Why Cohere’s ex-AI research lead is betting against the scaling race (techcrunch.com)

Sara Hooker, formerly VP of AI Research at Cohere and a Google Brain alum, has quietly launched Adaption Labs with Sudip Roy to challenge the industry’s scaling-first orthodoxy. Instead of pouring billions into ever-larger data centers and giant LLMs, the startup is betting on systems that continuously adapt and learn from real-world experience—efficiently and in production—rather than relying on massive pretraining or expensive fine-tuning/consulting. Hooker argues current approaches (pretraining, large LLMs and today’s RL) show diminishing returns and often can’t learn from mistakes once deployed. The move matters because it reframes where future performance gains might come from: online adaptation and environment-driven learning could be more compute- and cost-efficient than brute-force scaling, potentially democratizing who controls AI capabilities. Technical implications include shifting focus away from scaling compute toward methods that enable safe, continual learning in production (a problem current RL pipelines don’t solve well), and validating compact models that outperform larger counterparts on many benchmarks. Adaption Labs has been fundraising (reports of a $20–40M seed process) and is keeping methodological details close to the vest, but the startup’s thesis amplifies mounting evidence—academic and industry—that pure scaling is hitting limits and new learning paradigms may be the next frontier.