A conversation with Co-founders Sara Hooker and Sudip Roy of Adaption, with Emergence Partner Yaz El-Baba
Co-founders Sara Hooker and Sudip Roy started Adaption on a simple bet: that the next era of AI belongs to companies that can keep learning after they ship, not just the ones with the biggest training runs.
Sara spent years studying why good research ideas win or lose based on the hardware underneath them, and Sudip built the systems that make models fast enough to matter in production. That combination, research instinct paired with production discipline, is why we backed them, and why the team has moved quickly from a research thesis to a working product in Adaptive Data.
Partner Yaz El-Baba sat down with Sara and Sudip at a moment when the industry's assumptions about scale were already starting to wobble, to talk through where they think intelligence is actually headed, and why they built Adaption to get there.
The limits of model scale
Yaz: For most of the last decade, the story in AI was simple: bigger model, better results, whoever scales fastest wins. That story's been changing lately. Where are we right now in the cycle? Still scale-at-all-costs, or already shifting toward efficiency?
Sara: We are in an era of change, where innovation matters again. For the last ten years, scaling pre-training model size was enough. That created a concentration of power in a handful of providers. Returns have flattened, and all bets are off now.
At a trillion parameters, the economics break in both directions: models become too expensive to train and serve. This reckoning is a good thing. It means new ways of training models matter again. We are building models that learn continuously, become smarter than they were when first deployed, and are owned by companies around the world.
At a trillion parameters, the economics break in both directions: models become too expensive to train and serve. This reckoning is a good thing.
— Sara Hooker
Yaz: Sara, your paper, "On the Slow Death of Scaling," argues scaling laws are running out of road. What are the earliest signs you're seeing that it's already playing out?
Sara: No frontier lab will 5x the size of the training model this year. The returns from scale are in test-time compute, adaptive data, and real-world interactions, not in model size alone.
Sudip: We're seeing the same thing from the customer side. As AI adoption scales into production, users are increasingly cost and latency-conscious. They're realizing they don't need the full general capability of a massive model; they need specialized capability at lower cost and lower latency. That's exactly what Adaption is built to accomplish.
Open weights and ownership
Yaz: This summer, a wave of companies, including Nvidia and Microsoft, signed an open letter on open weights arguing there needs to be a balanced ecosystem between open and closed models. Adaption signed on too. Why?
Sudip: There's a narrative floating around the industry that open source models are somehow less secure. That's a narrow way of looking at things. Adaption's mission is to transfer ownership of intelligence across the full stack to the enterprises and communities we serve, and an open ecosystem is an important cornerstone of that mission.
Yaz: You've said before that open weights are safer because they can be inspected by anyone, similar to Linux and public encryption standards. Taking the flip side, when does that stop being true?
Sudip: I do think open weights are safer, because they can be probed by a much larger audience. But there's a key difference between how traditional open source software like Linux has been developed and how open-weight models are published today. With Linux, the entire process, including all the code, is public.
With an open-weight model, what's delivered is almost always the end outcome, not the process it came from or the original source data used to create it. A handful of open-weight models also publish their source data, but for the most part that process is still a black box. Still, it's a step in the right direction, and hopefully more players in the space move toward openness over time.
Hardware and inference
Yaz: Your work around the concept of “Hardware Lottery” argued that available hardware and software can determine which ideas succeed, independent of their merit. Open weights broaden access to models, but they don't equalize access to compute. How does the hardware lottery shape what you're building?
Sara: The hardware lottery has only narrowed in recent years. Nearly every chip built is optimized for one operation: the matrix multiply at the heart of deep learning. Almost nobody notices, because the winning ticket has been paying out for a decade. Compute access doesn't overcome the hardware lottery, but making intelligence as fast as possible accelerates innovation and improves overall access.
Yaz: You've spent a lot of energy rethinking how models are hosted for end users, especially given the compute constraints in the market today. Where do you see serving infrastructure heading?
Sudip: A few trends stand out. One is bespoke, custom hardware built specifically for inference, purpose-built ASICs rather than general-purpose chips. Another is decentralized inference that does not run everything in a data center. We're already seeing smaller models run on laptops, and we’ll soon see low-power devices like watches make local decisions that check in with a centralized system periodically.
The one I find most interesting from a systems perspective is that training and inference have always been treated as separate paradigms, optimized independently. That gap will erode as continuous learning systems mature and models update regularly, even during inference.
Adaptive Data and enterprise AI
Yaz: The highest-value internet data is largely tapped out at this point. It's not a coincidence that Adaption's first product launch was Adaptive Data. Why start there?
Sudip: Enterprises are sitting on massive volumes of data accumulated over decades in business; data that's never been seen by open or closed models. Enterprises are disincentivized to hand it over to frontier labs, because it's their moat.
We built Adaptive Data to make it seamless to connect that historical data to downstream AI systems, whether at training time or inference time. It's a key piece of transferring ownership of intelligence back to enterprises.
Yaz: As more ownership of intelligence shifts to the enterprise and the end user within it, where does adaptability go in the next year? What changes for the end user compared to the current experience with agents and LLMs?
Sudip: At one extreme, you could imagine a personal AI that knows your tastes and the context relevant to you well enough to make decisions almost as if you were making them yourself. We're far away from that. In the interim, adapting to a particular organization is a much more tractable step, and it's within reach over the next year: a system, not just a model, where the harness on top and the data feeding it are built specifically for that enterprise, instead of shipping something generic out of the box.
