The pitch Sara Hooker has been making to investors sounds almost heretical in an industry drunk on scale. Larger models, she argues, aren't the path forward—at least not the only one. And judging by the check she just landed, some of Silicon Valley's sharper minds are starting to agree.
Hooker and Sudip Roy, both former senior leaders at Cohere, have raised $50 million in seed funding for Adaption Labs, a San Francisco startup building AI systems designed to learn continuously rather than rely solely on massive upfront training runs. Emergence Capital led the February 4 round, with backing from Mozilla Ventures, Fifty Years, Threshold Ventures, Alpha Intelligence Capital, E14 Fund, and Neo.
Fifty million dollars. For a seed round. To put that in perspective, Carta pegged the median U.S. seed at around $2.5 million in 2024. Adaption's debut is roughly twenty times that figure—the kind of number typically reserved for later-stage companies with revenue and traction, not founders sketching architectures on whiteboards.
Walking Away From the LLM Gold Rush
The backstory matters here. Hooker wasn't exactly struggling at Cohere. As VP of Research, she ran Cohere Labs and spearheaded the Aya multilingual open-science project—work that earned her credibility in academic circles and beyond. Before Cohere, she spent years at Google Brain and DeepMind, the kind of résumé that opens doors in this business. Roy, meanwhile, led inference and platform engineering at Cohere after contributing to Google's Pathways and TFX systems—infrastructure that underpins some of the largest AI deployments on the planet.
Both left well-funded positions at a company racing to build ever-larger language models. That departure tells you something, perhaps more than the founders expected it would. Hooker has become increasingly vocal about what she calls the "slow death of scaling"—the idea that simply throwing more parameters and compute at models won't deliver the next breakthrough in performance or, crucially, economics.
It's a gamble, leaving the LLM gold rush to pursue a different bet entirely.
The Economics of Smarter, Not Bigger
So what exactly is Adaption building? The company describes its approach as "adaptive AI"—systems that learn at test time, employ gradient-free optimization methods, and dynamically allocate compute based on task complexity. Instead of pre-training gargantuan static models on everything, the goal is to create architectures that get smarter about when and how they learn.
If that sounds abstract, consider the financial pressure driving interest in this direction. Emergence Capital pointed out in its investment thesis that training costs for frontier models have been climbing roughly 2.4 times annually since 2016. By 2027, the firm projects, a single cutting-edge training run could top $1 billion. Those numbers make enterprises queasy. They also create an opening for startups promising efficiency over raw scale.
Mozilla Ventures, which joined the round with an eye toward trustworthy AI, published a blog post titled "Everything Intelligent Adapts" to explain its rationale. The firm sees Adaption's approach as aligned with building systems that evolve responsibly—AI that remains under meaningful user control rather than ossifying into black-box behemoths.
It's worth noting that Mozilla's participation adds a dimension beyond pure economics. The emphasis on adaptability and transparency speaks to growing unease about deploying AI systems that nobody fully understands, not even their creators.
A Syndicate of Believers
The angel roster, according to a LinkedIn post from Alpha Intelligence Capital, reads like a who's who of deep learning royalty: Jeff Dean, Thomas Wolf, Soumith Chintala, Karim Beguir. Adaption hasn't independently confirmed the full list, but the names circulating suggest the founders have credibility where it counts. These aren't check-writers chasing hype—they're researchers and engineers who've built the infrastructure that modern AI depends on.
That syndicate matters. In a market flooded with generative AI pitches, technical credibility separates signal from noise.
What's Being Built (and What Remains to Be Seen)
Adaption declined to share its valuation with Fortune, though some secondary sources have floated a $1 billion seed valuation. That figure remains unconfirmed, and frankly, at this stage, it's less interesting than what the company does with the capital.
The hiring spree is already underway. Open roles span research, platform engineering, and go-to-market functions across the U.S., Canada, the UK, India, the EU, and Latin America. The distributed footprint suggests Adaption intends to tap global talent pools from the jump—a pragmatic move for a company betting on architectural innovation.
The company's website outlines three pillars: adaptive data, adaptive intelligence, and adaptive interfaces. It's early—maybe too early to judge—but the thesis is that co-designing these elements will produce AI that's cheaper to run, faster to customize, and better suited to messy real-world deployments than today's monolithic models.
Whether Hooker and Roy can deliver on that vision is the multimillion-dollar question Emergence and its co-investors are now underwriting. The technical challenges are formidable. Test-time learning and dynamic compute allocation sound promising in theory, but making them work at scale, reliably, is another matter entirely.
Still, in an industry where the default playbook has been "more data, bigger models, higher bills," a contrarian bet on adaptability carries a certain logic. If nothing else, Adaption now has the runway—and the pedigree—to find out if the heresy is justified.
