Bonus Episode: Sam Altman (CEO, OpenAI) Talks GPT-4o and Predicts the Future of AI

In this cross-over episode, Sam Altman sat down with Logan Bartlett (Investor, Redpoint) on the day of the ChatGPT-4o announcement to share behind-the-scenes details of the launch and offer his predictions for the future of AI. If you've already listened to the episode, check out Pat and Jacob's debrief at the end of the episode! Altman delves into OpenAI’s vision, discusses the timeline for achieving AGI, and explores the societal impact of humanoid robots. He also expresses his excitement and concerns about AI personal assistants, highlights the biggest opportunities and risks in the AI landscape today, and much more.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:04Jacob Effron?

  This is Jacob Afaron, and today we have a special crossover episode of Unsupervised Learning. What we're going to hear is a conversation between Sam Altman and my partner Logan Bartlett. In keeping with our goal on the show to uncover the tools, principles, and views of the leading operators in AI, we figured it made sense to cross-promote this episode as it does just that. It was recorded on Monday, May 13th, the day of the announcement of GPT-4O, but the interview covers a bunch of topics from AGI, AI risk regulation, to OpenAI's future plans, and Sam's perspective on AI personal assistance. We hope you enjoy this special crossover episode of Unsupervised Learning. Definitely stay tuned for Pat and my debrief at the end, and without further ado, here's Sam Altman.

• 0:39Speaker 01

  Thanks for doing this.

• 0:40Speaker 02

  Yeah, of course.

• 0:41Speaker 01

  All right, I want to start off easy. What's the weirdest thing that's changed in your life in the last four or five years running OpenAI? Like, what's the most unusual shift that's happened?

• 0:53Speaker 02

  Um, I mean, quite a lot of things, but the sort of inability to just be like mostly anonymous in public is very, very strange. I think if I had thought about that ahead of time, I would have said, okay, this is like a weirder, this would be a weirder thing than it sounds like, but I didn't really think about it. It's like a much weirder thing. It's like a strangely isolating way to live.

• 1:15Speaker 01

  You believed in AI and the power of the business, so did you just not think through the derivative implications of running something that was that powerful?

• 1:23Speaker 02

  Yeah, I think I didn't think, there were all of these other things, like, oh, AI is going to be like really important, OpenAI is going to be a really important company. I didn't think I would like not be able to like go out to dinner in my own city.

• 1:34Speaker 01

  That's weird.

• 1:35Speaker 02

  That's weird.

• 1:36Speaker 01

  You made an announcement earlier, did I?

• 1:37Speaker 02

  We did.

• 1:38Speaker 01

  Multimodal 4O?

• 1:40Speaker 02

  Yeah.

• 1:41Speaker 01

  That's the Omega sign, right?

• 1:43Speaker 02

  The O, just the, like, O, like Omni.

• 1:45Speaker 01

  Yeah, Omni, okay, sorry. It works across text, voice, vision. Can you speak to why this is important?

• 1:52Speaker 02

  Um, because I think it's like an incredible way to use a computer. This fact, this, I, we've had like voice, the idea of like voice-controlled computers for a long time. You know, we had Siri, and we had things before that. They've never, to me, felt natural to use. And this one, many different reasons, what it can do, the speed, adding in other modalities, the inflection, the naturalness, the fact that you can do things like, say, hey, talk faster, or talk in this other voice, and that it's, the

• 2:29Speaker 02

  The fluidity, the pliability, whatever you want to call it. I just can't believe how much I love using it.

• 2:36Speaker 01

  Yeah, Spike Johnson would be proud.

• 2:38Speaker 02

  It's.

• 2:39Speaker 01

  Are there use cases that you've gravitated to?

• 2:42Speaker 02

  Well, I've only had it for like a week or something. But one surprising one is putting my phone on the table while I'm like really in the zone of working. And then without having to like change windows or change what I'm doing, using it as like another channel. So I'm like working on something I would normally like stop what I'm doing, switch to another tab, Google something, click around or whatever. But while I'm like still doing it, to just ask and get like an instant response without changing from what I was looking at on my computer, that's been a surprisingly cool thing.

• 3:15Speaker 01

  What actually made this possible? Was it an architectural shift or more compute?

• 3:20Speaker 02

  I mean, it was like all of the things that we've learned over the last several years, we've been working on audio models, we've been working on visual models, we've been working on tying them together, we've been working on more efficient ways to train our models.

• 3:36Speaker 02

  It's not like, okay, we unlocked this one crazy new thing all at once, but it was putting a lot of pieces together.

• 3:42Speaker 01

  Do you think you need to develop like an on-device model to decrease latency to the point for usability?

• 3:49Speaker 02

  For video? Maybe it would be hard to deal with network latency at some point. Like, like the, the, a thing that I've always thought would be super amazing is to put on someday a pair of AR goggles or whatever and just like speak the world in real time and watch things change. And that might get harder over network latency. But for this,

• 4:10Speaker 02

  you know, two, 300 milliseconds of latency feels super like feels faster than a human responding to me in many, many cases.

• 4:19Speaker 01

  Is, is video in this case, images?

• 4:21Speaker 02

  Oh, sorry. I meant video if you wanted like generated video, not pro, not, not input video.

• 4:26Speaker 01

  Got it, got it. So, so currently it's working with actual video as is?

• 4:30Speaker 02

  Well, like frame by frame of the video.

• 4:32Speaker 01

  Frame by frame. So it's, okay. You're related recently to ChatGPT, maybe not being, they're the next big launch, not being GPT-5. It feels like there's been sort of an iterative approach to model development that you guys have taken. Is it fair to say that's how we should think about it going forward? That it's not going to be some big launch, here's ChatGPT-5, but instead?

• 4:57Speaker 02

  We honestly don't know yet. I think that definitely one thing I've learned is that AI and surprise do not go well together. And although, you know, the traditional way a tech company launches products, we should probably do something different. Now, we could still call it GPT-5 and launch it in a different way, or we could call it something different. But I don't think we've figured out how to do the naming or branding for these things yet. Like, it made sense to me from like GPT-1 to GPT-4 at the launch. Now, obviously, GPT-4 has continued to get much better. We also have this idea that there's going to be like, you know, maybe there's like one underlying kind of like

• 5:40Speaker 02

  virtual brain and it can like think harder in some cases than others. Or maybe it's different models, but maybe these are just a care if they're different or not. So I don't think we know the answer to how we're going to like product market all of this yet.

• 5:52Speaker 01

  Does that mean maybe that the the the needs of the compute to make incremental progress on models might be less than what it's been historically?

• 6:03Speaker 02

  I sort of think we'll always use as much compute as we get. Now, we are finding incredible efficiency gains, and that's really important. One of the, you know, the cool, the cool thing that we launched today is obviously the voice mode. But maybe the most important thing is we were able to make this so efficient that we're able to serve it to free users. Like, best model in the world, by a good amount of people, look at that little thing, served to like anybody who wants to download ChatGPT for free. And it was a remarkable efficiency gain over GPT-4 and GPT-4 Turbo, and we have a lot more to gain there.

• 6:36Speaker 01

  I've heard you say that ChatGPT didn't actually change the world in and of itself, but maybe just changed people's expectations for the world?

• 6:46Speaker 02

  Yeah. Like, I don't think you can find much evidence in the economic measurement of your choice that ChatGPT really inflected productivity or whatever.

• 6:55Speaker 01

  Maybe customer support.

• 6:56Speaker 02

  Maybe some, maybe some areas.

• 6:57Speaker 01

  Summarization.

• 6:58Speaker 02

  If you look at like global GDP, you know, can you detect when ChatGPT launched? Probably not.

• 7:03Speaker 01

  Is there, is there a point that you think will be able to determine a GDP?

• 7:08Speaker 02

  Yeah, I don't know if you'll ever be able to say like, this was the one model that did it. But I think if we look at the graph, a couple of decades in the future, it'll be like, hmm, something changed.

• 7:18Speaker 01

  Yeah. Are there applications or areas you think are most promising in the next 12 months?

• 7:24Speaker 02

  I'm sure I'm biased just because of where, what we do here, but coding, I think is a, is a really big one.

• 7:29Speaker 01

  Kind of related to the bitter lesson. You spent some time recently talking about the difference between deeply specialized models trained on specific data for specific purposes versus generalized models that are capable of true reasoning.

• 7:41Speaker 02

  I would bet that it's the generalized model that's going to matter.

• 7:44Speaker 01

  And what is the most important thing there as you think about like someone that's focused singularly on a data set and all the integrations associated with something very narrow?

• 7:57Speaker 02

  If the model can do generalized reasoning, if it can like figure out new things, then if it needs to figure out how to work with a new kind of data, you can feed it in and it can do it. But it doesn't go the other way around, like a bunch of specialized models that I don't think a bunch of specialized models put together can't figure out the generalized reasoning.

• 8:16Speaker 01

  So the implications for that of coding specific models probably would be?

• 8:20Speaker 02

  I, I, I think a better way of saying this is I think the most important thing to figure out is the true reasoning capability. And then we can use it for all sorts of things.

• 8:27Speaker 01

  What do you think the principal means of communication between humans and AI is in two years?

• 8:34Speaker 02

  Natural language seems pretty good. I, I, I'm interested in this general idea that we should design a future that humans and AIs can sort of use together.

• 8:45Speaker 02

  Use them the same way. So I'm like more excited about humanoid robots than I am for other forms of robots because I think the world is like very much not designed for humans and I don't want that to get reconfigured for some more efficient kind of thing. I like the idea that AIs, that we talk to AI in language that's like very well human optimized and that they even like talk to each other that way, maybe. I don't know. But I think this is, I think this is generally an interesting direction to push.

• 9:13Speaker 01

  You said recently something to the effect of the models might ultimately get commoditized over time, but the most important thing would likely be the personalization of the models to each individual. First, do I have that right?

• 9:27Speaker 02

  I'm not certain on this, but I think it's like a thing that I would, that would seem like reasonable to me. Yeah.

• 9:33Speaker 01

  Then beyond personalization, do you think it's just normal business UI and ease of use that ultimately wins for end users?

• 9:41Speaker 02

  Those are, those will for sure be important. They always are. You know, I can imagine other things where there's like a sort of marketplace or a network effect of some sort that matters where it's, you know, we want our agents to communicate. There's different companies in an app store, but I sort of think that the rules of business kind of generally apply. And whenever you have a new technology, you're tempted to say they don't, but that's always like fake news and not always, usually fake news. And all of the traditional ways that you create during value will still matter here.

• 10:13Speaker 01

  When you see open source models like catch up to benchmarks and all of that, what's your reaction to it? Is that?

• 10:23Speaker 02

  I think it's great. Yeah. I mean, I think that there are, you know, like many other kinds of technology, there will be a place for open source, there'll be a place for like hosted models and it's fine. It's good.

• 10:37Speaker 01

  I'm not going to ask about any specifics related to this, but there have been press I might answer. There's been press reports related to looking to raise major amounts of money. Wall Street Journal, I think was a credible one to galvanize investment in fabs. Semi Industry, ATSMC and Nvidia have been ramping pretty aggressively to meet expectations of the need for AI infrastructure. You recently said that you think the world needs more AI infrastructure and then you said a lot more AI infrastructure.

• 11:08Speaker 02

  I do.

• 11:09Speaker 01

  Is there something you're seeing on the demand side that?

• 11:15Speaker 01

  Would require way more AI infrastructure than what we're currently getting out of TSMC and Nvidia?

• 11:20Speaker 02

  So first of all, I'm confident that we will figure out how to bring costs to deliver current systems way, way down. I'm also confident that as we do that, demand will increase by a huge amount. And third, I'm confident that by building bigger and better systems, there will be even, even more demand. We should all hope for a world where intelligence is too cheap to meter. It's just wildly abundant. People use it for all sorts of things and you don't even think about whether like, oh, you know, do I want this, do I, you know, do I want this like reading all my emails and responding to them for me? Or do I want this like curing cancer? Of course you pick curing cancer, but the answer is like, you'd love for it to do both things. And I just want to make sure we have enough for it to have that.

• 12:04Speaker 01

  I don't need you to comment on your own personal efforts here, although again, if you want to, please let me know. But humane and limitless and some of these like different physical device assistance, what do you think those have gotten wrong? Or where do you think the, the adoption maybe hasn't met user desires just yet?

• 12:25Speaker 02

  I think it's just early. I have been an early adopter of many types of computing. I had and very much loved the compact TC1000, like when I was a freshman in college. I thought it was just like so cool. And like, that was a long way from the iPad, a long, long way from the iPad. But, you know, it was directionally right. Then I got a trio. I was like the, I was very not cool college kid. I had like an old Palm Trio. And when it was like a, that was not a thing that kids had. And that was a long way from the iPhone, but we got there eventually. And, you know, these things feel like a very promising direction that's going to take some iteration.

• 13:09Speaker 01

  You mentioned recently that a number of businesses that are building on top of GPT-4 will be steamrolled, I think was your term, by future GPT. I guess, can you elaborate on that point? And second, like, what are the characteristics of AI-first businesses that you think will survive GPT's advancement?

• 13:27Speaker 02

  The only framework that I have found that works for this is you, you can either build a business that bets against the next model being really good, or a model that bets on that happening and benefits from it happening. So if you're Doing a lot of work to make one use case really work that was just beyond the capability of GPT four, GPT four, oh no. And then you get it to work, but then GPT five comes out and it does that and everything else really well. You're kind of like sad about the effort you put into that one thing to get it to barely work. But if you had somebody that just like kind of worked okay across the board and people were finding things to use for, but you didn't put in like tons of work to to make this one thing kind of possible.

• 14:11Speaker 02

  And then GPT five, or whatever we call it, comes along and it's just way better, everything you're like, you got the rising tide lifted, all your boats affect. You know what I would suggest is like, you're not building an AI business. In most cases, you're building a business and AI is a technology that you use. In the early days of the App Store, I think there were a lot of things that like filled in some very obvious crack. And then eventually Apple fixed that. And there wasn't, you know, you didn't keep needing like a flashlight app from the App Store, which is like part of the OS. And that was like going to happen. And then there were, I think, things like Uber that were enabled by having smartphones, but really built a very defensible, long-term business. And I think you just

• 14:56Speaker 02

  want to go for that latter category.

• 14:58Speaker 01

  I can come up with a lot of incumbent businesses that leverage you all that fit that framework in some ways. Are there any like novel types of concepts that you sort of think is in that example, the Uber, and it doesn't need to be, it could be a real company if you think of one, or even if it's a toy or just something that's interesting that you think is like enabled in that way?

• 15:23Speaker 02

  I would actually bet on the new companies for like many of these cases. A very common example people use is trying to build like the AI doctor, like the AI diagnostician.

• 15:35Speaker 02

  And people talk about, oh, I don't want to do a startup here because Mayo Clinic or take your pick is going to do it. And I'd actually bet it's a new company that does something like that.

• 15:44Speaker 01

  Do you have any advice for CEOs beyond that who are want to be proactive about preparing for these types of disruptions?

• 15:54Speaker 02

  I would say like, bet that intelligence as a service gets better and cheaper every year, and it is necessary, but not sufficient for you to win. So the big companies that take years to implement this, you can beat them, but every other startup is, you know, paying attention is going to do this too. And so you still have to figure out like, what's the long-term defensibility of my business? Now, the playing field is way more open than it's been in a long time. There's incredible new things to do, but you don't get a pass on like the hard work of building enduring value, even though you can now do it in more ways.

• 16:33Speaker 01

  Is there a job title or a type of job responsibility that you could envision existing or being mainstream in five years because of AI that like is maybe niche or nonexistent today?

• 16:49Speaker 02

  That's a great question, and I don't think I've ever gotten it before. People always ask like, what job is going to go away? The new one is a more interesting question. Let me think for a second. I mean, there's like a lot of things that I could talk about that I think are sort of less interesting or less huge. What I'm trying to do is like come up the areas of like, what will 100 million people do or 50 million people do? The broad category of new kinds of art, entertainment, sort of more like human-to-human connection, I don't know what that job title is going to be, but I think.

• 17:27Speaker 02

  And I don't know if this like we get there in five years, but I think there's going to be a premium on like human in-person like fantastic experiences. I don't know what we'll call that, but I can see that being like a very huge category of something new that we do.

• 17:40Speaker 01

  The most recent public tender of OpenAI was 90 billion or something in about there. Are there one or two things that you sort of look at as milestones? That will get OpenAI to be a trillion dollar company short of AGI?

• 17:57Speaker 02

  Uh, I think if we can just keep improving our technology at the rate we've been doing it and figuring out how to continue to make good products with it and revenue keeps growing like it's growing. I don't know about specific numbers, but we'll be fine.

• 18:12Speaker 01

  Is the the business monetization model today, the one that you think creates the one trillion dollar equity value?

• 18:20Speaker 02

  I mean, the chat GPT subscription model like really works well for us. Like surprisingly, I wouldn't have bet on that. I wouldn't have been confident it's going to do as well as it has, but it's been good.

• 18:29Speaker 01

  Do you think post AGI, whatever that term actually means? We'll be able to, I don't know, ask AGI what the monetization model is that might be different?

• 18:42Speaker 02

  Yeah, yeah, should be able to.

• 18:44Speaker 01

  I think we maybe saw in November, not to rehash, uh, that that the existing OpenAI structure left some things to be desired, which I don't think we need to rehash in total. You talked about it enough, I think, but um, you spoke into making changes along the way. What do you think the appropriate structure is going forward?

• 19:07Speaker 02

  Um, I think we're close to being ready to talk about that. Uh, we're like, we've been hard at work on all sorts of conversations and brainstorming there.

• 19:19Speaker 02

  I think like, hopefully in a, like this year, I think we'll be right talking this calendar year. You tell me. First. We'll see.

• 19:26Speaker 01

  When Larry and Brett Taylor got Battlefield promoted to board directors, I was waiting for, you know, my call never came through, but I, one of the interesting things I think about preconceptions around AI, to your point on the monetization model and all that, is I think we've all, I've heard you speak about it. Manual work, obviously first, followed by, you know, white collar, followed by creative. Obviously, it's proven to be kind of the opposite in some ways. Are there other things that are counterintuitive that you've looked at being like, well, I would have presupposed it to be this way, but it's actually proven to be the exact opposite?

• 20:00Speaker 02

  That's definitely the mega surprise to me, the one that you mentioned. There's other, like, I don't think I would have expected it to be so good at legal work so early, just because I think of that as like a very precise, complex thing. But, but no, definitely the big one is the, the observation of like, physical labor, cognitive labor, creative labor.

• 20:21Speaker 01

  For those that haven't heard, you make the point about AGI and why you dislike the term. Can, can you elaborate on, on that point?

• 20:28Speaker 02

  Because I know, I don't, I no longer think it's like a moment in time. I, I obviously have so many naive conceptions when you start any company. And, and particularly in a field that's like moving around as much as this one is. But my naive conception when we started is that we would like get to a moment where we didn't have AGI and then we did. And it would be a, a real discontinuity. And I still think there's some chance of a real discontinuity. But on the whole, I think it's going to look much more like a continuous exponential curve, where what matters is the pace of progress year over year over year.

• 21:10Speaker 02

  And you and I will probably not agree on the month or even the year that, or like, okay, now that's AGI. We can come up with other tests that we will agree with, but even that is harder than it sounds. And, you know, GPT-4 is definitely not over a threshold that I think almost anyone would call an AGI. And I don't expect our next big model to be either. But I can imagine that we're like only maybe one or two or some small number of ideas away and a little bit more scale from something where we're like, hmm, this is now kind of different. And I think it's important to.

• 21:44Speaker 02

  Stay vigilant about that.

• 21:45Speaker 01

  Is there a more modern Turing test? We can call it the Bartlett test, where that you think like, hey, when it crosses this threshold?

• 21:55Speaker 02

  I think when it's capable of doing better research than like all of OpenAI put together, even one OpenAI researcher, that is like a somehow very important thing that feels like it could or maybe even should be a discontinuity.

• 22:09Speaker 01

  Does that feel close?

• 22:13Speaker 02

  Probably not, but I wouldn't, I wouldn't rule it out.

• 22:15Speaker 01

  What do you think the biggest obstacles that you see to reaching AGI? It sounds like you think maybe the scaling laws have runway currently and holding for the next couple of years?

• 22:25Speaker 02

  Yeah, I think the biggest obstacles are new research. And, you know, one of the things I've had to learn shifting from like internet software to AI is research does not kind of work on the same schedule as engineering, which usually means it takes much longer, it doesn't work, but sometimes means it works tremendously faster than anyone could have predicted.

• 22:46Speaker 01

  What is that? Can you elaborate on that point that it's like not as linear in progress?

• 22:52Speaker 02

  I think the best way to elaborate on that is like historical examples. I'm going to get the numbers wrong here, but.

• 22:58Speaker 01

  I'm sure no one will try to correct you.

• 23:00Speaker 02

  Someone will.

• 23:01Speaker 01

  Yeah.

• 23:01Speaker 02

  I think the neutron was first theorized in the early 1900s. It was maybe first detected in the 10s or 20s.

• 23:13Speaker 02

  And the work on what became the atomic bomb started in the 30s and happened in the 40s. Like from, from not really having no idea that such like, that there was even the idea of a thing like a neutron to being able to like make an atomic bomb and just like break all of our intuitions about physics. That's like wildly quick. There are other examples that are sort of less pure science. Like there's the famous quote about the Wright brothers. Again, I'm going to get the numbers wrong here, but let's say it was like 1906. They said they thought flight was 50 years away and 1908 they did it, whatever, something like that.

• 23:56Speaker 02

  And then many, many other examples throughout the history of science and engineering.

• 24:01Speaker 02

  There's also plenty of things that we theorize that never happen or take decades or centuries longer than we thought, but sometimes it does go really fast.

• 24:09Speaker 01

  Interpretability, where are we on this path and how important is that long term for AI?

• 24:17Speaker 02

  There's different kinds of interpretabilities. There's the like, do I understand like every, what's happening at like every mechanical layer through the network? And then there's, can I like look at the output and say, there's a logical flaw here or whatever? I am excited about the work going on at OpenAI and elsewhere in this direction. And I think that interpretability as a broader field seems like promising and exciting.

• 24:42Speaker 01

  I won't pin you down. I assume you'll have a nice announcement when you're ready to say something. But do you think that that is going to be a requisite to mainstream AI adoption, maybe within enterprises or something?

• 24:59Speaker 02

  GPT-4 is like quite widely adopted at this point.

• 25:02Speaker 01

  Yeah, that's true. There's maybe a few things that I think you get asked questions about or maybe accused is too strong of a term, but that people are suspicious about, one of which is, I think there's this needle threading that exists between being excited about AGI, but also feels like you have a personal kind of apprehension about you say, um, OpenAI generally being the ones to harness it and unilaterally make decisions, which has led to some, you know, some body, some governmental structure where there's elected leaders instead of you making these decisions?

• 25:46Speaker 02

  Yeah, I think for like, I think it'd be a mistake to go regulate, heavily regulate like current capability models. But when the models, which I believe they will, pose significant catastrophic risk to the world, I think having some sort of oversight is probably a good thing. Now, there is some needle threading about where you set those thresholds and how you test for it. And it would be a real shame to sort of stop the tremendous upsides of this technology and letting people that want to go train models in their basement be able to do that. That'd be really, really bad. But, you know, if we have international rules for nuclear weapons. It's a good thing.

• 26:27Speaker 01

  The regulatory capture group, which I'm sure we can think of, which VCs fall into that bucket of accusatory around this regulation.

• 26:38Speaker 01

  What do you think they don't see about the potential risks inherent in AI?

• 26:46Speaker 02

  Well, I think they just don't get, I don't think they like, on the whole, seriously wrestled with AGI. These were also people who like, some of the loudest voices about AI regulatory capture were, you know, totally decrying it as a possibility not that long ago, not all. But I do have empathy for where they're coming from, which is like, regulation has not been really bad for technology. Like, look what happened to the European technology industry. Like, I get it, I really do. And yet, I think that there is a threshold that we are heading towards, above which we may all feel a little bit different.

• 27:18Speaker 01

  Do you think open source models themselves present inherent danger in some ways?

• 27:25Speaker 02

  No, current one does. But I could imagine one that could.

• 27:29Speaker 01

  I've heard you say that safety is kind of a false framing in some ways, because it's more of a discussion about what we explicitly accept, like airlines?

• 27:40Speaker 02

  Yeah, it's more like, safety is not a binary thing. Like, you are willing to get on airplanes because you think they're pretty safe, even though you know they crash once in a while. And what it takes for, to call an airline safe is like, a matter of some discussion. Some people have different opinions on.

• 28:02Speaker 01

  And it's a topical point right now.

• 28:04Speaker 02

  Topical point right now. They have gotten just unbelievably safe overall, like triumphantly safe. But safe does not mean no one will ever die in an airplane.

• 28:11Speaker 01

  Similarly, medicine, we really, the side effects and some people have adverse consequences around it. And then there's the implicit side of safety as well, like social media, right? Or things that have negative association. Is there something that you could imagine seeing on the safety paradigm that would?

• 28:32Speaker 01

  Cause you to act differently than pushing forward?

• 28:37Speaker 02

  Yeah, we have this thing called our preparedness framework that's sort of exactly that saying that, you know, in these categories at these levels, we'd act differently.

• 28:44Speaker 01

  I've had Eliezer on the podcast.

• 28:46Speaker 02

  How was that?

• 28:47Speaker 01

  It was wonderful. We sat for the longest podcast I've ever done. I think it was four hours of us.

• 28:53Speaker 02

  He has more free time than me, so I apologize. I can't go that long.

• 28:55Speaker 01

  I, we can, we can do multiple sessions. We don't need to do them all now. I think that his, his points, I think, stay fairly consistent.

• 29:02Speaker 02

  I'm grateful he exists.

• 29:03Speaker 01

  He's a very interesting guy to sit down with for four hours and talk. We went a bunch, a bunch of different directions. But I'd be remiss as a, as a friend of the pod to not ask a fast takeoff question. I'm curious, like, there's so many different fast takeoff scenarios and one of the constraints that I think we point to today is just a lack of AI infrastructure. Right? And I guess if there was some, some researcher developed a modification to the current transformer architecture, where suddenly the amount of data and hardware scale needed drastically reduced more like, I don't know, human brain or something like that.

• 29:45Speaker 01

  Is it possible we could see like a fast takeoff scenario?

• 29:50Speaker 02

  Possible, of course, and it may not even need a modification. It is still not what I believe is the most probable path, but I don't discount it. And I think it's important that we consider it in the space of what could happen. I think things will turn to be more continuous, even if they're accelerating. I don't think we're likely to go to sleep one day with like, pretty good AI and wake up the next day with genuine superintelligence. But even if, even if the takeoff happens over a year or a few years, that's still fast in some sense. There's another question about even if you got to this like really powerful AGI, how much does that change society on the next day versus the next year versus the next decade?

• 30:35Speaker 02

  And my guess is in most ways, it's not a next day or next year thing, but over the course of a decade, then the world will look quite different. I think the inertia of society is like a good helpful thing here.

• 30:46Speaker 01

  One of the things I think people also find they have suspiciousness around. I imagine the questions you don't love getting are Elon Equity and November Board Structure. Those are probably the three.

• 31:01Speaker 02

  Which one of those?

• 31:02Speaker 01

  I've answered them a lot of times. Which one of those do you like the least?

• 31:04Speaker 02

  I don't, I mean, I don't hate any of them. I just don't have any new to say on any of them.

• 31:09Speaker 01

  But, um, well, I guess I'm not going to ask the Equity one specifically because I think you've answered that in more than enough ways, although it is, people still don't seem to like the answer that enough money is a thing.

• 31:25Speaker 02

  Yeah, if I like made a trillion dollars and then gave it away, it would just, it would like fit with, I think, the expectation or the sort of way it's usually done.

• 31:34Speaker 01

  There was another Sam that thought about.

• 31:37Speaker 02

  Oh, that's true.

• 31:37Speaker 01

  Trying that in some ways.

• 31:39Speaker 02

  Yeah.

• 31:39Speaker 01

  Yeah, comparatively.

• 31:40Speaker 02

  No, I just mean like most people who make a ton of money.

• 31:42Speaker 01

  Yeah. Yeah. What, uh, what do you feel like your motivations, this pursuit of AGI, like outside of the Equity? I think most people take solace in the fact that like, oh, well, even if I have some higher mission, I still get paid for it in some ways. Like, what are your motivations now coming into work every day? Like, what's the most fulfillment derived from?

• 32:05Speaker 02

  Look, I tell people this all the time. I'm willing to make a lot of other life trade offs and sacrifices right now because I think this is the most exciting, most important, like, best thing I will ever touch. And it's an insane time and I'm happy it won't be forever. Like, you know, someday I get to go retire on the farm and I'll remember this fondly, but be like, oh, man, those were stressful, long, long stressful days. But it's also just incredibly cool. Like, I can't believe I'm, this is happening to me. It's like, this is like amazing.

• 32:38Speaker 01

  Was there a single moment, I guess we go back to the fame example of not being able to go out in your city or whatever, but has there been a single moment that was most surreal that like, oh, jeez, I don't know. Uh, I mean, you've done a podcast with Bill Gates. I'm sure you have your speed dial. If I took your phone right now, it would have a lot of very interesting people on it. Was there a single moment over the course of the last couple of years that you were like, this is a uniquely surreal moment?

• 33:03Speaker 02

  Kind of every day, there's something that's like, wow, if I could like, if I had like a little bit more mental space to step back, it's like, this would be crazy. Kind of a fish in water. But yeah, it is kind of like that effect. After all of that, like November stuff happened, you know, like that day or the next day, whatever. I got like, I don't know, 10, 20 texts, something like that from like major world, like presidents, prime ministers of countries, whatever. And that was not the weird part. The weird part was that happened and I was like, you know, kind of responding, saying like, thanks, whatever. And it felt like very normal. And then I, we had these like insane, super jammed, like four and a half days and just this like crazy state.

• 33:47Speaker 02

  And it was, it was just like weird, like not sleeping much, not really eating. Energy levels like very high, very clear, very focused, but just like your body was like in some weird like adrenaline charged state for a long time. And then it was like, all this happened the week before Thanksgiving and it was kind of crazy, crazy, crazy. Got resolved on Tuesday night. You canceled our podcast. Cancelled our podcast. Sorry, I don't usually cancel things. Anyway, then on that Wednesday, like now it's the Wednesday before Thanksgiving. Ali and I drove up to Napa and stopped at this diner. Gott's is very good. And on the drive up there, I realized I hadn't like eaten in like days.

• 34:27Speaker 02

  And then all of a sudden, like, kind of like normal, it was just like, okay, you know, this is like normally where we'd be doing on the weekend, heading out, like, whatever. And go to Gott's, order like four entrees, like heavy, like, you know, fried, like heavy. Entrees, like two milkshakes just for me and I was sat there and ate. And it was very satisfying. And as I was doing that, one of them, president of this one country, texted again and just said like, oh, I'm so happy it's all resolved, like great, whatever. And then it hit me that like, oh yeah, like all of these people had texted me and it wasn't weird. And the weird part was like realizing that that had like happened in the middle of it. And that that should have been this very weird experience and it wasn't. So that was like one that sticks out.

• 35:11Speaker 02

  Yeah, that is interesting. My takeaway is human adaptability to almost anything is just like much more remarkably strong than we realize. And you can get used to anything as the new normal, good or bad, pretty fast. And I kind of like, over the last couple of years, have

• 35:31Speaker 02

  Learned that lesson many times, but I think it says something remarkable about humanity and good for us and good as we stare down this like big transition.

• 35:40Speaker 01

  I remember post nine eleven, I'm sure you remember exactly where, but I was in Southern New Jersey in our town, you know, whatever, dozens of people passed away. And the, how close the town came together after a terrorist attack happened. And it seemed so normal, like that it was just the normalcy of it. Or I have friends in Israel right now and you talk to them about it. And they're like, no, it's, it's normal. I'm like, whoa, there's a war going like, it's gotta be surreal. And they're like, well, I mean, what are you gonna do? You go about your day, you go get your food, all that. And it's, it's amazing, these psychological impacting things. At the end of the day, we need to go get food and we need to, you know,

• 36:24Speaker 01

  talk to our friends and all this stuff. So it is, it is amazing how much that can happen.

• 36:28Speaker 02

  It really, like genuinely, that's been my big surprising takeaway to like feel it so this really.

• 36:36Speaker 01

  As you think about like models becoming smarter and smarter, what, you kind of touched on this a little bit earlier with the, the creative element, like what do you think remains uniquely human as models start doing more and more capabilities of what we used to consider?

• 36:55Speaker 02

  I think many, many years from now, humans are still going to care about other, other humans. I, you know, I was reading the internet just a little bit and everyone's like, oh, everyone's gonna fall in love with ChatGPT now. And everybody's, you know, like it's gonna be the ChatGPT girlfriend, whatever, whatever. I bet not. I bet we're, I, I think we're so wired to care long-term about other humans in all sorts of like big and small ways that that's gonna remain. Like our obsession with other people, sounds like you hear a lot of conspiracy theories about me, you probably don't hear a lot of conspiracy theories about AI. You might not care if you did hear one. I think we're like not gonna watch robots play each other in soccer probably as our like main hobby.

• 37:35Speaker 01

  As you run OpenAI, the company itself, and you, you built a lot of rules or frameworks at Y C on how to run businesses. And then you've, you've broken a lot of them.

• 37:51Speaker 02

  Some.

• 37:52Speaker 01

  Are there are there different types of people you hire for this business than you would have had you started a consumer internet company within the executive ranks or a B2B software company or something?

• 38:04Speaker 02

  Researchers are very different than like product engineers for the most part.

• 38:10Speaker 01

  But Brad or Mira or some of the executives like researchers are unique, but does OpenAI bring in a different type of executive or do you hire for a different trait?

• 38:21Speaker 02

  So I mostly have not like I am a sometimes you hire externally for executives, but I'm a big believer that if like you generally promote it's not it's probably a mistake to only promote people to be executives because that could reinforce a monoculture and you know, I think you want to bring in some new very senior people. But we mostly like homegrown talent here, and I think that's a positive, given how different what we do is from what you would do somewhere else.

• 38:47Speaker 01

  Is there a decision that you've made over the course of OpenAI that that felt the most important at the time of making it? And how did you go about making it?

• 39:04Speaker 02

  It'd be hard to point to just a single one, but the decision that we're going to do what we call iterative deployment, that we're not going to go build AGI in secret and then put it out into the world at once, which was the prevailing wisdom and the LAs or plan and others. I think that was like a quite important decision we made, and it felt like a really important one at the time.

• 39:26Speaker 01

  If if another company that

• 39:27Speaker 02

  Betting on language models was an important decision and felt like an important one at the time.

• 39:31Speaker 01

  I actually don't know the story of betting on language models. How did that come to be originally?

• 39:34Speaker 02

  Well, we were we had these other projects. We were doing the robot thing and video games and there was a very small effort. It started with one person looking at Looking at language modeling and Ilya really believed in it. Really believed in like the general direction became language models, let's say. And did GPT one, we did GPT two, we started to study scaling laws, scaled GPT three and and then we made a big bet, this was what we were gonna do. And it was not, it looks so, all of these things look so obvious in retrospect, they really don't feel that way at the time.

• 40:13Speaker 01

  One other thing you brought up recently was the, there's two approaches to AI, the replication of yourself and then the smartest employee.

• 40:21Speaker 02

  Oh, it's not, not AI itself, but like how you wanna use it. Like when you imagine using your personal AI.

• 40:28Speaker 01

  So there's a subtle distinction when you said it, but, but can you, can you expound on it? Cause it seemed like a fairly profound distinction of how at least Sam thinks about the, the future of AI use cases. So can you explain that point again? Cause clearly I misunderstood it.

• 40:45Speaker 02

  If you're gonna text me in, you know, five years in the future, I think you wanna be clear of whether you're texting me or my AI assistant. And then if it's my AI assistant that's gonna like, you know, bundle messages together and you'll get a reply later, or, or, you know, if it can easily do something you might ask my human assistant to do, then fine, you'll know that. I think there will be value in keeping what those things are separate and not that it's like, all right, the AI is truly just an extension of Sam. I don't know if I'm talking to Sam or Sam's AI ghost, but that's okay, cause it's the same thing. It's this merged entity.

• 41:29Speaker 02

  I think, I think there will be like Sam and Sam's AI assistant. And also I want that for myself. Like I don't wanna feel like this thing is just like this weird extension of me, but that it's a separate entity that I can communicate with across a barrier.

• 41:41Speaker 01

  You see it in, in music or creative where it becomes pretty easy to replicate a Drake or a Taylor Swift audio. We probably need some form of validation or some centralization that validates, say, this is actually the, the creative work of XYZ person. You're probably gonna want some version of that at a personal level too.

• 42:01Speaker 02

  Yeah, but it's like, you know, the, the way I think about like OpenAI is it's, I don't like, there's different people and I'm asking them to do things and they go off, or they ask me to do things and I go off. But it's not a single like board. And I think that's like a way we're all comfortable.

• 42:18Speaker 01

  And so, so what is that? Can you tie that back? Like the decentralization of letting individuals do their?

• 42:26Speaker 02

  Not, well, also that, but I meant more just kind of like, what is the abstraction of what my personal AI is going to be like?

• 42:33Speaker 01

  Got it.

• 42:33Speaker 02

  Like, do I think of that as this is just me and it's gonna like,

• 42:39Speaker 02

  take over my computer and do what's best and because it's me, that's gonna be totally fine. And it's answering messages on my behalf and it's, you know, gonna just like, I'm slowly gonna like take my hands off the controls and it's slowly gonna like be me. Or do I think of this as like, this is a really great person I work with, that I can say, hey, can you do this thing? Get back and then you're done. But I think of it as not me.

• 43:02Speaker 01

  As you think about the educational system and as we think about like the class of, college class of 2030 or 2035 or whatever, whatever, some group in the future,

• 43:15Speaker 01

  are there changes specifically that you think should be made within the college educational system to prepare people for the future we have?

• 43:28Speaker 02

  The biggest one is I think people should not only be allowed, but required to use the tools. There will be some cases where we want people to do something the old fashioned way, because it helps with understanding. You know, like I remember sometimes in math class or whatever, there'd be something you can't use.

• 43:44Speaker 01

  No calculators on the test.

• 43:45Speaker 02

  Yeah, but on the whole, like in real life, you get to use the calculator. And so you need to understand it, but then you gotta be proficient using the calculator too. And if you did math class and never got to use the calculator, you would be like a less, less good at the work you need to do later. You know, if all of the open AI researchers never got to use a calculator, open AI probably wouldn't have happened. Computers at least, you know? We don't try to teach people not to use calculators, not to use computers. And I think we shouldn't train people not to use AI either. It's just gonna be an important part of like doing valuable work in the future.

• 44:21Speaker 01

  Last one. In Planning for AGI and Beyond, you wrote, "The first AGI will be just a point along the continuum of intelligence," which we spoke about earlier. "We think it's likely the progress will continue from there, possibly sustaining the rate of progress we've seen over the past decade for a long period of time." Do you ever personally stop and?

• 44:39Speaker 01

  Process or visualize like what the future will look like in that? Or is it just too abstract to contemplate?

• 44:50Speaker 02

  All the time. I mean, I don't visualize it like, you know, we have these like flying cars in a Star Wars future city and I like that, but like definitely what it means when one person can do the work of hundreds or thousands of well-coordinated people. And what, what it means when, I don't want to say we can discover all of science, but kind of what it feels like, like what it would feel to us like is if we could discover all of science.

• 45:17Speaker 01

  Be pretty cool.

• 45:18Speaker 02

  Yeah.

• 45:19Speaker 01

  Sam, thanks for doing this.

• 45:20Speaker 02

  Thank you.

• 45:21Jacob Effron?

  Pat, that was a, that was a pretty cool conversation.

• 45:23Speaker 03

  Yeah, super cool. It was amazing to, to get his perspective on, on a bunch of things, I thought. And it's also really good timing, you know, with the, the announcement that they had at the beginning of this week and GPT four.

• 45:36Jacob Effron?

  Oh, there, there is a lot of luck in the podcast game. I think Logan had been trying to get this one on the books for like a year and then it just happens to fall the day after a big release. Though I know he was bummed because I think Ilya left the next day. And so I guess, I guess in open AI world, there's never a bad day to have a podcast scheduled. But what, what do you think?

• 45:52Speaker 03

  It was awesome. I, he's, he gives super direct and simple answers, which I really appreciated. And I think a couple of things that stood out to me. One, and you saw this in their launches, but how they're kind of pushing the narrative that will interact with AI the way that we do humans. And I think that really is, um, you know, going to be the future and talking to AI like they are, um, you know, a helpful agent or, um, you know, assistant in, in your life. And I think it's, it's really cool to see the, the demos that they shipped where, you know, you have much more interesting, human-like conversations where you can ask it questions and follow-up questions and interrupt it and it can change its tone. And I think that really changes the experience.

• 46:37Speaker 03

  And we use Alexa a lot, but it's, you know, it's good for certain things, setting a timer. And I think this is really not incremental, more of a fundamental change in how we interact with technology.

• 46:47Jacob Effron?

  Totally. And I feel like ChatGPT. Was one of those things that like if you knew it, you knew it was good and people were using it, but it wasn't like the most widely used feature. And so I think it's going to be incredible now. You know, I feel like the Pandora's box has been unleashed for a lot of these, you know, multimodal applications of audio, video, and just kind of excited to see what folks built, right? Like I think obviously the OpenAI team dreamed up a bunch of really interesting ways folks can use them. But I think what's so cool is this is now, you know, available. I think one of the most interesting parts of it is it's available for free, right? I think that's obviously a really cool, you know, part that they were able to both ship all of these really interesting capabilities, but then also get the cost low enough or eat the cost to get it out there. And it really does feel like the battle, you know, if OpenAI has always had these two kind of parts of the business,

• 47:30Jacob Effron?

  the consumer part and the enterprise part, really feels like these announcements thoroughly need to the consumer side, right? How do you become kind of the go-to personal assistant and kind of homepage and way you navigate things for an individual? And, you know, I think that's obviously really, really powerful and interesting real estate.

• 47:45Speaker 03

  He was saying, we launched all of this multimodal stuff, but one of the biggest announcements was the efficiency gains and that they could offer this, this stuff for free. And so I think we'll, we'll continue to see that. Yeah, it was, it was funny. He, I really liked his quote. He was like, you should bet that intelligence as a service gets better and cheaper every year.

• 48:02Jacob Effron?

  Yeah.

• 48:02Speaker 03

  And I thought that was kind of a good summary. And I also liked his kind of the simplicity of his framework where he's like, efficiency gains will make this way cheaper and demand is going to go way up.

• 48:13Jacob Effron?

  Totally. I also like the way he talked about like the fact that you can interact with these things, you know, via audio as, you know, potentially a way to stay in flow with what you're doing and how, you know, obviously some, there's just different times you're going to want to interact with these tools in different ways. And if you're kind of in the middle of something and, you know, you can just have like almost an AI that's looking over your shoulder with a desktop app and see what you're doing and opine on it. Like I think we're still in such early innings of finding out like, how do you figure, how do you kind of have these AI assistants, but in, in the flow of what you're doing? And that obviously seems like it's been such a key part of the success of products like GitHub Copilot or, you know, basically just getting the latency down to, or, or in this case, that providing a modality or the context to these models in the case of the desktop app in a way that you could just be in the flow of what you're doing and then very quickly get something that's really relevant. And I feel like.

• 48:59Jacob Effron?

  You know, after these announcements, it's so fun because there's probably a million people that were inspired to go build products. And so you and I in the next three, six months, we'll just hear like an endless amount of pitches of folk building on top of this. And I think you're going to build really interesting things.

• 49:10Speaker 03

  It was cool. He said he likes to put his phone on his desk and have it listening as he's kind of working. And then you can ask questions and use that to stay in flow. It was really cool to kind of compare that with the announcement that they had on Monday where they have the chat GPT voice agent slash desktop, almost like it's a gazillion times better, but kind of a Clippy like experience. And that thing is super cool. I haven't tried it yet, but the demo was amazing where you can send it screenshots and say, hey, what's, what's happening in this chart? Or I'm trying, what's happening in this block of code? And I think it's, it's a really cool experience to have this AI that kind of follows you around through your workflow. Yeah, it'll be, it'll be cool to see how people use it.

• 49:55Jacob Effron?

  Anything else that stood out to you in that conversation?

• 49:57Speaker 03

  I thought it was cool how Logan asked about GPT five and what that would look like. And if it would be a huge step function improvement or more incremental. And Sam was basically like, we don't know. And I thought that was really just honest and interesting. And it kind of gets to the point, there was the guy at Google that published the paper on how the models perform at a small scale is not indicative of how they perform at large scale. And you really need to push the envelope to understand what they're going to do. It's like small scale can't do math, large scale can do math. And I think it's, it's a really interesting human endeavor that we know that GPT four and GPT four O are really incredible models.

• 50:41Speaker 03

  But we don't necessarily know what's out there as we continue to push the scale. And I appreciated his honesty around that. And you know, that this really is the, the frontier of exploration where we're going to make them bigger and we're going to see.

• 50:53Jacob Effron?

  I also have to love the like competitive juices of it like obviously, you know, it's impossible to tell if it's coincidence or not, but the fact that they keep like getting one day ahead of Google and and and kind of releasing stuff relatively last minute with these events. And then, you know, Google obviously had an incredibly impressive demo day as well, but a lot of the oxygen was, you know, in the room was taken out by by OpenAI. And so I think the competitive dynamics between the two are really fun. But I mean, Google has gotten its game back like they're they're shipping some really impressive stuff. And obviously, it's interesting to see, you know, a lot of a lot of in a lot of in the same areas what what OpenAI is working on. What else stood out to you from the from the episode?

• 51:29Speaker 03

  Yeah, his take on AGI was interesting. I think he said that it was only, you know, a couple of decisions away and and some tweaks and but he also mentioned that he he no longer believes in this binary, we have AGI, we don't moment, it'll be more of an exponential continuous curve. And I definitely fall much more in the exponential continuous curve camp. I think that it'll get better and better and better. And then over time, we're gonna we're gonna look back a year or two years and say, wow, like, you know, I can't believe how much better the models have gotten. So I thought that was interesting.

• 52:03Jacob Effron?

  Totally.

• 52:04Speaker 03

  What about you? Anything else that stood out to you on?

• 52:06Jacob Effron?

  Yeah, I think like one thing he said, which really resonated is, you know, in a world in which, I mean, AI is going to be able to do so much for us. Like, I think he kind of he went back against the idea of these like AI girlfriends, right? And basically saying that, like, you know, yes, AI will be able to do crazy hyper personalization and crazy, you know, interactions. But like, actually, you know, we always value scarcity and and other human relationships. And so I think in in, you know, in a world of increasing AI, those human relationships actually will become, you know, even more important, right? And we can get flooded and bombarded with all sorts of other stuff. You know, I think both from a personal standpoint, and then also even from a business standpoint, the ability to kind of, I think there'll be a million AI coaches, but I feel like that human coach will still be like this, you know, premium or or thing that really has

• 52:49Jacob Effron?

  extra resonance. And so, but it's always interesting thinking about like, you know, what will, you know, if our intelligence is no longer what defines us, what what is? And then, you know, I also I obviously love any shout out to robots, given the the rabbit hole we've been on in the robotics. Space and I think, you know, yeah, I think Sam's obviously, you know, opening it's interesting, right? Obviously started doing a lot of robotics work and has recently partnered with a bunch of robotics companies. And I think that space is fascinating. And, you know, obviously what they're doing for the digital world, you know, feels like there's going to be an influence in the physical world.