← The a16z Show
The a16z Show
Steven Sinofsky: AI Doesn't Need New Rules Yet
00:29:16
Steven Sinofsky joins Theo Jaffee and Sofia Puccini for a conversation on AI regulation, open-source models, and what history can teach us about technological revolutions. Drawing on decades of experience leading products at Microsoft, Sinofsky argues that governments are rushing to regulate AI before they fully understand the technology, risking innovation in the process. They discuss the "precautionary principle," why open source has historically accelerated innovation, the role of regulation in emerging technologies, the AI competition between the U.S. and China, and why existing laws may already address many of the risks people attribute to AI.
Loading summary
Transcript39 lines
- [00:01]
Steven Sinofsky
The whole topic of regulation for me just seems completely backwards because it's starting before we even know what we're regulating. There's no reason why the AI companies should be against open source other than we just don't want our competition to exist and we don't want to bother to compete. We'd rather just compete with each other and not worry about that crazy open source competitor. Truth is, no one knows the future. What those assumptions mean are we should regulate this based on our own personal predictions of the future. But the history of being right about those predictions is pretty limited and so we should be really careful about that because those are all self serving.
- [00:45]
Podcast Host (Narrator)
How should governments regulate AI when the technology is still evolving? Theo Jaffe and Sofia Puccini are joined by Steven Sinofsky to discuss why he believes AI regulation is moving too fast. The case for open source models and what previous waves of technological innovation can teach us about building policy without stifling progress.
- [01:09]
Theo Jaffe or Sofia Puccini
We are live with Stephen Sinofsky, amazing tech analyst, author of Hardcore Software, longtime Microsoft leader, and we're so glad to have you back on Stephen. Welcome back. Hey there.
- [01:21]
Steven Sinofsky
Super fun to be here.
- [01:23]
Theo Jaffe or Sofia Puccini
Yeah, so, so much in AI. Like where do we even start?
- [01:28]
Steven Sinofsky
Well, our goal is to say stuff that we haven't said yet. So let's see if we can do that.
- [01:32]
Theo Jaffe or Sofia Puccini
We will try.
- [01:33]
Co-host or Panelist
Novel insights time.
- [01:34]
Theo Jaffe or Sofia Puccini
Yeah, novel insights. So open source, there's been much discussion of regulation of open source, Chinese open source models in particular on both sides of the pond. You know, China has considered export controls, but also Xi Jinping has said we're encouraging open source. And then in America we've had Scott Besson says we're going to be looking into IP theft and some people have considered import controls or requiring a license. So there's a lot going on right now. It's very fluid. What are your takes on regulation of open source models?
- [02:07]
Steven Sinofsky
Well, it just seems the whole topic of regulation for me just seems like completely backwards because it's starting before we even know what we're regulating. And if you go back in history, one of the best things about the 20th century in America was that so much experimentation was happening, so much innovation and that does have this potential downstream of like, oh gosh, this thing happened that we have to roll back. Like a super famous example was safety in cars and that car companies spent the first 50 years making cars, making cars. And then in the 1960s it sort of became clear that wow, we could actually make cars much, much safer. Now it took 60 years to understand that Cars were super dangerous. And it was all that evolution. And in all fairness, like it wasn't a design criteria. You know, seat belts and airbags and these things became a very long process to get them put into the upstream of making cars. And you could go through every sort of dimension. There's the antitrust law and the whole reason why, you know, oh my God, you can't be an oil company where you sell gasoline, produce gasoline and also pull it out of the ground. Or you can't make movies, have actors and actresses under contract and own the movie theaters and distribution and maybe that shouldn't be. And there, there are millions of examples. The challenges, the story of regulation is always told as though if they would have gotten in early, they would have prevented this stuff from happening. And you know, you look back and you're like, oh my, you know, what if they would have shown up at Henry Ford and said we need airbags. And Henry Ford is like, okay, we can't figure out how to make engines work. Like, let's, let's get that done. And you, you know, also these cars go 15 miles an hour and you know, we're just trying to make them go. And you know, of course air travel is like that. Every technology, innovation, it took a long time for two things to happen. One, for it to just work at all. And you know, you, today we're talking about AI and it can't be that. AI is both the most intelligent thing in the world and it generates gibberish hallucinations and the answers are raw and you can't rely on it. Like those both can't be true at the same time. You can look at social networking had the same effect where like, oh, it's a toy, it's a giant waste of time. You know, it's either that or it can topple governments. You know, it couldn't be. And they were, at the same time people were saying both of those things about social networks. And so it can't be. And now they've added this dimension that they, they are mind control devices on our nation's youth. And at the same time they're a huge waste of time and the content is all sloppy. Like all of these things can't be true. And so a lot of times it just this whole notion of regulating is this hindsight and you assume that only good would have happened if regulation happens. But it, that's just, you can't prove that. And so you're in this very tricky situation. And so today you have this, this world oops my. I just got lost on video. Hey, give me just a second and figure out what. What just happened? Oh, I just crashed.
- [05:38]
Theo Jaffe or Sofia Puccini
No worries.
- [05:39]
Co-host or Panelist
You're still here in voice form.
- [05:41]
Steven Sinofsky
So. Yeah, I got you in voice and. And I just cra. I like, literally crashed. Hang on.
- [05:47]
Theo Jaffe or Sofia Puccini
The spirit of Steven Sinofsky still wins.
- [05:49]
Co-host or Panelist
It's a beautiful gear icon.
- [05:52]
Steven Sinofsky
Yeah. Was that. I don't know. Oh, is it? Wow, that was super interesting.
- [05:57]
Co-host or Panelist
Yeah. We can pause the stream for a second while we figure it out.
- [06:00]
Steven Sinofsky
No, I, I like, I officially. And I am. I'm working. I am closer. Well, can we do with this? Is that working?
- [06:10]
Co-host or Panelist
It works. It works with audio. You just. Anonymous guest.
- [06:15]
Steven Sinofsky
Oh, hang on.
- [06:16]
Theo Jaffe or Sofia Puccini
Your newest anon guest is like suspiciously knowledgeable about the Microsoft Surface Pro.
- [06:20]
Co-host or Panelist
How about that?
- [06:21]
Steven Sinofsky
That works.
- [06:22]
Co-host or Panelist
You're back. You're so back.
- [06:23]
Steven Sinofsky
I'm. It's not the right video, but it is video. And so, so we have this thing where everybody. That was just a total crash, by the way of my, like, AV stack. And so we had this thing where people are just trying to rush to regulate and that it's. It actually came up. It happened so often that this guy came up with a. A term and it's called the precautionary principle. And the idea is regulators who, you know, look, you, you don't get a job in government to regulate and then say. And then don't regulate. Like you, you come to work to craft regulations and to come up with what you want to do. And so the precautionary principle was like, well, let's just get ahead and let's. Let's regulate before anything bad happens. But when you do that, what you really do is you. You constrain the solution set and I think back to the Internet. And what would have happened if we would have precautionarily regulated the Internet? Where would it have stopped? Like, would we all just be using AOL Instant messenger right now? Because that was like the Internet that everybody knew. And, and it had all these attributes that, that regulators liked. Like, it was one company and it was in Virginia. Like, literally that's even the best case because it was practically in the government already. And you know, you could, you could go to them and you could tell them to do stuff and, and it was a walled garden. If they didn't. If you didn't like something, you could turn it off and, and it was perfect. Or would it have been in like web 1.0 where everybody is using six tags in HTML. There was not yet really commerce and we all would have just looked things up on Yahoo and Excite, and they would have said, okay, these are the two companies we're blessing, Yahoo and Excite. And there's no video, there's no audio. You know, we're just, you know, and you just, you kind of go and you pick your time of when it was. We're all comfortable with, like, whatever it is that that's going on. And so that's, of course, what happens is at that moment, the companies that are really hot are the ones that the government seeks input from. And some people think that that's bad. I actually think it's good because they know, presumably more than anybody else, but what they also know is what they want to do. And so those companies, then they sort of push government to go in a certain direction. Historically, technology companies have said, whoa, how about this government, Hands off. We don't like regulation, like, at all. And that's a lot different than, say, what happens if the government's already regulating safety of chemicals? It's very hard for the pharmaceutical companies to say, leave us alone and don't regulate us at all. And so that was one of the first early regulatory bodies. But the tech companies this time around in AI have completely just blown my mind. Now, you have to understand, I spent like, I was a Microsoft employee for like, two years, and then the government started investigating Microsoft in like, 1992. Told you for what? You wouldn't even. You wouldn't even understand what I was talking about. But now we had, you know, two years ago, like, two years ago, what were people doing with AI? Like, chat bots were new, and you had the leaders of the AI companies showing up to Congress literally, please regulate us. Oh, we're begging you to regulate us. And that is just this crazy notion. And, and so what you had was they just played right into the government's view that they missed regulating technology, they missed regulating the Internet, they missed regulating the PC, they missed regulating the mainframe. And so this was their chance and they were being asked. And so they, like, rolled out the red carpet and the government knew what was going on, but because it solved a problem for them, which was how to really get into regulating big technology companies, they sort of went with it. And, and so the government normally, like, they kind of go, oh, they're, they're going to try to control us. They're going to lobby us and tell us what to do. And in fact, everybody knew that was what's happening. And that's called regulatory capture. And this is a long History. And most of the people in government that are savvy with government know that that's what's happening. That they know the companies are saying do this thing because it's good for everybody. And they always have good reasons. They always say, oh, the products will be better or the price will be lower or there'll be much clearer, you know, lines of service and the marketplace will be more fair. Like for example, AT&T, which is a fascinating example because they were actually a government created monopoly. Like literally the government created the phone company and they owned it essentially. But then when the government was saying, you know, we think we need to control your prices and do stuff, AT&T went to them and they said, okay, here's the deal. We can guarantee that every single address in America will have a telephone and it's called universal access. And that's why you can only let us do this. And so basically they made this deal that the AT&T was going to be the national phone company, even though it was a private company if they, they promised to deliver phones everywhere. And it got really tricky because with that came like a bunch of rules about using the telephone lines that today led to surveillance, you know, and, and all of this stuff. And because so what happened was Basically the TELE, AT&T, A private company, became really the United States phone company. And, and it was just a branch of the government. And this has happened time and again. J.P. morgan became the national bank of America, you know, the Standard Oil became the national oil Company of America. Even at some crazy extreme, Warner Brothers became the movie theater of America. You know, and, and every instant, it turns out, radio, television and telephone all sort of followed the same path where they were, were private companies that were basically beholden to the federal government. You saw a lot of this when the New Deal was created and FDR basically used the Federal Communications Commission to control the messaging of the New Deal by basically threatening the licenses of the radio broadcasters. And there's a lot of controversy of just how much they did and did it really cross some line. But the truth is they were going to, they were just in line. Whether they wanted to be or not, they were going to be in line. And the same thing happened with television. And so this whole idea of being against open source, it's really rooted in like no one. The government can't be against open source because if you get a government grant, you're required to release all your software as open source. Like that's the government way of doing computer science research. So how can all of a sudden the government say we don't want open source. It doesn't even make sense because none of the leading academic researchers will be government funded. And so it just makes no sense at all. To me and Microsoft we were a closed source. Bill Gates invented the closed source business model. And so we lived through the rise of HTTP and Apache and web and Linux and all of this other stuff. But there is a role for open source. And most importantly, the open source existed even when Microsoft existed, even when IBM existed. And so there's no reason why the AI companies should be against open source other than we just don't want our competition to exist and we don't want to bother to compete. We'd rather just compete with each other and not worry about that crazy open source competitor which I, I just, I can't even put words into how obnoxious that is. It's just not, it's just not Krishna. And they benefited enormously from the academic research community, all of which was open, published on, in open source. You know, it, it's just crazy. And that's how we got here and that's why we're even having this debate.
- [14:42]
Theo Jaffe or Sofia Puccini
Yeah, well I think that you know, most AI labs, AI companies would be like totally on board with this idea of like yeah, of course we don't want, you know, government centralization and regulation. But they think, you know, AI is not like the other technologies, it's not like the Internet, it's not like the PC. They're building towards, you know, artificial superintelligence. And this will be like a totally different thing and will require like a novel approach. So I think maybe different people in this debate are working on like different assumptions of like where the technology's headed.
- [15:14]
Steven Sinofsky
Well, those are, let's, let's, that, that's the most generous way to express it is different assumptions. The, the truth is no one knows the future. So what it, what those assumptions mean are we should regulate this based on our own personal predictions of the future. But the history of being right about those predictions is, the history of being right about those predictions is pretty limited. And, and so we should be really careful about that because those are all self serving. There's no, there's no reason and to say it's different. I, I genuinely believe there are people who work on AI who, who really believe in this asi AGI arc of super intelligence. I mean they, they went to Congress two years ago and they said literally there was fear in their eyes for the economy and for people dying everywhere at the same time, of course it didn't work and it was hallucinating and it was silly and it was making things up. So again, it's not. They're drawing some predictive arc. And even today people are fighting over is AI good enough to do stuff? What is it good at? So it just, you can't have all of these things being true at the same time. So what it really says is not now is not the time to regulate because we don't know where it's heading. So all you're going to do is kneecap innovation. And, and there are laws in place for a zillion, almost any scenario. In fact, I would say 100% of the scenarios that people say are problematic, there are already laws against them. Like there are already laws against, you know, putting bad drugs on the market. There are already laws against not giving people bank loans for their, their gender or their race or their other attributes. Any. There's already like Senator Warner today came out with a big four point plan about AI and he's like, well, we have to address the issue of non consensual nudity. And it's like, okay, but there are laws against this. Like you, you can't be showing nude pictures to children. Like that's already law. And you can't be showing nude pictures of children. It's already a law. And so like every one of these evil scenarios that they worry about, it's already illegal. Like I saw one state like, you can't, you can't be registered as a, an attorney if you're an AI. Well of course you can't be registered attorney. You can't even sign your name on the license. Like you, you know, like you there, you have to have a license to be a lawyer. You have to take a test, you have to do all of this stuff. And the AI doesn't do any of these things, nor can it like it can't register for college. Like people aren't making, they aren't being coherent because they're just so enamored with like getting their point of view and leaving off competition in some way or kneecapping big Tech in some way.
- [18:15]
Co-host or Panelist
Yeah. So if we have like the precautionary principle in mind, what would be the opposite of it? Like, what would be a very like responsible, like iterative way to start regulating AI knowing that like your viewpoint is like, we don't even know exactly what it's capable of. It's not even like, you know, like hashed out enough as a technology or as a sector to even know what to regulate.
- [18:38]
Steven Sinofsky
Well the first thing that's super easy is you can go through all the laws and say like, hey, in this state with this law for this thing, you know, did we make it clear that you can't be an optician if you're not an AI? Like should we just make sure that AI can't just go give eye tests? Like you can't set up a, an iPhone in the mall with like a camera and write lens prescriptions for people like okay, maybe you can do that because Ohio didn't word their law correctly and states are responsible for licensing opticians. So there is this step one. Okay, we have a whole new technology. Should we. Like you, this happened with EVs like where EVs weren't gas combustion engines and you they had to adhere to all the safety rules of cars except they had a whole bunch of different components in them. And all the rules for cars sort of got based on the gas combustion engine. So they need to go back and make sure that all the laws applied if you're, you had a battery operated car. And so how many cars had front trunks? Not very many. So were there laws about safety and crash zones and all of this and accounted for front trunks? Let's go check it out and make sure we're okay. And so there's a first step which is there are a lot of laws. There's 2 million laws on the books about everything. So do the right ones apply to AI? Like do a bunch of places need to go add and AI to their laws? And maybe that's csam, maybe that's spam, maybe that's surveillance. Like there's a bunch of stuff to just go do. And maybe today those are wrong. That's like that should keep everybody busy while we go figure out AI. And of course we had to do this with computers like you. You had to make sure that if you used a computer to do something that like you couldn't like you know that all the anti forgery laws applied if you happen to do output in a computer that, you know, I went through getting the legal system to use Microsoft Word and it turned out there were a whole bunch of features we had to go change in Word just to work correctly in the legal system as it stood, legal system didn't have to do anything. We had to go and just make sure we worked within the constraints of like here like crazy example like in the legal system having footnotes that are longer than a page. So footnotes have like a limit. They can only take up like a third of the page in a legal brief. But if your footnote is longer than that, you have to have like a feature and a word processor that continues the footnote on the next page. Like, we were like, wow, who thought of doing that? What a. And it's no big deal for using a typewriter because you haven't even gotten to typing that page yet. So it worked fine. So we had to go do that. So that's like a great step and that's an idea. Like, then we'll know what's missing and also while the technology evolves.
- [21:33]
Theo Jaffe or Sofia Puccini
Right.
- [21:34]
Co-host or Panelist
Yeah. I'm a big believer in like domain experts will sort of know the like in the case of the like opticians, well, it should be like opticians, registered opticians and people who are like, have expertise that are sort of like working with the states to, along with like some AI experts. And then it becomes like a perfect union. And that's how we like figure out where to go next. So. I totally agree.
- [21:55]
Steven Sinofsky
Yeah, there was an episode, there was an episode of the new TV show the Pit where one of the residents is using the. An AI system to sort of make it easier to do the patient notes. And then she makes an AI mistake. Like AI makes a mistake and. And there's no ambiguity. It's her note. Like she's in trouble. She can like go, oh my God, it's AI and it. There's just nothing you could do. It's her medical license. It's not the AI's medical license. Completely unambiguous. Now if she went and hurt the patient was super tragic, but she still hurt the patient and it's her accountability no matter who wrote the note.
- [22:33]
Co-host or Panelist
Yeah. Going back to open source. So something interesting of like the last week or so has been that like on one hand you have like China considering export controls on like chips and like open weight models. And then you also have the US government like considering maybe like some action against like Chinese regulating Chinese open source models. So it seems like on both sides, like there is an incentive to do that. Like, how do you see this playing out?
- [22:58]
Steven Sinofsky
Well, in this case, this is. This. There is an actual innovation war with China happening right now. There's not a cold war, there's sort of a trade war. But we're very mutually reliant on each other for trade. So it's hard to call it a trade war. But there's absolutely an innovation leadership war. And if you're the government and you're participating in innovation leadership war, There are two tools that you have. One of them is you bludgeon your international competitors with the regulations that are already in place. And then the other is you fund or nationalize companies to go compete. And those are the tools that you're seeing put in place. Now China is very pro, pouring money into the companies at a national level. This is something that we saw with Japan, the 1980s. In the 1980s, Japan was going to take over the technology world and they set up a huge ministry that sort of tried to shepherd the memory industry to dominate it. And if you look, they don't dominate the memory industry anymore. So it didn't really work, but they did for a long time dominate other aspects of tech. And they're very, very good at what they do. But the central planning, it, it's not. There's no track record where it really succeeds. These trade wars, this is where you get into what I think of as bank shots. The government's not going to go after AI directly. It's going to do these things on the side. Like you see, we're going to, oh, we're going to ban the chips. Okay, well, that may or may not slow them down, but it doesn't have anything to do with the model in the end. Like, okay, so it'll train longer or it'll, you know, there's no shortage of power in China, so they're not going to run out of power running the models longer or something like that. And so these bank shots are very popular with, with the government because also it doesn't look so rude like there's one going on with China or with Canada now over Canadian lumber. And we put a tariff on. I have no idea what Canadian lumber has done to us, but it's probably nothing. And this is over something that they did to us that isn't lumber because often you can't do the same thing back like it happened. You seen this in the brutality of war, where like one side bombs a school by accident. You don't bomb a school of theirs on purpose. You, you go and you bomb, you know, a factory that's important to the military or something. And, and so you get these indirections and signals. These are the tools of diplomacy. And you can fight all you want about how, how crazy they sound, but this is what they do. And they're going to come to work every day and, and participate in the innovation war with the tools that they have. And it's unfortunate for us that there are people helping them with, with their own bank shots. So, like When Anthropic says, hey, the best way to hurt China is to curtail open source, what that means is the best way to help Anthropic is to curtail our open source competitor, whether it's from China or the United States. And that's just, that is kind of un American. And so. And it's definitely untech. Like we've just. The tech industry is just never wanted this. It's very different than say the auto industry in the 1970s when I was a kid that, you know, the Detroit was just getting crushed by small Japanese cars because Detroit just couldn't make fuel efficient cars and they also couldn't make quality cars. Japan had had a reputation for low quality and through the 1960s really perfected making super high quality cars. They invented quality engineering. And then they are a very compressed city with a zillion people. So they made really small commerce that became immensely popular in the United States. And Detroit just convinced the administration that they were selling these cars below their cost, that they were dumping them. Does this sound familiar? This got used this week about, about, about Chemmy Down. And so the idea is that, oh, they're dumping it. Well, dumping turns out to be. It has to have a very specific definition about how much does it cost you and are you really offering it at less price? And even then, this is a bigger international trade diplomacy dialogue. It's not so cut and dry. You can't just say because this product exists, it's being dumped. And. And yet the Detroit was like dumping, dumping, dumping. You got to stop it. So what did Japan do? They actually built all their factories. And so you might remember, right, know that there's a bunch of Japanese car manufacturing plants in the US in, in the Carolinas and Alabama and Mississippi. And so that's what they got out of this. They still lost. Detroit doesn't even make cars anymore, so they still lost.
- [28:03]
Theo Jaffe or Sofia Puccini
Yeah.
- [28:04]
Steven Sinofsky
And so it didn't really help, right?
- [28:07]
Co-host or Panelist
Yeah.
- [28:07]
Theo Jaffe or Sofia Puccini
Well, it's been super interesting. Stephen, thank you so much for coming on mts. It's been great having you on.
- [28:15]
Podcast Host (Narrator)
Thanks for listening to this episode of the A16Z podcast. If you like 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 X16Z and subscribe to our substack @A16Z. Thanks again for listening and I'll see you in the next episode. 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 and A16Z or any of its affiliates. Information is from sources deemed reliable on the date of publication, but A16Z does not guarantee its accuracy.