Sriram Krishnan on Open Source AI's Biggest Week Yet

Sriram Krishnan joins Theo Jaffee and Sofia Puccini just after concluding his tenure as Senior White House Policy Advisor on AI to discuss one of the biggest weeks yet for open-source AI. They unpack the rapid release of models including Kimi K3 and Qwen, why open models are putting pressure on frontier labs, and what it means for pricing, competition, and the future of AI infrastructure. They also discuss AI policy, distillation, cybersecurity, the role of open-weight models, whether the U. S. should respond to China's growing AI capabilities, and how governments and frontier labs should navigate the next phase of AI development.

• 0:00Speaker 00

  You can bring it back to very business first principles. If you're providing a product of value, capitalism will find a way to make the supply chain work for you. So if you have an open weight model that is providing value, that means that every part of the stack underneath, whether it is a NeoCloud, the chip provider, somebody who provides gas turbines or fire suppression is going to orient itself to provide value. If you're providing a product of value, capitalism will take care of all the rest. If you go look at how the rest of the ecosystem is going, the growth is pretty strong and spectacular. And I think you're going to seeing that continue.

• 0:39Speaker 01

  Open source AI is moving faster than ever and the balance of power in the industry may be shifting. In this episode, Theo Jaffe and Sophia Puccini are joined by former White House AI policy advisor, Sriram Krishnan, to unpack with the latest wave of open models needs for frontier labs, AI policy, pricing, cybersecurity and America's position in the global AI race.

• 1:06Speaker 02

  We are back. We are live with Sriram Krishnan, who just finished his tenure at the Senior White House Policy Advisor on artificial intelligence. Previously, he was a general partner at Andreessen Horowitz and held senior roles at Microsoft, Meta, Snap and Twitter. So Sriram, we're so glad to have you on. Welcome to MTS.

• 1:25Speaker 00

  Thank you. Been a fan of everything you folks have been doing for the last few months. And excited to be here. I think this is the first time in about two years I've been able to do a video appearance without a suit and tie on. So I am so excited to be out of that.

• 1:44Speaker 02

  Yeah, amazing. So there's so much going on in open source. The last week, we just discussed we had Grok Build was open source. And then we had Thinking Machines and then Kimi K3 and then Quen 3.8. So you tweeted the other day, Kimi K3 is a big moment with multiple implications for the entire industry. Could you go into a little more detail on that? What are these implications?

• 2:05Speaker 00

  Yeah, so if you go back maybe four or five months, I think there was a moment in time when the only leading models were, I think, Opus 4.6 or 4.7 at the time, GPT-5.4 or 5.5 or wherever we were. And it felt like there was really no one else. And we were on this curve of self-improvement where the frontier labs were really going to draw really far away from everyone else. I think the last few weeks, if you are in the token consumption business, which I am, and I think many of you and your viewers are, it's been a great time because let's see, you had Elon and Michael at Cursor,

• 2:48Speaker 00

  you know, the SpaceX XAI team come out with Grok 4.5, which I've been using. It's a fantastic model. I think we forgot to mention this. We had Alex Wang and Meta come out with Muse Spark, which is also awesome. Last week, we had Mira Thinky come out with Inkling. I don't know their version number, but you know, the first version of that model, which I think is nearly Sota on many, many benchmarks. But I think the big news over the last three, four days was obviously Kimi K3 coming out, I think on Thursday or Friday. And then I think the last 24 hours, I haven't really played with it yet, but Quen coming out.

• 3:32Speaker 00

  So there's a lot of, just a lot of choices and alternatives coming out. And you know, what I was referring to is with Kimi K3 is the following. One is that it's just great to have choice in the ecosystem and to be able to point your harness of choice or you know, your agent of choice to multiple models. Second, I think we are in this really weird moment now where some of the American frontier models are constrained. For example, on cyber and on security. And I was talking to a friend of mine where this person was actually starting to do security work

• 4:15Speaker 00

  using Kimi K3 rather than Fable, because with Fable, he would run into these refusals and safeguards. So that seems like a very weird spot to be, which we can talk about for a second. I think it's also, you know, it's probably inevitable that if you are having choices from where you get your intelligence tokens from, that's going to put pricing pressure on the frontier models, which means I think you'll probably see the frontier labs have to drop token prices or find ways to, you know, find ways to match pricing, which means it's probably going to erode into their gross margin.

• 4:56Speaker 00

  It's probably great for the NeoClouds and every other layer of the stack. Because if you're a NeoCloud, like a Bastion or a Fireworks, or if you just have a bunch of black walls and you can power it and you can run that and you know, you can capture some of the economics, which is probably going to go to the frontier lab. So I think it's great as a consumer. I think it's great for the ecosystem. Lots of obviously questions on security, on distillation, on you know, how this is good. A very, very interesting moment.

• 5:28Speaker 03

  Totally. So I guess like the first question here would be, where do you predict the frontier labs? Like, how do you predict the frontier labs are going to react to this? Like now, you know, the ecosystem has been sort of changed. Like there's no going back. Like Kimi has been released and it's very close to the capabilities of Fable. So what do you think is next for like an Anthropic or an OpenAI?

• 5:49Speaker 00

  So a few things. I think it's very clear the frontier labs are going to push at the very, very frontier of, you know, the jagged performance we get. You know, as somebody who spent close to the last 18, 19 months trying to make sure America wins, I really want to see the American models, whether it's closed or open weight at the frontier, which, and so I think they'll continue to be that. What I suspect these open models could really start putting pressure on them is one on pricing. Because it may turn out that the number of tasks that you need absolutely frontier intelligence from is, you know, let's call it like one subset. But for a lot of other tasks,

• 6:34Speaker 00

  for example, like I have an agent which checks my email or I have an agent which quickly scans through my calendar. You know, you may not need frontier tokens. You may be able to get by with frontier minus one or your open weight token of choice. In which case, I think you're going to start seeing pricing pressure come on the frontier labs. You're probably going to see them, you know, try and make sure the absolute frontier models stay available and accessible. I think we've already seen Anthropic, and I have no inside confirmation about this was a motivation, but you've already seen them extend Fable. I think Fable is originally supposed to be available until, I don't know, like a week ago. It's already been extended. I predict there'll probably be more extensions just because otherwise you have an open model,

• 7:18Speaker 00

  which is very much at the frontier. I think the other interesting question is, where is the real moat? If you're a frontier lab, is it in the intelligence? Or is it in the harness? And I think, you know, Claude, you know, CoWork and, you know, Claude Code and Codex are amazing products. And I suspect you'll probably see like more effort on making them sticky. Because it might be that, you know, the actual intelligence itself, or at least the, the, the levels behind the actual frontier are much more of a commodity. So, so what does this mean for the ecosystem? You know, one, you know, this could put some revenue pressure on the frontier labs. Because I suspect that they might, you know,

• 8:02Speaker 00

  people might start pointing their harnesses at some of these other open weight models for any number of reasons, which we can get into. I think it's probably great, like I said, for the NeoClouds and anybody else with power and GPUs. And I know it's going to be a very, very interesting time.

• 8:20Speaker 02

  Right. So there's been, speaking of open source, there's been some chatter. Axios just reported this morning that the Trump administration is considering restricting Chinese open source models, either because they are Chinese and might be a national security issue or because they have reached like Fable level capabilities and agentic coding, and it could be a cybersecurity issue. So do you think the US government will do things that will crack down on open models? And do you think that there's anything that they should be doing in this respect?

• 8:53Speaker 00

  Well, I have like no inside information. As I like to tell people, I am no longer living off your American taxpayer money. But I have a lot of friends there and, you know, I think, and I think from everything I hear, look, if you go back to a year and a half ago, in the first few weeks, you know, the Trump administration said, hey, you know, we're going to come out of the AI action plan. And when we came out of the AI action plan, you know, it's right in the first page, if you kind of go to the first section of it, you're going to see the action plan talk about how important open source is. Now, I think the thing to think about is there are multiple different issues. Number one, I don't think it is great that the leading open weight models or open source models, you know, are not American, right?

• 9:38Speaker 00

  It's, you know, I think, you know, I think there's some great innovation happening with Moonshot and with DeepSeek and with Quen. But I would much rather prefer that the leading models are American. And by the way, there's lots of great efforts happening. I expect them to improve. You have, you know, Gemma from Google, you have NemoTron from Nvidia, you have obviously Thinky, you have, you know, newer startups like Reflection coming out. I think they will continue to be better, but we are in a moment of time when the leading models are Chinese, which I think, you know, which I think is not great. I would much rather prefer them to be American.

• 10:13Speaker 02

  Yeah.

• 10:14Speaker 00

  I think the second part of it is I, look, I grew up loving open source. Open source is a big part of, you know, how I got into computers, a big part of my career. And I was a big fan of Linus's law, as in Linus Torvalds of Linux fame's law. And his law was that given enough eyes, all bugs are shallow. And what I believe with that is that open weight models are inherently secure. Because, you know, when you download a model off Hugging Face, it means you have the entire world being able to take it apart, inspect it, you know, fine tune it, modify it, look at it in ways that you absolutely cannot if they are closed.

• 10:54Speaker 00

  So what I kind of believe that they bring a very, very different positive angle to security. That we're, I think, or the moment we are in, which I think is not great, is I think there was an incident with Hugging Face that got reported on earlier today, which, you know, I think there's an active tweet. And what Hugging Face seems to have found is that somebody was using, you know, essentially definitely an AI LLM agent to hammer them at multiple places and to kind of like, you know, find ways to break in. And, you know, and I think the way to counter that is to make sure American or Western or allies defenders

• 11:35Speaker 00

  have access to the best models to make one our software like just more secure.

• 11:40Speaker 02

  Yeah.

• 11:40Speaker 00

  So which is why, you know, I, like I said, I don't love that we are in a moment where it may be harder to look at your own code for exploits with, say, Fable than it is when you use a Chinese model. That's just like a weird part to be, and I think, you know, we should fix that. Now, one very interesting discussion, which has been happening a lot on X, is about distillation and what that means. So it's kind of a complex topic because there's a few things in there. So first of all, you know, like every model we have today distilled off of all human knowledge, right? Like, you know, if you go back to the original GPT or the original Claude,

• 12:24Speaker 00

  you know, they all had to derive crawling off the internet, you know, crawling off all of our, you know, blogs or tweets and content. And, you know, they kind of consume human knowledge and to bootstrap that. So distillation has always been a core part of how these models have been trained. Second, if you, you know, okay, let me ask you this. When I write a tweet these days, I am terrified of accidentally using like, you know, multiple hyphens or accidentally saying something which will cause Pangram to say this is AI generated. I once wrote a tweet recently, which I put it through Pangram to make sure it felt human or even though I knew I had written every human generated token.

• 13:08Speaker 00

  And, and so what we've all seen is, I would say, the rise of AI slop on the internet. So if you think about that, the internet has grown a lot over the last couple of years. A lot of that has been AI generated. You and I see content every single day where I go, there's definitely some tokens in there. And that also goes into the training of these models. So distillation of AI models is a core part of how these models are trained today. There's just no escaping that. That's number one. Number two, but I think the issue I think people are talking about is that it, you know, the, some of these models may be distilling at scale in kind of these industrialized ways

• 13:47Speaker 00

  where you have maybe a bunch of fake accounts or maybe you have people reselling, say, cloud subscriptions to somebody else and routing it. Any number of ways which are breaking terms of service and which are bad. So, so here's what I think. Number one is I think all the labs, you know, with the help of, you know, everybody, you know, from the government should be doing everything they can to make sure, you know, whether they're doing KYC, whether they are, you know, doing things to check the IP addresses where they're coming from. There are a lot more sophisticated things they do. They are making sure like, you know, these kind of this heavy industrial, you know, extraction of reasoning traces isn't happening. But, and I'm going to steal this from Dean Mayer of Sequoia,

• 14:30Speaker 00

  who had a fantastic post and please give him credit for this in his tweet. He wrote this yesterday. I think the situation which is bad today is that these, some of these models from other countries can train off American models. Whereas if you are an American open weight model, it may be really confusing or challenging on whether you can distill off of other American models, right? So if you kind of have a really uneven ecosystem here, where if you're a Chinese model, you could probably get a bunch of reasoning traces. But if you are, say, a new valley startup and you want to use some reasoning traces, you know, you don't know what the legal situation is. So one great idea, which I think came from him, I think Ben Thompson of Strategy had a similar idea today,

• 15:14Speaker 00

  was to basically say, how do we find a way to make distillation acceptable in any number of ways? Whether you are getting, you know, outputs of other models or sometimes it's more subtle. You know, if you look at any American open source model today, they are using Chinese models as a teacher or in a way as a part of the fine tuning process. And, you know, how do we make sure that is protected and enshrined? So if you're an American, you know, model company, you have the same level playing field as the Chinese models. Now, again, I think we're in this kind of this very interesting, unique moment in time. But overall, I think these open weight models are great for the ecosystem. It's more choice, it's, it's people competing on price. It's all the great things that innovation is supposed to bring.

• 15:59Speaker 02

  Right, totally. So another consideration that is going on right now is the idea of AI being able to automate, to improve first and then automate AI research. So, you know, Anthropic has written about this, OpenAI has talked about, you know, building in first an automated research intern and then working towards automating AI research. And many people believe that if that were to happen, the pace of AI capabilities progress would go very, very, very fast. So what kind of policies do you think the government, if any, should have around this? What do you think the government is going to do around a potential very rapid increase in capabilities of AI if that were to happen in the future?

• 16:44Speaker 00

  Well, again, I don't speak for the government anymore. You know, I don't have any inside information. I don't know whether the government has a real role in this. Like, look, whether RSI is real or not, where you are on the exponent is a much debated topic. You know, I, I've heard many, many schools of thought where, you know, they believe you're going to have automated AI researchers in a couple of years. But then there are others who think, well, there are fundamental improvements that you cannot crack. And yes, you're going to see improvements from models being able to train other models. But the exponent is going to be a lot more gradual rather than a very, very, you know, steep curve.

• 17:26Speaker 00

  So I, I think when I was in government, for me or for David, the entire focus was how do we have a fantastic ecosystem where there is, there are people competing to build products? You know, you are pushing innovation as fast as possible. And then when there are credible risks, you know, making sure when they appear, making sure you tackle them. So for example, with, you know, with cyber, you know, when you have a credible threat, when you know these models are capable of, for example, generating exploits in the latest firmware or the latest operating system, how do you then go tackle it? So when somebody gives me this question, I sometimes find it very theoretical.

• 18:07Speaker 00

  And I often find it a lot more useful to break it down into, okay, one, how do we make sure we have a bunch of competing choices? Two is what do we know to be credible? And, you know, how do we actually particularly tackle that? I think there are very credible threats on cyber, on biological advances, a couple of other topics. And I think there are definitely efforts to try and tackle just those. And then, you know, you, and I think the final part I would say is that in a lot of ways, I'm a big believer in using AI to help. So often, for example, the answer to helping with cyber is if you have, you know, using more AI to go scan your code base and making your code base more secure. So very often I think like, how do we then deploy to actually meet these threats as we go along?

• 18:53Speaker 02

  Totally, totally. So another point that Dean Ball raised actually was sort of this sentiment that like open weight models deter CapEx and they kind of like destroy the ability for frontier labs to monetize their capability leads. So how do we like promote a very healthy like open source ecosystem while, you know, ensuring that the frontier labs are still able to like do what they're best at?

• 19:18Speaker 00

  Well, I think at the end of the day, if you kind of bring it back to very business first principles, if you're providing a product of value, capitalism will find a way to make the supply chain work for you. So if you have an open weight model that is providing value, that means that every part of the stack underneath, whether it is a new cloud, whether it is a data center, whether it's a chip provider, whether it is somebody who provides gas turbines or fire suppression to a data center, every other part of the stack is going to orient itself to provide value because you, as an open weight model provider, you might be working with a bank because they don't want to work with a frontier model.

• 20:02Speaker 00

  They want to work with somebody that they can run in-house. Or if you go look at what somebody like Thinki is doing, they are fine tuning models for specific clients using internal data that they don't expose. So I kind of think of like, you know, if you're providing a product of value, the capitalism will take care of all the rest. And by the way, I think I'm being borne out. Like if you go look at the fraction of every other part of the stack, if you go look at how the inference clouds are doing, if you go look at how the rest of the ecosystem is going, the growth is pretty strong and spectacular. And I think you're going to continue seeing that continue.

• 20:44Speaker 02

  So to the extent that you can tell us, what are you working on now that you're out of government?

• 20:48Speaker 00

  Well, I'm going to be doing more, you know, live drop-ins, I think. And I'm going to be.

• 20:53Speaker 02

  Streamer.

• 20:54Speaker 00

  No, look, I think, well, look, I had this amazing life experience, which is untalled, and it was such a unique honor. And it's kind of really, you know, given me a sense of what countries and companies need to do to work with each other and to make sure we, everybody gets intelligence, whether it's from the frontier labs and open weight models or however that happens. And I want to try and make that happen in some shape or form. So I'm going to be annoyingly elusive, but I think that mission of making sure America, our allies, get access to AI at scale with having governments and these companies work together is a very important one. I've been working on that in many different ways.

• 21:38Speaker 00

  If you think about it, for many years now. So I want to continue to do just that. And, you know, write some banger posts and, you know, jump in on, you know, live streams from time to time.

• 21:50Speaker 02

  Yeah, well, we're excited. This has been so great. Thank you so much for coming on MTS.

• 21:55Speaker 03

  Yes, I'm sure we'll have you on again.

• 21:57Speaker 00

  Thank you. Thanks for having me. Much more to come.

• 21:59Speaker 02

  Absolutely.

• 22:02Speaker 01

  Thanks for listening to this episode of the A16Z Podcast. If you liked this episode, be sure to like, comment, subscribe, leave us a rating or review and share it with your friends and family. For more episodes, go to YouTube, Apple Podcasts and Spotify. Follow us on X at A16Z and subscribe to our Substack at A16Z.substack.com. Thanks again for listening and I'll see you in the next episode.

• 22:28Speaker 01

  This information is for educational purposes only and is not a recommendation to buy, hold or sell any investment or financial product. This podcast has been produced by a third party and may include paid promotional advertisements, other company references and individuals unaffiliated with A16Z. Such advertisements, companies and individuals are not endorsed by AH Capital Management LLC, A16Z or any of its affiliates. Information is from sources deemed reliable on the date of publication, but A16Z does not guarantee its accuracy.