Waking Up With AI

But, picking up on Katherine's podcast episode from last week, the release of DeepSeek and specifically the open source version — so that's the DeepSeek R1 model, which is different from the closed app version — put a real spotlight back on the role of open source in the AI ecosystem. So this is becoming a hot topic again, and I wanted to spend a little bit of time today talking about that. And specifically, if you go and you look online at commentators, articles, some are really heralding the release of the DeepSeek R1 model as a victory for open source AI. And what this boils down to is an argument that open source models may actually be starting to lead innovation in the foundation model space. Of course, there are others that dispute that, right? They basically say DeepSeek lags the performance of some closed models by several months. There's a lot of debate around this right now, you know, the performance of these different models and which ones are more advanced. I'm not going to try to resolve that on the podcast today.

But given that I'm sitting in Abu Dhabi today and there's so much focus on open source here, I thought I'd use some of our podcast time to talk about why people are excited about the promise of an open source AI ecosystem, and particularly for highly performative AI. Actually, as it turns out, luckily for me because I'm recording today, the UAE's Minister of State for Artificial Intelligence, Digital Economy and Remote Work Applications released a white paper just this morning called “Towards a Future of Responsible AI.” That paper is definitely worth a read in full, but I wanted to specifically call out a section on open source AI, which stresses that foundation models can serve as common building blocks and that we may see pressure to open source AI to create kind of a common technology infrastructure along the lines of Linux, which is used today, kind of all around the world as the backbone for a lot of consumer technology. And that white paper section on open source AI concludes with the sentence, “artificial intelligence, if it's to become basic infrastructure, could — and perhaps should — operate on the same open source basis.”

Now, this is really interesting. And I really wanted to press on this topic or this idea of open source as infrastructure because, for many of our listeners, the concept of AI as infrastructure may not resonate today. And that's because to date, many AI applications have been specific products that have been procured as standalone enterprise tools or downloaded as standalone consumer apps. They're things that we interact with, but they don’t really interact with each other. And we kind of just use them like, I can open my ChatGPT and ask it questions or take photos of things. It's kind of me interacting with the technology, but it's hard to see that infrastructure layer. But one thing that I've been spending a lot of time thinking about lately especially in our work, but me personally, is what it would take for us to collectively move to a world of agentic AI ecosystems. That is, a world in which people will have one or likely multiple agents that are able to kind of reason and plan and execute actions for them in the real world or the virtual world. Now, those agents might be personal to me as a user. I might have my own agent, like Anna's agent, but it could be provided by a company. So Paul, Weiss could have an agent, Unilever could have an agent, Delta could have an agent. And some of these agents are going to have to be able to talk to each other and work really seamlessly with each other. So we're going to need to have a world in which there's an ability for agents to have a handshake capability. And creating infrastructure for that kind of agentic AI ecosystem is going to require the ability for agents to interoperate. What I mean by that is work together. And that's exactly where open source tends to fit in and do really well. It can create common standards and tool sets for companies globally to create interoperable techniques and tools and products. So all these little agents that, you know, maybe I'm going to supervise like 10 agents or 100 agents. My agents can go and interact with Katherine's agents and maybe write our podcast scripts for us. Who knows? But anyhow, that is going to be a really important kind of underlying infrastructure level that I think we'll start to see emerge, maybe not immediately, but ultimately as the world of AI becomes like the world of agentic AI. That's kind of the direction of travel these days.

It's also important to think about that in a global perspective and the fact that open source can really accelerate innovation all across the world, particularly in countries that don't have access to just the kind of data or compute that they would need to train a full foundation model. And I think that was one thing that really excited people globally about DeepSeek. The concept that because it was cheaper to train, maybe — I mean, that's also debated — but probably cheaper to train and relied on maybe less highly performative chips — again, hotly debated — that could potentially be easier to replicate in countries like India, or other countries where folks are really excited and want to be experimenting and innovating and creating their own tech layer, infrastructure layer using AI. So that is a big piece of what is going on with the open source discussion and why there is this kind of buzz and excitement about open source, and we'll see how that pans out.

The tricky thing here is the regulatory response. So let's turn back briefly to DeepSeek.. A few days after the market reaction on Monday, January 27th, just as the dust was beginning to settle, Senator Josh Hawley introduced a bill in the US Senate that would limit Chinese access to AI technology. And I think it's worth just pausing on this bill in light of some of the open source points that we've been talking through. And the bill is called “Decoupling America's Artificial Intelligence Capabilities from China Act” of 2025. That is like a mouthful, and I'll unpack some of the implications quickly for you. In a nutshell, the bill proposes to, first off, ban the import from or export to or investment in any AI technology or IP developed in China. And then second, it would prohibit any collaboration or transfer of AI research with any, “entity of concern” connected to China. And this term, entity of concern, is defined in a really broadly way to include not just the Chinese government or military or corporations, but also colleges, universities, labs or individuals working on behalf of an entity of concern. So there's a lot to talk about in terms of those definitions and the breadth here, but it would really have very significant implications in terms of banning what we might think of as imports or exports to China of AI technology. And it's hard to imagine how an AI model could stay open source but then be closed off to people who would be encompassed within this law or covered by the law. But it also would really prevent the US from looking at or benefiting from Chinese open source models, as US companies wouldn't be able to download or learn from or adapt to the new technology coming out of China.

These are obviously a reaction to DeepSeek. Particularly, it's interesting that this purports to restrict AI imports without any carve-outs or kind of limitations based on model weights or codes or size. But we're seeing this kind of debate come up now in the policy space in the US around imports and exports that are kind of like sweeping in open source. And ultimately, the debate is going to be whether the pain of losing access to Chinese technology and losing China as a market to sell AI technology is worth the gain of making it harder for bad actors to steal technology or benefit from US technology. There's a lot there. And it can also really shape the future arc of AI innovation. So if we think about the fact that open models like DeepSeek will continue to be made available on a global basis, right? DeepSeek is available to anyone in the world. But if American companies, including startups, are precluded from leveraging those models in their product or model development, it could create other opportunities for countries to catch up to the US more quickly than they may have done without that access and with the US having that access. So it's really an interesting kind of geopolitical overlay on all of this. We're not sure if this bill is going to progress, but we'll keep you updated if it does. And with that, I think it's all we have time for today. I'm Anna Gressel, signing off from the UAE. Make sure to like and share the podcast if you've been enjoying it.