Thank you so much for joining me here, Ramin and Ion. And thank you guys for coming. Okay. We're going to talk about open weight efficient local models at a time of token maxing or minning, depending on your preference, and national security issues. There's a lot to unpack there. Ramin, you're the CEO and co-founder of Liquid AI, which is focused on making open weight efficient models that can bring AI into the real world. So, what does that mean exactly?
Yeah. So, at Liquid AI, our objective function has always been maximizing intelligence into the smallest unit of compute that you can actually get. So, the idea of efficiency, it's basically the DNA of our company, Liquid AI. We build efficient foundation models that can run on CPUs, GPUs, and NPUs. They can be outside of data centers. So, you can actually host the intelligence not just in the data centers inside of things, but you can also bring the intelligence into the physical world. And the best way to distribute this type of efficient technology is through open source. Why? Because open source touches developers. Developers get to know a new type of technology. Our core and fundamental technology is coming out of our research at MIT, what we have been doing for about a decade, finding alternative methods to build intelligent systems that can be more and more efficient compared to the transformer-based architecture that is out there. And we're distributing this technology to the world in an open source manner.
Ion, you're a legend in AI. You're co-founder and executive chairman of Databricks, and you're co-founder and chairman at Arena, and you're an advisor to Liquid. So, tell me a little bit about why, what was it about the company, what are they doing differently that made you interested in it and working with them?
Yeah, so let me start by saying that about open source. Obviously, I've been passionate about open source for a long time, and at Berkeley we build a lot of systems on the open source side for AI and for big data. And I think that obviously, like many of you here, believe that open source models are very important. And I think they are important because you want to have access, a reliable access to intelligence, so to speak. And I think that probably one of the most reliable ways to have access to intelligence is to run the models yourself. And I think that Liquid AI, because of the very small footprint relative to their power, they are kind of an ideal candidate to run for yourself. If you are an enterprise, if you are a small company, if you are a physical AI company, you can run these models on your own servers and still get tremendous power, and probably gain that's the most reliable and secure way to get access to intelligence.
Typically when we think about AI models now, we're still thinking of models that are far away in a cloud and they're being served. Talk to me a little bit about Liquid's models. How are they different? We know they're not large language models. They're a little bit different from that in the way that you're building them. But what is it that makes them really capable but also able to be served locally?
Yeah, so one of the things that we tried to solve as the company, so not just maximizing for as a foundation model company, we did not just want to maximize for capability and intelligence. We wanted to also work on the computational footprint of that intelligence. Imagine if your typical foundation model, the more information they process, exponentially the amount of energy that they actually consume is rising. This is because of the core and fundamental nature of the algorithms that are enabling those foundation models. So, we thought to change that algorithm, and we changed that algorithm to something that, instead of feeding in more and more information increasing exponentially, the increase in computation is actually closer to linear. So, we're talking about substantially reducing the computational cost of intelligence. The implications of it is that now you don't need giant GPUs or parallel data centers to really host this type of intelligence. You would be able to bring them even down to CPUs, smaller units of computation. And the origin of the whole thing came out of we are always inspired by nature, and we were inspired by brains. Imagine your brain is a very powerful intelligence tool that is running on 20 watts of power. So, if you can actually exhibit that amount of behavior, I think that the space of algorithms, efficient algorithms to explore is massive. There's no just one format to really get to intelligent behavior. Obviously, scale is king, and scale is going to unlock more and more capabilities for us, but there's matters of distribution, access to that technology, the energy cost, memory cost. There's so many other considerations that I think have to be studied right next to the capability of a model.
In the US, and Ion, you hinted at this a little bit before by talking about the power of reliability of running a model locally. In the US, we've been seeing government restrict access and request that access be limited to some of the most powerful AI companies, most powerful models, citing national security concerns. So, how do you see that kind of restriction from either from the US or from any country impacting the AI industry as a whole? And how does it impact Liquid?
So, I think, you know, as obviously everyone is seeing, things are very uncertain. Right? It's like compare the world today even with the world two months, one month ago. Right? What we've seen, obviously, the US administration regulating a little bit the access to the most powerful models from OpenAI and obviously Anthropic. You've seen, I think yesterday, some announcement coming out from China about also maybe regulating a little bit the access to even the AI models, the most powerful AI models outside China. So, I think that you see all of this, and for instance, enterprises, the one thing they don't like is uncertainty. Right? So, this is going to have an impact, and many of them are going to reassess their choices. And it was already, so it's like that many of the companies, and there's another aspect, another trend of this kind of token mixing, right? It was like three months ago everyone was kind of rewarding the employees for using more tokens, right? For their work because that was equating with increased productivity. Obviously, now everyone is raining on that. So, all of this basically saying that more and more organizations, even if they start with the most powerful models to prototype and evaluate AI for their businesses, then many of them actually are going to move to other small, smaller models which are cheaper, hopefully open source, because they're going to have more control and they are more reliable, they have more reliable access, which are good enough for their applications. And the way to look at that is you have four different kinds of applications, you have a certain threshold required for capabilities. And the most powerful models, what you see is the capabilities grow and grow, but once you cross that threshold, then any model will do it. And obviously, the open source models are also growing capabilities, so you are going to see for most applications and workloads, more and more open source models like Liquid open source models being more than good enough for those. So, the adoption will explode.
Do either of you worry that these kinds of restrictions, whether put in place by no matter who's putting them in place or how temporary they are, could soon extend to open source and open-weight models? Do you expect that to happen or are you concerned that that could also happen?
Restriction on open source? Yeah, I think certainly that is a clear danger there, right? If you think about the logic, right? Why models from Anthropic and OpenAI, GPT-5.6, were at least restricted for a period of time because they are, so to speak, too powerful. Okay? Which is very questionable, at least in my mind. What does that mean? Actually, there is no clear definition of what that means. And we know that open source models, the best of them, are a few months behind these most powerful models. So now, if you just think about, in a few months, you can make the same argument against the open source models. So what will happen then? Right? So I think it's anyone's guess, right? On the other hand, it's much harder to regulate the open source models. There are a bunch of ways you can download, right? So we'll see. It can be one of these things where the companies who are playing the, they're going to follow the instructions. They're not going to use it, and the other ones which are in the gray area will get benefit because they are going to use it, right? To use these kind of models. But I think that it is a real danger that we are going to see some kind of regulations, and it's going to be unfortunate.
