Maxwell Zeff argues in TechCrunch that Adaption Labs is betting against AI scaling

Maxwell Zeff argues in TechCrunch that Adaption Labs is betting against AI scaling

TechCrunch published Maxwell Zeff’s report on October 22, 2025. It describes Sara Hooker, Cohere’s former VP of AI Research, and co-founder Sudip Roy’s Adaption Labs as betting that scaling large language models has become an inefficient way to improve AI. Hooker says the startup aims to build systems that learn from real-world experience, but declined to explain its methods or whether it uses large language models. Hooker argues that models served the same way to everyone are expensive to adapt, and that systems could learn more efficiently from their environments. She says this could change who controls and shapes AI, and who the models serve. The report contrasts that bet with labs’ continued efforts to scale reinforcement learning and reasoning models, while noting earlier concerns about diminishing returns from pre-training and subsequent gains from reasoning models. Zeff’s report also describes Hooker’s focus on compact models and worldwide hiring. Its conclusion is conditional: if Hooker and Adaption Labs are right, adaptive learning could be more powerful and efficient than scaling. Whether their approach can deliver that result remains unknown; the report gives no details of the methods or evidence of their performance.

Publication: 1b40be42-6401-3b2c-c40d-e683ea33aa05

attribution — pass: The reading correctly attributes the report to Maxwell Zeff at TechCrunch and keeps Sara Hooker's claims and quotations attributed to her. The title's "argues" slightly overstates a news report's framing, but the headline itself frames the bet, so this is not misattribution. Marina Temkin's contributed reporting is omitted, which is minor.

That’s the bet Sara Hooker, Cohere’s former VP of AI Research and a Google Brain alumna, is taking with her new startup, Adaption Labs.

summary — uncertain: "Scaling" is the central term, but neither the reading nor the definitions list explains it. The source defines it as adding more computing power to existing training methods. Smallest repair: add that plain-language gloss where the first summary sentence introduces scaling, or add a term-library entry. The reading also flattens Hooker's force by dropping her verdict that scaling-pilled approaches are "attractive but extremely boring." It omits the source's concrete mechanism too: models already in use by customers cannot learn from their mistakes in real time. Restoring that quote and the toe-stubbing point would keep her voice and the stakes.

the idea that adding more computing power to existing AI training methods will eventually yield superintelligent systems capable of performing all kinds of tasks.

date — pass: October 22, 2025 matches originalPublished at day precision and the URL path. The dateModified is the same day. The "Most Popular" sidebar lists later stories, such as Anthropic's IPO, that are probably later page additions; the reading neither uses nor relies on them. The limitations state the retrieval date, and there is no hindsight framing. The source's conditional implication is reported as conditional, not as an outcome.

support — pass: Each claim matches a source passage: the inefficiency premise, the continuous-learning aim, the declined details, the claim about control and who models serve, and the conditional conclusion. The reading's statement that no methods or performance evidence are given is supported by Hooker declining to share details. The compact-models point is grounded in her past work and stated intent. The question follows from the evidence gap and is not a leading claim.

She declined to share details about the methods behind this approach or whether the company relies on LLMs or another architecture.

media — pass: No media was selected and no verified candidates were supplied, so there is no identity or relevance issue.