WEBVTT 00:00:00.000 --> 00:00:04.077 123123. We're going to do all that personal 00:00:04.160 --> 00:00:08.717 drama in 90 minutes. Oh yeah, oh don't you worry. Don't you worry. 00:00:08.800 --> 00:00:12.160 Is this part all going to be in the podcast? 00:00:12.880 --> 00:00:16.917 Bear with me through my slightly silly intro, but uh 00:00:17.000 --> 00:00:21.670 it's going to be okay. >> [music] 00:00:30.790 --> 00:00:32.437 [music] 00:00:32.520 --> 00:00:35.477 >> Welcome to the fortress of finance, the 00:00:35.560 --> 00:00:39.877 capital of capital. This is not that podcast. This is core 00:00:39.960 --> 00:00:43.837 memory. I am Ashley Vance. >> And I'm Kylie Robison. And I think we 00:00:43.920 --> 00:00:46.397 have a hell of an episode for you. 00:00:46.480 --> 00:00:50.837 Normally we do an introduction of people at this point, but 00:00:50.920 --> 00:00:55.117 that's perhaps unnecessary today. We have Sam Altman 00:00:55.200 --> 00:00:59.717 and Greg Brockman, the co-founders of OpenAI. You may have heard of it. Thank 00:00:59.800 --> 00:01:02.797 you guys for being here. Thank you for having us. 00:01:02.880 --> 00:01:04.517 >> [laughter] >> Thank you so much. 00:01:04.600 --> 00:01:07.837 I think this is the first time you guys have ever done a podcast together. That 00:01:07.920 --> 00:01:09.757 is amazing, but I think that's true. 00:01:09.840 --> 00:01:13.317 Certainly in a long time. In a long time. Maybe the first one. And I will 00:01:13.400 --> 00:01:18.437 not fight you for doing it on our show, but you did buy a podcast. Is there We 00:01:18.520 --> 00:01:20.197 just lucked out. You lucked out. 00:01:20.280 --> 00:01:24.317 >> Yeah. I'll take it. I'll take it. I was curious why you guys bought a podcast. 00:01:24.400 --> 00:01:26.757 It's not really You don't have to go deep on it, but did 00:01:26.840 --> 00:01:29.557 you have quick thoughts on that? 00:01:29.640 --> 00:01:34.277 I think that the people who who do TBPN are incredible. I think they're just 00:01:34.360 --> 00:01:38.157 very creative thinkers, and I think that in this world that we're moving to of 00:01:38.240 --> 00:01:41.557 building these AI systems that are so useful for people and helping people 00:01:41.640 --> 00:01:45.437 understand why why that's valuable for them in their personal lives and work 00:01:45.520 --> 00:01:48.397 lives. Like these are the kinds of people that I think could help tell that 00:01:48.480 --> 00:01:53.077 message. I've seen you on it. Have you been on TBPN? I have. Yeah, okay. Okay. 00:01:53.160 --> 00:01:55.557 Um I don't watch all the episodes, but It's 00:01:55.640 --> 00:02:00.037 a fun pod. Um well I thought, you know since this is you guys haven't done this 00:02:00.120 --> 00:02:04.037 together at least in a while, we would I was going to go down nostalgia lane just 00:02:04.120 --> 00:02:06.397 for a little bit at the beginning. 00:02:06.480 --> 00:02:09.157 Um, Kylie and I have gotten to know both of 00:02:09.240 --> 00:02:13.157 you over the years. You know, I was just I we were reflecting as we were 00:02:13.240 --> 00:02:16.437 preparing for this. We're a little bit past the 10-year 00:02:16.520 --> 00:02:20.597 anniversary. You guys are two of the remaining co-founders. I think Wojciech 00:02:20.680 --> 00:02:24.397 is the third. So you you know, you're this constant line that's been running 00:02:24.480 --> 00:02:27.557 through the company. >> company you started as an underdog. You 00:02:27.640 --> 00:02:30.397 ended up as the the top dog. It's all 00:02:30.480 --> 00:02:33.757 through a lot of drama, undulations. 00:02:33.840 --> 00:02:37.997 Um, I was I was just I was genuinely curious of 00:02:38.080 --> 00:02:42.677 how your relationship through all this has changed and how you 00:02:42.760 --> 00:02:45.997 guys have played off each other and if it's morphed over time. 00:02:46.080 --> 00:02:50.197 >> It is extremely nice in Look, we always wish there were less 00:02:50.280 --> 00:02:52.997 drama. We wish we just got to focus on the tech. 00:02:53.080 --> 00:02:55.237 But in a world of so much chaos and 00:02:55.320 --> 00:02:58.517 drama and you know, tension and fighting 00:02:58.600 --> 00:03:03.557 and power struggles, it has been unbelievably nice to have a relationship 00:03:03.640 --> 00:03:06.997 with someone that's got the full context. We have all this history and to 00:03:07.080 --> 00:03:11.477 really depend on each other in amazing times, very tough times. 00:03:11.560 --> 00:03:15.397 Uh, it has been one of the nicest things about OpenAI. You know, in many ways the 00:03:15.480 --> 00:03:16.717 very first 00:03:16.800 --> 00:03:19.397 moment of OpenAI was right after this 00:03:19.480 --> 00:03:23.837 dinner that we did in July of 2015 and 00:03:23.920 --> 00:03:26.517 Sam and I were driving back to the city together 00:03:26.600 --> 00:03:30.837 and we looked at each other and we were like, we have to do this. Right? There'd 00:03:30.920 --> 00:03:34.077 been all this conversation of is it too late to start a lab that could go after 00:03:34.160 --> 00:03:36.157 AGI and have have a positive impact. 00:03:36.240 --> 00:03:39.357 >> so ridiculous now that we were so worried about that. Yeah, it was too late. 00:03:39.440 --> 00:03:42.037 >> Yeah, that's the You missed it. You missed it. 00:03:42.120 --> 00:03:45.077 >> feeling like that when you guys started though. I was like, no. DeepMind's going 00:03:45.160 --> 00:03:46.477 to run away with this, you know, yeah. 00:03:46.560 --> 00:03:49.357 Yeah. And you know, the conclusion of dinner was it wasn't obviously 00:03:49.440 --> 00:03:52.157 impossible. And I think both of us just felt like, okay, this is just so 00:03:52.240 --> 00:03:54.277 important. We just have to do it. Yeah. 00:03:54.360 --> 00:03:58.277 And I think that spirit continues. And a lot of how we operate in the early days, 00:03:58.360 --> 00:04:01.517 like I remember I was unemployed at the time, so I was full-time on it the next 00:04:01.600 --> 00:04:05.157 day. Sam actually had a day job, um but we were constantly on the phone, like 00:04:05.240 --> 00:04:07.597 probably like five times a day. Yeah, exactly. 00:04:07.680 --> 00:04:11.637 >> Were you Were you guys already Were you like close friends already at that point 00:04:11.720 --> 00:04:15.797 or or not? We had known each other for a super long time. Or I felt like a super 00:04:15.880 --> 00:04:19.797 I don't actually Yeah. What What year was that? 2010? 2011? Whenever you 00:04:19.880 --> 00:04:22.877 started at Stripe. That's right. Yes. So we met through the Collisons. 00:04:22.960 --> 00:04:26.637 >> Yeah. And so we'd been kind of casual social friends. 00:04:26.720 --> 00:04:30.077 >> it wasn't as long as I thought. Yeah, maybe it was 2010 and this was now 2015. 00:04:30.160 --> 00:04:32.757 So five years. Time compresses. Yeah, 00:04:32.840 --> 00:04:37.797 and and and obviously I mean the being in the 00:04:37.880 --> 00:04:41.237 pressure cooker of all this doing this work, I mean I would imagine it only you 00:04:41.320 --> 00:04:43.997 guys have only got closer over time. 00:04:44.080 --> 00:04:47.477 Yeah, you you know, people use the word trauma bonding. I hate that. I've I've 00:04:47.560 --> 00:04:50.357 like I like the other things about like the people you're in the fox holes with 00:04:50.440 --> 00:04:55.157 the fox hole with, but the the one of the nice things about hard work, 00:04:55.240 --> 00:04:58.437 no matter what, but certainly hard work in um 00:04:58.520 --> 00:05:01.517 stressful times is you really like forge these relationships that 00:05:01.600 --> 00:05:03.277 I at least have not seen get formed any 00:05:03.360 --> 00:05:05.397 other way. Yeah. And I do I do think the 00:05:05.480 --> 00:05:08.597 way that Sam and I work and relate is 00:05:08.680 --> 00:05:12.597 maybe different from what you'd expect from a typical co-founder relationship. 00:05:12.680 --> 00:05:16.557 Like I think that we are just in constant contact, that five calls a day, 00:05:16.640 --> 00:05:19.597 two minutes, five minutes each, that kind of spirit remains. Like I think 00:05:19.680 --> 00:05:22.917 we're just in constant sync. And we don't always agree on everything, right? 00:05:23.000 --> 00:05:25.917 It's not like we come at the world from exactly the same point of view, but 00:05:26.000 --> 00:05:28.717 that's why we're so strong together, right? Is I 00:05:28.800 --> 00:05:32.477 think we have very complementary approaches, that Sam will say, "Here's 00:05:32.560 --> 00:05:36.157 an idea." I'll think about, "Well, maybe we could do this other way." Or what 00:05:36.240 --> 00:05:39.957 about if we approached it with this angle? Or how does this relate to this 00:05:40.040 --> 00:05:41.597 other thing that we're thinking about? 00:05:41.680 --> 00:05:45.717 And one thing I deeply appreciate about Sam is that I think he always sees these 00:05:45.800 --> 00:05:49.917 connections between different ideas, or just like keeps focused on here's the 00:05:50.000 --> 00:05:54.477 big picture that we need to get to and then together we figure out well how do 00:05:54.560 --> 00:05:59.277 we actually do it and I think connecting the grand ambition with the execution 00:05:59.360 --> 00:06:01.797 like that is what has always distinguished Open AI. Yeah, what are 00:06:01.880 --> 00:06:05.037 some of the points in these 10 years where you felt like it was really 00:06:05.120 --> 00:06:10.677 important that you guys diverged? Do you remember any key moments for you guys? 00:06:10.760 --> 00:06:13.637 I I think one of the things that Greg has done 00:06:13.720 --> 00:06:16.797 the best which is not my instinct is uh 00:06:16.880 --> 00:06:20.077 really just pushed to 00:06:20.160 --> 00:06:23.917 focus on the most important thing in his own work and also in what the company is 00:06:24.000 --> 00:06:26.357 going to do. Um so there have been times 00:06:26.440 --> 00:06:32.237 where I have wanted to do more things uh and Greg has just said, you know, is 00:06:32.320 --> 00:06:34.397 this the most important thing? Let's really just do this. Let's get the 00:06:34.480 --> 00:06:38.437 company focused. Uh and we've diverged on that and that's 00:06:38.520 --> 00:06:41.637 been like a very helpful spirit of Greg's throughout the company. 00:06:41.720 --> 00:06:46.197 Yeah, and I would also add I think even for example thinking about compute and 00:06:46.280 --> 00:06:48.477 just constantly raising the ambition. 00:06:48.560 --> 00:06:53.517 And sometimes I feel like okay, I kind of logically know that yes, like we're 00:06:53.600 --> 00:06:56.997 moving to this compute powered economy and yes, that demand is always going to 00:06:57.080 --> 00:07:00.757 outstrip supply, but like we've got all this like hard work to do and we already 00:07:00.840 --> 00:07:04.237 have all these big computers and we're operationalizing them and you know, you 00:07:04.320 --> 00:07:07.677 still have all of this like just physical infrastructure to build and you 00:07:07.760 --> 00:07:11.637 feel already swamped in it and swamped in it and Sam is like no, we need even 00:07:11.720 --> 00:07:14.357 more. And I think that that actually has 00:07:14.440 --> 00:07:18.837 been a very important thing to really not like sometimes it's easy to lose 00:07:18.920 --> 00:07:22.237 sight of the higher order bit of just the fact of this is going to be so 00:07:22.320 --> 00:07:25.757 important for not just the next 6 months, but this is 00:07:25.840 --> 00:07:29.397 what's important for the next 2 years, the next 5 years, 10 years and I think 00:07:29.480 --> 00:07:33.317 that the keep like you need this balance of sometimes 00:07:33.400 --> 00:07:36.917 swimming in the details, but you can't be swamped in the details and I think 00:07:37.000 --> 00:07:40.677 that that balance is something that I think again is what really has 00:07:40.760 --> 00:07:43.797 contributed to what Open AI is and where we are going to go. 00:07:43.880 --> 00:07:47.917 Is there There must be one product or 00:07:48.000 --> 00:07:51.357 strategy you guys have What's the one that you've disagreed about the most 00:07:51.440 --> 00:07:55.197 vehemently? >> I was just thinking when Greg was talking about this. This is not a 00:07:55.280 --> 00:07:58.637 a product, but it was this thing that came to mind is that I was going to say 00:07:58.720 --> 00:08:00.997 it before you asked that. Uh 00:08:01.080 --> 00:08:02.597 we used to talk a lot about how to talk 00:08:02.680 --> 00:08:06.477 about safety. We never disagreed on 00:08:06.560 --> 00:08:08.757 the extreme importance of safety 00:08:08.840 --> 00:08:11.637 and what it will mean to get this right 00:08:11.720 --> 00:08:15.917 or get this wrong, but the the field has had a strange relationship with how 00:08:16.000 --> 00:08:19.517 we've talked about safety, how we've used safety, and how much that becomes 00:08:19.600 --> 00:08:22.677 about power versus actually keeping things 00:08:22.760 --> 00:08:28.037 safe. And I I think earlier in our history, I got swept up more in the we 00:08:28.120 --> 00:08:31.157 got to really talk about this in a particular frame, 00:08:31.240 --> 00:08:35.757 and Greg was very disciplined about we're not going to fall into the 00:08:35.840 --> 00:08:39.437 traditional frame. We can't talk about that way. Now, even then, I think we 00:08:39.520 --> 00:08:43.757 have because this is so important, 00:08:43.840 --> 00:08:47.957 we I would ever have fallen into the trap of still talking more in the wrong 00:08:48.040 --> 00:08:49.837 frame than we should, but I think one of 00:08:49.920 --> 00:08:54.997 Open AI's greatest contributions to date has been 00:08:55.080 --> 00:08:58.197 finding a different way to talk about safety, not just in how we build the 00:08:58.280 --> 00:09:01.717 products, and how we talk about society needs to 00:09:01.800 --> 00:09:05.717 do, but like what how we deploy them, the idea whole idea 00:09:05.800 --> 00:09:12.640 of iterative deployment, and actually getting to 00:09:14.160 --> 00:09:17.077 maybe not the actually getting to a world where we're 00:09:17.160 --> 00:09:20.637 figuring out how to deploy products that get increasingly safe as the stakes go 00:09:20.720 --> 00:09:23.197 up, and Greg really held a line there 00:09:23.280 --> 00:09:25.357 that I think has been quite important to 00:09:25.440 --> 00:09:30.557 the company against extreme pressure not to do that. 00:09:30.640 --> 00:09:34.117 And I think it's been like quite quite important to our whole strategy, not 00:09:34.200 --> 00:09:37.157 just how we talk about things, but how we ship and build products. Yeah, and if 00:09:37.240 --> 00:09:41.157 you look at for example the OpenAI Foundation 00:09:41.240 --> 00:09:43.237 which is the nonprofit that governs 00:09:43.320 --> 00:09:46.677 OpenAI and has a very large chunk of of 00:09:46.760 --> 00:09:51.517 equity one of its pillars is AI resilience. And 00:09:51.600 --> 00:09:55.317 what that really means is thinking about how do we make AI be something positive 00:09:55.400 --> 00:09:59.437 for the world? And the answer is not any one intervention, right? It's not you 00:09:59.520 --> 00:10:04.357 have chain of thought monitoring and now you've achieved the mission. It's really 00:10:04.440 --> 00:10:09.197 a whole deep sequence of different ways that 00:10:09.280 --> 00:10:12.797 society should orient around this technology. And I think that this 00:10:12.880 --> 00:10:15.877 perspective of you're not going to solve 00:10:15.960 --> 00:10:18.637 AGI going well for the world in a paper. 00:10:18.720 --> 00:10:21.637 It has to be a worldwide effort from 00:10:21.720 --> 00:10:25.957 contributions from society, from many different people, from many different 00:10:26.040 --> 00:10:29.277 ways of really understanding what this technology is, how it will affect 00:10:29.360 --> 00:10:31.077 people, how it will affect the world. 00:10:31.160 --> 00:10:35.237 And this this is something that was not, I think, at all appreciated or 00:10:35.320 --> 00:10:38.877 understood when we were starting out 10 years ago because it's very easy just to 00:10:38.960 --> 00:10:41.757 fix it on, you know, we're technologists. We're, you know, building 00:10:41.840 --> 00:10:44.877 technology. That's the only problem we need to solve. I'm not saying anyone 00:10:44.960 --> 00:10:47.997 explicitly said it that way, but I think sometimes you can fall into a mental 00:10:48.080 --> 00:10:51.357 trap of thinking about it that way. And so a lot of what I think we have spent 00:10:51.440 --> 00:10:55.477 time on is being first principles thinkers and really thinking about how 00:10:55.560 --> 00:10:58.957 do you operationally deliver transformative technology to the 00:10:59.040 --> 00:11:03.117 world in a way that is going to actually help people in their daily lives. And 00:11:03.200 --> 00:11:06.637 one thing you realize is, well if you have a very powerful piece of technology 00:11:06.720 --> 00:11:10.717 that will change things, probably it's going to go better if you've had a less 00:11:10.800 --> 00:11:13.837 powerful piece of technology you've already helped change things in a 00:11:13.920 --> 00:11:17.837 positive way. And so if you just think about it that way, you start to really 00:11:17.920 --> 00:11:20.997 be pulled down this road of thinking about resilience, thinking about 00:11:21.080 --> 00:11:25.677 iterative deployment. And again, I think this is a lot of the dynamic within 00:11:25.760 --> 00:11:28.557 OpenAI with between the two of us that we're always thinking about these 00:11:28.640 --> 00:11:31.637 questions of how do we actually achieve this mission and make it go better. 00:11:31.720 --> 00:11:34.117 Yeah, I I feels like just yesterday I 00:11:34.200 --> 00:11:37.237 saw you at South by 2022 2023 and it was 00:11:37.320 --> 00:11:41.397 a completely different discussion about AI than what we hear today and it feels 00:11:41.480 --> 00:11:45.717 like that's something that's changed in 10 years of like how you discuss safety 00:11:45.800 --> 00:11:47.877 and alignment and I'm wondering how you 00:11:47.960 --> 00:11:52.877 reflect on that now like what would you have changed in those panels and those 00:11:52.960 --> 00:11:58.120 news hits and like what have you learned about talking about safety? 00:11:58.240 --> 00:12:02.557 Even before we get to safety, uh I think we have fallen in the frame as tech 00:12:02.640 --> 00:12:07.717 nerds of talking about we're going to build superintelligence and 00:12:07.800 --> 00:12:11.357 dot dot dot it's going to be great for you. And we've not filled in enough of 00:12:11.440 --> 00:12:15.077 the dot dot dot. Like we talk about we're building this amazing technology, 00:12:15.160 --> 00:12:19.277 it's going to do all these wonderful things and 00:12:19.360 --> 00:12:23.117 there is like a sense in the world now of okay 00:12:23.200 --> 00:12:28.317 looks like you were right, you are going to build this thing. Um why like 00:12:28.400 --> 00:12:32.597 why do we want that? What's that going to do for us? And the thing that a lot 00:12:32.680 --> 00:12:35.717 of the field has said of oh it's going to you know cure cancer and you'll be so 00:12:35.800 --> 00:12:38.517 happy or like that's clearly not quite resonating. A 00:12:38.600 --> 00:12:39.757 lot of people are like sure cure cancer 00:12:39.840 --> 00:12:42.837 that'd be wonderful. Um 00:12:42.920 --> 00:12:47.397 I think what people really want is prosperity 00:12:47.480 --> 00:12:52.237 agency that they're going to continue to have meaningful work to do. Um there I 00:12:52.320 --> 00:12:55.077 saw an incredible post the other day that really stuck with me which was like 00:12:55.160 --> 00:12:59.077 a a right to adversity. People people actually you want some 00:12:59.160 --> 00:13:01.357 challenges in life. You don't want every day to be perfect and everything done 00:13:01.440 --> 00:13:05.357 for you and there there's a fear with AI that like mm let's say you're right, 00:13:05.440 --> 00:13:07.717 let's say you build it, let's say it like makes all this money and does all 00:13:07.800 --> 00:13:09.357 the work and whatever like what do I do? 00:13:09.440 --> 00:13:11.757 What's my kid going to do? What's life going to be like? Where's the growth 00:13:11.840 --> 00:13:18.600 going to come from? What's the what are people going to strive for? Um and 00:13:19.000 --> 00:13:21.317 I think we've talked as a field and as 00:13:21.400 --> 00:13:26.197 OpenAI and as Greg and I we've talked a lot about the amazing technology and 00:13:26.280 --> 00:13:30.037 what it can do and what it like the technological marvel 00:13:30.120 --> 00:13:34.677 and we have not connected the dots enough on 00:13:34.760 --> 00:13:37.997 here's what the future's going to be like. Here 00:13:38.080 --> 00:13:42.717 is, you know, when when 00:13:42.800 --> 00:13:46.837 when I talk to parents with kind of school-age kids the most common question 00:13:46.920 --> 00:13:48.637 is like what what should my kids study? 00:13:48.720 --> 00:13:52.037 What's the future going to be like? What what will still have economic value and 00:13:52.120 --> 00:13:56.077 I I and I realize that that's not quite the question they're they're really trying 00:13:56.160 --> 00:13:58.877 to ask. It's like how is my kid going to have a fulfilling life in this new 00:13:58.960 --> 00:14:02.157 world? You know, you can answer it economically if you want, but 00:14:02.240 --> 00:14:04.877 that doesn't actually settle people and I think it's because there's there's a 00:14:04.960 --> 00:14:08.037 deeper thing here. I was just going to say and I think I 00:14:08.120 --> 00:14:12.477 think we actually have a lot of perspective on this and I think that 00:14:12.560 --> 00:14:14.477 for example 00:14:14.560 --> 00:14:17.517 within ChatGPT, we have so many people 00:14:17.600 --> 00:14:22.277 who say that my life or the life of a love loved one was saved through 00:14:22.360 --> 00:14:26.477 information I got through ChatGPT, right? Whether it's there's someone who 00:14:26.560 --> 00:14:28.677 had a kid who had 00:14:28.760 --> 00:14:33.637 blinding headaches. They were denied an MRI. They used ChatGPT to try to 00:14:33.720 --> 00:14:37.397 research the symptoms and used it to argue to be able to get insurance for 00:14:37.480 --> 00:14:40.477 the MRI. Turned out there was a brain tumor. They were able to intervene and 00:14:40.560 --> 00:14:43.557 save his life. And just navigating that 00:14:43.640 --> 00:14:47.597 experience that that family says, "We have no idea how we would do this 00:14:47.680 --> 00:14:53.157 without ChatGPT." And that's just one story. There are so many of those and I 00:14:53.240 --> 00:14:56.557 think people are really understanding that this technology can help not just 00:14:56.640 --> 00:15:01.157 abstractly society, but can help them, right? It can help them make money. 00:15:01.240 --> 00:15:04.157 Right now we're starting to see this wave of entrepreneurship. I think that's 00:15:04.240 --> 00:15:07.557 going to be a huge theme throughout this year. And I think we have a lot of 00:15:07.640 --> 00:15:11.557 perspective on well we're going to build these AIs that are 00:15:11.640 --> 00:15:15.237 able to make rather than you contort yourself to computer, right? You think 00:15:15.320 --> 00:15:18.597 about the way we do work. Yeah. It's not natural, right? It's not It's not kind 00:15:18.680 --> 00:15:21.557 of what we were designed to do. Um 00:15:21.640 --> 00:15:26.077 and instead the computer's going to do work for you. And the question of well, 00:15:26.160 --> 00:15:29.757 what is the work? What is good? Like, what is the thing that you 00:15:29.840 --> 00:15:33.957 actually want to be done? That's going to be something very deeply human, 00:15:34.040 --> 00:15:35.997 right? This agency, this empowerment. 00:15:36.080 --> 00:15:40.237 And so, I think there is this very optimistic, and not blindly optimistic, 00:15:40.320 --> 00:15:43.837 but I think this very positive change that this technology is bringing right 00:15:43.920 --> 00:15:48.357 now, but it's so much easier to just notice what's going to go away. Like, 00:15:48.440 --> 00:15:52.837 what is the thing that you thought was solid and stable that's going to change? 00:15:52.920 --> 00:15:55.237 But, it's much harder to see well, what's coming? Like, what's the new 00:15:55.320 --> 00:15:59.117 thing that you get? What is the the benefit that that comes as a part of 00:15:59.200 --> 00:16:01.957 this transformation? And I think that's something that we're increasingly 00:16:02.040 --> 00:16:05.637 realizing that we also have to really spell out in addition to spelling out 00:16:05.720 --> 00:16:09.277 the other sides of it. All right. What do we do at Core Memory? We cover 00:16:09.360 --> 00:16:11.477 innovative, fast-moving, 00:16:11.560 --> 00:16:16.837 forward-thinking companies, which is why Core Memory is sponsored by Brex, 00:16:16.920 --> 00:16:21.477 because Brex is the intelligent finance platform for 00:16:21.560 --> 00:16:25.837 many of these companies. 30,000 companies from startups to the world's 00:16:25.920 --> 00:16:30.157 largest corporations rely on Brex's technology for their 00:16:30.240 --> 00:16:32.477 finances. They've got smart corporate 00:16:32.560 --> 00:16:35.957 cards, high-yield business banking, and 00:16:36.040 --> 00:16:38.557 expense automation tools that are 00:16:38.640 --> 00:16:41.757 fantastic. I hate doing my expenses, and 00:16:41.840 --> 00:16:45.997 Brex's AIs and software run right through those expenses, figure out where 00:16:46.080 --> 00:16:48.317 we're spending money, and take care of 00:16:48.400 --> 00:16:50.157 so much stuff for you, so you don't have 00:16:50.240 --> 00:16:53.797 to waste your time on it yourself. Go to 00:16:53.880 --> 00:16:58.077 brex.com/corememory to learn more, and just, you know, get 00:16:58.160 --> 00:16:59.837 with the program. Let's get going. Let's 00:16:59.920 --> 00:17:02.877 get out of this archaic finance software 00:17:02.960 --> 00:17:07.677 and move toward the future. Core Memory and Brex. 00:17:07.760 --> 00:17:11.357 I have a couple questions based off what both of you were saying. Um 00:17:11.440 --> 00:17:17.117 So, I just went to Oberlin College with my son, and we were in this room 00:17:17.200 --> 00:17:21.677 um with I don't know, 600-700 people and it was fascinating to me. I mean it's in 00:17:21.760 --> 00:17:25.797 Ohio and people had come from all over the country and 00:17:25.880 --> 00:17:30.197 the president was up there giving a talk and then opened it up to Q&A and so many 00:17:30.280 --> 00:17:33.637 of the questions were about AI but they were 00:17:33.720 --> 00:17:36.797 it reinforced for me that 00:17:36.880 --> 00:17:41.597 how much of a bubble we live in here because they were what I would consider 00:17:41.680 --> 00:17:44.317 I mean it's not trying to like put anyone down or they they were pretty 00:17:44.400 --> 00:17:48.277 basic question I I just my my overwhelming sense of the room was like 00:17:48.360 --> 00:17:53.837 wow you guys kind of like don't know what's unfolding and about to to come 00:17:53.920 --> 00:17:58.277 across all of you and I don't know it was I don't know it was like alarming it was 00:17:58.360 --> 00:18:00.517 just it was eye-opening for me you know 00:18:00.600 --> 00:18:05.597 and and they were asking um They were obviously asking stuff about 00:18:05.680 --> 00:18:08.837 like well how are you going to use AI in class how are you going to stop kids 00:18:08.920 --> 00:18:12.437 from just relying on it but then it was like you know they kind of advanced a 00:18:12.520 --> 00:18:15.917 little bit beyond there but I felt like this room 00:18:16.000 --> 00:18:19.477 of pretty smart people were not not really in touch with what's going on. So 00:18:19.560 --> 00:18:24.237 I mean if this is a problem that people don't know these examples of of what 00:18:24.320 --> 00:18:26.157 could be done and we don't even know 00:18:26.240 --> 00:18:31.517 what shape this could take exactly you just I don't see how we solve this in 00:18:31.600 --> 00:18:35.237 sort of um kind of prepare people in some way. 00:18:35.320 --> 00:18:39.157 >> Well I I do have one sort of positive take on this which is 00:18:39.240 --> 00:18:42.517 that I think that if you just read about AI 00:18:42.600 --> 00:18:46.917 it gives you one impression right it's again it's like very it just feels like 00:18:47.000 --> 00:18:49.517 you're trying to wrap your mind around this new technology what is it but when 00:18:49.600 --> 00:18:53.517 you use it it's so intuitive. Right and that is the purpose of AI in 00:18:53.600 --> 00:18:56.477 many ways it's like you think about back to 00:18:56.560 --> 00:19:01.077 how we design computers for the past you know 70 years it's really about the 00:19:01.160 --> 00:19:04.437 machine doesn't quite get you right it's like you have these goals in your head 00:19:04.520 --> 00:19:07.757 you have to break it down into the machine's language whether that's 00:19:07.840 --> 00:19:11.397 writing assembly code or okay fine now we get higher level languages now you 00:19:11.480 --> 00:19:14.677 can kind of talk to your computer, but even if you think about how ChatGPT 00:19:14.760 --> 00:19:18.117 works, you still have to really understand concepts of language models, 00:19:18.200 --> 00:19:21.957 right? The fact like why do you have to create new conversations, right? New new 00:19:22.040 --> 00:19:25.277 new tabs. Like, why can't it just be a thing you talk to that remembers 00:19:25.360 --> 00:19:29.117 everything, right? Why there's there's these limitations that are techno 00:19:29.200 --> 00:19:32.157 technological limitations, but we're improving them. And so, I think that 00:19:32.240 --> 00:19:35.397 what we're building is the most intuitive technology. We are building 00:19:35.480 --> 00:19:39.237 something that can torch the machine to you. And when people use it, they 00:19:39.320 --> 00:19:42.997 realize, "Wow, this is what I can do with it." Like, some of the stories that 00:19:43.080 --> 00:19:47.637 I love hearing are someone in the Midwest who I you know, one one 00:19:47.720 --> 00:19:51.957 of my friends is who his sister was telling him 00:19:52.040 --> 00:19:55.437 about this app that she wanted to to see someone she wished someone had created. 00:19:55.520 --> 00:19:57.317 She described it in a detailed way. 00:19:57.400 --> 00:19:59.957 While he was listening, he was like, "Uh-huh. Uh-huh." typing into Codex 00:20:00.040 --> 00:20:03.077 exactly what she was saying. Pushed enter. 00:20:03.160 --> 00:20:05.677 Few hours later, he shows her the result. 00:20:05.760 --> 00:20:07.437 App that's exactly what she described. 00:20:07.520 --> 00:20:09.997 And she's like, "What the hell is this? 00:20:10.080 --> 00:20:14.157 Who built this? Like, this is exactly what I wanted." And he said, 00:20:14.240 --> 00:20:17.637 "You built it." Right? And that is the kind of aha 00:20:17.720 --> 00:20:21.677 moment that I think that everyone is going to go through. And once you're 00:20:21.760 --> 00:20:25.157 once you experience what AI can do for you and the fact that 00:20:25.240 --> 00:20:29.397 you are now empowered, like you have a image in your head that you want to see 00:20:29.480 --> 00:20:32.437 in the world. If you look at some of the upcoming 00:20:32.520 --> 00:20:35.757 some of the new image models we have, absolutely incredible. Like, you can 00:20:35.840 --> 00:20:38.637 create in a way that was never possible before. 00:20:38.720 --> 00:20:42.837 And the fact that if you are like like another story that I think 00:20:42.920 --> 00:20:45.277 about. So, my grandmother had 00:20:45.360 --> 00:20:48.837 dementia Alzheimer's, and that 00:20:48.920 --> 00:20:52.197 it was extremely tough for for everyone. 00:20:52.280 --> 00:20:54.877 But, one thing that really stuck out to me was she was actually able to use her 00:20:54.960 --> 00:20:59.117 Alexa to play music, and she could still remember 00:20:59.200 --> 00:21:01.477 the lyrics and sing along to songs. And 00:21:01.560 --> 00:21:02.997 it was just a way that she was connected 00:21:03.080 --> 00:21:07.757 to who she was through technology. And 00:21:07.840 --> 00:21:10.517 that really stuck stuck out to me. It's like if you can build interfaces that 00:21:10.600 --> 00:21:15.277 are intuitive that anyone can connect to them in ways that 00:21:15.360 --> 00:21:19.157 right now it feels like we we have pieces of technology that are targeted 00:21:19.240 --> 00:21:22.957 to specific individuals that you have to build all the skill that's not really 00:21:23.040 --> 00:21:26.637 related to deeply what you want, but it's just somehow specific to the tech 00:21:26.720 --> 00:21:30.837 itself. Three things came up for me during 00:21:30.920 --> 00:21:34.317 what Greg was saying, um which I I agree with all of. 00:21:34.400 --> 00:21:39.197 One before we launched ChatGPT, we used to try to talk to people 00:21:39.280 --> 00:21:40.757 and say, "This AI thing is coming. 00:21:40.840 --> 00:21:43.517 You've got to pay attention. It's going to change everything. It's really It's 00:21:43.600 --> 00:21:47.517 It's really important." And it got kind of no attention. We wrote these 00:21:47.600 --> 00:21:51.237 beautiful blog posts. We sort of did all these incredibly impressive 00:21:51.320 --> 00:21:54.757 feats. We won, you know, video game competitions. We had a robot that could 00:21:54.840 --> 00:21:58.717 do a Rubik's Cube with a hand. We like did all this amazing stuff. And we kind 00:21:58.800 --> 00:22:02.437 of felt pretty cool and we kind of said, "Oh, you know, 00:22:02.520 --> 00:22:05.277 the New York Times wrote about this. We must be doing great." But really truly 00:22:05.360 --> 00:22:06.437 no one cared. 00:22:06.520 --> 00:22:09.157 It got no actual impact. And then we 00:22:09.240 --> 00:22:12.757 launched ChatGPT, which was by far not 00:22:12.840 --> 00:22:16.557 the most impressive technological thing we had done. By far by far. 00:22:16.640 --> 00:22:20.157 And as soon as people could feel it, they're 00:22:20.240 --> 00:22:24.277 like, "Okay, I understand it." I think that was the moment probably most that 00:22:24.360 --> 00:22:27.917 the world has said maybe this AI thing is real. Like collectively all at once. 00:22:28.000 --> 00:22:30.317 Because people can use it and they can get value out of it. They develop their 00:22:30.400 --> 00:22:33.997 own sense. It's very different than hearing about it. That's happened again, 00:22:34.080 --> 00:22:38.717 as Greg said, with coding models. Um but those have really been the two kind of 00:22:38.800 --> 00:22:42.437 like large significant moments where I think 00:22:42.520 --> 00:22:44.677 the world has updated and said, "Okay, there's this thing happening." There 00:22:44.760 --> 00:22:49.597 will be more in the future. But so far the fact that you can 00:22:49.680 --> 00:22:54.917 ask a computer anything and get an answer or have a computer do anything anything 00:22:55.000 --> 00:22:58.757 with code for you. Um Tho- those have been really powerful and 00:22:58.840 --> 00:23:02.357 I I think that's how the world's going to update. Us saying superintelligence 00:23:02.440 --> 00:23:04.237 is coming is going to change everything. 00:23:04.320 --> 00:23:06.757 You know, maybe the people who listen to this podcast will sort of say, "All 00:23:06.840 --> 00:23:08.237 right, sounds sounds reasonable. 00:23:08.320 --> 00:23:09.837 Probably should pay attention to that." 00:23:09.920 --> 00:23:13.877 But it won't it won't have that big impact on the world. Uh 00:23:13.960 --> 00:23:17.157 So, I think the most important thing we can do to help that audience of people 00:23:17.240 --> 00:23:20.517 think not just about like what does this mean for people not cheating in class, 00:23:20.600 --> 00:23:24.237 but like what is this going to mean for the world is to ship great delightful 00:23:24.320 --> 00:23:27.877 products that create a lot of value and are easy for people to use. 00:23:27.960 --> 00:23:30.397 And we will continue to do that. 00:23:30.480 --> 00:23:35.397 The The second thing is we have seen again and again in our history that when 00:23:35.480 --> 00:23:39.717 we put something out that's pretty good, people say, "I can't really imagine what 00:23:39.800 --> 00:23:42.197 happens if this gets better. It can't get much better." 00:23:42.280 --> 00:23:44.917 This image model that Greg was mentioning, which we'll launch soon, was 00:23:45.000 --> 00:23:48.357 a real example of that for me. I kind of roughly thought image generation was 00:23:48.440 --> 00:23:51.157 solved. I was like, "Yeah, it's really good. Like, I don't need it to get any 00:23:51.240 --> 00:23:54.197 better. There's a lot of" And then this new thing has been a real reminder of 00:23:54.280 --> 00:23:59.117 like, "Wow, this can go so much further and there's so much more I can do." What 00:23:59.200 --> 00:24:02.717 does it do? It makes ridiculously great images. Uh 00:24:02.800 --> 00:24:06.477 and Has it solved text? Because I tried to make core memory merch mock-ups Try 00:24:06.560 --> 00:24:09.917 again very soon. >> [laughter] >> Uh the the team really did a great job at 00:24:10.000 --> 00:24:14.397 this one. But even with ChatGPT, like when we put GPT-4 in ChatGPT, I remember 00:24:14.480 --> 00:24:17.917 a lot of people, like sophisticated friends of mine, saying, "This is it. 00:24:18.000 --> 00:24:20.877 This is AGI. Like, if the model got smarter, I don't care. Just make it 00:24:20.960 --> 00:24:23.717 cheaper. Like, I can't This is amazing." 00:24:23.800 --> 00:24:27.677 Um and if you go back and use that whatever 00:24:27.760 --> 00:24:32.037 it was, uh I guess like March of 2023 version of 00:24:32.120 --> 00:24:35.437 GPT-4, you'll be like, "This was terrible." 00:24:35.520 --> 00:24:38.357 But at the time people were like, "It's solved. It's solved. I mean, it's beat 00:24:38.440 --> 00:24:40.797 the Turing test. It's done. It can't get any, you know." 00:24:40.880 --> 00:24:44.957 And and then it gets better and better and it like you can just 00:24:45.040 --> 00:24:48.637 keep raising the expectations and possibility of what you can do, to say 00:24:48.720 --> 00:24:51.517 nothing of then you're going to do reasoning models and do code. I'm just 00:24:51.600 --> 00:24:55.077 talking about how much better ChatGPT got in that time period. So, I think we 00:24:55.160 --> 00:24:59.117 see this again and again, too, where the world is like, "Okay, you made this 00:24:59.200 --> 00:25:03.117 amazing thing. That's as good as it's going to get. That's as good as I need." 00:25:03.200 --> 00:25:06.837 And then, month by month, or at least quarter by quarter, expectations and 00:25:06.920 --> 00:25:09.437 ability ratchet up in a huge way. 00:25:09.520 --> 00:25:13.840 And then, the third thing that Greg was saying there is 00:25:14.520 --> 00:25:18.317 these models are still quite dumb relative to what they will 00:25:18.400 --> 00:25:22.277 be, but more than that, they have quite limited awareness of your life. 00:25:22.360 --> 00:25:27.357 You are still having to like massage them and 00:25:27.440 --> 00:25:30.277 cajole them and try to get the thing that you want. 00:25:30.360 --> 00:25:33.237 Uh we are not no longer that far away 00:25:33.320 --> 00:25:37.997 from a model that just knows all of your context. It knows 00:25:38.080 --> 00:25:40.357 about you. It knows about your life. It knows what you're doing. It knows what 00:25:40.440 --> 00:25:43.157 you care about. It knows about the people in your life. It has access to 00:25:43.240 --> 00:25:46.437 your computer and your browser and, if you want, of course, in the ways you 00:25:46.520 --> 00:25:49.157 want, and it has access, maybe increasingly over time, to what's 00:25:49.240 --> 00:25:52.117 happening in in the real world around you. 00:25:52.200 --> 00:25:55.237 That is going to be a complete change to what 00:25:55.320 --> 00:25:59.437 it feels like to use a computer and what it feels like to use AI. 00:25:59.520 --> 00:26:04.757 And I am tremendously excited about that, but I don't think even we have a good 00:26:04.840 --> 00:26:07.717 intuition yet for what that's really going to feel like. Yeah. And to to that 00:26:07.800 --> 00:26:11.437 point, yeah, you think about how much time you spend right now just explaining 00:26:11.520 --> 00:26:15.397 to chat or whatever tool you're using what's going on. And if you think about 00:26:15.480 --> 00:26:18.477 how frustrating that is, like, if it's like you got a coworker you're 00:26:18.560 --> 00:26:21.157 constantly trying to explain to them, "No, this is kind of what I want. This 00:26:21.240 --> 00:26:24.377 is what's going on." I have projects I come back to and it's like that. 00:26:24.460 --> 00:26:26.197 >> [laughter] >> Exactly, right? Like, it's it's just not 00:26:26.280 --> 00:26:29.957 it's not really how you want to how you want these systems to behave. And and 00:26:30.040 --> 00:26:33.637 and just to say one more thing on on what Sam was saying is that I think that 00:26:33.720 --> 00:26:37.197 one of the biggest technical challenges we have at OpenAI is too much 00:26:37.280 --> 00:26:41.597 opportunity. Like, AI is this like field of boundless opportunity. No matter what 00:26:41.680 --> 00:26:44.717 dimension you expand on, there's going to be something new and unprecedented 00:26:44.800 --> 00:26:49.597 and amazing. And so, the important thing is having a vision for where can we 00:26:49.680 --> 00:26:51.717 really focus, where we're going to get 00:26:51.800 --> 00:26:56.917 the most returns and the most benefit from this you know having multiple 00:26:57.000 --> 00:27:00.357 different efforts that all add up to something. And I think that when it 00:27:00.440 --> 00:27:04.117 comes to we have these coding systems now, but I think it's going to clearly 00:27:04.200 --> 00:27:08.157 expand to just all computer work. It's and it's funny by the way that you know 00:27:08.240 --> 00:27:13.037 you think about just the degree to which you have all of this like the work that 00:27:13.120 --> 00:27:16.157 you do it's like very seamlessly interlinked between you're having 00:27:16.240 --> 00:27:19.997 in-person conversation, you're typing this on your computer and trying to 00:27:20.080 --> 00:27:22.117 really figure out how to get the context in there. It's going to be such an 00:27:22.200 --> 00:27:25.397 important problem. All these things matter, but then also in your personal 00:27:25.480 --> 00:27:29.637 life you really want this and we're starting to call where we want to go the 00:27:29.720 --> 00:27:34.277 personal AGI, right? This AI that really knows you has the context that you it 00:27:34.360 --> 00:27:37.437 you can trust it, right? That you ask it for you ask it questions related to 00:27:37.520 --> 00:27:40.997 finances just or health and it can give you trustworthy information. All these 00:27:41.080 --> 00:27:43.917 things are important and it needs that context as well. And so you start to 00:27:44.000 --> 00:27:48.357 really see that there's this blurry line between an AI that's used for this deep 00:27:48.440 --> 00:27:52.277 computer work and an AI that's used in your personal life at a technological 00:27:52.360 --> 00:27:55.757 level even if you want different systems cuz you want one that's really just 00:27:55.840 --> 00:27:59.077 knows your work context, one that knows your personal context. And so you can 00:27:59.160 --> 00:28:03.317 pursue those in parallel and build on the same 00:28:03.400 --> 00:28:06.837 technological foundations. And the thing that I find most amazing about the 00:28:06.920 --> 00:28:10.317 technology we're building on is it's all at the core 00:28:10.400 --> 00:28:13.917 one neural net, right? It's still deep learning. You're scaling it up. You're 00:28:14.000 --> 00:28:16.717 building one system and you're applying it to all of these different amazing 00:28:16.800 --> 00:28:20.477 applications. I I Can I ask a quick Uh I want to ask a 00:28:20.560 --> 00:28:23.757 question sort of along those lines. I mean I promised myself I wouldn't do 00:28:23.840 --> 00:28:27.997 this to you. Um but I mean I was kind of like I wasn't like a non-believer, but I 00:28:28.080 --> 00:28:30.557 like I could see AI doing incredible 00:28:30.640 --> 00:28:35.557 things, but I was I was kind of like skeptical until relatively recently on 00:28:35.640 --> 00:28:37.877 on >> What what updated you? Well, a couple 00:28:37.960 --> 00:28:42.077 things. I mean one was I started playing with agents a lot more and really 00:28:42.160 --> 00:28:46.557 shaping them to do what I wanted and kind of feeling what both of you were 00:28:46.640 --> 00:28:50.277 talking about. I was like, "Oh, this is actually saving me a lot of time. It's 00:28:50.360 --> 00:28:53.637 doing my bidding. It's doing it quite well." And the other is, you know, I 00:28:53.720 --> 00:28:58.437 cover biotech so much and just seeing some of the results that are 00:28:58.520 --> 00:29:01.997 coming out is is It strikes me that like coding and 00:29:02.080 --> 00:29:06.077 biotech seem like the most obvious places right now where it's really 00:29:06.160 --> 00:29:08.597 going. But then, you know, I was reading it I was talking to you the other day at 00:29:08.680 --> 00:29:12.437 dinner and reading some of the things that you've written about, 00:29:12.520 --> 00:29:16.717 you know, I see this clear path based off LLMs to going to what you 00:29:16.800 --> 00:29:21.197 describe. I still have this part of me though. I think it's because I deal in 00:29:21.280 --> 00:29:25.357 language so much where the writing is still 00:29:25.440 --> 00:29:27.997 not great where I 00:29:28.080 --> 00:29:32.877 This is nothing other than like anecdotal and intuitive. It's just It 00:29:32.960 --> 00:29:34.717 strikes me that 00:29:34.800 --> 00:29:37.197 the feeling I get back still is like, 00:29:37.280 --> 00:29:40.637 "No, this is not a super intelligence." 00:29:40.720 --> 00:29:43.357 We we are not there yet on personality. 00:29:43.440 --> 00:29:47.757 >> But like does LLM Right. But LLMs will get us there. 00:29:47.840 --> 00:29:51.317 I was just thinking of it as it's very jagged, right? And I mean, it's Okay, 00:29:51.400 --> 00:29:54.477 for example, just in the past couple days 00:29:54.560 --> 00:29:58.717 our AI solved this long-standing British problem. 00:29:58.800 --> 00:30:02.157 So, this mathematical mystery that's been, you know, of great interest for 00:30:02.240 --> 00:30:05.997 for a long time. A mathematician spent a lot of time working on this, many years 00:30:06.080 --> 00:30:10.157 thinking about this problem, posted about his his views of of what was what 00:30:10.240 --> 00:30:13.917 what the contribution was. Terence Tao also saying that this looks like there's 00:30:14.000 --> 00:30:18.037 like maybe connection between different fields of mathematics that this AI has 00:30:18.120 --> 00:30:21.997 discovered. And we're starting to see real beauty come out of these machines. 00:30:22.080 --> 00:30:23.477 Now, that's a particular domain, right? 00:30:23.560 --> 00:30:26.917 Mathematics is very different from creative writing. And I think that the 00:30:27.000 --> 00:30:30.877 fact of these AIs being able to have interesting insights and being able to 00:30:30.960 --> 00:30:35.277 help like it's you know, there's so much that these mathematicians are now 00:30:35.360 --> 00:30:39.717 saying, "Well, think about what more we can do." And so, I I think that we have 00:30:39.800 --> 00:30:43.357 a jagged frontier in terms of what these AIs are good at. We know how to keep 00:30:43.440 --> 00:30:46.797 pushing that frontier back, but I think that the the thing that we're really 00:30:46.880 --> 00:30:50.837 looking for, if you think about AlphaGo move 37, right, that was something that 00:30:50.920 --> 00:30:54.437 not just it was this deep insight, but it really changed people's understanding 00:30:54.520 --> 00:30:58.157 of the game of Go, and more people play Go now, right? So, that it really 00:30:58.240 --> 00:31:00.797 increased the meaning of what people were doing. And so, I think that what 00:31:00.880 --> 00:31:03.437 we're going to do is we're going to make something that is going to you're going 00:31:03.520 --> 00:31:06.797 to have fewer complaints about it, but I think you're also going to be just like 00:31:06.880 --> 00:31:11.197 way more able to do the thing that you want in writing than you can imagine 00:31:11.280 --> 00:31:15.117 today. I'm sorry to harp on this, but I was told GPT-5 was going to be really 00:31:15.200 --> 00:31:17.597 good at writing. And my like update when 00:31:17.680 --> 00:31:22.797 I became less cynical was reasoning models, and when Claude delivered like 00:31:22.880 --> 00:31:26.837 really good research, and then you guys had a research feature, that was for me 00:31:26.920 --> 00:31:28.677 like, "Oh my god, this is amazing." 00:31:28.760 --> 00:31:32.597 Saved me a lot of time. But, I was promised before that I would have good 00:31:32.680 --> 00:31:35.837 writing by now. >> The writing still there's no there's no 00:31:35.920 --> 00:31:39.397 soul, you know? There's there's something missing. Well, I'll I'll tell 00:31:39.480 --> 00:31:42.877 you. So, from a technical perspective, the 00:31:42.960 --> 00:31:46.597 technology we have is you can train a model 00:31:46.680 --> 00:31:50.237 in this unsupervised way, right? So, it really we look at all the publicly 00:31:50.320 --> 00:31:53.837 available data, and it learns to predict what comes next, and it's really about 00:31:53.920 --> 00:31:56.277 putting it in new situation, trying to figure out what's a reasonable thing to 00:31:56.360 --> 00:32:00.237 do. Then we do a reinforcement learning step, where it actually tries out ideas, 00:32:00.320 --> 00:32:03.677 and it gets rewards and punishments based on how well it did. You know, just 00:32:03.760 --> 00:32:07.277 positive signal, negative signal, not not You're not beating it. Exactly. Not 00:32:07.360 --> 00:32:12.717 at all. >> [laughter] >> Yeah, just just you know, this this signal. Um and then 00:32:12.800 --> 00:32:16.277 the hard part is, well, how do you judge? How do you decide if something 00:32:16.360 --> 00:32:21.557 was yep, thumbs up or thumbs down? And so, in math and science, much easier 00:32:21.640 --> 00:32:25.037 than in some of these more open-ended fields. But, we also have AI's that are 00:32:25.120 --> 00:32:28.397 getting much smarter and much more able to provide that kind of reward signal. 00:32:28.480 --> 00:32:32.277 And so, I think that the part of the challenge has been how do you expand the 00:32:32.360 --> 00:32:35.557 task the set of tasks that can be graded, and that's been a lot of the 00:32:35.640 --> 00:32:39.477 focus. And uh there's actually very interesting like a lot of this stuff 00:32:39.560 --> 00:32:42.197 dates back to the very beginning of Open AI. We kind of had a picture this is how 00:32:42.280 --> 00:32:44.877 it would go. But I would just say I think we're getting there. I think we 00:32:44.960 --> 00:32:48.197 definitely have a lot more progress to make. But yeah, hopefully you can keep 00:32:48.280 --> 00:32:51.957 giving us feedback and we'll we'll be able to improve it for you. 00:32:52.040 --> 00:32:56.557 One embedded challenge in that is that the the writing you want is very 00:32:56.640 --> 00:33:00.597 different than the writing that most people want. And right now we have to 00:33:00.680 --> 00:33:04.477 you know make a model that about a billion people use and all kind of kind 00:33:04.560 --> 00:33:08.517 of like. And what we'd like to do is to get the model so good at personalization 00:33:08.600 --> 00:33:10.797 that you think it's a great writer. 00:33:10.880 --> 00:33:14.237 And some other person that has a very different kind of life and set of needs 00:33:14.320 --> 00:33:17.317 than you also thinks it's a great writer for what they want. And those are very 00:33:17.400 --> 00:33:19.277 different things. >> I think that's what I'm kind of asking 00:33:19.360 --> 00:33:23.877 though. Like for the personal AGI you would want it to have that sparkle, 00:33:23.960 --> 00:33:26.117 that soul, that sort of that magic, 00:33:26.200 --> 00:33:29.517 right? And and I just I 00:33:29.600 --> 00:33:33.277 if it's not And I saw you know obviously Jan LeCun's 00:33:33.360 --> 00:33:36.557 been arguing against this for a long time. I I caught this Dennis thing very 00:33:36.640 --> 00:33:40.237 quick where he seemed to be kind of aligned with you guys but you know that 00:33:40.320 --> 00:33:43.677 LLMs will get us there but he seemed to be saying we still need a couple more 00:33:43.760 --> 00:33:46.277 ticks maybe of something special to to 00:33:46.360 --> 00:33:48.437 but you guys seem so confident. 00:33:48.520 --> 00:33:53.157 >> can solve we haven't tried to make it be you know a what Ashley Vance will think 00:33:53.240 --> 00:33:56.877 is a great writer. We have tried to make it you know solve open math problems 00:33:56.960 --> 00:34:00.157 that the smartest mathematicians in the world can't solve. And I don't want to 00:34:00.240 --> 00:34:03.557 say anything about your intelligence relative to these mathematicians but I I 00:34:03.640 --> 00:34:07.477 think >> Writers Writers are smarter But I was going to say like you know 00:34:07.560 --> 00:34:10.917 solving math is also hard. So if we can do that 00:34:11.000 --> 00:34:14.997 I'm pretty confident this approach can also learn what you think is great 00:34:15.080 --> 00:34:16.877 writing and figure out to do it for you. 00:34:16.960 --> 00:34:20.797 But but but by the way we also have some new new models coming on on this 00:34:20.880 --> 00:34:23.637 dimension as well. So after this podcast comes out 00:34:23.720 --> 00:34:27.277 >> wait to judge it. Let us know. Let us know if it looks better. writing I think 00:34:27.360 --> 00:34:30.477 personality and I think we're always improving like all the 00:34:30.560 --> 00:34:33.517 capabilities is moving up the jagged frontier. and so I think in this field 00:34:33.600 --> 00:34:37.557 it's always really important to judge the not the current position but what 00:34:37.640 --> 00:34:41.357 has that slope been and also fit it to an exponential. And so if you think 00:34:41.440 --> 00:34:44.557 about how does the writing stay compared to a year ago? Sorry if you're 00:34:44.640 --> 00:34:48.837 disappointed with 2.5. Don't worry we're we're we're very motivated to you know 00:34:48.920 --> 00:34:52.997 to to to really deliver. But I think that I think we really have a line of 00:34:53.080 --> 00:34:56.197 sight for how to improve it for for every application that people want. 00:34:56.280 --> 00:34:58.677 Since day one the core memory podcast 00:34:58.760 --> 00:35:01.197 has been supported by the fine people at 00:35:01.280 --> 00:35:03.357 E1 Ventures. They are a young and 00:35:03.440 --> 00:35:06.157 ambitious VC firm in Silicon Valley 00:35:06.240 --> 00:35:08.437 investing in young and ambitious 00:35:08.520 --> 00:35:11.077 companies and people. Thank you so much 00:35:11.160 --> 00:35:13.717 to E1 Ventures for all your support. In 00:35:13.800 --> 00:35:18.157 the world that you guys have outlined um 00:35:18.240 --> 00:35:22.557 this technology okay, you know, say everything is working great and this you 00:35:22.640 --> 00:35:26.037 know, this is uh it's a quite optimistic scenario 00:35:26.120 --> 00:35:30.077 um diseases are being cured um resources 00:35:30.160 --> 00:35:33.957 are becoming more abundant, problems are being solved. So so humanity as a whole 00:35:34.040 --> 00:35:38.357 is being lifted up. It's I still don't see I mean it seems like very very some 00:35:38.440 --> 00:35:42.117 of the world's smartest people are developing this technology. It strikes 00:35:42.200 --> 00:35:44.797 me it will disproportionately benefit 00:35:44.880 --> 00:35:46.477 them 00:35:46.560 --> 00:35:50.237 in almost all scenarios even though like 00:35:50.320 --> 00:35:53.717 everyone might get lifted up a bit that 00:35:53.800 --> 00:35:58.837 um you know, things really get more extreme because every time um 00:35:58.920 --> 00:36:01.517 what you guys are talking about about how you would manipulate these tools use 00:36:01.600 --> 00:36:05.797 these I I just feel like I feel like this 00:36:05.880 --> 00:36:09.037 things will go even more extreme and the have the haves will be 00:36:09.120 --> 00:36:12.517 >> the permanent underclass and that sentiment of like seeing these really 00:36:12.600 --> 00:36:16.037 powerful tools and feeling like oh my gosh, what what am I even here for then 00:36:16.120 --> 00:36:17.917 and feeling very disenfranchised. 00:36:18.000 --> 00:36:20.797 >> you get have all this time to do fun creative things. I just feel like other 00:36:20.880 --> 00:36:22.997 people will be manipulating this in such 00:36:23.080 --> 00:36:27.837 extreme fantastic ways that >> This This is at the core of the OpenAI 00:36:27.920 --> 00:36:31.517 mission. Like this is really why we started this place. 00:36:31.600 --> 00:36:34.597 Because you see this powerful technology coming. It's going to be the most 00:36:34.680 --> 00:36:39.157 important technology ever created. How do you ensure it benefits everyone? Like 00:36:39.240 --> 00:36:42.837 truly all of humanity. Like that is in our mission. Ensure AGI benefits all of 00:36:42.920 --> 00:36:47.077 humanity. And we really mean it. And if you look at our corporate structure, 00:36:47.160 --> 00:36:51.197 we've tried multiple different iterations at trying to in the structure 00:36:51.280 --> 00:36:54.397 codify some of our values of how do we ensure that this is something that 00:36:54.480 --> 00:36:58.797 really does benefit people. You see it in the product choices we make, right? 00:36:58.880 --> 00:37:03.077 We decided to launch ChatGPT because we really believe this is technology that 00:37:03.160 --> 00:37:06.477 we need to be able to put into people's hands. And that was very controversial 00:37:06.560 --> 00:37:09.477 by the way, that there was a different school of thought saying that the way to 00:37:09.560 --> 00:37:13.517 do it is you have to build it in secret, that you can't get people access, that 00:37:13.600 --> 00:37:17.757 you need to do it this other way. And I think that when we look at the how do we 00:37:17.840 --> 00:37:20.837 actually make society resilient, how do we actually benefit people, it all kind 00:37:20.920 --> 00:37:23.597 of points to the direction that we we have been going. And there's a lot of 00:37:23.680 --> 00:37:25.757 nuance to it. Um but I think that the 00:37:25.840 --> 00:37:30.757 way that we're we're headed is very much that I think the floor for everyone is 00:37:30.840 --> 00:37:33.597 not going to go up just a bit. Like I think we are going to head to this world 00:37:33.680 --> 00:37:36.197 where I mean you think about even having 00:37:36.280 --> 00:37:38.237 a doctor in your pocket that's better 00:37:38.320 --> 00:37:43.317 than the best medical team that anyone in the world could get today, 00:37:43.400 --> 00:37:47.677 accessible to anyone who has a smartphone. Like that is coming and that'll be free. 00:37:47.760 --> 00:37:50.957 Right? That is a wild wild fact. That's not a small It's going to be a little 00:37:51.040 --> 00:37:55.597 bit better. That is fundamentally raising the floor in this massive way. 00:37:55.680 --> 00:37:58.557 Now I think the question of also raising the ceiling, I think that's fair too, 00:37:58.640 --> 00:38:01.877 right? I think that the question of exactly how the distribution goes. And 00:38:01.960 --> 00:38:05.717 we think about this a lot. You look at the OpenAI Foundation. That's That has, 00:38:05.800 --> 00:38:09.397 you know, somewhere between, you know, 25%, 30%, somewhere around there of 00:38:09.480 --> 00:38:13.837 OpenAI equity, right? That's like more than 150 billion dollars. If it is the 00:38:13.920 --> 00:38:16.837 case that OpenAI becomes very successful, that all of that value 00:38:16.920 --> 00:38:20.357 locked up in the nonprofit, like that is something that will be for really 00:38:20.440 --> 00:38:23.877 benefiting the world and we think that's a good thing. Well, we were at dinner 00:38:23.960 --> 00:38:27.197 the other day, It's real quick, and I was listening to you talk about what you 00:38:27.280 --> 00:38:30.557 do with agents, and I've known you I mean, look, you're smart as 00:38:30.640 --> 00:38:35.317 and I was just like, "Jesus, man, Greg, with this 00:38:35.400 --> 00:38:38.877 machine I like even though I feel like I'm doing 00:38:38.960 --> 00:38:43.437 interesting things that help me out when I heard you describing your life and 00:38:43.520 --> 00:38:46.117 what you were doing, I was just like, "Oh my god, man. I mean, there's just 00:38:46.200 --> 00:38:49.157 going to be super people running around." 00:38:49.240 --> 00:38:51.877 I don't know, it's like kind of intimidating, really. I mean, it was 00:38:51.960 --> 00:38:55.317 cool, but it was also yeah, quite intimidating. I'm like, "I don't know 00:38:55.400 --> 00:38:58.677 how to use this like you do at all." Can 00:38:58.760 --> 00:39:03.760 I Can I try a less sanitized version of Yes, please. 00:39:04.000 --> 00:39:05.317 I hope I don't get in trouble for this. 00:39:05.400 --> 00:39:08.677 Uh I I can see three 00:39:08.760 --> 00:39:11.157 futures of the world. Um 00:39:11.240 --> 00:39:14.357 I can see one where, as Greg was saying, 00:39:14.440 --> 00:39:18.757 the floor comes way up. You know, everybody gets subjectively like 10 00:39:18.840 --> 00:39:21.277 times richer, just the sort of materially 00:39:21.360 --> 00:39:25.997 material abundance, the prosperity is crazy huge. People are like, "Man, 00:39:26.080 --> 00:39:30.597 relative to my life a decade ago, I'm doing great." But also in that world, 00:39:30.680 --> 00:39:33.677 these people who really learn how to use agents and get a lot of compute together 00:39:33.760 --> 00:39:36.237 and whatever, we have some trillionaires, you know, maybe 10 00:39:36.320 --> 00:39:41.157 trillionaires, whatever. So, like, the floor comes way up, 00:39:41.240 --> 00:39:45.877 like dramatically up, but because this is a lever that people can really use, 00:39:45.960 --> 00:39:49.317 uh the uh the the sort of like most capable, 00:39:49.400 --> 00:39:52.077 most ambitious that people who already started rich and have access to a lot of 00:39:52.160 --> 00:39:54.557 compute, inequality gets worse. 00:39:54.640 --> 00:39:56.957 So, that's that's one world I can see. 00:39:57.040 --> 00:40:01.597 Um I can see another world where we, through 00:40:01.680 --> 00:40:05.997 many different kinds of things that could happen, um the the floor doesn't 00:40:06.080 --> 00:40:09.437 come up as much, we don't generate as much total prosperity, but also there's 00:40:09.520 --> 00:40:13.437 like less inequality. So, maybe people in a decade feel like 00:40:13.520 --> 00:40:16.757 twice as rich, but inequality has come down. 00:40:16.840 --> 00:40:22.437 Um and actually, I'll just like stop and talk 00:40:22.520 --> 00:40:25.317 to those two because I think this is like the real crux of the issue. 00:40:25.400 --> 00:40:28.277 >> Is the third much scarier? >> No, but it it's like a distraction from 00:40:28.360 --> 00:40:32.837 the point I want to make. I think a lot of people in the world 00:40:32.920 --> 00:40:35.717 uh like we think it's obvious. Greg and I think it's obvious that people should 00:40:35.800 --> 00:40:39.557 prefer the first. I assume you also agree with that. 00:40:39.640 --> 00:40:43.277 Um but emotionally, it's not where a lot of people are. 00:40:43.360 --> 00:40:47.117 Wait. Well, I was going to say I don't I don't want to claim what what is obvious 00:40:47.200 --> 00:40:52.037 or not. I but I think it's I think for me personally, I just see so much 00:40:52.120 --> 00:40:56.037 potential in this technology and I think it's just so important from a societal 00:40:56.120 --> 00:40:59.437 level, just thinking about even American competitiveness. 00:40:59.520 --> 00:41:02.317 Like you look at robotics. Like I don't think that we are ahead in robotics at 00:41:02.400 --> 00:41:04.517 all. We are not. But you're not. 00:41:04.600 --> 00:41:08.757 With the software side, we do have a lead we do have this opportunity and I 00:41:08.840 --> 00:41:12.437 think that is something all of these factors together in my mind are are 00:41:12.520 --> 00:41:16.397 worth considering at once. Y- yeah. I mean, clearly I mean, 00:41:16.480 --> 00:41:18.837 actually, people who disagree on a lot of other things I think do agree that 00:41:18.920 --> 00:41:22.957 America is needs to be competitive on chips and robots and AI and everything 00:41:23.040 --> 00:41:26.677 else. But but there's going to be this huge question about how we organize 00:41:26.760 --> 00:41:28.957 society and the economy and do we push 00:41:29.040 --> 00:41:32.397 for maximum prosperity and accept the 00:41:32.480 --> 00:41:37.437 inequality that will come with that or do we say like, "Actually, we're going 00:41:37.520 --> 00:41:41.077 to constrain it because we're going to be more focused 00:41:41.160 --> 00:41:46.117 on the relative nature and the fear that well, if we don't do something, yeah, 00:41:46.200 --> 00:41:48.237 maybe the world will get more prosperous, but the people who are 00:41:48.320 --> 00:41:51.717 really good at using this are going to have all this power." And 00:41:51.800 --> 00:41:55.117 you know, emotionally, I think intellectually, it's kind of seems clear 00:41:55.200 --> 00:41:56.597 to me. Emotionally, 00:41:56.680 --> 00:41:58.357 I really get why it's not clear. And I 00:41:58.440 --> 00:42:02.757 really get the the fear that if we let 00:42:02.840 --> 00:42:05.757 inequality run here, the compounding 00:42:05.840 --> 00:42:10.637 nature of this tool is something that we don't yet understand. Now, the last 00:42:10.720 --> 00:42:13.437 thing I'll say on this is 00:42:13.520 --> 00:42:17.997 kind of no matter what I think everyone should want much more 00:42:18.080 --> 00:42:21.757 compute, much more infrastructure uh and the cheapest possible access to 00:42:21.840 --> 00:42:26.157 AI because otherwise I think you really exacerbate inequality if there's a 00:42:26.240 --> 00:42:29.157 limited amount of this and the price goes up because of supply and demand and 00:42:29.240 --> 00:42:32.397 only the rich people have it. Yeah, I want I wanted to to just build on that 00:42:32.480 --> 00:42:35.877 because I think that to some extent 00:42:35.960 --> 00:42:39.597 and this also plays back to to earlier just how Sam and I work together, how we 00:42:39.680 --> 00:42:42.757 think, right? Because I think that laying out these two options I think is 00:42:42.840 --> 00:42:46.077 actually really good point where it's like, well, are those the only two 00:42:46.160 --> 00:42:48.997 options, right? Is there something that we're not seeing in the space? And I 00:42:49.080 --> 00:42:52.477 think that the last thing that Sam said to me is an unlock in terms of thinking 00:42:52.560 --> 00:42:57.637 about so AI is really opportunity. AI is opportunity for everyone if you have 00:42:57.720 --> 00:43:00.877 access, right? If you have compute. If you don't have compute, you can't, 00:43:00.960 --> 00:43:03.557 right? No matter how good you are with agents, if you don't have the computer 00:43:03.640 --> 00:43:07.277 to run them, you're not going to be able to do much. And so I think a world where 00:43:07.360 --> 00:43:10.637 if everyone does have access to compute and 00:43:10.720 --> 00:43:13.717 you know, I think for for my generation growing up, we were much better at using 00:43:13.800 --> 00:43:17.637 computers than our parents, right? We grew up with it. We were, you know, 00:43:17.720 --> 00:43:22.157 native. And I think that the generation growing up now are going to be as good 00:43:22.240 --> 00:43:25.277 as I am at using agents, I think they're going to be 10 times better. It's wild 00:43:25.360 --> 00:43:28.677 to watch that. It is crazy. I mean, I wrote this is like the first massive 00:43:28.760 --> 00:43:33.317 story on core memory was that dude who built the nuclear fuser with with 00:43:33.400 --> 00:43:36.557 Claude. I just remember it was the the fuser didn't even sort of matter as much 00:43:36.640 --> 00:43:40.037 as like what I saw him doing with his computer. I'm like, you just use this 00:43:40.120 --> 00:43:41.637 thing different. And you know, he's like 00:43:41.720 --> 00:43:44.477 22 and he has eight AI applications up 00:43:44.560 --> 00:43:48.957 and Yeah. And if you can get compute to that kid, if you can get compute to 00:43:49.040 --> 00:43:53.357 every kid then it is really going to be the most 00:43:53.440 --> 00:43:57.037 extreme version of the American dream, right? Where it's really anyone who has 00:43:57.120 --> 00:44:01.157 the desire, the willpower to really play with this technology, to really lean in, 00:44:01.240 --> 00:44:04.477 to build with it, to get the most out of it, I think you're going to see this 00:44:04.560 --> 00:44:07.197 outperformance and I think you're going to see people have all sorts of 00:44:07.280 --> 00:44:11.677 mobility. And so, I think that to some extent the right way of answering this 00:44:11.760 --> 00:44:15.077 question of option one versus option two is to say 00:44:15.160 --> 00:44:19.197 Neither Exactly. And something else. And I think it's possible. I feel like we're 00:44:19.280 --> 00:44:22.517 doing some broad stuff and and we have a bunch of specific questions that I'm 00:44:22.600 --> 00:44:26.917 going to I'm going to be bad and ask one more broad one. It's okay. Just based on 00:44:27.000 --> 00:44:29.757 what we're talking I mean, you guys probably know, maybe you don't. You 00:44:29.840 --> 00:44:35.077 know, I just cover hardware so deeply and um I think I'm pretty 00:44:35.160 --> 00:44:37.877 I would venture to say I've probably been to more factories and hardware 00:44:37.960 --> 00:44:41.717 startups in the US than just about anyone and keep an eye on China. I mean, 00:44:41.800 --> 00:44:44.837 the thing that always plays through my head is that yeah, we're ahead on 00:44:44.920 --> 00:44:48.717 software. We're ahead on AI. Software's been this 00:44:48.800 --> 00:44:53.077 the story of the US for the last 40 or 50 years of our our strength. But, you 00:44:53.160 --> 00:44:55.757 know, when I think about how 00:44:55.840 --> 00:44:58.077 this technology manifests itself in the 00:44:58.160 --> 00:45:00.717 physical world, I do not see a future 00:45:00.800 --> 00:45:05.517 where the US is competitive in any way, shape, or form. Not just in robotics, 00:45:05.600 --> 00:45:08.957 but you know, just all of the componentry that goes in to these 00:45:09.040 --> 00:45:11.877 hardware systems. I go to El Segundo. 00:45:11.960 --> 00:45:15.677 I see 30 startups. It's great. There's like this flourishing of ideas and 00:45:15.760 --> 00:45:19.637 people trying things, but it's like such small potatoes compared to what I see in 00:45:19.720 --> 00:45:23.797 China. Um it just it just strikes me that 00:45:23.880 --> 00:45:28.317 there obviously will be a physical manifestation of all this in the world 00:45:28.400 --> 00:45:32.877 and and the US just seems like extraordinarily disadvantaged to to win 00:45:32.960 --> 00:45:36.037 that to me. Well, I think of people who are working 00:45:36.120 --> 00:45:37.477 incredibly hard to try to change that. 00:45:37.560 --> 00:45:42.397 I'd say Sam is Like on robotics and well, I know 00:45:42.480 --> 00:45:43.597 through through all the investments in 00:45:43.680 --> 00:45:49.720 different hardware companies or No, no, we're trying to figure out how 00:45:50.800 --> 00:45:53.597 Robots are an obvious part of this and we're trying to figure out how to be 00:45:53.680 --> 00:45:56.397 very successful at robotics. The I think 00:45:56.480 --> 00:46:00.637 if you could pick one thing to make the US competitive at manufacturing and the 00:46:00.720 --> 00:46:03.997 world of atoms in general, you would say we need a lot of robots that can build a 00:46:04.080 --> 00:46:06.277 lot lot more robots. But like we can't 00:46:06.360 --> 00:46:11.077 even make a actuator. You know, like going to figure that out. We we will 00:46:11.160 --> 00:46:16.157 Open AI. Okay. Like through your robotics? It would have to. We have to. 00:46:16.240 --> 00:46:19.477 Uh the But there's other parts of this that 00:46:19.560 --> 00:46:23.317 really matter. If you think about like the number of gigawatts of power that 00:46:23.400 --> 00:46:26.317 the world is going to need just to support AI workloads, the number of 00:46:26.400 --> 00:46:30.677 chips we're going to have to fabricate, the the boring stuff about like how 00:46:30.760 --> 00:46:34.117 we're going to just get enough racks assembled and enough network cables 00:46:34.200 --> 00:46:39.277 strung and made in the first place. Uh the US is extremely behind here. I think 00:46:39.360 --> 00:46:42.837 robotics will be the solution if you just look at how fast this has to happen 00:46:42.920 --> 00:46:46.117 and what the US is currently good at and how much infrastructure we have to build 00:46:46.200 --> 00:46:52.157 up and how long that would take to do by hand. Um but 00:46:52.240 --> 00:46:54.797 you are right in the diagnosis and the 00:46:54.880 --> 00:47:00.397 criticality of it and I think luckily we have like a new 00:47:00.480 --> 00:47:01.557 uh 00:47:01.640 --> 00:47:04.197 like a new piece on the chessboard. 00:47:04.280 --> 00:47:08.157 Through Open AI or the United States? 00:47:08.240 --> 00:47:10.517 >> Okay, tell me I mean what do you mean? 00:47:10.600 --> 00:47:14.837 I I I have a different feeling. I I I feel like we're play acting 00:47:14.920 --> 00:47:18.877 hardware and we're going to get completely If host. Well, on the current 00:47:18.960 --> 00:47:21.157 trajectory, yes. No, I we totally agree with that. 00:47:21.240 --> 00:47:24.677 Uh the If we can make true general purpose 00:47:24.760 --> 00:47:27.557 robots, if we can have a 00:47:27.640 --> 00:47:32.797 Codex equivalent power thing for robots that you can say like go configure a 00:47:32.880 --> 00:47:35.997 factory this way and make me more of this kind of robot or go figure out a 00:47:36.080 --> 00:47:38.397 mine and refine this kind of thing. 00:47:38.480 --> 00:47:42.797 >> then the playing field changes. Okay, but you said the US and not Open AI 00:47:42.880 --> 00:47:45.397 specifically. But yeah, I feel like you're pointing at something that Did 00:47:45.480 --> 00:47:49.597 you know though? Well, I don't think the US has a credible plan other than this 00:47:49.680 --> 00:47:53.037 kind of AI plus robotics to catch up fast enough. I didn't know what the 00:47:53.120 --> 00:47:55.917 chess piece was. Did something You saw something recently. 00:47:56.000 --> 00:47:59.357 >> Oh, I just I mean I meant general purpose artificial intelligence. 00:47:59.440 --> 00:48:02.157 >> I feel like there's there's a real chicken and egg that's existed in 00:48:02.240 --> 00:48:06.117 robotics where if you don't have the robotic hardware, 00:48:06.200 --> 00:48:08.557 it's kind of hard to develop the software. And if you don't have the 00:48:08.640 --> 00:48:11.437 great brain, it's kind of hard to be motivated to develop the hardware. And 00:48:11.520 --> 00:48:14.717 we really saw this. We had a robotics project back in 2018. Do you remember 00:48:14.800 --> 00:48:17.317 the robotic hand? Yeah, yeah. Super cool. 00:48:17.400 --> 00:48:19.477 The thing about that hand to know, so we trained it through reinforcement 00:48:19.560 --> 00:48:22.557 learning. Actually, the exact same algorithm we used to solve competitive 00:48:22.640 --> 00:48:26.277 video games, right? Wild. Single piece of technology doing both. The one 00:48:26.360 --> 00:48:30.637 difference is that this hand would run for 20 hours before it had these strings 00:48:30.720 --> 00:48:34.197 as its tendons, and they would snap. And then you'd have downtime, mechanical 00:48:34.280 --> 00:48:37.157 engineer would come in and fix it, you know, maybe be wake them up, have them 00:48:37.240 --> 00:48:40.717 come into the office. And you realize that you cannot 00:48:40.800 --> 00:48:44.677 do ML that way. You just cannot. So, actually, we ended up canceling that 00:48:44.760 --> 00:48:48.037 project. That team went on to go work on what became GitHub Copilot, you know, to 00:48:48.120 --> 00:48:51.037 contribute to that effort. And so, you just realize how much faster software 00:48:51.120 --> 00:48:54.917 world has been moving. But I think we're at a point now where we have these 00:48:55.000 --> 00:48:57.677 amazing general-purpose algorithms that can be used in all sorts of different 00:48:57.760 --> 00:49:02.077 ways. And we see how if you start to be able to apply those to physical world, 00:49:02.160 --> 00:49:05.877 then I think you are going to have a very different story for how the 00:49:05.960 --> 00:49:10.197 hardware development and how the hardware tuning to the co-design is 00:49:10.280 --> 00:49:13.477 going to happen. So, I think that there's like real potential for change, 00:49:13.560 --> 00:49:16.997 but we as a nation need to have the willpower to really do this. But I think 00:49:17.080 --> 00:49:20.877 we really agree with you that without something like that, the current 00:49:20.960 --> 00:49:25.077 trajectory looks terrible. We're so He's convinced that it's really bad. 00:49:25.160 --> 00:49:27.317 It's pretty bad. It's really bad. 00:49:27.400 --> 00:49:31.197 >> Taking a huge step back, you know, robotics, models, these are things that 00:49:31.280 --> 00:49:34.557 you guys are interested in building and really care about, but you guys have 00:49:34.640 --> 00:49:38.517 recently, it's been reported that you consolidated and focused, and I'm really 00:49:38.600 --> 00:49:41.997 curious, I'm sure like the audience wants to know, like, what's on the table 00:49:42.080 --> 00:49:45.797 now? What got cut? What do you guys care about? And why did you make those cuts? 00:49:45.880 --> 00:49:49.237 Why was it important? So, before I let Greg answer, before he does, Greg has 00:49:49.320 --> 00:49:53.477 taken over really figuring out what our cohesive product offering and the 00:49:53.560 --> 00:49:55.277 research to support that is going to be. 00:49:55.360 --> 00:49:59.117 Um and it's been amazingly uh joyful inside of the company. Like I it's going 00:49:59.200 --> 00:50:01.277 to take a little bit longer for all the stuff to ship. He's only been in the 00:50:01.360 --> 00:50:06.477 role for a few weeks, maybe, something like that. But uh it the the energy and 00:50:06.560 --> 00:50:09.597 excitement and the sort of like enthusiasm about what Greg is doing here 00:50:09.680 --> 00:50:12.557 is unbelievable. So, you can say what it's going to be. Dude, like can you 00:50:12.640 --> 00:50:16.397 expand on so like a few weeks ago you came in and because some of these cuts 00:50:16.480 --> 00:50:19.357 were already taking place, right? I mean, so but you've come in and like 00:50:19.440 --> 00:50:23.477 assessed everything. Uh so I've I've always been very involved behind the 00:50:23.560 --> 00:50:27.317 scenes with, you know, many parts of OpenAI and so I think that kind of 00:50:27.400 --> 00:50:29.957 taking a a foreground uh role is 00:50:30.040 --> 00:50:34.837 relatively recent uh in this particular area. Although, fun fact is I actually 00:50:34.920 --> 00:50:36.757 built the very first version of the API. 00:50:36.840 --> 00:50:40.357 So, like I've been doing products since product existed at OpenAI and always 00:50:40.440 --> 00:50:43.717 deeply cared about it. And there's a bunch of things I could say 00:50:43.800 --> 00:50:46.677 there. Like I think that how core it is to our mission uh was something we 00:50:46.760 --> 00:50:49.197 didn't appreciate before we started building products and afterwards you 00:50:49.280 --> 00:50:53.717 realize how important it is. So, that's why I've always been so close to it. 00:50:53.800 --> 00:50:57.517 So, the place that we're at now is that we are 00:50:57.600 --> 00:51:02.237 clearly at a moment of transition to agents. 00:51:02.320 --> 00:51:05.157 No question, right? People in software engineering you've been feeling this 00:51:05.240 --> 00:51:07.557 for, let's say, the past 6 months. 00:51:07.640 --> 00:51:11.477 And the you know, at the course of 2025 I think there was a transition from 00:51:11.560 --> 00:51:14.277 yeah, it's like kind of autocomplete to okay, yeah, you have kind of a sidebar 00:51:14.360 --> 00:51:17.477 in your editor and then you'll start talking mostly over there, but you're 00:51:17.560 --> 00:51:20.437 still mostly doing the same kind of software development you were doing 00:51:20.520 --> 00:51:24.957 before to now it's like actually you want a tool like Codex that is really an 00:51:25.040 --> 00:51:27.757 agent management platform. And the agents are going to take care of all the 00:51:27.840 --> 00:51:31.277 details. The agents are going to do all the the nitty-gritty work. And there's 00:51:31.360 --> 00:51:35.237 still probably 20% of how you put things together and the the way that you will 00:51:35.320 --> 00:51:39.237 structure your code and exactly some of these these higher-level things that you 00:51:39.320 --> 00:51:41.877 would normally put in an architecture document that the human still really 00:51:41.960 --> 00:51:45.837 cares about and wants to manage, but the details of exactly the code, like nope, 00:51:45.920 --> 00:51:50.757 that's what agents are for. And so, the question that we had is, first of all, 00:51:50.840 --> 00:51:52.277 how do we really rise to this moment? 00:51:52.360 --> 00:51:54.037 Cuz it's not just software, right? We 00:51:54.120 --> 00:51:56.357 see line of sight for every single 00:51:56.440 --> 00:52:01.077 vertical, right? For law, for finance. 00:52:01.160 --> 00:52:03.797 Some of the mechanical skills there of writing, you know, creating spreadsheets 00:52:03.880 --> 00:52:06.917 and presentations. How do we make our models extremely good at those, right? 00:52:07.000 --> 00:52:10.317 It's like you work with domain experts, you produce evaluations, you produce 00:52:10.400 --> 00:52:14.837 training data, you have the AI actually take its great domain knowledge and 00:52:14.920 --> 00:52:18.117 apply it in these verticals and gets experience and get those, you know, yep, 00:52:18.200 --> 00:52:19.677 you did a good job in order to figure 00:52:19.760 --> 00:52:22.557 out what what good looks like. So, we 00:52:22.640 --> 00:52:27.077 have the mechanical the the vision of exactly how to do this, but we need to 00:52:27.160 --> 00:52:31.277 make sure that we're building the right product surface to unlock all of it. And 00:52:31.360 --> 00:52:35.077 one thing that we have found is that the models have shifted from being the 00:52:35.160 --> 00:52:38.837 product to being a part of the product, right? They used to have these very thin 00:52:38.920 --> 00:52:41.757 layers of software on top of them, and you didn't have to like think that hard 00:52:41.840 --> 00:52:45.317 about how it was architected. But now, it's a very fat layer, right? That you 00:52:45.400 --> 00:52:49.917 have things like skills, connectors, you have exactly how you hook up to computer 00:52:50.000 --> 00:52:53.237 use, how you manage context and memory, and all of these things. And so, there's 00:52:53.320 --> 00:52:57.277 just this deep layer of software, which is almost you could think of it as the 00:52:57.360 --> 00:53:00.157 AI. It's kind of like we have this brain in the form of the model, and now we're 00:53:00.240 --> 00:53:03.397 building the body. Both are hard. They have to be co-designed together. And so, 00:53:03.480 --> 00:53:06.877 a lot of what we're focusing on is, number one, getting together an amazing 00:53:06.960 --> 00:53:10.837 agentic platform. Like, that is the number one focus that we are delivering 00:53:10.920 --> 00:53:14.397 on. We have teams that are executing extremely well on this. I'm super 00:53:14.480 --> 00:53:18.637 excited about what we'll be releasing over upcoming weeks. Um the second is, 00:53:18.720 --> 00:53:21.677 where do you actually want to apply these agents? And that our priority 00:53:21.760 --> 00:53:24.677 there is really towards computer work, right? So, that And I I use that term 00:53:24.760 --> 00:53:27.677 very deliberately, by the way. People like to talk about knowledge work, but 00:53:27.760 --> 00:53:29.877 no one thinks of themselves as a knowledge worker, right? Like, that's 00:53:29.960 --> 00:53:33.357 not a thing that people do. It's kind of a term that's like almost removed from 00:53:33.440 --> 00:53:36.757 what the actual thing is. But the thing I like about computer work is it's like, 00:53:36.840 --> 00:53:37.997 I don't really want to do computer work. 00:53:38.080 --> 00:53:40.237 That doesn't sound like the thing that I want. But you realize how much of your 00:53:40.320 --> 00:53:44.637 time you do spend doing it, like chain behind your desk, you know, typing away, 00:53:44.720 --> 00:53:47.797 hunching over your shoulders, getting your carpal tunnel, all of those things. 00:53:47.880 --> 00:53:52.997 And so, we are focusing on that, bringing Codex that exists today, not just for 00:53:53.080 --> 00:53:56.397 software engineers, but really making Codex be for everyone. And that's 00:53:56.480 --> 00:53:57.957 something that's coming very quickly. 00:53:58.040 --> 00:54:02.477 We'll have some updates even coming today as of this podcast filming. And 00:54:02.560 --> 00:54:04.717 there's a lot of exciting things that are still in the pipeline for that 00:54:04.800 --> 00:54:08.597 direction. And then the third thing is really thinking about personal AGI, 00:54:08.680 --> 00:54:13.237 which is about Thinking about ChatGPT right now, used by a billion users. And 00:54:13.320 --> 00:54:16.957 every single person on the planet is going to want an AI that represents 00:54:17.040 --> 00:54:21.277 them, has their context, that they have built trust with, that it's not just 00:54:21.360 --> 00:54:24.437 something that you build, or it's not just something that 00:54:24.520 --> 00:54:27.317 you talk to one-on-one, but it can be out there doing things for you. For 00:54:27.400 --> 00:54:30.557 example, maybe it knows that you like a specific musician, and that musician's 00:54:30.640 --> 00:54:33.957 in town, and it knows it proactively, and tickets have just become available. 00:54:34.040 --> 00:54:36.797 There's some great tickets that are very cheap that you you get for you. It just 00:54:36.880 --> 00:54:39.997 goes and buys them, right? And maybe it knows it's built trust with you, so it 00:54:40.080 --> 00:54:42.557 kind of knows you're allowed to to do this without getting approval, or maybe 00:54:42.640 --> 00:54:46.837 it realizes, "I should probably I'm not quite sure. I should check." And so, 00:54:46.920 --> 00:54:50.997 we're building that as well. And if you think about these things, 00:54:51.080 --> 00:54:54.557 they all are kind of expressions of something that fits together into 00:54:54.640 --> 00:54:58.677 cohesive whole. Like in the end, you fast forward to where we're going, you 00:54:58.760 --> 00:55:02.037 really just want an AGI, right? You don't want a language 00:55:02.120 --> 00:55:05.397 model, you don't want threads, you don't want any of these like details. You just 00:55:05.480 --> 00:55:09.157 want something that is helping you, that is operating on your behalf, that is 00:55:09.240 --> 00:55:11.997 able to help you solve problems, that knows what your goals are and achieve 00:55:12.080 --> 00:55:14.717 those in work context, personal context. 00:55:14.800 --> 00:55:18.877 And so, this is what we are prioritizing and building. Yeah. And so, you know, to 00:55:18.960 --> 00:55:21.957 me one of the important questions, you know, people are asking, "How do we 00:55:22.040 --> 00:55:25.717 think about consumer? How do we think about enterprise?" And the answer is, 00:55:25.800 --> 00:55:29.277 if you take the definitions of these words as they exist today, we care a lot 00:55:29.360 --> 00:55:32.437 about consumer, we care a lot about enterprise. But I think that the meaning 00:55:32.520 --> 00:55:37.077 of these words will change and blur, because what we're doing is we are going 00:55:37.160 --> 00:55:40.237 to So, we're going to unlock this wave of entrepreneurship. Again, I I think 00:55:40.320 --> 00:55:41.757 we're seeing the leading edges of it. 00:55:41.840 --> 00:55:44.597 Small companies be able to get tons and tons of revenue that was not possible 00:55:44.680 --> 00:55:47.317 before. And that's been a trend for a while. 00:55:47.400 --> 00:55:50.517 It's just going to really accelerate. Is that enterprise? 00:55:50.600 --> 00:55:54.837 Is that consumer? It's kind of neither, right? And so, I think we are really 00:55:54.920 --> 00:55:59.237 focused on solving goals across all context. And that is the lens that we 00:55:59.320 --> 00:56:03.717 that we look at things through. And so, that has mean that has that has meant we 00:56:03.800 --> 00:56:07.477 need to deprioritize other things that are also amazing on their own. 00:56:07.560 --> 00:56:10.757 >> So, what got cut? >> Yeah. >> Well, So, Sora is the most Sora is the 00:56:10.840 --> 00:56:14.037 most obvious one. >> And why? Well, because it's a different 00:56:14.120 --> 00:56:16.957 branch of the tech tree, right? So, if you look at the models that actually 00:56:17.040 --> 00:56:19.077 power Sora, that they're not unified 00:56:19.160 --> 00:56:22.037 with the core GPT series. And secondly, 00:56:22.120 --> 00:56:25.277 the use case is not quite unified 00:56:25.360 --> 00:56:29.397 either, right? It doesn't fall as far under this goal like there's creative 00:56:29.480 --> 00:56:32.357 expression. There's something very important there. I think it was like an 00:56:32.440 --> 00:56:34.917 incredible model. The team does incredible work. And I think that 00:56:35.000 --> 00:56:38.757 technology will live on for other applications. But what we were really 00:56:38.840 --> 00:56:42.877 focusing on was this the product suite that we want to be delivering. What do 00:56:42.960 --> 00:56:46.637 we want to be doing over the next 3 months, 6 months, 12 months? And the 00:56:46.720 --> 00:56:50.557 thing I described, by the way, is just step one. Because we also see line of 00:56:50.640 --> 00:56:52.597 sight to much more powerful models. 00:56:52.680 --> 00:56:55.797 Like, you look at what we're doing right now in mathematics. It's Okay, it's 00:56:55.880 --> 00:57:00.157 actually kind of mind-blowing. Where this result that I just mentioned of 00:57:00.240 --> 00:57:04.117 solving the new Erdős problem that seems actually really significant. 00:57:04.200 --> 00:57:07.157 That was just someone using GPT-4 Pro. 00:57:07.240 --> 00:57:09.997 Like, 2 years ago >> Yeah, we used to like train our model We 00:57:10.080 --> 00:57:12.157 had a team of 20 people to try to train 00:57:12.240 --> 00:57:17.117 our models to go solve a computing Olympiad, and we got a bronze medal. A 00:57:17.200 --> 00:57:20.037 team of 20 people for like 2 weeks and lots of compute. And now it's just this 00:57:20.120 --> 00:57:21.797 model that we trained very casually. 00:57:21.880 --> 00:57:24.837 Someone is able to just point it at problems and get this kind of result. 00:57:24.920 --> 00:57:28.917 What if you point that at drug discovery? What if you do take that team 00:57:29.000 --> 00:57:32.237 of 20 people and all that compute and And really try to push it for scientific 00:57:32.320 --> 00:57:36.077 discovery. And that's something no one is pricing in right now. So, I think 00:57:36.160 --> 00:57:40.037 it's rising to the moment of agents really trying to make sure that the 00:57:40.120 --> 00:57:44.557 product investments we're making are sort of well-structured, that we think 00:57:44.640 --> 00:57:47.517 about how all these pieces fit together, that we have connectors that work really 00:57:47.600 --> 00:57:50.957 well, and each of these pieces can be composed, and really also build an 00:57:51.040 --> 00:57:54.477 ecosystem because it's not just about what we built, right? That we want to 00:57:54.560 --> 00:57:58.557 build some example agents. You think of Codex almost as an example agent, but it 00:57:58.640 --> 00:58:01.917 should be that if you're a developer, if you're someone with a creative idea, you 00:58:02.000 --> 00:58:04.997 can build your own agent, right? That you can build it for your application, 00:58:05.080 --> 00:58:07.837 for your purpose. You care about the specific math problem, you should be 00:58:07.920 --> 00:58:10.477 able to apply the agent to your that math problem. And so, we're we're 00:58:10.560 --> 00:58:14.557 enabling all of that. But you I mean, okay, just bear with me for 1 second 00:58:14.640 --> 00:58:18.117 just to go quick through a couple things around this. I mean, so we I mean, part 00:58:18.200 --> 00:58:22.317 of this was like you had to probably cut Sora for compute to get 00:58:22.400 --> 00:58:24.517 compute, right? So, that compute. 00:58:24.600 --> 00:58:27.997 >> reported that it was taking up a lot of compute. I mean, every everything in 00:58:28.080 --> 00:58:31.197 this field that is successful is going to take a compute. I mean, the two of 00:58:31.280 --> 00:58:34.757 you in particular, and I think Ilya and some others, you know, kind of famous 00:58:34.840 --> 00:58:39.077 for going all in on where you were going to direct the compute in the early days, 00:58:39.160 --> 00:58:43.077 and now you have to make these very difficult decisions. You have to serve 00:58:43.160 --> 00:58:46.157 all these customers that you have. You got to make some money cuz you're 00:58:46.240 --> 00:58:48.477 spending a lot of money, and you know, 00:58:48.560 --> 00:58:51.797 so you two are both wired to take the 00:58:51.880 --> 00:58:54.637 very biggest bet possible, but you're 00:58:54.720 --> 00:58:59.677 constrained now by the business into some degree. So, 00:58:59.760 --> 00:59:02.157 that just seems very difficult. I feel like it's not in 00:59:02.240 --> 00:59:05.037 your nature. Does it You're looking skeptical. 00:59:05.120 --> 00:59:09.037 It's a It's a strange phrasing because I don't feel constrained by the business. 00:59:09.120 --> 00:59:12.837 I feel enabled by the business because the business is what has really allowed 00:59:12.920 --> 00:59:17.797 us to say we can scale compute, right? I remember when we launched ChatGPT, and I 00:59:17.880 --> 00:59:20.037 remember we were talking right afterwards, and we're trying to figure 00:59:20.120 --> 00:59:24.037 out how much compute to buy. And I was just like, we got to buy it all. 00:59:24.120 --> 00:59:27.157 We just got to we got to do it cuz it's just so clear there's so much demand. 00:59:27.240 --> 00:59:30.797 And think that that has been a huge unlock in our ability to get lots of 00:59:30.880 --> 00:59:34.717 compute for mission critical >> this incredible revenue machine, we 00:59:34.800 --> 00:59:38.597 would not be able to convince anyone that we should get all this compute. 00:59:38.680 --> 00:59:41.597 >> Get get this. But then the impression obviously when you see the stories about 00:59:41.680 --> 00:59:45.357 Stargate and things like that is that you guys are somehow pulling back on 00:59:45.440 --> 00:59:48.877 infrastructure. I don't know where that's coming from. Like there you know, 00:59:48.960 --> 00:59:51.317 there will be like a site here and there we say okay, you know what this 00:59:51.400 --> 00:59:54.197 particular site maybe it only has air cooling and so this is not as valuable 00:59:54.280 --> 00:59:57.677 to us as this other site and that's like there's specifics and people really want 00:59:57.760 --> 00:59:59.477 to write the story of like pulling back. 00:59:59.560 --> 01:00:02.797 But very soon it will be again like open AI is so reckless. How can they be 01:00:02.880 --> 01:00:06.477 spending this crazy amount? So like the media will go 01:00:06.560 --> 01:00:09.477 will flip out either way just cuz you all need something to write about I 01:00:09.560 --> 01:00:14.277 guess. Um but we will keep building out as much compute as we possibly can. One 01:00:14.360 --> 01:00:19.477 one one one thing that I think is also worth thinking about is that compute for 01:00:19.560 --> 01:00:24.397 us is not a cost center. It's a profit center, right? When you deploy it in the 01:00:24.480 --> 01:00:28.277 product. And so in many ways our business is extremely simple, right? We 01:00:28.360 --> 01:00:32.517 rent or buy compute and then we resell it at a margin. 01:00:32.600 --> 01:00:36.357 And as long as we have some positive margin on it, then it's scalable, right? 01:00:36.440 --> 01:00:41.317 Because the demand is just unlimited. So okay, so data center hardware still 01:00:41.400 --> 01:00:44.637 all systems go? Or you mean like our own chip? Yeah, own 01:00:44.720 --> 01:00:49.557 chip, networking Yeah, very very excited about our our chip. 01:00:49.640 --> 01:00:51.117 We have an incredible team there. 01:00:51.200 --> 01:00:55.997 >> One Titan one? Uh no, we're not talking about timelines of course, but I'll just 01:00:56.080 --> 01:01:00.317 say that like I spend a lot of time with that team and I think it's so So they're 01:01:00.400 --> 01:01:04.157 doing great. So that's our all systems go. Robotics sounds like it's all systems 01:01:04.240 --> 01:01:07.197 >> to be a while till we have something to where you're like ah this is the chat 01:01:07.280 --> 01:01:10.277 should be team moment, but But but you haven't pulled back on that program. 01:01:10.360 --> 01:01:12.757 Social net social network 01:01:12.840 --> 01:01:17.437 Robots are not the current thing, but are going to be so clearly important in 01:01:17.520 --> 01:01:20.357 the future. >> Okay. And social network? Not doing that. 01:01:20.440 --> 01:01:21.877 >> Not not doing that right now. 01:01:21.960 --> 01:01:25.757 And then clearly like the super app, the browser, all that stuff still still going. 01:01:25.840 --> 01:01:28.517 >> Yeah, and one thing to realize about super app because I feel like it's one 01:01:28.600 --> 01:01:32.117 of these things where it's like a catchy word, but I thought you guys like came 01:01:32.200 --> 01:01:35.997 up with this. I mean but You guys used this like internal short Exactly. This 01:01:36.080 --> 01:01:38.997 is the thing to realize is that like sometimes we're communicating to to our 01:01:39.080 --> 01:01:42.157 team and then of course it ends up being communications the world and these are 01:01:42.240 --> 01:01:45.957 not intended to be the same thing. Super app in many ways is an iceberg, right? 01:01:46.040 --> 01:01:49.477 It's like yes, we're going to have a an app and you'll see updates to what the 01:01:49.560 --> 01:01:53.037 Codex app is today to make it so it's Codex is for everyone and I think that 01:01:53.120 --> 01:01:56.597 ultimately like what that becomes that we have we have a lot of steps to get to 01:01:56.680 --> 01:02:00.077 where we want to be, but it's really about saying we're building this unified 01:02:00.160 --> 01:02:04.197 identity infrastructure that I described and that is I think going to be a huge 01:02:04.280 --> 01:02:08.237 unlock for every single thing that people want to do. Okay. Um 01:02:08.320 --> 01:02:11.317 You and I have talked a little bit about yeah, 01:02:11.400 --> 01:02:12.797 you guys have had lots of drama. It's 01:02:12.880 --> 01:02:16.157 slowed you down at times. You still have 01:02:16.240 --> 01:02:18.597 drama sadly. Um there's a lawsuit 01:02:18.680 --> 01:02:24.317 coming. Uh It'll be fun. Yes, yes. Well, let's talk about that in a second. Um 01:02:24.400 --> 01:02:25.557 Like which company do you think has 01:02:25.640 --> 01:02:28.997 actually executed better over the last 01:02:29.080 --> 01:02:33.637 two years? Open AI or Anthropic? 01:02:33.720 --> 01:02:36.197 Look, I think that this is a hard thing 01:02:36.280 --> 01:02:39.757 to say from just the current moment, 01:02:39.840 --> 01:02:44.117 right? Because I think that the view in my mind is you got to step back 01:02:44.200 --> 01:02:47.997 and really think about what is it that we're all here to do, right? And I think 01:02:48.080 --> 01:02:51.357 that from the Open AI perspective, the things we've been talking about in some 01:02:51.440 --> 01:02:55.437 ways are about it's very clear that yes, selling to enterprises like figuring out 01:02:55.520 --> 01:03:00.237 how to really deliver these coding tools and it's been really it's been actually 01:03:00.320 --> 01:03:04.037 I think competition can really help elevate your own thinking and realize 01:03:04.120 --> 01:03:07.717 that hey, we need to focus on this. And one example in coding was that I think 01:03:07.800 --> 01:03:11.357 we got late to the game of not just building models that were good in the 01:03:11.440 --> 01:03:14.117 abstract of coding. Like we always had the best numbers on programming 01:03:14.200 --> 01:03:18.437 competitions, but you also need to apply them to messy repos, real-world data, 01:03:18.520 --> 01:03:20.717 those kinds of things. And that was something that I think we had 01:03:20.800 --> 01:03:22.757 appreciated later than Anthropic did. 01:03:22.840 --> 01:03:26.877 So, I think that's That's kudos to them, but also something that's helped elevate 01:03:26.960 --> 01:03:30.117 our own execution. And now, head-to-head, Codex versus Claude, I 01:03:30.200 --> 01:03:32.997 think that that that we get very favorable results. 01:03:33.080 --> 01:03:36.157 Um and I think we've done an incredible Our teams have done an incredible job 01:03:36.240 --> 01:03:39.157 across the whole company to really build a product that's not just competitive, 01:03:39.240 --> 01:03:41.437 but actually ahead in many, many ways. 01:03:41.520 --> 01:03:44.917 Um but, I think that the the core it's 01:03:45.000 --> 01:03:47.397 never about the ups and downs of the 01:03:47.480 --> 01:03:52.997 news cycle. It's about progress towards AGI, towards it benefiting everyone. And 01:03:53.080 --> 01:03:55.597 that is something that I think we've been extremely focused on. The team's 01:03:55.680 --> 01:04:00.237 executing extremely well. And so, I just want to say that you can't always tell 01:04:00.320 --> 01:04:03.357 from the outside, but if you think about the things we 01:04:03.440 --> 01:04:07.437 talked about, that there is this focus, but on many timescales that all add up 01:04:07.520 --> 01:04:11.437 to where we need to go. Yeah, there's a million questions I want to ask with the 01:04:11.520 --> 01:04:15.757 time we have left. But, one of them we we touched on this, like making sure a 01:04:15.840 --> 01:04:18.517 bunch of models are in the hands of a bunch of people and they have that 01:04:18.600 --> 01:04:21.757 powerful technology. But, we've reached a point where now some of these models 01:04:21.840 --> 01:04:25.197 are too powerful, we're being told, and they're they're gated for only certain 01:04:25.280 --> 01:04:29.637 companies. Claude and Mythos have really made a lot 01:04:29.720 --> 01:04:33.957 of headlines and I think a lot of a lot more fear. And I'm wondering what you 01:04:34.040 --> 01:04:37.157 guys think of this new moment. Are we getting to a point where like we need to 01:04:37.240 --> 01:04:40.797 start keeping these powerful models behind the scenes rather than in the 01:04:40.880 --> 01:04:44.717 hands of everybody? There are people in the world who 01:04:44.800 --> 01:04:48.037 for a long time have wanted to keep AI in the hands of a smaller group of 01:04:48.120 --> 01:04:53.077 people. Um you can justify that in a lot of different ways. And some of it's real, 01:04:53.160 --> 01:04:56.157 like there are going to be legitimate safety concerns. 01:04:56.240 --> 01:05:01.077 Um but, I expect But, if what you want is like, we need control of AI just us 01:05:01.160 --> 01:05:04.237 cuz we're the trustworthy people, I think the the fear-based marketing is 01:05:04.320 --> 01:05:05.677 probably the most effective way to 01:05:05.760 --> 01:05:08.717 justify that. Um that doesn't mean it's 01:05:08.800 --> 01:05:13.757 not legitimate in some cases. Uh but, it is 01:05:13.840 --> 01:05:17.317 you know, clearly incredible marketing to say, "We have built a bomb. We are 01:05:17.400 --> 01:05:19.877 about to drop it on your head. We will sell you a bomb shelter for $100 01:05:19.960 --> 01:05:22.637 million. You need to like run across all your stuff, but only if we like pick you 01:05:22.720 --> 01:05:25.637 as a customer." And it 01:05:25.720 --> 01:05:29.317 the way that we view balancing these new 01:05:29.400 --> 01:05:33.157 capabilities that are going to come with uh 01:05:33.240 --> 01:05:36.917 still our our belief that the world needs to get 01:05:37.000 --> 01:05:41.117 and use and understand and come up with new ideas for this technology is not 01:05:41.200 --> 01:05:45.077 always easy. We We have had um cybersecurity in our preparedness 01:05:45.160 --> 01:05:48.797 framework for a long time, and we have been building mitigations to figure out 01:05:48.880 --> 01:05:51.877 how we release this, um how we put these models in the hands 01:05:51.960 --> 01:05:55.517 of a trusted access program, how how we 01:05:55.600 --> 01:05:57.717 then put more capable models in the 01:05:57.800 --> 01:06:02.597 hands of everybody. Um but, there will be a lot more rhetoric 01:06:02.680 --> 01:06:05.237 about models that are too dangerous to release. There will also be very 01:06:05.320 --> 01:06:08.197 dangerous models that will have to be released in different ways. 01:06:08.280 --> 01:06:12.077 But, to the point Greg was making about 01:06:12.160 --> 01:06:16.077 the goal here is to benefit everybody. 01:06:16.160 --> 01:06:20.317 And also to I don't want to say market this in a 01:06:20.400 --> 01:06:23.997 way, but but like get the world to come along on this journey with us where 01:06:24.080 --> 01:06:27.397 where it's like we are going to give you more powerful technology. There's going 01:06:27.480 --> 01:06:29.877 to be responsibility that goes along with that. We are going to help set up 01:06:29.960 --> 01:06:32.317 the world for success as much as we can. 01:06:32.400 --> 01:06:36.477 Um but, we will try to avoid the fear-based 01:06:36.560 --> 01:06:39.677 marketing as much as we can. So, I'll I'll just ask you directly. Do you think 01:06:39.760 --> 01:06:42.917 Mythos is just a lot of marketing and not I'm sure it's a great cybersecurity 01:06:43.000 --> 01:06:46.197 like I We've been talking about this for a long time. Like this has been in our 01:06:46.280 --> 01:06:50.157 in our model, but there's there's a way of saying like like our version of this is 01:06:50.240 --> 01:06:53.717 to say these models are going to get much better at cyber. We have this 01:06:53.800 --> 01:06:56.837 preparedness framework category. Here's our plan for how we deploy this into the 01:06:56.920 --> 01:06:59.237 world. Here's how this trusted access program looks like. Here's the 01:06:59.320 --> 01:07:01.517 mitigations we put on the models. 01:07:01.600 --> 01:07:04.957 Um I think I'm talking about how it is not how it is one of the categories in their 01:07:05.040 --> 01:07:08.477 preparedness framework. Uh so, I'm sure Mythos is a great model 01:07:08.560 --> 01:07:12.357 uh for cybersecurity. But, I think we have a plan we feel good 01:07:12.440 --> 01:07:16.437 about for how we put this kind of capability out into the world. When the 01:07:16.520 --> 01:07:19.957 Anthropic stuff was going down with the Department of War, I mean, from my 01:07:20.040 --> 01:07:22.757 perspective, it looked like 01:07:22.840 --> 01:07:25.437 you had sort of David Sacks and people 01:07:25.520 --> 01:07:29.837 with long ties to Elon. I could see putting pressure to make some of these 01:07:29.920 --> 01:07:33.797 things happen and have the government focus it on that. 01:07:33.880 --> 01:07:37.357 You guys kind of came in pretty quick after and and had an announcement of 01:07:37.440 --> 01:07:39.877 your own that was more favorable. 01:07:39.960 --> 01:07:40.957 Um 01:07:41.040 --> 01:07:44.037 I mean, you know, Elon's 01:07:44.120 --> 01:07:47.437 aggressive against you guys as well. 01:07:47.520 --> 01:07:50.637 And some of this is funny to me that I mean, there is a world where like you 01:07:50.720 --> 01:07:55.397 guys and Anthropic are actually on this this other side and and it feels like 01:07:55.480 --> 01:07:59.477 Elon and even Zach to some degree, my understanding, 01:07:59.560 --> 01:08:03.117 kind of pressuring things. As a side, yeah, I 01:08:03.200 --> 01:08:06.997 mean, did you did you feel like Anthropic was treated fairly in that? Do 01:08:07.080 --> 01:08:13.160 you feel um No. >> [snorts] >> Um I don't 01:08:13.360 --> 01:08:15.197 Well, I think there was like a lot of bad behavior to go around there, but I 01:08:15.280 --> 01:08:17.157 don't think Anthropic was treated well. 01:08:17.240 --> 01:08:18.837 Can you explain more? Like what was 01:08:18.920 --> 01:08:24.360 standing out to you about what was particularly wrong? Um 01:08:25.760 --> 01:08:30.797 I don't like look with the things that have happened since with, 01:08:30.880 --> 01:08:34.477 you know, these models reaching this cybersecurity threshold that is clearly 01:08:34.560 --> 01:08:36.997 in the national security interest. I I think it all has a little bit of a 01:08:37.080 --> 01:08:40.157 different flavor to it, but 01:08:40.240 --> 01:08:43.397 you know, like threats of 01:08:43.480 --> 01:08:46.597 using the DPA and actually using supply 01:08:46.680 --> 01:08:49.197 chain risk designation, like this is 01:08:49.280 --> 01:08:50.877 this is not the relationship that I 01:08:50.960 --> 01:08:53.757 think our government and our AI efforts 01:08:53.840 --> 01:08:58.237 need to have. Uh We really care about supporting the US 01:08:58.320 --> 01:09:00.917 government. I think that's going to become increasingly important as these 01:09:01.000 --> 01:09:04.957 models get more capable. And I certainly don't think it's like a 01:09:05.040 --> 01:09:08.797 good stance for uh labs to say, you know, we have the 01:09:08.880 --> 01:09:11.557 super weapon by the way we're not going to work with you to help help you defend 01:09:11.640 --> 01:09:17.960 the country. Um but I also don't think it's good for the government to be like 01:09:18.400 --> 01:09:21.197 fighting this stuff out in the press and using the big hammer that the government 01:09:21.280 --> 01:09:23.077 has that needs to be used very rarely. 01:09:23.160 --> 01:09:27.837 So, I A thing about OpenAI is we generally try 01:09:27.920 --> 01:09:32.077 to be like moderate and centrist and reasonable. And that is what we've tried 01:09:32.160 --> 01:09:34.837 to do here with the US government, but I 01:09:34.920 --> 01:09:38.197 see no future no good future where 01:09:38.280 --> 01:09:42.437 leading AI efforts don't assist the US government. Uh if everything we say is 01:09:42.520 --> 01:09:45.877 right, we believe the things we're saying about where the models are going 01:09:45.960 --> 01:09:50.997 to go, which I certainly do, then the government needs our help and we are 01:09:51.080 --> 01:09:54.637 honored to get to provide it. You know, when I am old as a tech reporter, I 01:09:54.720 --> 01:09:57.397 mean, when I was first writing about tech, you know, the big 01:09:57.480 --> 01:10:00.797 war was it was basically Microsoft versus everybody else and it was they 01:10:00.880 --> 01:10:05.397 were the big evil um proprietary software company. You had the Yeah. open 01:10:05.480 --> 01:10:08.757 source companies counter. You had Mark Andreessen with Netscape as you know, 01:10:08.840 --> 01:10:11.517 and and you know, like things would play out in the press and it would be a 01:10:11.600 --> 01:10:16.077 little vitriol here and there, but you know, and then people had their 01:10:16.160 --> 01:10:17.917 philosophical and religious camps. 01:10:18.000 --> 01:10:20.357 Obviously, like this stuff where we are 01:10:20.440 --> 01:10:23.677 in AI land is insane. Um 01:10:23.760 --> 01:10:27.757 that so much of this seems tied to the personalities of 01:10:27.840 --> 01:10:32.317 you, Elon, um Zuck, Dario, and and Demis and the 01:10:32.400 --> 01:10:34.717 various animosities and world views. 01:10:34.800 --> 01:10:37.997 Like you wrote about this the Shakespearean 01:10:38.080 --> 01:10:42.117 drama of it. Like how do we possibly get past this? I mean, you guys are all so 01:10:42.200 --> 01:10:46.597 um I feel I don't see a path out and and obviously 01:10:46.680 --> 01:10:49.437 like we mentioned there's a lawsuit coming that will only Wouldn't all the 01:10:49.520 --> 01:10:51.517 hands on stage? I'm so sorry to say. 01:10:51.600 --> 01:10:56.637 >> Some some of the people involved only 01:10:56.720 --> 01:11:01.077 trust themselves to get it right. And because they think the stakes are 01:11:01.160 --> 01:11:04.837 infinite and because they don't, you know, for different reasons, uh they 01:11:04.920 --> 01:11:07.717 don't think anybody else can 01:11:07.800 --> 01:11:09.117 do it right or they don't want anybody 01:11:09.200 --> 01:11:13.557 else to do it. Um that leads to 01:11:13.640 --> 01:11:16.677 I think some very toxic behavior. We can't control the behavior of other 01:11:16.760 --> 01:11:19.597 people. Uh but we will continue to 01:11:19.680 --> 01:11:22.557 advocate for doing this as a 01:11:22.640 --> 01:11:26.837 collective project that humanity has to get right together and not something 01:11:26.920 --> 01:11:28.357 where it should be about 01:11:28.440 --> 01:11:30.757 one person or one ideology winning or 01:11:30.840 --> 01:11:34.157 losing or you know, having a 01:11:34.240 --> 01:11:38.117 sole victory or something like that. I mean, we'd be remiss and I know it's 01:11:38.200 --> 01:11:41.837 super sensitive, but I mean, it you know, if you look at what happened 01:11:41.920 --> 01:11:46.357 to you personally over the last week, there's tons of science fiction books 01:11:46.440 --> 01:11:50.277 that have been written about what happens when um 01:11:50.360 --> 01:11:51.837 the factions that really don't want to 01:11:51.920 --> 01:11:54.877 see technological progress happen get um 01:11:54.960 --> 01:11:57.997 you know, things get more extreme. Um 01:11:58.080 --> 01:12:02.117 I don't know. Like there's an argument to be made that this this switch is now 01:12:02.200 --> 01:12:05.277 flipping. Along those lines. With people that want 01:12:05.360 --> 01:12:09.077 to like stop. Yeah, I mean, you know, I've I've read Richard Clark's book 01:12:09.160 --> 01:12:13.437 years ago where a Bill Joy-like figure decides AI can't happen and he's running 01:12:13.520 --> 01:12:16.357 around destroying data center. You know, this stuff has been out there forever. 01:12:16.440 --> 01:12:19.317 It just feels like, God, I mean, I don't see how things could get like more 01:12:19.400 --> 01:12:23.157 heightened and so um it doesn't seem like it's getting better. 01:12:23.240 --> 01:12:27.317 I I assume it will go up and down, but directionally more heightened. 01:12:27.400 --> 01:12:30.797 This must be like you wrote about. I mean, clearly this must be horrifying. 01:12:30.880 --> 01:12:33.077 Horrifying. Yeah, yeah, I mean, I don't really have anything like deep to say 01:12:33.160 --> 01:12:35.520 here. 01:12:36.280 --> 01:12:40.120 That was a crazy way to wake up. Um 01:12:40.720 --> 01:12:44.437 The the first day I was sort of in this light kind of adrenaline shock about it 01:12:44.520 --> 01:12:46.957 and just trying to like figure out logistics. 01:12:47.040 --> 01:12:50.557 And then the day after I was just like, you know, 01:12:50.640 --> 01:12:53.357 there's going to be more stuff like this and it's incredibly disheartening and I 01:12:53.440 --> 01:12:54.757 went through a real depressive cycle 01:12:54.840 --> 01:13:00.720 about it. Um, but it's very scary and I 01:13:04.480 --> 01:13:07.077 Yeah, I don't think I have anything super deep to say. I think the doomerism 01:13:07.160 --> 01:13:10.477 talk hasn't helped. Uh, 01:13:10.560 --> 01:13:14.197 I think the way certain other labs talk about us hasn't helped. I actually don't 01:13:14.280 --> 01:13:16.437 want to make like a a side statement. I think the way Anthropic talks about 01:13:16.520 --> 01:13:21.720 OpenAI doesn't help. Um, and 01:13:22.880 --> 01:13:24.837 you know, I I I like 01:13:24.920 --> 01:13:26.837 I hope that cooler times will prevail. 01:13:26.920 --> 01:13:31.437 One One thing I do want to say, just to you know, the point of we can't control 01:13:31.520 --> 01:13:32.717 other people, we can control ourselves. 01:13:32.800 --> 01:13:37.077 Like one thing I I've been just really impressed by and astounded by with Sam 01:13:37.160 --> 01:13:39.637 is just how throughout this time, like 01:13:39.720 --> 01:13:43.197 that very day, he was there doing things 01:13:43.280 --> 01:13:47.917 that absolutely required him. Like he just like keeps pushing on the mission. 01:13:48.000 --> 01:13:52.397 And I think that is something that I don't take for granted at all. Like I just 01:13:52.480 --> 01:13:56.597 think that the degree of resilience that is represented by Sam is is extreme and 01:13:56.680 --> 01:13:58.637 I think very underappreciated. 01:13:58.720 --> 01:14:03.157 The I mean I've seen the same thing. I don't know 01:14:03.240 --> 01:14:05.677 what people will be like, oh, you're taking it easy or something like that. I 01:14:05.760 --> 01:14:09.077 don't know. I I am It's hard to think of a like another 01:14:09.160 --> 01:14:14.317 individual over the last 3 years who's been through more dramatic business and 01:14:14.400 --> 01:14:16.957 and personal good >> The narrative has been quite dramatic. 01:14:17.040 --> 01:14:19.077 >> be sympathetic cuz they're like you're rich. 01:14:19.160 --> 01:14:21.957 >> be sympathetic cuz they're like And that's fine. Um, 01:14:22.040 --> 01:14:25.437 I I I said this the night before 01:14:25.520 --> 01:14:27.437 that this thing happened at my house. 01:14:27.520 --> 01:14:31.877 Um, I had some people over for dinner and we were talking like uh some work 01:14:31.960 --> 01:14:33.957 colleagues and we were talking about the next phase and they're like, uh, it's 01:14:34.040 --> 01:14:36.757 been like a you know, a brutal time for you in the press, but I I guess it's 01:14:36.840 --> 01:14:39.437 like good in some sense and I was like, uh, you know, at least no one tried to 01:14:39.520 --> 01:14:40.477 kill me. 01:14:40.560 --> 01:14:42.637 Um, You said that the night before? 01:14:42.720 --> 01:14:46.560 Uh, and 01:14:46.720 --> 01:14:50.557 and it but it really did put into perspective that like, you know, people 01:14:50.640 --> 01:14:53.557 can say all the mean things they want and as long as they only say the mean 01:14:53.640 --> 01:14:56.557 things, it's like it's really not that big of a deal. You keep going. 01:14:56.640 --> 01:14:58.917 >> You always told me your dream is to like 01:14:59.000 --> 01:15:03.597 retire out in that someday. So like why not We got a lot of work left. But like 01:15:03.680 --> 01:15:06.557 somebody's going to get It's clear somebody's going to get to AGI now. We 01:15:06.640 --> 01:15:09.997 got like five companies that could reasonably do this. 01:15:10.080 --> 01:15:12.917 >> Which company would you most like to get there first? Is that Is that what it 01:15:13.000 --> 01:15:15.997 comes down to? >> cuz look, I think that the point cuz the 01:15:16.080 --> 01:15:19.037 the missing thing is this point of how 01:15:19.120 --> 01:15:23.477 do we help people really understand what it is that this technology can do for 01:15:23.560 --> 01:15:25.757 them. And I think that's something that every 01:15:25.840 --> 01:15:29.117 company can contribute to, right? That I think is something we have a perspective 01:15:29.200 --> 01:15:32.557 on. We've talked a lot about it in this podcast. And fundamentally, I think that 01:15:32.640 --> 01:15:35.957 that is something that is kind of the 01:15:36.040 --> 01:15:39.997 responsibility and on all the people who are trying to create this technology is 01:15:40.080 --> 01:15:44.077 also to show its benefits and why people should want it. Why should people be 01:15:44.160 --> 01:15:48.397 protecting and defending the ability to create it? Like why does America need to 01:15:48.480 --> 01:15:52.277 have leadership here? Like why is all this good for not just a country, but 01:15:52.360 --> 01:15:55.677 you personally, your future, the future of your kids? And that's something that 01:15:55.760 --> 01:16:00.037 we wake up thinking about like like I would say every day. Like it I don't 01:16:00.120 --> 01:16:02.917 think it's an overstatement. Like this is something we talk about constantly. 01:16:03.000 --> 01:16:05.997 We think about whether it's weird creative, you know, 01:16:06.080 --> 01:16:09.237 legal structures for a company. We've done that many times 01:16:09.320 --> 01:16:12.957 because we're trying to solve for this mission, for this thing that we think is 01:16:13.040 --> 01:16:16.117 so important. And that's something where if other people want to contribute to 01:16:16.200 --> 01:16:18.717 that, then more power to them. Like that's something we should all be doing, 01:16:18.800 --> 01:16:23.077 but it's something that drives us uniquely. We so believe that if we can 01:16:23.160 --> 01:16:26.477 deliver technology that enables prosperity for everybody, if we can give 01:16:26.560 --> 01:16:30.717 people more agency over the future, uh that will 01:16:30.800 --> 01:16:35.037 through some fits and starts lead to a better world. I don't think all 01:16:35.120 --> 01:16:37.957 people working in the field believe that, but anyone who does, we are 01:16:38.040 --> 01:16:41.757 delighted to work with and that is the the that, you know, we we want to move 01:16:41.840 --> 01:16:45.797 the world towards. Go Going back Well, I think we have time for like a couple 01:16:45.880 --> 01:16:47.477 questions. >> Actually, I have wish list questions. 01:16:47.560 --> 01:16:50.517 >> I've got Can I'm going to I'm going to give you two answers real fast. 01:16:50.600 --> 01:16:53.637 >> I actually want to end on the one that I'm about to ask. So, I'll do this one 01:16:53.720 --> 01:16:56.837 and then [laughter] All right, we're ready. We're ready. I 01:16:56.920 --> 01:16:59.317 mean, how you 01:16:59.400 --> 01:17:04.717 There is like a How existential do you view this trial? 01:17:04.800 --> 01:17:08.477 I I actually think it's a real opportunity for us to tell our story. 01:17:08.560 --> 01:17:13.637 Because if you look at the course of OpenAI, we have really let 01:17:13.720 --> 01:17:17.117 the other sides, when there's splits, tell the story. And we really have done 01:17:17.200 --> 01:17:19.597 our best not to sort of say, well, well, 01:17:19.680 --> 01:17:24.717 that's not really what happened. Like, let's talk about the truth. And the In 01:17:24.800 --> 01:17:27.557 this case, we finally have no choice, right? 01:17:27.640 --> 01:17:30.397 Because we have to defend ourselves. We have to tell the truth. We have to tell 01:17:30.480 --> 01:17:32.837 what happened. And I'm extremely proud. 01:17:32.920 --> 01:17:35.637 Like, I've spent a lot of time looking back at the history at a bunch of 01:17:35.720 --> 01:17:38.397 different messages. And of course, there's always things that you can like 01:17:38.480 --> 01:17:40.837 try to like cherry-pick, be like, "Aha, you said this thing." 01:17:40.920 --> 01:17:44.557 >> Your diaries are are famous. I know, right? But the thing is, that they're 01:17:44.640 --> 01:17:48.517 not because And first of all, incredibly personal documents. Extremely painful to 01:17:48.600 --> 01:17:52.597 have something so personal be sort of taken from you and then, you know, 01:17:52.680 --> 01:17:56.157 attempted to be weaponized. But those particular, you know, sentences are kind 01:17:56.240 --> 01:18:00.597 of the worst thing that the opposition could find. You're like, "Really?" And 01:18:00.680 --> 01:18:03.357 the thing is, they're all taken out of context, right? The question of, you 01:18:03.440 --> 01:18:07.277 know, "Hey, like, we're in the middle of this negotiation, right? We've all 01:18:07.360 --> 01:18:11.957 agreed the only path forward for OpenAI is a for-profit." Sam, Ilya, Greg, Elon, 01:18:12.040 --> 01:18:15.317 we all agreed on this. We all said this is what we got to do. This is literally 01:18:15.400 --> 01:18:18.837 the thing for the mission. And now you're in this crazy negotiation, right? 01:18:18.920 --> 01:18:20.677 Elon's like, "You need majority equity. 01:18:20.760 --> 01:18:24.117 You need to to be CEO. You need full control." And 01:18:24.200 --> 01:18:26.597 we got so close. It's like, "Okay, fine. 01:18:26.680 --> 01:18:29.597 Okay, we're not going to be equal partners. You need this this massive 01:18:29.680 --> 01:18:33.077 amount of equity. If that's what you say you need, like, 01:18:33.160 --> 01:18:36.797 we would like you to be involved. We can get there. Okay, fine. You know, Sam 01:18:36.880 --> 01:18:40.037 will be CEO. Elon will be CEO. He needs it so that everyone knows he's in 01:18:40.120 --> 01:18:43.957 charge. Fine. But absolute control. 01:18:44.040 --> 01:18:47.557 Absolute control over OpenAI. Even if you say, "Well, I'll dilute down. I'll 01:18:47.640 --> 01:18:50.917 give it up in the future." And you're like, "What is our mission? 01:18:51.000 --> 01:18:54.957 Like, do we really believe in our mission? Do we really care about this picture of 01:18:55.040 --> 01:18:57.997 we want this technology to benefit everyone, and there shouldn't be one 01:18:58.080 --> 01:19:00.797 person person in charge of the whole future. 01:19:00.880 --> 01:19:04.077 Doesn't matter who that person is. Like, that was the breaking point. That was 01:19:04.160 --> 01:19:06.197 the thing that caused us to say no. 01:19:06.280 --> 01:19:11.197 And so, we've never told that story for years. But now we will. And so, I I 01:19:11.280 --> 01:19:14.117 think that it is a real opportunity for people to understand what truly 01:19:14.200 --> 01:19:18.037 motivates us, what we truly stand for. I think it's insane that he's doing this, 01:19:18.120 --> 01:19:20.957 but I'm kind of my fear at this point is he decides to 01:19:21.040 --> 01:19:23.517 like drop the case right before the trial, and we don't get to do all this. 01:19:23.600 --> 01:19:26.317 But I am happy to like explain all this to the world and have this chapter 01:19:26.400 --> 01:19:30.317 behind us. Okay. So, my my question is rounding back to the beginning of this 01:19:30.400 --> 01:19:34.477 podcast. Like, I was talking about this narrative I feel like I as an AI 01:19:34.560 --> 01:19:37.557 reporter grew up with hearing of, you know, 01:19:37.640 --> 01:19:40.157 "We're all going to die, and we're all going to lose our jobs." I feel like I 01:19:40.240 --> 01:19:44.597 heard that time and time again, and now portions of your personal notebook are 01:19:44.680 --> 01:19:48.037 public. Um I'm curious, again, how would you have 01:19:48.120 --> 01:19:49.957 talked about this differently with 01:19:50.040 --> 01:19:53.237 hindsight being 20/20? 01:19:53.320 --> 01:19:57.357 Well, I think that the way that we think about it today 01:19:57.440 --> 01:20:01.117 is the way that I wish we had talked about it in some sense. But I don't know 01:20:01.200 --> 01:20:04.197 we could have with the knowledge we had at the time. For example, even some of 01:20:04.280 --> 01:20:06.437 it is related to the technology itself. 01:20:06.520 --> 01:20:09.917 We had a picture in 2017, 2018, the way 01:20:10.000 --> 01:20:13.877 we would build AGI was through a competitive multi-agent simulation. 01:20:13.960 --> 01:20:17.717 Like, imagine an island of like a thousand agents that all are in a battle 01:20:17.800 --> 01:20:21.197 to survive and replicate. And you could see that, if you put a ton of compute 01:20:21.280 --> 01:20:25.357 into it, maybe it would build something very smart. But that smart thing 01:20:25.440 --> 01:20:28.117 that would not be connected to human values at all. You'd have to have a 01:20:28.200 --> 01:20:31.277 separate step of figuring out, how do you even talk to this thing, right? 01:20:31.360 --> 01:20:34.277 Didn't grow up at all in the real world, doesn't have any concept of language, 01:20:34.360 --> 01:20:38.757 doesn't have any connection to our reality. It's just smart and powerful. 01:20:38.840 --> 01:20:42.797 That's a very scary system, right? You start from a place of 01:20:42.880 --> 01:20:46.037 trying to think about how could you possibly align it. But instead we have 01:20:46.120 --> 01:20:50.037 this language model route, which is rooted in our values, which is rooted in 01:20:50.120 --> 01:20:54.317 understanding humans, and we have chain of thought that actually you can 01:20:54.400 --> 01:20:57.637 monitor, that actually if you approach things right in a technical sense, you 01:20:57.720 --> 01:21:01.237 have a path to make that be faithful. So it really represents what is truly 01:21:01.320 --> 01:21:04.877 motivating the AI. And you just realize we have a totally different 01:21:04.960 --> 01:21:08.917 technological path to get to the outcome we're talking about, and it's a much 01:21:09.000 --> 01:21:13.157 more optimistic one. And so I think that there was some technical learnings we 01:21:13.240 --> 01:21:17.077 had to have for what is this technology truly going to be, how will it be 01:21:17.160 --> 01:21:20.477 created, what are the ways in which you make it useful. And that is something I 01:21:20.560 --> 01:21:23.677 think we could not have appreciated, but we do today. 01:21:23.760 --> 01:21:25.997 Um it's been a fascinating discussion. 01:21:26.080 --> 01:21:28.397 It's um really kind of you guys to both 01:21:28.480 --> 01:21:31.477 show up and do this together. Um 01:21:31.560 --> 01:21:35.797 we we have been a humble podcast. We're honored um for for you guys to spend 01:21:35.880 --> 01:21:39.237 this time. And I guess I mean the obviously we covered a lot of ground, 01:21:39.320 --> 01:21:42.197 but the headline is in relatively short 01:21:42.280 --> 01:21:46.837 order, new new models Yeah. are really good new models. Yeah. Um well, thank 01:21:46.920 --> 01:21:48.077 you guys. Thank you guys so much. 01:21:48.160 --> 01:21:50.507 >> Thank you for having new models useful for everyone. 01:21:50.590 --> 01:21:55.360 >> [laughter] >> Thank you guys. Thank you. Thank you. 01:21:56.080 --> 01:21:58.397 The Core Memory podcast is hosted by me, 01:21:58.480 --> 01:22:01.717 Ashley Vance, and or Kylie Robinson, or 01:22:01.800 --> 01:22:04.717 both of us together. It is produced by 01:22:04.800 --> 01:22:06.877 me and David Nicholson. [music] 01:22:06.960 --> 01:22:09.277 Our theme song is by James Mercer and 01:22:09.360 --> 01:22:12.757 John Sortland, and the show is edited, 01:22:12.840 --> 01:22:16.797 always, by the John Sortland. Thank you 01:22:16.880 --> 01:22:21.077 so much to Brex and Combinator Ventures for all [music] your support, and thank 01:22:21.160 --> 01:22:23.157 you most of all to everybody for 01:22:23.240 --> 01:22:26.037 listening or watching. We love you. 01:22:26.120 --> 01:22:28.277 Please leave us a like, a review, a 01:22:28.360 --> 01:22:31.797 subscribe, all those tremendous things. 01:22:31.880 --> 01:22:33.087 Thank you and we'll see you again. 01:22:33.170 --> 01:22:35.190 [music]