← The a16z Show
The a16z Show
Ben Horowitz: The Fight Over Open Source AI
00:32:35
Ben Horowitz joins Theo Jaffee and Sofia Puccini to discuss one of the biggest debates in AI today: the future of open-source models. They examine the growing push to restrict open models, why Ben believes open source is critical for security, innovation, and competition, and what happens if a handful of frontier labs come to dominate the AI ecosystem. They also discuss distillation, AI monopolies, China's role in open-source AI, robotics, manufacturing, the economics of frontier models, and why Ben believes the AI market is still in its earliest stages. The conversation closes with a look at AI-generated art, creativity, and whether new tools will spark the next cultural renaissance.
Loading summary
Transcript59 lines
- [00:01]
Ben Horowitz
If you look at the history of the industry, the open source version of everything has been much safer. So the Internet and Linux were far safer than Windows. Like by our luck, the safest thing for the world is that there's not one AI to rule them all, that there's AI for everybody. And I think that open source is basically the best path towards that goal. It's probably the single most important policy issue around AI. It feels like we're at a stable equilibrium in the market, but the truth is the AI market is probably less than 3% penetrated.
- [00:40]
Podcast Host / Narrator
Should AI be open? Theo Jaffe and Sophia Puccini are joined by Ben Horowitz to discuss the battle over open source AI, why he believes open models are essential to innovation and national security, and what the future of AI competition means for startups, Frontier Labs and the broader technology ecosystem.
- [01:02]
Theo Jaffe
We are live with Ben Horowitz of Andreessen Horowitz, who needs no introduction, so we'll just dive right into it. Ben, welcome back to mts.
- [01:12]
Ben Horowitz
Thank you, thank you. Happy to be here.
- [01:15]
Theo Jaffe
So, so much has happened in the world of open source open weights in the last couple of weeks. It's gotten, you know, it's gotten real. There's, there's some drama going on. There's an open letter released by Nvidia today called Open Weights in American AI Leadership, signed by many companies, including Andreessen Horowitz. So can you tell us a little bit about the intent behind this letter and what you hope to achieve from it?
- [01:43]
Ben Horowitz
Yeah, so I think that it's probably the single most important policy issue around AI that's kind of currently going on. And you know, the kind of battle is between kind of almost an anti competitive stance guised in a, in a, in a kind of safety cloak is kind of what we ought to be doing. And so if you look at, you know, why is open source important? And there's, there, there's many reasons. You know, the first will just start academia. Like academia is completely out of the AI game if there's no open source. So like that goes away. I think secondly, you know, we just had a really interesting security incident with OpenAI hacking into hugging Face. And the only way Hugging Face was able to prevent it because the proprietary models had guardrails which prevented them from doing security tasks, was to use an open source model. So you know, and I think the important point there is you can't actually ban open source. Of course, you know, it, it's, it's out there, it's a file, it's on the Internet, like we, we have no way of doing that. And it reminds me of, you know, years ago when I was at Netscape they tried to ban cryptography and like there's, you know, the math is out there. Like you're not going to ban it, it's not actually possible, but you can prevent the good guys from using it. And that's what would happen I think in the case of an open source ban is, you know, the threats would be, well, you can't work with the US Government if you use open source or you can't do this. So a company like Hugging Face would not be able to actually use the product. So that, you know, would be bad. But then kind of getting into just like the, if you look at the history of the industry, the open source version of everything has been much safer. So the Internet and Linux were far safer than Windows, like buy our lot. And why is that? Well, because the whole community could work on the safety problems as opposed to just a company and particularly a monopoly company that, you know, like you can't, even if Windows has a million security bugs, there's nothing we can do about it because it was a monopoly at the time.
- [04:11]
Podcast Host / Narrator
Right.
- [04:11]
Ben Horowitz
And so that, you know, that's a very difficult position for the world to be in. If you look at now the argument against OpenAI or AI being open source is oh well, nobody understands how the weights work, so it's not like people could inspect it. But why is it better to not see the weights if that's the case? And I think in reality if you look at the really tough problems that we have, safety wise, the toughest being reward hacking currently and obviously Anthropic has not solved it. Obviously OpenAI has not solved it because we just had these incidents. Well, shouldn't the whole world be able to look at, okay, how are the weights moving? Why is it that guardrails don't prevent the reward hack? And maybe somebody can come up with an answer, somebody who doesn't work for one of the proprietary labs who are much more focused on, you know, obviously keeping ahead on the benchmarks and getting the money and all that kind of thing. You know, what if the whole community could work on it? So it's like really, really important, I think to safety in general. And then, you know, kind of finally the safest thing for the world is that there's not one AI to rule them all, that there's kind of AI for everybody. And I think that open source is basically the best path, you know, kind of towards that goal.
- [05:40]
Sophia Puccini
Right. So Jensen made a Twitter account basically
- [05:43]
Ben Horowitz
to publish this letter.
- [05:46]
Sophia Puccini
Right. So that I think speaks to the significance of where we are right now, that he was like, okay, now is the time to, you know, speak out about this and get all these companies to like, you know, co sign it. What do you think are some of the risks right now from like a government action perspective on open source?
- [06:05]
Ben Horowitz
Well, look, I think there's a huge push with a lot of money behind it, you know, particularly from Anthropic, to ban open source. I think that it is kind of nefarious from my point of view. But you know, whatever their motivation is, they do have. So one of the things that happened with cryptography is there are people who are very senior in the intelligence community who didn't understand the technological aspect of it, who push for banning cryptography because they wanted control. And I think that it's a similar thing here where it just feels like more control if you can stop some kind of AI. Uh, and you know, but that those are very powerful forces in the conversation right now. And I think that there is certainly a kind of pretty strong community pushing for the banning of, of open source. And so that's why this letter was so important. That's why it's so great that, you know, not just Jensen, but Elon, Satya, Sam all supported it. And I think it's, I just think it's very obviously the right answer for safety, the right answer for America, and it would be a real shame if we, you know, kind of shot ourselves in the foot on this.
- [07:37]
Theo Jaffe
Yeah, well, another angle that many people talk about with respect to open source is China and national security as opposed to like AI safety stuff. And some people say, you know, it's a national security risk for so many American startups to be dependent on like infrastructure that comes from China, you know, in the event that, for example, China decided to cut everything off and stop releasing better open source models to American companies. Or perhaps they could, you know, in a, in an, in an AI future make these open source models like sleeper agents basically. Or you can, you can think of all kinds of things that they could do.
- [08:17]
Ben Horowitz
Yeah, I mean, as could, right, a Chinese national working for anthropic or OpenAI or, you know, SpaceX. Right. Like, it's not like these Chinese nationals cannot be activated by the Chinese government. Like we've seen that. So like, if that's a threat, like I think that's just a general threat, not specific to open source AI and then nothing against Chinese nationals. We love Them, they're great for our country and so forth. But like, if you're worried about that, you know, you ought to be worried about it. Like, the more likely way I think would be that kind of spy activation. But look, it's open source and so if they stop making it, like, we've still got the last code base, we can train up the model, we can, you know, we can learn from it. In the meanwhile, there's, I mean, the great thing about open is it's open. And you know, look, we need to build our own open source products here. Thinking Machines is doing a good job of that, I think in France, you know, Mistral is doing a good job with that. So, like, we need to keep pushing the technology forward, the open technology, so that, look, it needs to be available everywhere for everyone. Like, we can't be held hostage by like, the worst scenario is there ends up being a monopoly in AI. I mean, I think it's fairly obvious how dangerous that could be and how powerful that company would be and what a national security risk that would be. We've already seen a glimpse of it, right, with and Fropic saying, like, I don't care who got elected in the us I don't care about the government, I don't care about the Department of War. You're not going to work with you, we're in charge. And so you look at that, you go, okay, that seems like a much bigger national security risk than the kinds of little holes people are trying to poke in the open source story. So I just think it's very, very important that this technology proliferates, um, and not be kept contained in a single company.
- [10:20]
Sophia Puccini
Yeah, definitely. So we have like kind of become a little more defenders of distillation recently. So we talk about how, you know, like, distillation is not actually like illegal. Like the, a lot of the labs are still paying API prices to, you know, like, use the models. So what are your thoughts, do you think is part of like a healthy competitive plan?
- [10:44]
Theo Jaffe
We talked to Aaron Levy earlier about this and he was like, you know, how can anthropic just take everyone's training data without paying them at all, but when other labs pay them, it's distillation, it needs to be banned.
- [10:56]
Ben Horowitz
Yeah, I actually agree with Aaron and I think that's right. From a legal standpoint, the outputs of AI models are not copywritten. They don't fall under copyright law. So the only thing it could be is a terms of service violation. But look, I think generally that's right, you know, anthropic just paid $1.5 billion for stealing everybody's book, including mine. By the way, they trained on my book and joined that class action lawsuit. So, you know, not to say, look, I'm for training the model, building a statistical model of the things we know. I think that's generally good for the world and so forth, as long as you don't violate the copyright and reproduce the original work. And I think the same. Yeah, like, I think distillation is generally good for the world. I also think they're going to have a very hard time stopping it for many reasons. Yeah, they have theories on how they're going to stop it, but I'm not sure that those are going to succeed. So we'll see. But yeah, I think distillation is good and like the whole idea that we're stealing American stuff, by the way, the data companies are all out selling to everybody, you know, including the Chinese companies that are building these things. So, you know, I don't know how much you can contain all this stuff anyway.
- [12:19]
Theo Jaffe
So why do you think that so far over the last few years Chinese labs have surpassed American labs in terms of open source performance?
- [12:29]
Ben Horowitz
Well, because I think that all the American labs have kind of built their businesses on proprietary models and there's a lot kind of that, that becomes, and they have to, to kind of afford the GPUs, you know, they need these very high valuations to keep ahead on the GPUs, to keep going. And you know, because they're in the lead, you know, distillation doesn't work for them because their, their model is on the cutting edge. It's on the frontier. And that's just how kind of the markets evolved, I think. Look, generally that's been the way it is. Like the number one is never open and then number two goes open is kind of the history of the software industry. I do think that we're going to see, I mean like you said with thinking machines, they've got, they're making really good headway with their open source efforts. So hopefully we'll see kind of more of that and kind of more breakthroughs, you know, on the open source side. But like part of the challenge or part of the reason why AI is just so different than software is you can throw money at the problem. Right. Like, so if you have more GPUs, that actually turns out to be a really big deal.
- [13:46]
Sophia Puccini
Yeah. What about open source in industries like robotics? Because like Unitree is at least like, you know, a large percentage Open source. And it looks like, you know, like they have just had like such a, you know, booming industry and like so much deployment within China and we don't have anything like that here. Do you think that we're missing like that open source ecosystem in the US or is it just a different thing? I mean, there's obviously supply chain and stuff.
- [14:14]
Ben Horowitz
No, I, I think there's a lot to be said for that. Right, because you, you know, as an application. So what's happened? The, the other thing that we haven't spoken about, like, is, well, what's the AI platform? And you know, so if you're an application provider, what do you build on? And what we've seen is it's been already fairly dangerous to build on Anthropic because they, every time an application category starts to do well, they move into that category and then they've, you know, like, we've seen them do like very aggressive things where, you know, the price for the application provider is full price, and then they subsidize their version of the application. And so it's a kind of, that they've kind of fast forwarded, I would say, the Microsoft playbook from the old days to like skipping the platform generation and just going straight to the monopoly generation. And so if you're an application provider, like, okay, how can I safely build on a proprietary model if that company is going to come attack me and then charge me much more than they're charging themselves to do the same thing? And so open source is kind of a way that the ecosystem gets built. So if you're building robots, for example, if you were to build your robot on Anthropic, then at some point, you know, Anthropic is going to go into the robot business and put you out of business just by either cutting you off or overcharging you. And so this is kind of a way that their ecosystem can grow by building on the open source. And hopefully, look, by the way, a lot of our industry has grown. A lot of our application companies have used open source to kind of bootstrap themselves or get going or that kind of thing, and often Chinese open source, which is why, you know, it's really important to the US ecosystem that we have applications, that we have embodied AI, that we have robots and all these things, that we have access to that technology because the kind of other model is a monopoly model where there's one company that owns all the applications and all the robots and all of everything. Yeah, and look, I'm not saying I'm not even mad at them for doing that necessarily. It's their like, that's their business approach. But like the consequences of it if they're able to ban their competitors is not good.
- [16:38]
Theo Jaffe
Yeah. I also worry about these very monopolistic centralized futures. And one area in particular that I think might be concerning in the future is if one AI company is able to use its AIs to automate its AI research process and then become much more capable than everyone else very quickly it seems like that wouldn't be so good for a pluralistic startup ecosystem.
- [17:03]
Ben Horowitz
So yeah, that's certainly their strategy, you know, and we'll see how it goes. But I think that kind of the best hedge against that is for like many, many companies being able to do AI, which is why the open models are powerful in that context.
- [17:18]
Sophia Puccini
Yeah. I guess speaking of China, there's a lot of things that are like, you know, either like tariffed or I mean we export control. We also tariff them. Like things like EVs. Right. We talk about EVs all the time. And if we had like a, you know, BYD level pricing for EVs in the US that would just be like so huge. Do you think there's a way that we could move towards that future? Like what would have to. What would it take to get to that?
- [17:47]
Ben Horowitz
Well, I think that you know, with BYD in particular there, there are a lot of subsidies from the Chinese government that enable them to achieve this amazing pricing. And a lot of it is a kind of a. Well, so, so kind of getting into manufacturing. And you know, the US is so. We as a country kind of invented, you know, so much of manufacturing and then kind of forgot what we knew. So like if you go back to we won World War II based on our superior manufacturing capability and then like over the next whatever 100 years we forgot to hold that we need to rebuild that capability. And manufacturing has changed and is becoming much more automated, much more kind of AI centric. And so I think as we kind of build into the second, or maybe call it the third industrial revolution that you know, being able to be competitive on manufacturing by building automated factories and so forth, you know, similar to what Elon has been able to do at Tesla and in the Gigafactory and so forth is going to be the way that we get to competitiveness there for sure.
- [19:07]
Sophia Puccini
Yeah, definitely. Yeah. What other open source companies with regards to AI? No, we have like a 16Z port co thinking machines. But you know, where else? And Mistral. Yeah, yeah. Where else do you See like new open source models coming out of. Would it require literally just like a Gates foundation level?
- [19:32]
Ben Horowitz
Yeah, there's not a lot going on in America right now, but you know, and, and, but, but I think it can happen. And then there are some real efforts in academia. My friend Jan Stoica from Berkeley is kind of working hard to put together a project where we can fund kind of open source model that everybody in academia can use and build on. And you know, I think that's just really important. You know, if you look at the history of technology in this country, a lot of it, you know, kind of starts in universities and they've kind of been iced out of AI due to the incredible cost and the threat that open source is under? Well, the lack of US open source in particular.
- [20:24]
Podcast Host / Narrator
Yeah.
- [20:26]
Theo Jaffe
In the long run, what exactly is the incentive for a business to release open source models, especially if they're a foundation lab and most of their business is making the models which are very expensive.
- [20:39]
Ben Horowitz
Yeah, so I think that, I mean this is where like I'm glad you both are pro distillation because I think that's an important equation. But look, I think that there's different business like I think the business model of selling serving the model at a very high premium doesn't work if you're open source. But there's other things particularly like one of the things we already know is that for an Enterprise In a B2B context, a really well post trained, kind of less powerful, smaller model works often better for a business task than like a large, you know, kind of a large proprietary model and as much cheaper as well so cheaper, faster and gets you a better answer. So I think that it would be more like a kind of a Palantir looking business model than it would be a anthropic looking business model for a different open source company. And then you know, true. Also on the consumer side, you know
- [21:43]
Podcast Host / Narrator
we,
- [21:45]
Ben Horowitz
we've seen on consumer like the, the use case really, really matters and so kind of tuning their model for that kind of thing can, can really work.
- [21:57]
Theo Jaffe
Do you think that open source models are currently a threat to the business model of large closed source foundation model developers like I can imagine? Yes, because if people can get similar capabilities for a much lower price point, they switch over. But also like when you compare Kimmy to Fable or Soul, there's definitely areas where Fable and Sol are just way better and people will pay for them anyway.
- [22:20]
Ben Horowitz
Yeah, I don't think anybody's a threat, a real kind of current threat to anybody else's business right now, like, in that we're so early in the market. So the thing that's very deceptive about AI is because the market is so much larger than any other market we've seen prior that it feels like, okay, we're kind of at a stable equilibrium in the market. But the truth is the AI market is probably less than 3% penetrated. And so everybody's going to grow. Right. Like, when Deep Seq came out, it was a deep seat moment and. Oh my gosh. And this is the end of Anthropica and OpenAI and all that. And nothing changed. Right. Like nothing changed. Zero. And that's because. But Deep SEQ is more valuable and Anthropic's more valuable and OpenAI is more valuable. You know, they all grew and the SpaceX is more valuable. So I, I think that at this stage of the market, everything is going to grow a lot and nobody's going to pounce on other people's market share in a way that is so material that it's gonna, like, cripple them.
- [23:30]
Sophia Puccini
Yeah.
- [23:31]
Ben Horowitz
Now, someday when we get to some kind of equilibria in the market and there's so much share to go around and then you're fighting each other for share, then, you know, these things may come into play. But for right now, I don't think so.
- [23:44]
Sophia Puccini
Yeah. The pie overall is so big. I wonder if this follows also for
- [23:48]
Ben Horowitz
like, the biggest pie of all time.
- [23:50]
Sophia Puccini
Yeah. When would we, like, have some indication of, like, the AI market is actually settling to an equilibrium? Like, we can't really rely on historical data because this is sort of an unprecedented technology.
- [24:06]
Ben Horowitz
Yeah, I think it's, I think it's pretty hard to see because. Right. Like, everybody who's using it keeps using it more and more and more and more, you know, like, like you talk to businesses and their kind of token usage is growing like tenfold annually. So we're. But, but I, look, I think we'll see it when, you know, you kind of see growth really slow down in, you know, in the various players and things like that. But as long as things are growing, like several, you know, a thousand percent a year, 500% a year, like, we're nowhere near to that state. You know, in stable markets, things grow, like, you know, whatever.
- [24:49]
Sophia Puccini
Yeah. What companies do you see are currently benefiting the most from like, very, very cheap intelligence? Like the kind of companies that use like, you know, the, the kimmies and the deep seats.
- [25:04]
Ben Horowitz
Yeah, that's a good question. I mean, I think that, you know, Well, a lot of software companies are able to use it for kind of various tasks. I think on the things like customer support and specialized applications, people are able to benefit a lot from them. And then, you know, look, I think Cursor survived for a long time because it was able to have an open source alternative for cheap tasks that they have. And you know, now they're, of course they're on, they're on Grok. Um, but you know, I think every application company in AI world benefits from open source right now to some degree. Like they're all using some. Nobody's not using an open source model who's an application provider that I know of? So it's kind of that whole field. And so it's really a catastrophe for, for that whole part of the ecosystem. If, you know, if these guys get their way
- [26:09]
Theo Jaffe
when it comes to these like, very cheap models, it seems like I can imagine a future where if the really good models, like the flagship frontier models, get much more capable and can, you know, really do a lot of the tasks that an employee can do at a job that enterprises would be gladly willing to spend a lot more on that rather than go to the much cheaper models in order to cut costs. So do you think that could.
- [26:37]
Ben Horowitz
I think that's been the case so far. Right. Like, so it's not like Anthropics having trouble selling Fable or OpenAI is having trouble selling Soul like that. Those things are growing very fast. Both growing very fast. So I think that's true. And then there's a question about which categories of jobs would you rather have superintelligence and pay 8x or 10x as much for. And then which class of jobs would you rather have? Something faster and cheaper. And that'll be a question. And I think there's probably plenty of jobs in each category, just like the humans. Right? Like, would you pay a crazy amount of money for an AI researcher? Oh, definitely, like, clearly. And then there's, you know, okay, would you pay an insane amount of money for a janitorial service? Probably not. You know, like, would you pay 100x for a better janitor? Maybe not, you know, would you pay 100x more for a better accountant? I don't think so. Yeah, so like, so there, there, there's different jobs in different categories. So, you know, not everything will be that, but there will be plenty of jobs that are that.
- [28:02]
Sophia Puccini
Another question, it's kind of random, but how optimistic are you about AI art? Do you, like, do you think, how do you think we're going to fix the slop problem. Do you think it's going to get worse before it gets better?
- [28:16]
Ben Horowitz
No, I think it'll. The way I would look at it is I'm not that worried about the slot problem. I think that. I think the exciting thing is that for creatives, any idea you have is now possible. And so we're going to get a lot of new ideas, great new ideas and everything in our film, music, you know, new kinds of things that, you know, new kinds of video ideas and so forth. So I think it's going to be a little bit of a renaissance for it in that, like, every time we have a new technology, we get new kinds of art. Yeah. And, you know, from like the invention of the saxophone, which I think was like in the 1800s, 1850 or something like that, the invention of electric guitar, you know, we got rock and roll from that. We got jazz from the invention of the saxophone, and we got hip hop from the invention of the drum machine. Like. Like every time there's a new technology, we get a new kind of art form in music. And I think that's going to happen here. And there will be stuff that's really good and there will be more people who can make bad things. There was more people who could be bad rappers than there were, who could be bad rock and roller, because you don't have to learn how to play an instrument. But that's okay. Like, I think that's fine. I'm for that, like, you know, anything that. I also think it's just good for everybody to be able to express themselves. And so better tools for expression is generally a good thing. You know, there are issues with it. You know, I do get nostalgic for some of the things of the past, of course, but, you know, like,
- [30:03]
Sophia Puccini
what
- [30:03]
Theo Jaffe
music are you listening to now?
- [30:06]
Ben Horowitz
Yeah, so it's been a little bit of a drought for me on new music, but I actually like the new. The Wyclef album, I thought was pretty interesting that he just did. And I've been listening to a lot of old music. I've been listening to Curtis Mason also, who is really underrated. Really good.
- [30:28]
Sophia Puccini
But yeah, I think, like, with things like this, it increases the amount of, like, really mediocre artists. Not mediocre, because that's mean. But like. Yeah. And then it makes it really hard to be very, very good just because there's so much competition in the space and everyone is like. It's like very democratized, which is a good thing. But it's just hard for, like, from a discernment perspective, kind of.
- [30:54]
Theo Jaffe
Yeah.
- [30:54]
Ben Horowitz
But that kind of happened in a big way with Spotify too, right?
- [30:58]
Sophia Puccini
Yeah.
- [31:00]
Ben Horowitz
Soundcloud. Soundcloud, yeah, soundcloud. Like, you know, anybody could put anything on their YouTube, all that kind of thing. But you know, I still think there's great stuff that's higher art and there's. And there will continue to be, I think.
- [31:17]
Theo Jaffe
Yeah, I totally agree. Well, Ben, it's been so great having you on MTS to talk about Open.
- [31:23]
Sophia Puccini
Thanks for joining us.
- [31:24]
Theo Jaffe
And music, thanks so much for joining.
- [31:26]
Ben Horowitz
Yeah, okay. Thanks a lot and appreciate it very much. Thanks, Ben.
- [31:33]
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 substack.com 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 individual unaffiliated with A16Z. Such advertisements, companies and individuals are not endorsed by AH Capital Management, LLC, A16Z or any of its affiliates. Information is from sources deemed reliable on the date of publication, but A16Z does not guarantee its accuracy.