Inside White House AI Policy

Episode transcript:

Note: This transcript is generated from a recorded conversation and may contain errors or omissions. It has been edited for clarity but may not fully capture the original intent or context. For accurate interpretation, please refer to the original audio. 

JOHN QUINN: This is John Quinn, and this is Law, disrupted. And today we have a super interesting guest, and that’s Sriram Krishnan, who until very recently, about a month ago, Sriram, you left the White House?

SRIRAM KRISHNAN: A little less than a month, yeah,

JOHN QUINN: A little less than a month ago, until he left the White House, he served as the Senior White House Policy Advisor on Artificial Intelligence.

And there he was very involved in architecting the US AI Action Plan. We’ll talk about that a little bit. In that role, he worked closely with the Department of Commerce on AI export control policy and semiconductor supply chain strategy, and engaged directly with allied governments and industry on questions of compute access, advanced chip flows, and sovereign AI capability.

I mean, before he went into government service, he was a general partner at Andreessen Horowitz, a firm that we are very proud to represent, where he led the firm’s international expansion and invested across AI infrastructure and consumer technology. Before that, he held senior product leadership roles at Microsoft, Meta, and Twitter, including early work building Windows Azure and scaling Meta’s mobile advertising platform.

So Sriram thanks so much for joining us for this show. I mean, you’ve been involved in sort of at the cutting edge of AI regulation policy in this country. Very much a hot topic, a topic on which people have strongly held different points of view. Let me just pose to kick this off.

Some people might say that the Trump administration started out being very, you might say, anti-regulation, wanted to preempt all state regulation of AI, rescinded the Executive Order 14110, which the Biden administration had proposed. So there was kind of an anti-regulation attitude, it seemed, in the Trump administration.

But more recently, we have the Trump administration adopting policies, you know, stopping for example blocking temporarily the release of the Claude Mythos-5 AI, restricting OpenAI’s GPT 5.6. Now we have this voluntary review program. What would your response be to that critique, that the Trump administration has kind of flip-flopped in its attitude towards regulation of AI?

SRIRAM KRISHNAN: First, John, before we get there, I’ll get that in a second. It’s an honor to be here. You just told me that there’s been over 200 episodes, so congratulations. I just wanna say for all the listeners, when John was first introduced to me by a mutual friend, this mutual friend said, “John is the most scary and fearsome person in all of the legal practice, and you do not want him as an enemy.”

JOHN QUINN: I’m not scary.

SRIRAM KRISHNAN: I just wanna say I’m here under duress, and, you know, and I’m here so I don’t piss John off. So this is not of my own free will. Now, with that disclaimer out of the way, no, thank you so much. And I sort of dispute all of those characterizations. I think, and just to be very clear, I am now a private citizen.

I am no longer living off your taxpayer money, John..

JOHN QUINN: About time.

SRIRAM KRISHNAN: Hope you feel you got a good deal. So I don’t speak for the administration anymore, but I have a lot of friends over there. I talk to them all the time, and I was obviously very involved for the 18, 19 month stretch until a few weeks ago.

So just to kind of set a little bit of groundwork here, the Trump administration on AI, in which I and David Sacks were a big part of, we came in with a very strong worldview that the, what the Biden administration had done around AI, which purely focused on existential risks, which purely, you know, made it very difficult for our allies to get access to our technology, was absolutely the wrong way to handle this AI race with China.

I remember my very first day on the job, even before I got my job, I was called into the White House. This was, you know, before I’d gotten sworn in. It’s got a few hours to go, so I was not officially a government employee yet. And, you know, I was like, “Hey, there is this new Chinese model that has come out called DeepSeek, and people in the White House, all the way to the Oval Office, including the President, want to know what it is about, why is it doing well.”

And if you remember the question was, are they able to do things very cheaply? And this was literally even before I’d gotten the job, and I remember getting, having to go on a guest pass to go brief everybody. So that was my… You know, that was how I got started. And we came in, and our whole worldview then, now has been how we win this AI race with China how we make sure as this technology progresses, we stay ahead.

So every single thing that you mentioned, for example, the first couple of days, we rescinded a bunch of executive orders by the way, I will keep saying we. I haven’t gotten used to the idea of not, you know, the administration. So, you know, give me some grace over there. It’s been a couple of weeks.

But in the first few weeks, the administration came in, the president rescinded the executive order, which made it very hard for our allies to get access to our technology. President Trump also gave us four to five months to come up with the new plan, the AI action plan. And you are right, and all of it was with this idea that America needs to win. And the way we need to win is to make sure to quote David Sacks, “We let Silicon Valley cook.”

We make sure our innovation ecosystem all of the wonderful entrepreneurs and companies I know you work with, but several of them, you know, they do what they do best because that is how America wins. Now, this does not mean, and I think this is where the crux of what you’re asking is, that we are blind or not cognizant of the risks posed by this technology.

If you go back to the action plan,it’s a riveting document. If any of you are like, “Hey, I, you know, I need some dramatic thriller reading,” it’s 30 pages. I’m very proud of it. I highly recommend it. There is a lot in there about how we are cognizant about some of the risks posed by this technology across many, many levels.

And, yeah, and behind closed doors, we had many conversations about how to handle that. So the first thing I would say is our worldview was the way we win is through less red tape, less bureaucracy but it being incredibly cognizant of the novel risk the, these may pose. And our view was that the best way to handle these risks is through more technology, more innovation, more getting the flywheel of the ecosystem going.

Now, that’s kind of the baseline. We can kind of dig into that. Now the question I think you’re asking is: How does this then converge or match all the events of the last three to four months? And what you are referring to for those of you who may not have been paying close attention every single day, is about, I guess, three months ago now, Anthropic came to the administration.

They said that we have a cyber model, a model that has advanced cyber capabilities, especially, you know, for in layman’s terms, hacking capabilities, being able to discover and exploit..

JOHN QUINN: Yeah, discover vulnerabilities in just about any security system right? 

SRIRAM KRISHNAN: 

Exactly, and second, they had a consumer version of that, Fable, which I think all of us can use, where you know, there were safeguards in place to make sure that regular people like you or me could not access these capabilities.

And the incident you’re referring to and I think I will talk, I’ll point to what David Sacks said, is the government basically felt at the time that they were caught by surprise, and they wanted to make sure that our critical national infrastructure, whether it is a power plant, whether it is, you know, our armed forces the you know, the vice president would always say, “How do I…

You know, I have faith that the big tech companies like Microsoft or Apple will take care of themselves. Who’s going to protect the water treatment plant in Indiana, which may not have a huge security budget or may not have the most advanced capabilities?” So our view was, how do we make sure that our critical systems are safe and to have the time to do so?

Because ultimately the government is responsible. So I keep repeating this phrase, and I’ll end on this. Two things can be true at the exact same time. It can be very true that you know, we don’t want bureaucratic red tape. We don’t want,.. And by the way, you know I’m gonna piss off a lot of the lawyers listening..

We don’t want a huge legal process where you need a bunch of lawyers to submit an application to a government agency before you can launch a model. Nobody wants that. You know, sorry for those among you, you know, who have to do that today. On the other hand, that doesn’t mean that if there is something which threatens national infrastructure, the government is not going to go, “Oh, wait, wait, let’s make sure everything is safe.”

Both those can be true. And just to finish, the one thing I said this recently in the Financial Times is we always say we will never have a huge bureaucracy for AI. And what I mean by that is if you look at certain other regulated sectors, look at nuclear. Before this administration, for thirty plus years, not a single reactor was approved.

If you go look at the FDA for the last several decades and all these stories of the bureaucracy involved, we will never have in the Trump administration a world where you need months of approval, a lot of lawyers involved. We will never have that. You know, that is not what this government will ever wind up doing.

JOHN QUINN: Yeah. Okay, well let’s talk about I mean, before we get into the latter part, you know, the government wanting to check and make sure that national systems, et cetera, are safe, the kind of thing that caused the pause on, um, Mythos and Fable 5. Let’s talk about what type of regulation that’s sort of de minimis that might make sense in your view, in the Trump administration’s view while you were there.

If you look at the 50 states, and we’ve done these surveys of what AI regulation exists in the different states, the most common issues are the use of AI in making employment decisions and sort of deep fakes, use of AI in controlling you know, fake news and impersonation and the like. In your view, in the Trump administration’s view, what was the government’s attitude, just by way of example, towards regulation of those two types?

SRIRAM KRISHNAN: Before I get to that, I wanna ask you a question. You’re obviously very, very plugged into the AI regulation landscape. I follow your excellent newsletter on the topic, and every once in a while I’ll say, “Well, I disagree,” or, “Well, I know that’s like, you know, how, not how I think about it.” What do you think the right AI regulatory approach per state and per country should be?

JOHN QUINN: Look, I’m not a big fan of this patchwork of regulation in the different states. I know we have that. You know, Illinois has this biometric secure- you know, Personal Biometric Information Act which trips up companies all the time. I think this is an area where we do need some national law. I would say the same thing about privacy, by the way.

We’ve never been able… It’s kind of a stunning fact that we’ve never been able to get a national privacy law, so you have this patchwork, and companies have to comply with all these different rules in different countries. I guess I, you know, I kind of see the point about the use of AI in employment. 

I maybe don’t know enough about whether to have an opinion about whether the regulation is actually effective. Like in New York, for example, New York City, they have an ordinance which apparently every year you have to audit your system and show that it is not biased. Now, data scientists will tell you that’s easier said than done.

What does that mean to say, you know, your system is not biased? I fear we have lawmakers enacting rules like that which may not be possible to enforce or, you know, kind of might be a dead end but, you know, aspirationally, I understand the value of that. Obviously, we don’t want AIs that are biased, that are trained on subsets of the population with the result that you’re gonna have biased outcomes.

If we can do something with AI to combat deepfakes and, you know, some way of you know, ensuring or identifying the authenticity of posts and information and the like, I do think that’s a threat to democracy. You know, fake news and the like. But again, it seems to me this is not an individual state’s problem. This is a national problem.

SRIRAM KRISHNAN: Great. And then just to draw you out a little bit on this before, because I think we are in violent agreement, what would you then do nationally, especially when working with Congress is not often the easiest task?

JOHN QUINN: Look, you know, I referred to the absence of a national privacy law. For years and years, we have, there’s been an effort to enact one. They haven’t been able to report it out of committee in Congress. So I mean, that’s just an unfortunate fact about our government these days. So the president has been, you know, he’s taken the initiative with executive orders. I’m not sure we have any choice except to do things by executive order.

SRIRAM KRISHNAN: Exactly. And I think so I sort of wanted to ask you that because you’re so plugged in, just to walk you down the path of how we acted. So just to get back to what you first said, the patchwork of law is a terrible place to be because what happens in practice is the big states, New York, California, can really lay down the law for how the country acts.

JOHN QUINN: Yeah, in this country that’s called the California effect. In Europe it’s called the Brussels effect. You know, which becomes global. The Brussels… Everybody, nobody’s gonna write off Europe, so of course you’re gonna comply with the GDPR, for example, which I don’t, I don’t think has improved anybody’s lives. Like, check the box, accept all cookies. I don’t think that’s helped any of us.

SRIRAM KRISHNAN: I feel so much safer every time I hit I accept in the half a millisecond, you know, on every website. I feel so much more safe. But you’re absolutely right. When Brussels does that, it’s to set a rule for the rest of the country, and that happens in California. So a couple of things.

One is I think nobody wants our startups and entrepreneurs to have to think about, “Okay, how do I launch the new model, the new ChatGPT, the new Claude, the new Muse Spark, the new Grok, to 50 different states and go, ‘Okay, great. I’m now going to hire a bunch of lawyers in a compliance department and go figure out how to launch this across 50 different states.'”

I’ll tell you this. I was in Europe for a summer after I left the job and I was traveling in Europe. I could not use Meta’s Muse Spark because Meta basically has said, “I’m gonna give up, and I can’t launch this in Europe because it’s too complicated.” So nobody wants that across 50 states So then you have two options.

One is to work with Congress, which we are going to do. The administration has basically said that we have come out with a national legislative framework, which I’ll come to in a second, and that we are going to work with Congress to figure out how to pass that. Now I’m not going to sit here and say that is the easiest task.

I have learned that working with Congress has a lot of machinations, but I am hopeful that this administration is going to try and get some consensus on the Hill. But what we did was we said we do not want a regime of 50 different states, especially New York and California, setting the rules for the road.

Just to give you a couple of examples, California, a couple of years ago, with SB 1047, essentially tried to ban open source across the country. If that had passed, and if Governor Newsom hadn’t vetoed it, open source models would have been de facto illegal because the CEOs of these companies would have personal criminal liability for what would have happened.

New York, as you said, has a bunch of different things which I think are not healthy for models. So we basically came and said, “Okay, we are going to have an executive order. We are going to make sure that communities across America do have protections,” and we call it the four Cs. One, on creators and IP, which I think goes to what you talked about.

If you make a song and you pretend to be Taylor Swift, right? Like, that should not be allowed, and you should be protected from that. Number two is community. You know, how do communities have rights about data centers that are close to them, and how do they get protected? Three are children. Lots of topics about how do we make sure that when kids talk to a model, how are they protected, you know, and essentially when the model is maybe telling them to do things which are not healthy for them.

And four is censorship, which is how do we actually remove bias from all of our models and you know, for any number of reasons. So we said we’re going to… We call it the four Cs, and these are going to be protected. And what the administration said was if you have onerous state laws, and trust me, there’s a bunch across the country, if you have onerous state laws, then the administration is going to try and stop it.

And if you look at the last few weeks, there has been a lot of progress. The administration has been working with many, many people on the right and some on the left to try and get this national framework passed. Now, do I know whether it’ll pass for sure? I am very hopeful, but you and I both know that Congress is a very interesting entity.

JOHN QUINN: Yeah. So that would basically preempt individual state laws and replace that with a uniform national law. So we would not have this patchwork.

SRIRAM KRISHNAN: Exactly. That’s the idea. And the idea is that until we get that national law, this executive order basically lays out all the four things that we protect people from, which I think capture all the issues that you talked about and some others and gives it the tools to basically say, “Hey, if there is a state which is pushing a very onerous piece of legislation, we have a tool set to try and stop it.”

JOHN QUINN: Yeah. So, I mean, let’s talk, let’s move to the other subject about the government sort of pausing and wanting a first look at new models because, you know, potential threats to national security and the like. Did I understand you to say that Anthropic actually, when Mythos, before it was released, Anthropic actually came to the government on its own initiative and said, “Hey, we have this new model that you ought to know about this”?

Is that how that went down?

SRIRAM KRISHNAN: Yes, and, and this has been publicly reported in multiple places. Anthropic, um, I’m not sure which part of the government they approached first or how it happened, but the administration definitely knew I would say maybe a couple of weeks or a week before Project Glasswing was announced, that there is a new model which has novel capabilities on cyber.

And I think the question from the government was, okay, this is amazing for America because we now have mechanisms to basically make our stuff safer. Whether you’re running a network, whether you’re running software, we can, you know, build a fortress across America, American software and all of it, which is amazing.

But the issue was that how do we make sure that if our adversaries get access to this, and our adversary could be a 17-year-old script kiddie somewhere in Eastern Europe, or it could be a nation state actor, how do we make sure that they don’t use this to cause harm or to cause mischief, either deliberately or not deliberately, right?

And because we do have a responsibility to make sure our critical systems are secure. So finding that balance and essentially making sure our critical systems are safe, whether it’s the banking system, the global financial system, or healthcare or power or, you know, you can go down the list. How do we make sure they are safe and protected?

So that was the real focus.

JOHN QUINN: Yeah. So now we… It seems that the government is preparing to formalize some type of voluntary pre-release review program, which has now been in the works. There’s been some announcements about what that might look like, and I think, if I understand correctly, it’s going to be formalized and maybe in an executive order or something.

Can you tell us a little bit about that?

SRIRAM KRISHNAN: Well, I think we are recording this on July 28th, and there has been a lot of reporting about things the government might be doing and which I’ll let people go read on their own. I think what the government has been trying to do, um, and you should go back, everyone here, and look at the executive order which was published perhaps, I think forty-five days ago or so, which basically said the US gov.. We will never have a regulatory body or a licensing regime for AI.

I forget the exact text, but there are a couple of lines which basically say this in, um say this directly. But we want to make sure that when the people who are tasked with keeping our country’s system safe, they get to see this maybe a few days before the people who wish us harm do. Which I think is a perfectly reasonable thing to do, so.

Now, the other part of this is that this only really matters to a very, very small subset of companies and for a very small set of capabilities. We are talking purely about models which can exceed or meet certain high cyber and, you know, other capabilities and so if your model is just amazing at, I don’t know, something else, the government doesn’t care.

It is purely about if your model is at the frontier or exceeding this capability, then, you know, the government just wants these companies to say, “Hey, give us a heads up. Work with us so we can protect our systems first before the whole world has access.” Which I think is a perfectly reasonable thing to do.

JOHN QUINN: Yeah, but is there a potential that the government could block it? Could look at it and say, “Wait a second we don’t want this released.” Is that in the cards?

SRIRAM KRISHNAN: I think the government has always been that we are not to quote Susie Wilde’s tweet, “We are not in the business of picking winners and losers. We are not in the business of, you know, putting down red flags.” I actually think it’s a bit of a moot discussion because in my experience, all the labs and the government are super collaborative.

They all have the same goals. They all want to keep our system safe. Nobody wants to have a model out there which causes something bad to happen. Nobody wants that. So I kind of find the whole thing theoretical and moot because in practice, it has been a very collaborative, positive set of engagements

JOHN QUINN: Yeah. So this voluntary pre-review program is something that the large model companies aren’t really opposing?

SRIRAM KRISHNAN: I think, you’re talking about, I think AI and Anthropic?

JOHN QUINN: Yeah

SRIRAM KRISHNAN: I think they have basically, if you look at the last two months, they have been hey, this is a very productive collaboration with various parts of the US government, whether it is the the people who focus on security, whether it is the Pentagon, there’s a bunch of other folks involved, that it’s a very positive collaboration with them because I will bring it back to this, everybody has the same goal.

Like, no company wants to go out and say, “Well, you know, you folks might think this is dangerous, and, you know, there’s a bunch of vulnerable systems,” but he’s gonna go out because they don’t want to see really bad things happening. And again, I would say this is very narrow. This is for a very, very small set of models which can meet or exceed capabilities only on a couple of dimensions.

And the idea is we, the good guys, get access to this and use it to fix our stuff to find vulnerabilities before the bad guys do.

JOHN QUINN: Right. 

SRIRAM KRISHNAN: By the way, John, I wanna ask you this. What is your read on that? Do you agree or disagree? What would you do?

JOHN QUINN: I mean, it seems perfectly reasonable to have a pre-government review rather than something that’s really cutting edge, bleeding edge maybe that may have some risks associated with it just released into the wild without somebody having a look at it first. So I think that makes perfect sense

SRIRAM KRISHNAN: Great. I’m just writing a note here. John Quinn agrees with America, you know, President Trump’s policy, right? That should be the headline, you know, for this podcast.

JOHN QUINN: No, on these subjects, I think I do have a lot of agreement. But let’s change the subject a little bit and talk about, you know, the potential for international regulation of some kind. You know, President Xi of China has been talking about this, that there should be some international bodies or standards or controls or cooperation.

Demis Hassabis has just come out with what seems to be a thoughtful proposal for a US-led global AI standard watchdog of some kind where there would be, I guess, models when they’re released, they would first be subject to some 30-day review period by this international body.

And, you know, remarkably in Silicon Valley, there’s been a lot of endorsements of his idea from people like Satya Nadella, Sam Altman, and Elon Musk agreeing on this. What are your thoughts about this subject? You know, the analogy is often drawn to after the Second World War, after Hiroshima and Nagasaki, Oppenheimer and many other scientists were calling for international regulation.

That never happened, in part because neither the US or Russia would cooperate in that. I mean, what’s your view on the international regulatory body? Is there a role for that? Can that happen? What might it look like?

SRIRAM KRISHNAN: I think you’re conflating two different things here. I think you’re conflating the role for various international bodies on AI regulation and what the specifics of the Demis proposal are. So on the former, what I would say is the America is in a unique moment, right? Think about the leading labs, think about the leading chip companies, think about data center capacity, think about every part of it up and down the stack.

We are in the lead, right? And this race is ours to lose. Everyone else is behind, starting with China. So first off, right, like, you know, I, I think our first worldview is how do we protect that and how do we make sure America and our allies, right? We have a lot of, you know, allies around the world.

How do we make sure they get access to American technology? And by the way, this is one of the things, just to take a little tangent here, I think the previous regime got it wrong. My first two weeks on the job, I went to Paris for the AI summit, the previous one, which happened in Paris, and this was a little bit of, you know, the first time we had gotten a chance to talk about AI policy.

The vice president did a big speech but after the speech, I spent the entire day talking to various ambassadors of various countries. And I remember getting so many ambassadors basically being so angry about the previous regime’s posture about you can’t have access to our technology. And I don’t want to name them, but they would go say, “We are your allies.

We’re helping you in X or Y situations, but you won’t sell us something that a teenager in America can get access to,” right? And so first off, the Trump administration, I think, has done a lot of amazing work in making sure our allies in Europe, our allies in, you know, across the Five Eyes, many, many other parts of the world, they get access to the American stack from chips on upwards.

And I can tell you, you know, as I talked to them over the last 18 months, so many of our…these other countries are really, really happy to work with us because they want access to it. That’s number one. Number two, I have a lot of suspicion when someone says international bodies, because going back to the Brussels effect, this feels to me often like, okay, how do we get the 2026 version of accepting cookies and banners across the world sitting far away from America?

And I think it would be a very bad idea to have this amazing lead that we have, thanks to Silicon Valley, thanks to these entrepreneurs, thanks to the financial markets, to basically be handed over to a bunch of bureaucrats sitting, you know, far, far away who do not have America’s best interests at heart.

So that’s number two. Now, to Demis’ proposal, I think you misread it. I think what Demis talks about is having an SRO. And when I think about it, I think about something like the MPAA in Hollywood, where instead of having… You know, if you go back to the era of Jack Valenti and how the MPAA came about you know, nobody wanted governments to come in and to go watch every single movie and be like, “Well, I don’t know, like, that’s too much skin over there, too much sex, you know, too much swearing,” or whatever the case may be.

JOHN QUINN: So they set their own rules. They set up an organization to set their own rules for the industry.

SRIRAM KRISHNAN:  Exactly, right? And because they felt that they were more qualified, they could set up the process, and it was essentially a little bit of, we will work on this and you hold us accountable. I think when I look at Demis’ proposal, I think it is about how to, again, in a very narrow set of cases where you have a bunch of well, a very small set of models that have advanced cyber capabilities or bio capabilities, how does the industry…

Because the industry, by the way, is best positioned. Often they have the talent, they have the compute, they have the technology skills. How do they come together as an industry to hold, you know, make sure they’re staying safe? Whether you are a hyperscaler, whether you are a lab, whether you are a, you know, a chip company, how do you work together on this?

And I think that’s why you’ve seen such a positive reaction because I think it’s a wonderful proposal. There’s a lot of great elements in it because a lot of people are like, “Yes, this is our problem to solve collectively instead of asking somebody else to solve for this.” So in some ways, I think about this as a modern version of APA.

And I, I do think, by the way, we need to figure out a way to have our allies involved. There are some amazing model companies in the UK, in Europe, and I think it would be great to have them involved, but I do not want to see a world where we hand over control of American AI to a bunch of bureaucrats in another country

JOHN QUINN: Yeah. No, I get that. I understand the distinction you’re drawing there, and it’s a very important one. I mean, you’ve made the point a couple of times about how we’re in this race with China, and this race is ours to lose. We’re ahead across the board. But let me just raise the question: Are we? I mean, you now have this: what is it?

Kimi 3 model released by Moonshot, which apparently is amazing. I mean, whether or not it was distilled from a US model or not, it’s incredible. And I’ve read that they’re about to, they’re working on, they’re saying the next one will be even bigger. Many, many more parameters. China seems to be ahead in open source, isn’t it?

And open source is important.

SRIRAM KRISHNAN: I think you’re saying a couple of things here. Are we ahead in the race? Let me ask you this. What models do you use every single day? Okay. What models do you see your clients using?

JOHN QUINN: Claude. Similarly, Claude a lot, OpenAI as well

SRIRAM KRISHNAN: Okay. What do you think are the most capitalized companies in the world on AI? Who do you have? OpenAI, Anthropic, NVIDIA, there’s a few others. So I’m just trying to put up a little bit of a scoreboard here, you know..

JOHN QUINN: Of course..

SRIRAM KRISHNAN: where I’m gonna- 

JOHN QUINN: But you do read about US companies you know, I do read these stories about how they’re using, they find these Chinese models more efficient, they’re less expensive to use. They are getting traction in US industry, at least from what I read

SRIRAM KRISHNAN: I think so I’m trying to say two things. America is absolutely in the lead in the race, but we don’t have a big lead. I think the lead is small because it turns out that it is very, very possible to build open weight models or open source models, you can say that interchangeably, which are very, very close to the frontier.

And currently, the Chinese models, whether they are from Moonshot, Kimi K3, Zhipu, um, DeepSeek might have a new one coming out soon. There are others, you know, Seesaw, for example, on the video side or the multimodal side. There’s a bunch of others. They are very, very good. And I would argue, you know, our American open weight models are really good.

We have, for example, one from Reflect Thinking Machines. We have Nemotron from NVIDIA. We have Gemma, right? Like I would say they are near the Chinese models, but probably not ahead. So it is absolutely important that America finds a way to win on open weight models..

JOHN QUINN: Why is that so important? 

Because open weight models have a very critical role to play in the ecosystem. I’m sure folks have seen the comment from Satya Nadella or Alex Karp over the last few weeks, where a lot of companies go, “I want to maybe use the frontier model for things that require frontier intelligence, but I want either control, or I want to fine-tune my model, or I want to optimize for cost.

And so I don’t want to pay  frontier token margins.” And this is where open weight models have a lo-role to play. I think in my mind, I think about control, the ability to customize and fine-tune, and cost as sort of the drivers for open weights. I also think there is a huge security element because when you have an open weight model, you can look at it, you can fine-tune it, you can customize it in any number of ways and in a way that you may not be able to do with a closed model.

So I think it is absolutely important for open weight models. I’ll give you a story. Again, as we are recording this, one of the hot stories from last week was about OpenAI having an incident where one of their models, you know, broke out of their network and wound up attacking Hugging Face which is a, you know, which is a little bit like the GitHub of open weight models, if you will.

And what I think one of the most notable parts of the story for me is that when Hugging Face used various closed models for defense, where they were like, “Okay, how do we make sure our systems are protected? There’s obviously an LLM trying to attack us. We need to protect it.” They found that they could not use a closed model, all the leading closed models.

Because it turns out all the leading closed models, they said, “This looks like insecure stuff to me. We don’t want to deal with it.” So what do they have to do? They had to turn to a Chinese open source model, which actually did a great job. And so just think about this for a second, right? We have maybe one of the first ever model escapes containment issues, you know, that people have long talked about.

But it turns out that it is the closed model which is maybe not being helpful or negative, and it’s the Chinese open weight model which actually was critical for defense. This is not a good situation to be in. I’m just laying out a fact pattern here, right? Like and this is not a good situation to be in.

JOHN QUINN: What was the advantage that the Chinese open weight model had in playing defense?

Because all the American op- closed models refused to engage in security. They basically threw up their hands and they said, “This looks insecure. We can’t deal with it.” I think this is a, this is, I think a little bit of a pattern where some of these models have way too aggressive refusal rates because, you know, they are either trying to play it safe or there are some, you know, or they’re not being, They aren’t getting the balance between attack and defense right. 

But I think this is one of the advantages of open weight models, because if you have an open weight model, right, you are free to do with it as you want, and in this case play a defense. So for me, this is a wonderful advertisement for open weight models, that an open weight model, you know, played defense and saved the day, right?

It was a knight in shining armor. However, look, I don’t wanna give credit to some of these Chinese companies. I think they do really smart work. But as somebody who is in the job of trying to win the race for America, I would have really preferred to have seen an American open weight model play in that role.

And my hope is, whether it is from Thinking Machines or from Nemotron, or there are a couple of other companies like Reflection in the, you know, waiting in the wings, we will soon see a future where it’s American open weight models that are used.

JOHN QUINN: Yep. Sriram, what’s next for you? I mean, you left the White House a little less than a month ago. What are you up to now?

SRIRAM KRISHNAN: What do you think I should be doing? Give me career and life advice, John.

JOHN QUINN: Well, you and I have talked a little bit, so I know some of the things that are going on in your head

SRIRAM KRISHNAN: John’s a very scary figure, so I just listen to whatever he tells me. I think for me it is very hard to convey to people how special the experience was to work in the White House and to be a very, very tiny part of, you know, just serving the country.

And I feel very fortunate. And in a lot of ways it has given me a front row seat in the Oval, in the Situation Room to see what we as a country and other countries need to do to win on AI. So, you know, I’ll be a little bit mysterious, but I’ll say I wanna find ways to continue that work. And because I think the next two years or 18 months are going to be critical.

I’m very much a believer that we are riding this exponent and the models are going to continue to get faster. I mean, I’m gonna continue to get better faster. And that does have all sorts of downstream implications that countries are only now starting to wake up to, starting with America. So for example, whether it means how do we get our data centers, whether it means how do we have onshoring in place, any number of topics, I want to be helpful.

So that’s what I plan on working on. I’m gonna be mysterious and vague, but stay tuned

JOHN QUINN: No, I mean, look, Sriram obviously you have a lot to offer, and I have no doubt that you’re gonna be able to make, continue to make important contributions for this country and internationally in thinking about AI, the future. Thanks so much for appearing on the podcast. This is John Quinn.

This has been Law, disrupted.

---

Published: Aug 7 2026