The End of Monolithic AI and What Comes Next

The days of monolithic AI are over. So what does that mean?

If you look at, um, the evolution of, um, top-tier models, you can see that they're almost, like, converging to, uh, the same common point, that you get marginal improvements after spending, like, millions and millions of dollars. There's a big gap between how fast, uh, the energy infrastructure can grow and how fast, uh, the AI is actually demanding, um, the compute growth. And in order to bridge that gap...

Welcome back to the Nebius for Startups podcast. My name is Marques Patrice, and I'll be your host today. Joining me is Sudeep Roy. He's the co-founder and CTO of Adaption Labs, the frontier lab focused on extremely efficient and adaptable intelligence. I kind of want to start with how you stumbled upon the idea of building adaptable intelligence because for so long the conversation has been scaling laws.

Yeah.

And you guys are actually challenging that.

Yeah.

So I'd love to kind of hear how did you stumble upon th- this problem and, and, and what made you kind of go all in and want to start solving it?

Throughout our careers, we have been focused on, um, doing AI efficiently from, um, I guess, different perspectives. Sara has been more focused on the AI and the algorithmic side of things. I have been more focused on the system side of things. And, uh, throughout the last 10 years, we have also, uh, looked at how the industry has, um, evolved, uh, especially in the last couple of years, and what challenges people usually run into in using AI. And, uh, one of the challenges that we saw was people limiting themselves to just have access to APIs, uh, for proprietary models, and itching to exercise more control, but being limited to only techniques like prompt engineering. We thought that solving that problem actually poses a lot of fundamental challenges, which, um, was the, I guess, the genesis of Adaption.

At Adaption Labs, the tagline here is mo- the days of monolithic AI are over. I'd love to hear from you, what does that mean?

Many people want to exercise control over their AI stack, but the only thing that they're limited to is changing the prompt, because that's the only thing that you can do through an, um, API. But we want to enable, um, a much wider form of control, whether it is in the data space and the intelligence space, or even in the interface, how people actually interact with the system itself and how people provide feedback that can be then used to continuously learn and adapt over time, and, um, make, um, AI much more customizable, adaptable, and, um, I guess in the extreme, personalized as well, um, as opposed to just this one big, uh, I guess, black box, um, that you have right now.

I'd love to kind of hear a little bit more about your real-time, um, adaptation approach. Like, how are you guys actually infusing real-time adaptive intelligence into the models?

There's a lot of interesting inference time scaling strategies which are, which can be implemented extremely efficiently so that the adaptation, uh, doesn't take months or weeks, as is the case with, like, the current generation of models, but is, um, almost instant in the order of you send a request and you can