Sam Altman Returns to OpenAI. Now What? — With Aaron Levie

0:00Sam Altman is back at OpenAI. But does that mean everything is right in the Microsoft and OpenAI universe? And what happens next to the AI field? Aaron Levy joins us to break it all down right after this. Welcome to Big Technology Podcast, where we've somehow become a daily show about the travails and the next chapters of the OpenAI saga. Maybe this is the final one. We're here today with Vox CEO Aaron Levy, who seems to always be here when things are going insane in the AI world. And maybe that's just a function of this industry, but we definitely have a live one to discuss today. Aaron, welcome to the show.

0:40Thanks. Good to be on. And I do like that I get to talk about all the drama with you all the time. Yeah, we love having you on and we appreciate you being here in a very hectic moment in the tech world. Let me start with this. Is this good for AI? Yes, it is. Why? I think, and I want to be very clear that anything I'm saying is purely on reading the outside reports on what's going on. So some of this is a bit conjecture. But if you look at kind of what has been talked about as what led to the events where Sam had to depart OpenAI, I think actually, I think this conversation on, you know, sort of AI safety and doomerism actually does need to come to a head eventually.

1:35And I think we have to get to a point where we're talking about AI safety in much more, I think, reasonable, practical terms. I think we have let, this is all just sort of my personal view. I think we've let sort of the conversation move more toward, you know, really extreme, very like just insanely low probability existential risk type conversations. And when you have that at a board level, and especially if there's multiple board members that are starting to or always have believed that, you know, AI represents an existential risk to humanity. and they're in this position where they're going to save humanity if they stop something from happening, that is clearly going to lead to just undesirable outcomes in an organization whose mission is to build AI and advance the state of AI.

2:31You eventually will not be able to reconcile those views at some point. And so I think to some extent this was probably always going to have to blow up at some stage, and this was just the moment that it did. And I think it's going to lead to some much more healthy conversations about how do we advance AI, do it in a safe way. But let's start to get a little bit more grounded in how we talk about some of these topics. Do you think OpenAI let some of this safety stuff get out of hand on its own? I mean, Sam was talking recently about how it's good that the board can remove him if they think he's taking this in the wrong direction.

3:08I think that, you know, it's interesting, right? So I think actually that's a very good sentence to say because it's showing that there's accountability and that you are not some kind of – you can't run the organization in a dictatorial way and have full autonomy and there's no oversight. Like that's actually – like you would want a CEO to say that about their organization. I think what we're learning, though, is that it wasn't a board that was diverse enough in opinions and experienced enough in some cases on how you operate a large-scale organization, how you make thoughtful decisions at a board level.

3:50and that actually turned out to be the problem. And he probably, at the time of saying that, didn't predict that that was something that would be plausible as a scenario. I think he was probably imagining saying that as like, if I truly am kind of running out of control and going crazy, the board has the ability to remove me. And that's a safety valve to the extent that people think there's risk out there. I don't think that his intent was they would do that you know, preemptively by 10 years before we've sort of shown any evidence that we're doing things in an unsafe, you know, type way. I've come down pretty hard against the doomers.

4:28Even before this, I wrote a story, maybe I think it was last week about how the AI doomers are finally getting their comeuppance. And I did get some feedback that said, I think it was kind of interesting to hear it that said, basically, like, what if you're discounting something that serious? and of course I don't think that I personally don't think that we're like heading towards this AI doom scenario but like you do hear the way that Sam has been talking about the stuff that they've developed let me and I gave this I spoke about this on one of our emergency shows as it was going on but to Lorene Powell Jobs he said I think people have in their mind how much better the model will be next year and it's remarkable how much different it is so I'm kind of curious Like, do you feel like these these fears are completely overblown?

5:16And given the acceleration and the trajectory that the technology is moving, is it possible that the folks who are concerned about safety might have a point? I think that it's first of all, I like the you know, I don't really like this term so much and how and how it generally gets used. But but I like the marketplace of ideas. Like, I actually think I actually think we should be constantly debating, you know, AI and AI safety and where the risk is and whatnot. I think it's a very interesting intellectual type of conversation. And I have friends that are on extreme ends of the AI safety debate.

5:54But then that's still different from board governance. And you're literally on the board of directors of a company whose mission it is, and a nonprofit whose mission it is, to advance the state of AI. And so then at some point, we have to live in the real world and be practical about the implications of that. Like, don't join that board if your views are what you're doing is that harmful to society. If you really deeply have that innate fear, like, that's just the wrong board of directors to be on. So it and so like like, you know, I like there's an alternative universe where on Thursday night, the board members that that were thinking that open AI was moving too fast, they should have just resigned in protest, done a public done a public post about why they resigned.

6:41And we should debate about how fast open AI is moving. But you don't think that this that this that we're like at a point where the technology is potentially threatening to humanity? I can't imagine how we get from what we're looking at today to threatening to humanity. And so we are like complete step function, discontinuous breakthroughs away from that. We are at the stages of like we have some productivity gains across a few sectors right now with generative AI. and we're in the very early stages of trying to figure out the implications of how to incorporate this into our software in a meaningful way.

7:25There's simply nothing that relates to what we're looking at now and something that is sort of humanity ending. Right. So looking ahead, Sam's back at OpenAI right now. Is this really the end of this? We don't have a board seat from Microsoft. We don't have the results of whatever investigation is going to happen. The board is apparently like it consists of like it's a hilarious board right now. What is it like Larry Summers, Adam DeAngelo and Brett Taylor? And and so so do you think this is really the end? I mean, they're supposed to fill it with nine more people. It's amazing. You even got back in there.

8:00What do you think about it? Well, it's it's almost by definition not over. But I don't I think that the dramatic part period has I think if you like looked at the Google Trends graph of like how much drama and how many people are following the OpenAI saga, I think we're going to see a dramatic, you know, sort of, you know, drop from this point, simply because I think between Brett and Adam and Larry, you know, their charter is very clear at this point. It's, you know, again, this is from what I'm reading, build out a strong board that can govern, you know, OpenAI. And you can almost just like already predict the kinds of names of people that would be on that board.

8:41It'll be operators, it'll be policy people, and, you know, it'll be a, you know, I can only predict a very well, you know, constructed board of, you know, strong leaders, and then their job will be to provide accountability and sort of oversight of the organization, and Sam and Greg and the team will go and build AI, and that seems like, you know kind of almost everything's back to normal but probably with a hopefully a clearer charter of of the organization you know on a go forward basis yeah you and i had emailed uh when this was all getting started and you know talking about having you on here when things resolve and when sam agreed to go to microsoft i was like is it time to email aaron it just felt like the ball was still up in the air and i that's when once he decided he's back like i was like okay maybe this is the time.

9:34I hope this is really it. This feels it. I mean, the fact that you have the either forced or voluntary resignation of the board members that appear to be at the center of a lot of this, I think we're now entering a new chapter of this. And I think it was very clear in the process, the quote unquote leverage that Sam had, where basically the entire organization was willing to go over to Microsoft. So at some point, you know, it becomes kind of semantics, which is like, well, you know, what is OpenAI if there's only 20 people left and, you know, a chat-gbt domain, if everything moves over to Microsoft and they can just go replicate the same thing, then you kind of lose all your leverage because all you've really done is, you know, you've gotten rid of a nonprofit organization essentially.

10:22And so you don't, there's not a lot of, you don't have a lot of negotiating leverage in that case. So I think we're kind of, now that we've seen all the pieces on the board, I think it becomes pretty, a much more boring saga at this point, which is a good thing. Right. Okay, so I want to spend the bulk of our time together talking about what's next. Because obviously, like, this is this seismic moment, and there's like a very tidy narrative that a lot of people have wrapped up that it's great news for OpenAI and Microsoft. and I'm questioning that a little bit. It seems like a lot of people are now kind of evaluating their relationship with OpenAI, not saying they're leaving, but saying they're like, wow, have I locked my company into the OpenAI API and what can I do to make sure that I am protected in case something like this happens?

11:15What do you think? Yeah, so first of all, I think other than a very fast-moving startup where you just don't have any time to have multiple paths that you're investing in, I think most organizations at scale have already imagined a world where they have to have optionality on where their AI comes from. Any CIO we talk to, any CTO of, let's say, a SaaS company above a couple hundred million in revenue, and you talk to them about their AI strategy, either they've already done a mix of investments in AI. They're doing something with Llama. They're playing with Anthropic. They've done something with Google.

11:57Or they've at least been building an architecture that supports that. So I think that optionality is already existing. But OpenAI has the really important feature, which is they have the most advanced models at the lowest cost per token. And that is a competitive weapon and a competitive advantage that does not seem to be slowing down at this point. And so despite all of the events of the past week, that advantage is just an objective advantage. And these are not at this moment yet thus far commoditized components that you can just sort of swap in and out. Maybe they will get there in five years from now.

12:42I actually would prefer that they do because we always like to be able to have flexibility with our underlying suppliers. But they're not swappable right now. And OpenAI does have the most advanced and the lowest cost per token models from our testing. And so I think the implication is simply just that, you know, hopefully there will continue to be leapfrogs by other vendors, Google and Amazon and Anthropic and Meta. But the reason why, at least I personally felt like it was super important to have the open AI piece resolved is they do have the best technology. And so it would be very unfortunate to lose that and have to go to something that is more, less advanced and then thus produce a less high quality experience for our customers.

13:28So that's why I really like the resolution with what we landed on. Right. So let me know if I'm getting this wrong, but you do have this OpenAI integration where people can effectively chat with their documents, right? Yep. And so you feel today that you can't really swap that out for an Anthropic, for instance. We totally could. I'm talking about very small, subtle differences, but in the most objective measurement, OpenAI is in the lead. And so we could, of course, introduce slightly more degraded functionality in a pinch. But I prefer an outcome where OpenAI continues to execute as we've seen.

14:12Right. So are you building now or are you preparing now to be in a place where you're effectively model agnostic or are you there already where you can very quickly switch? We are model agnostic. Okay. What are you hearing from? I mean, Anthropics out there, Inflections there, they must be like right now making a very serious sales pitch to try to get others on. I mean, what are you hearing from them? Yeah. So I mean, because it's only been a couple of days, I probably can't sort of yet synthesize what their pitches have done or how that's changed. I would say, you know, interestingly, in something like Anthropic, and I have a lot of respect for their advancements and their models are incredibly strong, to be very clear.

15:00In the case of Anthropic, you know, interestingly, you know, they are a safety first organization, which is fantastic. Again, in that sort of idea of marketplace of ideas, you want different companies trying out different things. but it's not 100 % obvious that they would not have the same kind of board type event as OpenAI at some point if they've – Right. They could have the same thing. Yeah. Like for the first time ever, like I actually am like, oh, actually I would want to see your board of directors before telling all of our customers to go rely on your product. Like I didn't realize this was a sort of an area of risk previously.

15:43And now it's obviously very, very apparent that it is in this space. I had this CEO on LinkedIn drop a comment on a post that I wrote today, and I thought it was very interesting. He said, we have a proprietary AI engine for internal corporate data plus open AI for public data. I changed that strategy last week after the fiasco emerged to look at other point solutions in this space, such as Anthropic, which looks OK for now. I've changed and I will not be changing back. I'm disappointed as we put a lot of effort and cost into OpenAI. But you cannot run a solution stacked faced with the circus and held ransom every time something goes wrong or people don't agree.

16:20It is really bad and I will not solely rely on OpenAI. It's too risky. You think that's going to be the mentality for others? I think that, again, I think that's a good mentality. There's upper limits of what – of then how that actually works because you can be model agnostic, but if one model is literally superior for your use case, then you're still going to default to the best model. And then really what you're saying is that in a kind of nuclear scenario, you can downgrade to some other model, but that's not like your default path. And that's actually just good business continuity in general.

17:03I think even when you're thinking about purely your infrastructure, if one data center zone goes down, you go to another. And that could actually be an entirely different vendor in some cases. So I think AI will have a similar characteristic as you incorporate AI into your products. But there are going to be subtle differences of then the product experience as you kind of downshift to another AI model. Yeah. Okay. And now I'm going to get to the open source thing. So I did post this on Substack, you know, my thoughts about the move to model agnosticism. And someone took it one step further, which is that, and they said it this way, they said, if you're a company that fine tuned open AI models for your product, this weekend or this past week may lead you to switch to open source foundation models.

17:51Basically, that if you needed any level of customizability, you may kind of say, well, screw it. Like, we're not going to risk any company with any board, kind of like what you pointed out. We're going to go open source. We're going to, you know, tune the model that works for us. And that indicates that there'll be a rise of open source development. What do you think? That actually, you know, so open source is the best antidote to and the best, you know, counterbalance to, you know, any kind of private provided technology. So I'm in general always a favor of open source. I think what we've seen from open source is still not as advanced as some of the proprietary services.

18:37But this event is just one of like an infinite set of reminders of why it's always more ideal if you can control your technology. Nothing can ever be ripped out from under you under any circumstance. you know, kind of with some asterisks. And so I would always, if I always had the choice of the same technology, closed or open source, I will always choose open source. Today, that is not yet a choice that is realistic, just because, again, where the state of the models are. And that's where you kind of get into the talent question, right? Opening, I had always been able to attract these top flight researchers because it had this dual promise of pursue AGI, but do it like in a way that you could feel safe with, you know, consciously because of this special board.

19:27Now the board's going to look a lot more like a standard corporate board and it might give companies that are trying to develop open source models a leg up because they could be like, hey, listen, like at least this is like publicly available open source. You're contributing to the greater good. In fact, already today, Jan LeCun, who's the chief AI scientist at Meta, is amplifying this message and basically, you know, all but telling people to come over to meta and that meta could be the big winner here. What's your perspective? Yeah. So, um, so I think, you know, uh, uh, I, I think in the history of, of software and technology, we always have, we always have some yin and yang of, of a great proprietary platform.

20:07And then somebody who counter positions that with an open platform. And we had it with Android and iPhone. We've had it with Linux and, you know, Windows and other platforms. We have it in Oracle versus MySQL. We've had it in buying kind of proprietary infrastructure and then Facebook's open data center, you know, design effort. So there's almost inevitably always a closed versus open source, you know, battle in technology. And in each kind of product category or class of product, There's sometimes a different winner. So in my personal life, I use mostly Apple products because that vertical integration that they provide means that you usually have better security or better user experience.

20:54But in our data center design, we almost always want open source infrastructure services because of our ability to control that, control upgrades, get the power of the community developing on it. And so, you know, AI is one of these technologies, which is how much does it benefit from the vertical integration? That's an open question. How much does it benefit from the pure kind of scale and flywheel effect of the company that gets strongest, can spend the most on GPUs, which can train the biggest models, which can then have, you know, the better AI, which then gets more revenue, and then that becomes a flywheel?

21:35You know, right now, you know, Facebook is sort of doing this. They're somewhat subsidizing the whole effort without kind of like obvious monetization. Whereas when Google did this with Android, they were going to monetize it through effectively search and advertising. And so, I mean, I love that Meta is investing so much in open source AI. They obviously benefited from probably their ad algorithms or whatnot. But it's probably hard for them still to compete with OpenAI on some dimensions because OpenAI has a tremendous amount of focus. They have a tremendous amount of resources to attack this problem.

22:15And clearly they have some of the best talent in the world working on it. But I think, again, if you look at this in 10 years from now or 20 years from now, you'd almost always believe that open source would be able to catch up. just because we've seen that in almost every other era of technology. But there could also be some idiosyncrasies about this where the commercial version is able to always sort of stay ahead for some set of reasons. And so I would just say this one is too confusing to predict at this point. Where's Google in all this? I mean, we keep hearing about this Gemini model. It's nowhere to be seen.

22:51Yeah, so Google is – Google I would normally think of as in the – It's actually interesting because Google usually will do the counter strategy to whatever the incumbent strategy is. That's sort of their typical path of disruption. So, you know, in theory, they should be the meta of this situation doing the open source. But they obviously haven't pursued that path. So, which kind of leaves them, you know, somewhere in between these approaches. And I think we're all, you know, equally waiting on the Gemini launch. and we'll just see kind of what that looks like and where it's at. Sam Altman's sort of made the rounds in Washington and around the world and been a very effective lobbyist for the regulation and the safeguards that he wants because he's had this company at his back that's been this nonprofit AI safety focused company.

23:41Can he still do that under a new structure? I mean, he's going to lose some of the sheen and credibility you'd imagine that he brought beforehand. Well, first of all, I don't know that they've changed the nonprofit element. Do you know if that's the case? Well, not, no, the corporate structure isn't going to be different, but like, we all know that it's going to be very different, a very different board. Yeah. You know, in terms of the people on there and like the Microsoft association, no more surprises. Yes. I guess I never, I never, I mean, I don't know how like, you know, policy engines think, but like I never saw, I never saw that their board was the reason why I could trust that organization.

24:22I trusted them more because they have a working product with leaders that I generally understand their motivations. They have the most advanced technology. They have the best researchers. And so as a result of this change, I would say – I would actually say – I mean actually, if anything, Sam's whole sort of commentary is probably even more credible because we've – They did fire him. They did fire him. So like he clearly is – he works at the mercy of the board. But I would probably trust the decisions to fire him much more with a much more stable, organized, thoughtful board of directors. Do you know what the story is behind this war between effective altruists and these E-accelerationists or effective accelerationists?

25:17For the life of me, Aaron, I'm trying to piece it together. It's bananas. I think I understand it. What's your take? I don't know. I mean, so here, I'll just give my understanding. I think there are these, you know, these effective altruists, which tend to like, you know, want to be very cautious about the development of AI. And then the accelerationists who are like build as fast as possible. Like it's basically like, you know, the open AI board of last week against Mark Andreessen's techno optimist manifesto. Is that right? Like and how big of a factor are these factions within Silicon Valley?

25:51So, well, first of all, I think, like, you know, in a very simplistic way, I think you just described it. So I think you do fully understand it. Relief. And, like, the subtleties are, like, I think there's one camp, and I don't want to, like, I'll be label-free for a second. So I don't, you know, step into landmines. I think there's a camp that sort of subscribes to, hey, like, let's slow down AI progress. and that'll give us time to figure out how we can make it more safe. And then there's another camp which says, like A, it's not even really clear if that makes any sense. Like how does one decide what's the right speed?

26:35Like should we type the code slower? Should we just arbitrarily put fewer GPUs to work? Like it's sort of like antithetical to everything else we do in technology progress, which is when you get to the next step, you learn more information. So you actually want to get to the next step as quickly as possible. So then you get to the next step and so on. And so then that other camp is sort of like, well, actually, AI is actually a good thing for society. So let's actually accelerate as much as we can to be able to get there. And, oh, by the way, along the way, we'll learn what's not safe or what's not working.

27:10And the ecosystem will continue as we always have in the past, you know, many decades of technology. we'll continue to build the right safeguards as we go along. But, oh, by the way, like, we don't need to be fearful for our AI future. So that's also why we can just have this underlying sort of rate of acceleration. And that's, you know, how I generally view the two camps. And I think that the, you know, it's interesting. I actually, probably a week ago, I would have sort of said, hey, yeah, these are like academic conversations that like talk about at dinner. and these are just the intellectual fun debates of Silicon Valley.

27:49Obviously, it turns out that these things are much more significant and real at the board level of these AI labs. And I would not have predicted that. Just one thing about AI safety. One of the things that I saw, I mean, I had an AI founder who was basically DMing me and said, economics always win. That even if we had an existential risk with AI, if OpenAI just stopped producing their thing, their technology, like the people that invested in them and the employees with money on the line would find a way to override whatever safety was built into their structure? What do you think about that?

28:28Well, the reason why this is hard is my brain has not yet been able to click. My brain hasn't been able to click into what is the actual specific fear that people are worried about. Right. And and then why that fear, if that fear even gets remotely true, we're not even like talking about like capitalism, like level, like like, you know, intervention. We're talking about government intervention. Like if I literally is doing things that kill people, then then like, you know, Josh Kushner will not be able to override that, you know, you know, because he did a tender round. That's like literally where the FBI is going to show up and they're going to be like, what have you guys been building?