Why Cohere’s ex-AI research lead is betting against the scaling race
AI companies are pouring billions into building massive data centers, some as large as Manhattan and consuming as much power as small cities. This relentless push stems from one belief — that more compute equals smarter AI. Yet not everyone agrees. A quiet rebellion is brewing, and why Cohere’s ex-AI research lead is betting against the scaling race is now one of the most talked-about shifts in the AI world.
Image Credits:Chris Behroozian
Sara Hooker, former VP of AI Research at Cohere and Google Brain alumna, believes the era of endless scaling is hitting its limits. Her new startup, Adaption Labs, challenges the notion that simply throwing more GPUs at large language models will yield the next leap in intelligence.
Breaking away from the scaling dogma
Hooker and co-founder Sudip Roy, another Cohere and Google veteran, are steering Adaption Labs toward a different frontier — building AI that learns and adapts continuously. Instead of static training on vast datasets, their goal is dynamic, real-world learning.
She announced the venture quietly in October 2025, sharing that her team is tackling “the most important problem: building thinking machines that adapt and continuously learn.”
Hooker’s belief marks a pivotal shift from the compute-heavy race led by giants like OpenAI, Anthropic, and Google DeepMind.
Why Cohere’s ex-AI research lead is betting against the scaling race — and what it means for AI
Hooker argues that scaling large language models is reaching a plateau of diminishing returns. Each new model requires exponentially more data, energy, and money for smaller gains in capability.
“There’s a turning point now where it’s very clear that the formula of just scaling is no longer sustainable,” she told TechCrunch. Instead, Adaption Labs is focusing on efficiency, adaptability, and learning from experience — three traits Hooker says will define the next generation of AI systems.
From Cohere to Adaption Labs: A new kind of intelligence
Adaption Labs aims to build AI that thinks more like humans — not by memorizing more data, but by forming connections, learning continuously, and applying knowledge flexibly.
Hooker hasn’t revealed whether her startup’s technology is based on large language models or something entirely new, but she emphasizes its focus on real-world learning efficiency. This could represent a major departure from today’s GPU-driven training paradigm.
Betting on adaptation over acceleration
While major players like OpenAI, Meta, and Google double down on trillion-parameter models, Hooker’s team is betting on adaptability over acceleration. It’s a contrarian move that could reshape the industry if proven effective.
The AI community has long debated whether scaling alone can lead to general intelligence. Hooker’s approach suggests that the future lies not in bigger models — but in smarter learning architectures capable of evolving after deployment.
A quiet revolution in AI research
Adaption Labs joins a small but influential group of startups questioning the sustainability of AI’s current growth path. From data scarcity and energy consumption to environmental impact, the downsides of scaling have become impossible to ignore.
By prioritizing adaptive systems, Hooker’s vision represents a return to the roots of AI research — focusing on learning, reasoning, and long-term evolution rather than brute-force computation.
Why Cohere’s ex-AI research lead is betting against the scaling race could redefine the next AI era
If Adaption Labs succeeds, it may signal the start of a new phase in AI — one where continuous learning and adaptability trump raw computational power. Investors and researchers alike are watching closely, wondering whether Hooker’s gamble could inspire a broader industry pivot.
The question now isn’t how big AI models can get, but how well they can learn to think for themselves.
Post a Comment