continuously and adapt at inference time — an explicit bet against the scaling-law orthodoxy. Her background spans five African countries, Ireland, and multiple academic and corporate research institutions; her work consistently returns to the question of who AI research serves and what it systematically overlooks.

Early Life & Education

Hooker’s parents — an Irish mother and British father who met in Sudan — relocated to Lesotho when she was four years old. She grew up across sub-Saharan Africa, living in South Africa, Mozambique (where she attended middle school with instruction in Portuguese), Lesotho, Eswatini, Kenya, and Liberia until the age of nineteen. The experience of navigating multiple languages and educational systems as a child — and witnessing the uneven distribution of access to infrastructure — would later animate her research focus on efficient and multilingual AI.

B.A., Carleton College (Class of 2013)
Hooker studied at Carleton College in Northfield, Minnesota, where she developed her early interest in computer science and quantitative methods.

Ph.D., Computer Science — Mila – Quebec AI Institute
Hooker completed her PhD at Mila, supervised jointly by Hugo Larochelle and Aaron Courville, two of the central figures of the Montreal deep learning school. Her doctoral work focused on gradient flows, neural network training dynamics, and the properties of sparse and compressed networks — research directions that would feed directly into her later applied work at Google Brain.

Career

Delta Analytics — Founder (2014–present)

In 2014, while completing her PhD, Hooker founded Delta Analytics, a nonprofit that builds technical capacity for nonprofits and social sector organizations through data science projects. The organization has continued to operate independently of her subsequent roles.

Google Brain / Google DeepMind — Research Scientist (2017–2022)

Hooker joined Google Brain in 2017, where she worked on model interpretability, compression, and sparsity. Several of her key papers originate from this period:

Her interpretability benchmark (NeurIPS 2019) proposed a rigorous evaluation protocol — ROAR (Remove and Retrain) — for testing whether saliency methods genuinely identify the features a model relies on, highlighting significant unreliability in standard methods. Her compression and pruning work examined what neural networks forget when compressed, showing empirically that pruning disproportionately degrades performance on underrepresented and long-tail data — an early and important linkage between model compression and algorithmic fairness. “The Hardware Lottery” (2020) consolidated this period’s research into an influential conceptual essay on how hardware constraints systematically favor certain algorithmic families over others, shaping the field’s trajectory in ways researchers rarely make explicit.

In 2019, Hooker was a founding member of Google’s first AI research office on the African continent, located in Accra, Ghana — a role consistent with her longstanding interest in expanding where AI research happens geographically.

Cohere Labs / Cohere For AI — VP of Research (April 2022–2025)

Hooker joined Cohere in April 2022 to build and lead Cohere Labs (also known as Cohere For AI), the company’s research arm, positioned as a nonprofit-adjacent research lab pursuing open science alongside Cohere’s commercial activities. Under her leadership the lab launched several notable initiatives:

Aya Project (2023–2024): A large-scale collaborative effort coordinating over 3,000 researchers from 119 countries to build multilingual instruction-following datasets and models. The resulting Aya model covered 101 languages — roughly doubling the number of languages supported by existing open generative AI systems — and was released openly. The project also generated research into multilingual evaluation and the specific failure modes of compressing models trained on non-English languages. Hooker described the mission in personal terms: having grown up in environments where Portuguese, Sesotho, Swazi, and Swahili surrounded her, she connected AI language access directly to cultural connection.

Aya Expanse (2024): High-performance 8B and 32B multilingual models designed to narrow the capability gap between English-centric systems and models serving the world’s less-represented languages. Released openly alongside the original Aya work.

Cohere For AI Scholars Program: Hooker launched a structured scholars program to extend research mentorship and collaboration to researchers outside major AI hub institutions.

“The Leaderboard Illusion” (2025): Co-authored while still at Cohere Labs, this preprint argued that popular model evaluation leaderboards — particularly Chatbot Arena and similar anonymous crowd-preference platforms — are systematically gameable and produce rankings that reflect superficial stylistic features and evaluation artifacts rather than underlying model capability. The paper prompted significant discussion about benchmark validity and attracted coverage for its critique of how leading AI labs were managing their public-facing evaluations.

Adaption — Co-Founder and CEO (October 2025–present)

In October 2025, Hooker publicly announced Adaption (also referred to as Adaption Labs) alongside co-founder Sudip Roy, previously director of inference computing at Cohere. The company’s core thesis — that AI progress will be limited less by model scale than by the inability of static, frozen systems to adapt to deployment conditions — represents a direct extension of Hooker’s decade-long critique of compute-maximalism. Adaption is developing systems focused on three pillars: adaptive data (learning from interaction rather than curated pretraining sets), adaptive intelligence (gradient-free inference-time learning), and adaptive interfaces (user experiences beyond the standard chat bar). In February 2026, the company raised a $50 million seed round led by Emergence Capital Partners, with co-investors including Mozilla Ventures, Fifty Years, Threshold Ventures, Alpha Intelligence Capital, E14 Fund, and Neo. In May 2026, the company launched AutoScientist, its first public product: an automated fine-tuning system that raised win rates from 48% to 64% against its own researchers’ configurations in internal evaluations.

Key Contributions

“The Hardware Lottery” (Communications of the ACM, 2021; arXiv 2020) — Hooker’s most widely cited solo-authored piece argues that the dominant approaches in AI research are those that happened to align with the architectures GPU and TPU hardware was designed to run efficiently. Ideas that require irregular computation — such as symbolic reasoning, capsule networks, or sparse-activation models at their time of introduction — are systematically disadvantaged regardless of their potential. The paper reframed a structural bias in the field and has become a standard reference in discussions of AI research sociology and infrastructure dependence.
ROAR Interpretability Benchmark — Introduced in “A Benchmark for Interpretability Methods in Deep Neural Networks” (NeurIPS 2019, with Erhan, Kindermans, Been Kim), ROAR provides a model-agnostic protocol for evaluating saliency methods by measuring the degradation in model performance when features identified as important are removed and the model is retrained. It revealed that many popular interpretability methods perform little better than random baselines under this test.
Compression and Algorithmic Fairness — Hooker’s research on what compressed and pruned models forget (arXiv 2019 and subsequent work) was among the first systematic demonstrations that model compression disproportionately harms performance on underrepresented groups and rare examples — a finding with direct practical implications for how efficient AI is evaluated and deployed.
Aya Model and Dataset — Led the Aya initiative at Cohere Labs, producing an openly released multilingual instruction dataset and model covering 101 languages. The project is one of the largest coordinated multilingual AI research collaborations on record, and its open releases have been widely used by researchers working on low-resource language models.
Aya Expanse — Open 8B and 32B multilingual models (2024) designed to close the performance gap between English-dominant systems and underserved language communities; released openly to advance multilingual AI access.
“The Leaderboard Illusion” (arXiv 2025) — Co-authored critique of popular LLM evaluation leaderboards, documenting how anonymous crowd-preference systems can be gamed and how current ranking methodologies fail to distinguish genuine capability from presentation effects; contributed to broader field-level discussion about evaluation methodology integrity.
“On the Limitations of Compute Thresholds as a Governance Strategy” (arXiv 2024) — Analysis of the limitations of using floating-point operation counts as a proxy for AI risk in governance frameworks, arguing that compute thresholds are a poorly calibrated regulatory instrument.
Delta Analytics — Founded in 2014, this nonprofit has delivered data science capacity-building to social sector organizations for over a decade, a sustained institutional expression of Hooker’s view that technical expertise should be distributed beyond profit-driven contexts.
Underrated ML (podcast) — A podcast co-hosted with her brother Sean Hooker discussing underappreciated ideas and research in machine learning; an expression of her commitment to accessible scientific communication.

Awards & Recognition

TIME100 AI — Most Influential People in AI (2024)
Fortune — Top 13 AI Innovators (2023)
World Economic Forum — Council on the Future of Artificial Intelligence (ongoing member)
Kaggle ML Advisory Research Board (ongoing member)
MLC Research Group (member)

Key Relationships

Hugo Larochelle — PhD advisor at Mila; now Scientific Director of Mila and one of Hooker’s most prominent professional connections; the Montreal deep learning tradition Larochelle represents directly shaped her approach to training dynamics and compression.
Aaron Courville — PhD co-advisor at Mila; Courville’s work on generative models and neural network theory provided a second formative intellectual strand alongside Larochelle’s focus on practical deep learning.
Sudip Roy — Co-founder of Adaption; former director of inference computing at Cohere; the operational and technical counterpart to Hooker’s research leadership in building the company.
Been Kim — Google Brain collaborator; co-author of the ROAR interpretability benchmark; their joint work on saliency method evaluation is among the most-cited in interpretability research.
Shakir Mohamed — Google DeepMind researcher and DeepIndaba co-founder; shares Hooker’s focus on African AI development and community-building; appears among her most prominent professional connections.
Ian Goodfellow — Among Hooker’s most prominent followers; Google Brain period connection reflecting the broader Google research network she was embedded in.
Andrew Ng — Prominent connection in her professional network; shared focus on AI access and education.

Personal Style

Hooker’s writing is distinctive for its willingness to critique structural conditions in the AI field itself — hardware constraints, benchmark methodologies, geographic concentration of research talent — rather than focusing solely on technical problem-solving within those conditions. Her essays tend to operate at the intersection of empirical observation and institutional analysis, asking not just what a model can do but what conditions shaped the research choices that led to this model. She is particularly direct about the costs of the scaling paradigm, having committed her current company to the thesis that the next decade of AI progress requires adaptive efficiency rather than brute-force pretraining — a position she frames as a scientific bet rather than a commercial hedge. Her upbringing across multiple African countries and linguistic environments has given her a personal stake in the multilingual AI question that coexists with the technical framing; she has described language in AI as “super personal” and connected it explicitly to the question of whose inner world current AI systems can reach.