AI政策の中間選挙連合、Anthropicの安全法案、VCの利益相反

The models that OpenAI and Anthropic have internally are a lot better than the ones that are currently deployed. And so kind of seeing what the kind of the future of this looks like and how it operates is important. Most of the people who can lead in government tend to be older and older tends to be, by definition, you know, almost way out of date. My hope, as opposed to my fear, it's the dual mandate. How do we deploy AI so both America continues to lead the world in artificial intelligence, making it American intelligence? But I think my American intelligence is better than Trump's super intelligence. Before we get into it, if you're watching on YouTube, give this one a like or subscribe to our channel. It's the best way to help more people find these conversations. Now, back to the episode. Great. So lovely to be here in person with you for this episode of Read Rift. And it's great that it's in Silicon Valley here in the Greylock offices. Absolutely. Silicon Valley, my second favorite place after New York. So as we head into the midterms, no surprise, there are several competing ideas about what we should do with AI, how we should legislate it.

So we have Bernie Sanders and Greg Sizar, and they want to ban superintelligence altogether, and they want to pause advanced AI development until we can have federal guidelines. Then we also have a bipartisan push from Chris Murphy, Democrat Chris Murphy and Republican Josh Hawley, and they have proposed holding AI developers responsible if their AIs are the ones doing the hacking. Then we have the president. Trump has been sending mixed signals. He wants to call it super intelligence. He called it a hoax. It's unclear at best where he stands, I would say. So looking ahead to 2027, we're going to have a new Congress, a new Senate. What do you think the political coalitions around AI are going to look like? So unfortunately, the naturally most aggressive coalitions are essentially probably anti-AI, therefore anti-the future, anti-American intelligence, et cetera. And I think one of the things we all need to work about is to say, no, it should be AI as American intelligence.

AI is things that are helping America. And so like a pause or a stop is a terrible idea because it's like, no, no, we should be shaping it in this direction. It's like, oh, no, no, we're only pausing or stopping it so that we can run a congressional process to figure out what it is. And if you kind of said, OK, running a congressional process to make anything in technology potentially work at all, work better, not be a catastrophic outcome, Running it through the congressional process would strike me as like maybe almost the worst possible thing you could do. And then similarly, look, I do think that figuring out kind of accountability and accountability laws is an important thing. And we have this thing like right now, like if you look at just kind of the hugging face incidents and say, well, where does accountability lie here? It's almost like, well, that's weird because it's almost like product liability or something because it's coming out of a product. It's the only kind of legal category we have for that right now.

But it's like, okay, so where should the AI model developer, where should the AI model user, where should the, you know, as an organization, where should the individual, you know, what does it relate to prompting? What does it relate to, you know, the AI agent itself? And where is all this? We're going to just really need to kind of rethink this. And I think that the right place to start is to say, hey, we should commission some smart work about thinking about what that should be now. It should be a dynamic framework for evolving because I think that the – as you know, one of the things I think about AI native organizations is they're building for the initiation of AIs of the future, which is hard for everyone, even for technologists, especially hard for non-technologists. But it's the path we're on. It's like things are changing. Like one of the things that I think people are not really taking as the right kind of generalization from like OpenAI's release of the math proofs is the models that OpenAI and Anthropic have internally are a lot better than the ones that are currently deployed.

And so like the kind of seeing what the kind of the future of this looks like and how it operates is important. My hope, as opposed to my fear, would be that the coalitions would be how do we deploy – it's the dual mandate. How do we deploy AI so both America continues to lead the world in artificial intelligence, making it American intelligence? But I think my American intelligence is better than Trump's super intelligence, you know, as a kind of a vector. and the 60 plus percent of people who are struggling in this economy, in Trump's economy, whether it's oil prices or anything else, it's helping them too. And it's not, wait, oh, in a number of years it'll get to you. It's as closely linked as possible in developing that. And that's what I would want coalitions around. Now, for decoding what Trump might actually sign in the law, I think there's probably two parts to that decoding.

One is I think that the administration is kind of robustly in competition with China. So something that was oriented in winning at China is, I think, a good path. And then the other is, you know, you kind of look at a lot of the administration's actions, and it's either grifters or clowns. And, you know, so, you know, maybe, you know, parts of it can be bought. I mean, we saw the ambassador to Greece, you know, is one explicit. But, you know, there's also freedom coin. Maybe if you buy a certain amount of freedom coin, that will enable a certain kind of AI regulatory framework, you know, et cetera, et cetera. So, like, if you say, what will Trump sign, you know, if it's not competition with China, maybe it's something in those arenas. Right, not going to be what we want. Well, so, I mean, sort of bipartisan work is seemingly more and more impossible, but you could imagine that a sort of wing of the progressive Democrats get together with the populist Republicans as the ones who oppose AI.

Could you see a similar scramble with some Republicans coming together in a bipartisan way with Democrats to promote not necessarily unfettered pro-AI, but sort of pro-AI, pro-business, pro-economics? Like, is that even possible? I think that's fairly unlikely, unfortunately, because I think that if you look at every action that Trump has taken over his political career, it's divisive versus unitary. Like he never said And so it very hard for more than call it you know one or two or three brave Republicans to break from the pack because they know that Trump will go really intensive because he doesn want any collaboration across the aisle He's not doesn't care about collaboration for the better of American society. He wants to only have, you know, people following himself. Right. Right. And so I think that the coalition to to kind of kind of American intelligence, I think, will be hard to to make work.

He won't want to give Democrats any wins. He'll want to stop it again. Like you said, he can only sign something if there's some grift in it for him or his family. Or and that was part of the reason I was saying maybe maybe geopolitics and the China competition. So that's the government side of things. But obviously, the labs are also stepping up and have their own ideas, both because they need to protect their businesses. But also, we believe they genuinely care about what are the outcomes here. And they want to make sure that AI is best poised for the future. And so Reuters reported this week that Anthropic has been privately pitching lawmakers on its own frontier AI safety bill. And they posited that frontier models would undergo independent evaluations. evaluations have to report imminent risks in 72 hours. But they also proposed sort of some safeguards that define the limits of regulators' authority. They said that suspending a model would require the Commerce Department to seek a federal judge's approval with thresholds involving imminent risks of more than a billion dollars in damage or more than 50 deaths. And so Anthropoc has also committed $40 million to public first action,

but they said that those donations are restricted to public education and policy advocacy, and they cannot fund election activity. Meanwhile, you saw there was a lot of talk about Greg Brockman and leading the future. There's a lot of people at OpenAI who were upset with that donation. They were saying this undermines the work that OpenAI is trying to do. And so recently, both OpenAI and Greg Brockman have sort of walked back those donations, and they aren't as involved. You've spent a lot of time in this world, the intersection of technology and government. And sort of previously, I feel like, you know, the technology world sort of left government alone. But now it is so intertwined that it's clear that these companies want to have a voice in Washington. How do you think these AI companies, specifically the Frontier Labs and Thropic Open AI, should engage with government right now? There's a whole set of things. So obviously is kind of what I earlier said that the Trump administration is a combination of crifters and clowns. you know, for example, take the insanity of the Iranian war and what it's doing for American and global oil prices.

I think they posited that the average American family will have to spend an extra $2,000 this year because of the Iran war. So just the kind of a set of different things there. Now, that being said, I think it's really important that the kind of all-scale tech companies work with the American governments, in all of our institutions because I think that that's actually, in fact, it's working for the American people. So I think that the tendency where people say, oh, we're really frustrated, we run around it, whether it's a red, blue, purple, whatever government, I think it's very important to work with governments. Now, obviously, the line that you have to navigate here is this kind of question of working for the people, not working for your industrial interest. part of what I think is good about what, for example, the anthropic, you know, you know, proposals doing, I think open AI is motivated the same way as they're trying to figure out how it may have different theories about how AI is good for America, but it's, it's the, how do we enable

that kind of American intelligence? How do we enable that kind of, of, of pro society basis? Now, classically, there's, there's unfortunately over time, economic interests tend to do regulatory capture. They tend to do other kinds of things as ways of doing this. And we really need to try to limit that, avoid that, et cetera. But being in a dialogue really matters. I mean, it's apparent to everybody who, you know, can, you know, has, you know, some, any attention is whatsoever that AI is going to be a massive impact on society, on work, on industries, on geopolitics, on national defense. And so having people involved and expert and articulating what's possible. And, you know, when you look at, for example, Anthropic and OpenAI, they're not just, you know, kind of saying, hey, here's all the things that can be good. They're also saying, here's some of the things that we think are dangerous to navigate and are trying to make that happen. And so, for example, I think that the Anthropic proposal of saying, hey, this is a good thing to have some, you know, regulatory predictability on, but let's go

through the rule of law system versus kind of, you know, autocratic, you know, kind of whimsical things from, you know, kind of any administration, including this one, I think is actually a good effort and sketch of this. So I think that being involved is good. Now, Now, you know, classically, a lot of the, you know, kind of more lefty progressives think, you know, industry involvement whatsoever can only lead to bad. And the thing that they don't track is that part of what it makes, I think, all democracies work is a true pluralism, right? Like you actually have to have, you know, experts, not just from academia, but also from industry involved. And by the way, industry saying we know what some of the things that are going on, and so we can argue that is a good thing. Now, spending money is one of the things you have to be, I think, transparent about, one of the things I think you have to be public about in terms of what you're doing. and I think that part of when you say both the kind of public education

that Anthropics is doing and also the shift from open AI, I think that's a response to transparent, open things where they're getting public pressure both from public and from their employees. I think that's part of the reason why the transparency is important because it's shifting around kind of what the way that you slot in to that, you know, humist, pro-American future. And I think that's what's really important. What do you think we're going to need to do to get, you know, you're talking about we need the partnership between government and industry because this is how we make innovation happen. But the government right now doesn't have the expertise. How do we solve for that? Is it more elected officials who have that background? Is it getting some folks who used to be at Anthropic or OpenAI to go into government? How do we build up that government's expertise? or is that just impossible because the economics on the other side are just too great? So I think government will trail massively industrial expertise for a couple of structural a couple of structural reasons yeah one structural reason is um that the people who tend to go into government

are not the people who you like you only learn this from being an industry right and like very very few people shift from industry and not just from economic bases it's just like hey i'm leading a successful career starting companies you know investing in things etc and then there's like to do that is very, very hard to get. And there's a lot of antibodies. And so there tends to be a kind of go into it. So that's one. Another one is most of the people who can lead in government tend to be older. And older tends to be, by definition, almost way out of date technically. Right, not on the cutting edge. Right. So and even when you get smart people like Mark Warner or Gina Rolando and so forth, they have to bring in other folks to consult on that. So that's the reason why I tend to think it's how do you assemble the kind of essentially working groups? Like how do you have NGO groups? How do you have, you know, council advisors? How do you have, like, for example, you know, it's part of the reason why, like, the Presidential, you know, Council on Science and Technology is so important.

It's like – and how do you make more of those organizations work? And then, for example, I think you have to include a strong industrial participation, but then you try to include that strong industrial participation in ways that is transparent about economic interest, is transparent about contribution, and is also selected to have people who go, look, I might be working at Company X, but I'm sitting here amongst a council of other people. And there's other folks of industry, not just like perhaps competitors to my company, but other companies as well. And then a set of other people. And then you can kind of reason to good outcomes. And I think that's what you're essentially looking for. And I think all of the kind of Western democracy governments need to do that. Now, it's great if more technologists, technology experts. You know, I tend to think that, you know, one of the things that probably like part of the thing I've I've supported in various ways is, you know, Stanford and Harvard have various ways that they try to educate both Congress people and staffers on things, whether it's cybersecurity.

You know, Condoleezza Rice does a great job with the Spogli Institute and a bunch of other folks here. Or, you know, Jennifer Lerner at Harvard does decision making and the science of decision making. I think obviously some of those things for technology and in particular for like technology broadly, but also, you know, AI technology, I think will be very important to make happen. Because even though they won't become experts, even just like getting informed about what the basic concepts are and how fast it's evolving is very important. I mean, I think we saw in the Obama administration, a lot of people in Silicon Valley were called to serve. You know, we obviously are good friends with DJ Patil, who said, no, I have to go serve this government. And I think in 28, if we have candidates who have that inspiration, innovation, friendship with Silicon Valley, we'll be able to see more technologists call to serve and sort of be experts in AI with the future. And just so people aren't too alarmed, actually, just connectivity. Yeah. Right. It doesn't have to be friendship. Right. It just, like, channels of communication and discussion.

It doesn't have to be like an alliance of any sort as much as there's a lot there that can be leveraged to the American people's benefit. So make sure that's happening, even if part of the role might be to, you know, to make sure that, you know, certain companies or certain industries are accountable to some interests that they may not otherwise be naturally given the market doing. All right. So shifting gears away from politics and government. One piece of a growing debate we're seeing is sort of the open versus closed debate happening, and specifically in the business community. You have a lot of companies that aren't Frontier Labs. They don't have their own models. But increasingly, a huge portion of their company relies on these models working well. And so when you have a company that is so reliant on an external provider, how do you make sure that, one, you're retaining IP, that these models that are coming into your company have the security, have the safeguards. I mean, Satya Nadella recently argued that companies should treat both closed and open

weight models as potential insider risks. And he said that businesses need to control their own permissions. They need to monitor actions and independently verify the results throughout these systems. And in his words, we need to separate the supply of intelligence from the authority over it. This is going to be an increased question for all companies, but specifically big companies in the Fortune 500 that are relying on Anthropoc, OpenAI, Microsoft, et cetera. So how should these companies think about their own sovereignty as it relates to their partnerships with Frontier Labs? Well, not surprising since I'm still on the Microsoft board and Satya and I, you know, talk regularly. I'm a strong endorser of Satya's, you know, kind of, call it lens, you know, attention, focus on this as a key issue. And obviously, I think that the question is, say, well, how do we maintain the integrity of our intelligence, of our capabilities, et cetera? And we don't want to treat, you know, kind of the quasi-new AI agentic employees as a bleed of could be business intelligence.

You could imagine the kind of worst parts of this might be, does it bleed to the stock market and kind of insider trading? Does it bleed to competitors that then know things? Does it bleed to suppliers that then know that they can hold you over a barrel, et cetera? So there's real reasons. It's not just kind of like a cybersecurity or a ransom hacking. There's like real reasons why that IP really matters in all these fronts. And so the question is, is, you know, where can you trust the provisioning of these models? In addition to, I think that he's right that the right way is to kind of build frameworks that, you know, can be harnesses or genetic frameworks and things that, you know, have, you know, independently by the company and by the company's auditors validation of things that are happening, which is very complex to do with models. It's one of the reasons why you tend to abstract for models. And like even an open weights model, you don't really know what's fully going on there. And so you so like treating it as like, well, we're running the model on our premise doesn't necessarily mean that mean that it has those things.

So that that attention to how do you maintain that integrity of intellectual property you know kind of company internal secret you know trade secret data and processes I think really important Now I do think that one of the things that will also be his way of doing this is do you trust the model provider? Right. And so I think that in cases of companies like Microsoft, which have had decades of our first thing is each individual customer. And so people say, well, why does Microsoft products sometimes trail and it's because it's like, well, it's because we are first and foremost making sure that this is for our corporate customers in terms of doing that. And that can be an important way of navigating that. I think there's also like one of the things I think will happen is happening already, but will happen more is that companies will get specifically post-trained models for them.

Like I think, you know, a number of different companies, but like you get a, okay, I get a post-trained model from me. I trust the model provider, not just the model that the company may have just built for everybody, but also built a model for me and there's a contract for it. So I think there's a broader range than just the harnesses and control. I think the strategic goal is exactly right, as Satya says, but I think that there will be an evolving range of set of how to navigate towards that strategic goal. All right. Thinking about the new world of business with AI, it feels like there was perhaps unwritten rules in the future that when you're a VC, you don't invest in competitors. If you're in Uber, you don't invest in Lyft. You have to make sure that you're – and feel free to disagree with the premise. But at TechCrunch Disrupt this week, Benchmark Partners said that AI companies' rapid expansion is making these conflicts between investments increasingly difficult to avoid. There was an explosion of activity on X recently about a public dispute between Factory and Cognition over an advisor who moved from one company to another.

They raised concerns. Is that advisor taking confidential information? Vinod Khosla jumped in, and he publicly sided against Factory, even though his firm backs both companies. He called Factory a struggling second-tier competitor and accused its CEO of having no decency or sense of proper behavior. So these are some fighting words. you've been a founder, you've been an investor. What's your take? Can you back competitors? Is it increasingly difficult in the AI world? How should investors parse that? I think it is increasingly difficult in the AI world to keep channels of potential competitive conflict. First, AI companies evolve fast in different directions. Second, they evolve very expansively. Third, AI agents' broad capabilities themselves as they empower these things mean that you can grow into spaces. Fourth, seeking what are the most important strategic assets, whether it's data or other things or coding capabilities, then I'll draw a lot of AI

companies into similar kind of zones. In short, I think the increasing competition is inevitable. So what does that mean as an investor? I think that means the investors say, look, you still need to navigate with integrity, which means don't bleed information between companies of any sort, even the questions or the struggles or other kinds of things they're doing. Make sure that companies are aware of where you might have conflicts, either as an individual or as a firm, and say, hey, you know, just be aware, we're also an investor in cognition or we're also an investor in fathom as a kind of way of navigating. And, you know, the questions around, you know, I think for employees or for anyone who is a member of company one, If you're going to company two, you have to be extra careful and clear and transparent about how you're not taking any company IP.

Now, the hardest ones are there's all this like water cooler conversation that isn't like, oh, we have a like it's not like taking a model or taking a data source. But it's like, hey, I've got this clever idea that's kind of more like tradecraft and how one might construct better coding or what kinds of people are constructing coding in kind of interesting ways. And like, well, that's not like collectible IP. And I may have thought of it. Now, I shouldn't sit down and say, hey, company two, let me tell you what company one is doing. but it will inform how I think and that's part of how people evolve and you have to have some expectation that that will happen but you try to be as transparent of it as possible and for example if I'm a company one and another person had this idea I don't go say oh I have an idea for you because it wasn't my idea and and I don't try to push people toward but obviously as I'm thinking about it I might come up with my own version or different and I might say hey I've

got an idea. And that's kind of the neat, how humans operate. I mean, that's one of the reasons Silicon Valley has been successful. Exactly. But if you're a founder and you find out that one of your investors, or they tell you to your point of transparency that they've invested in a competitor, does that change what you as a founder will share with that investor, with the board, et cetera? Well, ideally you have a close enough relationship with that investor that you have high trust to the walls of integrity. If you don't, then for example, something I've helped founders do sometimes say, well, I want a different partner from the firm. Um, I want to, I want to limit certain kinds of information that I'm sharing. You have to have that integrity of the boardroom. And so, um, I think it's possible. I do this myself. I've been on companies where there is conflicting lines and, And, you know, what you want to be is you want to have such a depth of relationship. People are like, I have 100% confidence that there is zero bleed in either direction.

And I think that's what's important to operate. And then, by the way, that's the kind of investors and board members founders should seek. I was about to say, there's probably a bigger problem with that board member if you don't have the trust in them, if you're questioning them anyway, if you feel like you can't have those open conversations. Reid, delighted to be here with you today. Pleasure. If you enjoyed this conversation, follow the show wherever you get your podcasts and consider leaving us a review. It really does help audiences find us. Possible is produced by Pallet Media. It's hosted by Ari Finger and me, Reid Hoffman. Our showrunner is Sean Young. Possible is produced by Tanasi Dilos, Katie Sanders, Spencer Strasmoor, Imozu, Amon Suri, Danny Garrison, Trent Barboza, and Tafadzwa Nima Rundwe. Special thanks to Surya Yalamanchili, Saida Sepieva, Ian Ellis, Greg Beato, Parth Patil, and Ben Rallis.