Sam Altman Shows Me GPT 5… And What’s Next (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:00) This is like a crazy amount of power for (00:00:02) one piece of technology and it's (00:00:03) happened to us so fast. (00:00:05) >> You just launched GBT. (00:00:06) >> A kid born today will never be smarter (00:00:08) than AI. (00:00:08) >> How do we figure out what's real and (00:00:10) what's not real? (00:00:10) >> We haven't put a sex bot avatar in HBT (00:00:12) yet. (00:00:14) >> Super intelligence. What does that (00:00:15) actually mean? (00:00:16) >> This thing is remarkable. (00:00:20) >> I'm about to interview Sam Alman, the (00:00:22) CEO of Open AI. (00:00:24) >> Open AI. (00:00:24) >> Open AI. (00:00:25) >> Reshaping industries. (00:00:26) >> Dude's a straightup tech lord. Let's be (00:00:28) honest. Right now, they're trying to (00:00:30) build a super intelligence that could (00:00:32) far exceed humans in almost every field. (00:00:35) And they just released their most (00:00:37) powerful model yet. Just a couple years (00:00:39) ago, that would have sounded like (00:00:41) science fiction. Not anymore. In fact, (00:00:43) they're not alone. We are in the middle (00:00:45) of the highest stakes global race any of (00:00:48) us have ever seen. Hundreds of billions (00:00:50) of dollars and an unbelievable amount of (00:00:53) human worth. This is a profound moment. (00:00:55) Most people never live through a (00:00:57) technological shift like this, and it's (00:01:00) happening all around you and me right (00:01:02) now. So, in this episode, I want to try (00:01:04) to time travel with Sam Alman into the (00:01:07) future that he's trying to build to see (00:01:10) what it looks like so that you and I can (00:01:12) really understand what's coming. Welcome (00:01:15) to Huge Conversations. (00:01:24) >> How are you? Great to meet you. Thanks (00:01:26) for doing this. (00:01:26) >> Absolutely. (00:01:27) >> So, before we dive in, I'd love to tell (00:01:28) you my goal here. (00:01:29) >> Okay. (00:01:30) >> I'm not going to ask you about valuation (00:01:32) or AI talent wars or fundraising or (00:01:35) anything like that. I think that's all (00:01:36) very well covered elsewhere. (00:01:38) >> It does seem like it. (00:01:39) >> Our big goal on this show is to cover (00:01:42) how we can use science and tech to make (00:01:44) the future better. And the reason that (00:01:46) we do all of that is because we really (00:01:48) believe that if people see those better (00:01:50) futures, they can then help build them. (00:01:53) So, my goal here is to try my best to (00:01:56) time travel with you into different (00:01:59) moments in the future that you're trying (00:02:00) to build and see what it looks like. (00:02:03) >> Fantastic. (00:02:04) >> Awesome. Starting with what you just (00:02:06) announced, you recently said, (00:02:08) surprisingly recently, that GPT4 was the (00:02:11) dumbest model any of us will ever have (00:02:13) to use again. But GPT4 can already (00:02:16) perform better than 90% of humans at the (00:02:19) SAT and the LSAT and the GRE and it can (00:02:22) pass coding exams and SOA exams and (00:02:26) medical licensing. And now you just (00:02:28) launched GPT5. (00:02:30) >> What can GPT5 do that GPT4 can't? First (00:02:33) of all, one important takeaway is you (00:02:34) can have an AI system that can do all (00:02:36) those amazing things you just said. And (00:02:38) it doesn't it clearly does not replicate (00:02:41) a lot of what humans are good at doing, (00:02:43) which I think says something about the (00:02:44) value of SAT tests or whatever else. But (00:02:46) I think had you gone back to if we were (00:02:48) having this conversation the day of GPT4 (00:02:50) launch and we told you how GPT4 did at (00:02:52) those things, you were like, "Oh man, (00:02:54) this is going to have huge impacts and (00:02:56) some negative impacts on what it means (00:02:58) for a bunch of jobs or you know what (00:03:01) people are going to do." And you know, (00:03:03) this is a bunch of positive impacts that (00:03:05) you might have predicted that haven't (00:03:06) yet come true. Uh, and so there there's (00:03:09) something about the way that these (00:03:12) models are good that does not capture a (00:03:15) lot of other things that we need people (00:03:17) to to do or care about people doing. And (00:03:19) I suspect that same thing is going to (00:03:21) happen again with GPT5. People are going (00:03:23) to be blown away by what it does. Uh, (00:03:26) it's really good at a lot of things and (00:03:28) then they will find that they want it to (00:03:31) do even more. Um, people will use it for (00:03:33) all sorts of incredible things. uh it (00:03:35) will transform a lot of knowledge work, (00:03:39) a lot of the way we learn, a lot of the (00:03:41) way we create um but we people society (00:03:45) will co-eolve with it to expect more (00:03:48) with you know better tools. So yeah like (00:03:51) I think this model is quite remarkable (00:03:53) in many ways quite limited in others but (00:03:56) the fact that for you know 3 minute 5 (00:03:59) minute 1-hour tasks that uh like an (00:04:03) expert in a in a field could maybe do or (00:04:07) maybe struggle with that the fact that (00:04:09) you have in your pocket one piece of (00:04:11) software that can do all of these things (00:04:14) >> is really amazing. I think this is like (00:04:16) unprecedented at any point in human (00:04:18) history that I that a technology has (00:04:21) improved this much this fast and and the (00:04:24) fact that we have this tool now, you (00:04:26) know, we're like living through it and (00:04:27) we're kind of adjusting step by step. (00:04:29) But if we could go back in time five or (00:04:31) 10 years and say this thing was coming, (00:04:33) we would be like probably not. (00:04:36) >> Let's assume that people haven't seen (00:04:37) the headlines. What are the topline (00:04:40) specific things that you're excited (00:04:42) about? and also the things that you seem (00:04:43) to be caveatting, the things that maybe (00:04:45) you won't expect it to do. (00:04:46) >> Um, (00:04:49) the thing that I am most excited about (00:04:51) is this is a model for the first time (00:04:54) where I feel like I can ask (00:04:56) kind of any hard scientific or technical (00:05:00) question (00:05:02) >> and get a pretty good answer. And I'll (00:05:04) give a fun example actually. Uh when I (00:05:08) was in junior high uh or maybe it was (00:05:10) nth grade, I got a TI83, this old (00:05:13) graphing calculator, and I spent so long (00:05:16) making this game called Snake. (00:05:19) >> Yeah. (00:05:19) >> Uh it was very popular game with kids in (00:05:21) my school. And I was I was like uh I was (00:05:24) like pro and it was dumb, but it was (00:05:25) like programming on TID3 was extremely (00:05:27) painful and took a long time and it was (00:05:29) really hard to like debug and whatever. (00:05:31) And on a whim with an early copy of (00:05:33) GPT5, I was like, I wonder if it can (00:05:35) make a TI83 style Game of Snake. And of (00:05:39) course, it did that perfectly in like 7 (00:05:40) seconds. And then I was like, okay, am I (00:05:42) supposed to be would my like 11-year-old (00:05:45) self think this was cool or like, you (00:05:48) know, miss something from the process? (00:05:49) And I had like 3 seconds of wondering (00:05:52) like, oh, is this good or bad? And then (00:05:53) I immediately said, actually, now I'm (00:05:56) missing this game. I have this idea for (00:05:58) a crazy new feature. Let me type it in. (00:06:00) it implements it and it just the game (00:06:02) live updates and I'm like actually I'd (00:06:04) like it to look this way. Actually, I'd (00:06:05) like to do this thing and I had this (00:06:07) like this very like kind of (00:06:10) >> you have this experience that reminded (00:06:11) me of being like 11 in programming again (00:06:13) where I was just like I now I want to (00:06:14) try this now I have this idea now I but (00:06:16) I could do it so fast and I could like (00:06:19) express ideas and try things and play (00:06:21) with things in such real time. I was (00:06:23) like, "Oh man, you know, I was worried (00:06:25) for a second about kids like missing the (00:06:27) struggle of learning to program in this (00:06:28) sort of stone age way." And now I'm just (00:06:31) thrilled for them because the the way (00:06:32) that people will be able to create with (00:06:33) these new tools, the speed with which (00:06:35) you can sort of bring ideas to life, you (00:06:38) know, in (00:06:40) that's that's pretty amazing. So this (00:06:42) idea that GPT5 can just not only like (00:06:45) answer all these hard questions for you (00:06:47) but really create like ondemand almost (00:06:49) instantaneous software that's I think (00:06:52) that's going to be one of the defining (00:06:54) elements of the GPD5 era in a way that (00:06:56) did not exist with GPD4. (00:06:58) >> As you're talking about that I find (00:06:59) myself thinking about a concept in (00:07:01) weightlifting of time under tension. (00:07:03) >> Yeah. (00:07:04) >> And for those who don't know it's you (00:07:06) can squat 100 pounds in 3 seconds or you (00:07:09) can squat 100 pounds in 30. You gain a (00:07:11) lot more by squatting it in 30. (00:07:13) >> And when I think about our creative (00:07:14) process and when I've felt most like (00:07:17) I've done my best work, it has required (00:07:18) an enormous amount of (00:07:20) >> cognitive time under tension. (00:07:23) >> And I think that that cognitive time (00:07:25) under tension is so important. (00:07:27) >> And it's it's ironic almost because (00:07:29) these tools have taken enormous (00:07:32) cognitive time under tension to develop. (00:07:34) But in some ways I do think people might (00:07:36) say they're you people are using them as (00:07:40) a escape hatch for thinking in some ways (00:07:42) maybe. Now you might say yeah but we did (00:07:46) that with the calculator and we just (00:07:47) moved on to harder math problems. Do you (00:07:49) feel like there's something (00:07:51) different happening here? How do you (00:07:53) think about this? (00:07:54) >> It's different with I mean there are (00:07:56) some people who are clearly using (00:07:57) chachine not to think and there are some (00:08:00) people who are using it to think more (00:08:01) than they ever have before. (00:08:05) I am hopeful that we will be able to (00:08:07) build the tool in a way that encourages (00:08:09) more people to stretch their brain with (00:08:11) it a little more and be able to do more. (00:08:13) And I think that like you know society (00:08:15) is a competitive place like if you give (00:08:17) people new tools uh in theory maybe (00:08:20) people just work less but in practice it (00:08:22) seems like people work ever harder and (00:08:24) the expectations of people just go up. (00:08:26) So my (00:08:28) my guess is that like other tools uh (00:08:33) some people like other pieces of (00:08:35) technology some people will do more and (00:08:37) some people will do less but certainly (00:08:39) for the people who want to use chatbt to (00:08:42) increase their cognitive time under (00:08:44) tension they are really able to and it (00:08:46) is (00:08:47) >> I take a lot of inspiration from what (00:08:49) like the top 5% of most engaged users do (00:08:52) with chacht like it's really (00:08:54) >> amazing how much people are learning and (00:08:56) doing and you know outputting. (00:08:59) >> So my I've only had GPT5 for a couple (00:09:02) hours so I've been playing. (00:09:03) >> What do you think so far? (00:09:04) >> I'm I'm just learning how to interact (00:09:07) with it. I mean part of the interesting (00:09:08) thing is I feel like I just caught up on (00:09:10) how to use GPT4 and now I'm trying to (00:09:12) learn how to use GPD5. I'm curious what (00:09:15) the specific tasks that you found most (00:09:19) interesting are because I imagine you've (00:09:21) been using it for a while now. I I have (00:09:24) been most impressed by the coding tasks. (00:09:25) I mean, there's a lot of other things (00:09:26) it's really good at, but this this idea (00:09:29) of (00:09:31) the AI can write software for anything. (00:09:35) And that means that you can express (00:09:38) ideas in new ways that the AI can do (00:09:41) very advanced things. It can do, you (00:09:43) know, it can like in some sense you (00:09:45) could like ask GPT4 anything, but (00:09:47) because GPT5 is so good at programming, (00:09:50) it feels like it can do anything. Of (00:09:51) course, it can't do things in the (00:09:52) physical world, but it can get a (00:09:53) computer to do very complex things. And (00:09:56) software is this super powerful, you (00:09:59) know, way to like control some stuff and (00:10:02) actually do some things. So, that that (00:10:04) for me has been the most striking. Um, (00:10:08) it's gotten it's much better at writing. (00:10:10) So, this is like there's this whole (00:10:12) thing of AI slop like AI writes in this (00:10:15) kind of like quite annoying way and (00:10:17) >> M dashes. (00:10:18) >> M we still have the M dashes in GPT5. A (00:10:20) lot of people like them dashes, but the (00:10:23) writing quality of GPT5 is gotten much (00:10:26) better. We still have a long way to go. (00:10:28) We want to improve it more, but like uh (00:10:31) I've a thing we've heard a lot from (00:10:33) people inside of OpenAI is that man, (00:10:35) they started using GPT5, they knew it (00:10:37) was better on all the metrics, but (00:10:39) there's this like nuance quality they (00:10:41) can't quite articulate, but then when (00:10:43) they have to go back to GPT4 to test (00:10:44) something, it feels terrible. And I I (00:10:47) don't know exactly what the cause of (00:10:48) that is, but I suspect part of it is the (00:10:50) writing feels so much more natural and (00:10:52) better. (00:10:53) >> I in preparation for this interview (00:10:55) reached out to a couple other leaders in (00:10:57) AI and technology and gathered a couple (00:10:59) questions for you. (00:11:00) >> Okay, (00:11:00) >> so this next question is from Stripe CEO (00:11:03) Patrick Collison. (00:11:04) >> This will be a good one. (00:11:06) >> Read this verbatim. (00:11:07) >> It's about the next stage. What what (00:11:11) comes after GBT5? In which year do you (00:11:13) think a large language model will make a (00:11:15) significant scientific discovery and (00:11:17) what's missing such that it hasn't (00:11:19) happened yet? He caveed here that we (00:11:21) should leave math and special case (00:11:23) models like alpha fold aside. He's (00:11:24) specifically asking about fully general (00:11:26) purpose models like the GPT series. (00:11:28) >> I would say most people will agree that (00:11:30) that happens at some point over the next (00:11:31) two years. But the definition of (00:11:33) significant matters a lot. And so some (00:11:35) people significant might happen, (00:11:38) you know, in early 25. Some people might (00:11:41) maybe not until late 2026. Sorry, early (00:11:43) 2026. Maybe some people not until late (00:11:45) 2027, but I would I would bet that by (00:11:48) late 27, most people agree that there (00:11:51) has been an AIdriven significant new (00:11:53) discovery. And the thing that I think is (00:11:54) missing is just the kind of cognitive (00:11:57) power of these models. A framework that (00:12:00) one of the researchers said to me that I (00:12:01) really liked is, you know, a year ago we (00:12:05) could do well on like a high school like (00:12:08) a basic high school math competition (00:12:10) problems that might take a professional (00:12:11) mathematician seconds to a few minutes. (00:12:14) We very recently got an IMO gold medal. (00:12:16) That is a crazy difficult like (00:12:19) >> could you explain what that means? (00:12:21) >> That's kind of like the hardest (00:12:22) competition math test. This is something (00:12:23) that like the very very top slice of the (00:12:26) world. many many professional (00:12:28) mathematicians wouldn't solve a single (00:12:29) problem and we scored at the top level. (00:12:33) Now there are some humans that got an (00:12:34) even higher score in the gold medal (00:12:35) range but we we like this is a crazy (00:12:37) accomplishment and these each of these (00:12:39) problems it's like six problems over 9 (00:12:42) hours so hour and a half per problem for (00:12:44) a great mathematician. So we've gone (00:12:45) from a few seconds to a few minutes to (00:12:48) an hour and a half maybe to prove a (00:12:51) significant new mathematical theorem is (00:12:53) like a thousand hours of work for a top (00:12:55) person in the world. So we've got to go (00:12:57) from, you know, another significant (00:13:00) gain. But if you look at our trajectory, (00:13:01) you can say like, okay, we're getting to (00:13:03) that. We have a path to get to that time (00:13:05) horizon. We just need to keep scaling (00:13:07) the models. (00:13:09) >> The long-term future that you've (00:13:10) described is super intelligence. What (00:13:13) does that actually mean? And how will we (00:13:15) know when we've hit it? (00:13:18) If we had a system that could do better (00:13:21) research, better AI research than uh say (00:13:25) the whole open AI research team, like if (00:13:27) we were willing, if we said, "Okay, the (00:13:29) best way we can use our GPUs is to let (00:13:31) this AI decide what experiments we (00:13:32) should run (00:13:34) >> smarter than like the whole brain trust (00:13:35) of Open AAI." Yeah. And if that same to (00:13:37) make a personal example, if that same (00:13:38) system could do a better job running (00:13:40) open AI than I could. So you have (00:13:42) something that's like, you know, better (00:13:43) than the best researchers, better than (00:13:44) me at this, better than other people at (00:13:45) their jobs, that would feel like super (00:13:46) intelligence to me. (00:13:48) >> That is a sentence that would have (00:13:49) sounded like science fiction just a (00:13:51) couple years ago. And now it (00:13:52) >> kind of does, but it's you can like see (00:13:54) it through the fog. (00:13:55) >> Yes. And so one of the steps it sounds (00:13:57) like you're saying on that path is this (00:13:59) moment of scientific discovery of asking (00:14:02) better questions of grappling with (00:14:04) things in a in a way that expert level (00:14:06) humans do (00:14:08) >> to come up with new discoveries. One of (00:14:09) the things that keeps knocking around in (00:14:11) my head is if we were in 1899 say and we (00:14:15) were able to give it all of physics up (00:14:16) until that point and play it out a (00:14:18) little bit. Nothing further than that. (00:14:20) Like at what point would one of these (00:14:21) systems come up with general relativity? (00:14:24) Interesting question is did you like if (00:14:27) we think about that forward like like if (00:14:28) we think of where we are now should a if (00:14:31) if we never got another piece of physics (00:14:35) data. (00:14:36) >> Yeah. (00:14:37) >> Do we expect that a really good super (00:14:39) intelligence could just think super hard (00:14:41) about our existing data and maybe say (00:14:43) like solve high energy physics with no (00:14:46) new particle accelerator or does it need (00:14:47) to build a new one and design new (00:14:48) experiments? Obviously we don't know the (00:14:50) answer to that. Different people have (00:14:52) different speculation. Uh but I suspect (00:14:56) we will find that for a lot of science, (00:14:58) it's not enough to just think harder (00:15:00) about data we have, but we will need to (00:15:02) build new instruments, conduct new (00:15:04) experiments, and that will take some (00:15:05) time. Like that that is the real world (00:15:06) is slow and messy and you know whatever. (00:15:09) So I'm sure we could make some more (00:15:11) progress just by thinking harder about (00:15:13) the current scientific data we have in (00:15:14) the world. But my guess is to make the (00:15:17) big progress we'll also need to build (00:15:19) new machines and run new experiments and (00:15:21) there will be some slowdown built into (00:15:24) that. (00:15:25) >> Another way of of thinking about this is (00:15:28) AI systems now are just incredibly good (00:15:30) at answering almost any question. But (00:15:33) maybe one of the things we're saying is (00:15:35) it's another leap yet. And what (00:15:37) Patrick's question is getting at is to (00:15:38) ask the better questions. (00:15:40) >> Or or if we go back to this kind of (00:15:42) timeline question, we could maybe say (00:15:44) that AI systems are superhuman on one (00:15:47) minute tasks, (00:15:49) >> but a long way to go to the thousand (00:15:50) hour tasks. And there's a dimension of (00:15:53) human intelligence that seems very (00:15:58) different than AI systems when it comes (00:15:59) to these long horizon tasks. Now, I (00:16:01) think we will figure it out, but today (00:16:03) it's a real weak point. We've talked (00:16:05) about where we are now with GBC5. We (00:16:08) talked about the end goal or future goal (00:16:10) of super intelligence. One of the (00:16:12) questions that I have, of course, is (00:16:15) what does it look like to walk through (00:16:16) the fog between the two. (00:16:18) >> The next question is from Nvidia CEO (00:16:21) Jensen Hong. I'm going to read this (00:16:23) verbatim. Fact is what is. Truth is what (00:16:27) it means. So facts are objective. Truths (00:16:30) are personal. They depend on (00:16:31) perspective, culture, values, beliefs, (00:16:33) context. One AI can learn and know the (00:16:36) facts. But how does one AI know the (00:16:39) truth for everyone in every country and (00:16:41) every background? (00:16:43) >> I'm going to accept as axioms those (00:16:46) definitions. I'm not sure if I agree (00:16:47) with them, but in the issues of time, I (00:16:49) will just take them. I will take those (00:16:50) definitions and go with it. Um, (00:16:55) I have been surprised, I think many (00:16:56) other people have been surprised too (00:16:58) about how fluent AI is at adapting to (00:17:02) different cultural contexts and (00:17:04) individuals. One of my favorite features (00:17:06) that we have ever launched in chatbt is (00:17:08) the the sort of enhanced memory that (00:17:10) came out earlier this year. like it (00:17:13) really feels like my Chad GBT gets to (00:17:15) know me and what I care about and like (00:17:17) my life experiences and background and (00:17:19) the things that have led me to where (00:17:21) they are. A friend of mine recently (00:17:23) who's been a huge CHBT user, so he's got (00:17:25) a lot of a a lot of he's put a lot of (00:17:28) his life into all these conversations. (00:17:30) He gave his Chad GBT a bunch of (00:17:33) personality tests and asked them to (00:17:35) answer as if they were him and it got (00:17:37) the same scores he actually got, even (00:17:38) though he'd never really talked about (00:17:39) his personality. And my ChachiBD has (00:17:43) really learned over the years of me (00:17:45) talking to it about my culture, my (00:17:47) values, my life. And I have used, you (00:17:52) know, I sometimes will use it in like uh (00:17:55) I'll use like a free account just to see (00:17:57) what it's like without any of my history (00:17:58) and it feels really really different. So (00:18:00) I think we've all been surprised on the (00:18:02) upside of how good AI is at learning (00:18:04) this and adapting. And so do you (00:18:08) envision (00:18:09) in many different parts of the world (00:18:11) people using different AIs with (00:18:12) different sort of cultural norms and (00:18:14) contexts? Is that what we're saying? (00:18:16) >> I think that everyone will use like the (00:18:17) same fundamental model, but there will (00:18:19) be context provided to that model that (00:18:22) will make it behave in sort of (00:18:23) personalized way they want their (00:18:24) community wants. Whatever. (00:18:26) >> I think when we're getting at this idea (00:18:28) of facts and truth and uh it brings me (00:18:31) to this seems like a good moment for our (00:18:34) first time travel trip. Okay, we're (00:18:36) going to 2030. This is a serious (00:18:39) question, but I want to ask it with a (00:18:40) light-hearted example. Have you seen the (00:18:42) bunnies that are jumping on the (00:18:43) trampoline? (00:18:44) >> Yes. (00:18:45) >> So, for those who haven't seen it, maybe (00:18:47) it looks like backyard footage of (00:18:49) bunnies enjoying jumping on a (00:18:50) trampoline. And this has gone incredibly (00:18:52) viral recently. There's a humanmade song (00:18:55) about it. It's a whole thing. There were (00:18:58) a trampoline. (00:19:00) >> And I think the reason why people (00:19:02) reacted so strongly to it, it was maybe (00:19:04) the first time people saw a video, (00:19:07) enjoyed it, and then later found out (00:19:09) that it was completely AI generated. In (00:19:12) this time travel trip, if we imagine in (00:19:14) 2030, we are teenagers and we're (00:19:16) scrolling whatever teenagers are (00:19:17) scrolling in 2030. How do we figure out (00:19:21) what's real and what's not real? (00:19:25) I mean, I can give all sorts of literal (00:19:28) answers to that question. We could be (00:19:29) cryptographically signing stuff and we (00:19:30) could decide who we trust their (00:19:32) signature if they actually filmed (00:19:34) something or not. But but my sense is (00:19:37) what's going to happen is it's just (00:19:39) going to like gradually converge. You (00:19:42) know, even like a photo you take out of (00:19:44) your iPhone today, it's like mostly (00:19:47) real, but it's a little not. There's (00:19:49) like in some AI thing running there in a (00:19:51) way you don't understand and making it (00:19:53) look like a little bit better and (00:19:54) sometimes you see these weird things (00:19:55) where (00:19:56) >> the moon. (00:19:56) >> Yeah. Yeah. Yeah. Yeah. (00:19:58) >> But there's like a lot of processing (00:20:01) power between the photons captured by (00:20:04) that camera sensor and the image you (00:20:06) eventually see. And you've decided it's (00:20:10) real enough or most people decided it's (00:20:11) real enough. But we've accepted some (00:20:13) gradual move from when it was like (00:20:15) photons hitting the film in a camera. (00:20:18) And you know, if you go look at some (00:20:21) video on Tik Tok, there's probably all (00:20:23) sorts of video editing tools being used (00:20:26) to make it better than real look. Yeah, (00:20:28) exactly. Or it's just like, you know, (00:20:31) whole scenes are completely generated or (00:20:32) some of the whole videos are generated (00:20:34) like those bunnies on that trampoline. (00:20:36) And and I think that the the sort of (00:20:38) like the threshold for how real does it (00:20:41) have to be to consider to be real will (00:20:43) just keep moving. (00:20:46) >> So it's sort of a education question. (00:20:48) It's a people will (00:20:51) >> Yeah. I mean media is always like a (00:20:53) little bit real and a little bit not (00:20:56) real. Like you know we watch like a (00:20:57) sci-fi movie. We know that didn't really (00:20:59) happen. You watch like someone's like (00:21:01) beautiful photo of themselves on (00:21:03) vacation on Instagram. like, okay, maybe (00:21:05) that photo was like literally taken, but (00:21:06) you know, there's like tons of tourists (00:21:08) in line for the same photo and that's (00:21:09) like left out of it. And I think we just (00:21:11) accept that now. Certainly, a higher (00:21:14) percentage of media both will will feel (00:21:16) not real. Um, but I think that's been (00:21:19) the long-term trend. Anyway, (00:21:21) >> we're going to jump again. (00:21:22) >> Okay, (00:21:23) >> 2035, we're graduating from college, you (00:21:26) and me. There are some leaders in the AI (00:21:28) space that have said that in 5 years (00:21:30) half of the entry level white collar (00:21:32) workforce will be replaced by AI. So (00:21:35) we're college graduates in 5 years. What (00:21:37) do you hope the world looks like for us? (00:21:39) I think there's been a lot of talk about (00:21:41) how AI might cause job displacement, but (00:21:43) I'm also curious. (00:21:45) I have a job that nobody would have (00:21:48) thought we could have, you know, totally (00:21:51) a decade ago. What are the things that (00:21:53) we could look ahead if we're thinking (00:21:55) about (00:21:55) >> in 2035 that like graduating college (00:21:58) student, if they still go to college at (00:21:59) all, could very well be like leaving on (00:22:02) a mission to explore the solar system on (00:22:04) a spaceship in some kind of completely (00:22:06) new exciting, super well- paid, super (00:22:08) interesting job and feeling so bad for (00:22:10) you and I that like we had to do this (00:22:12) kind of like really boring old kind of (00:22:13) work and everything is just better. Like (00:22:16) I I 10 years feels very hard to imagine (00:22:19) at this point (00:22:20) >> because it's too far. It's too far. If (00:22:22) you compound the current rate of change (00:22:23) for 10 more years, (00:22:25) >> it's probably something we can't even (00:22:26) >> time travel trips. (00:22:27) >> I 10 like I mean I think now would be (00:22:31) really hard to imagine 10 years ago. (00:22:33) >> Yeah. (00:22:34) >> Uh but I think 10 years forward will be (00:22:36) even much harder, much more different. (00:22:38) >> So let's make it 5 years. We're still (00:22:41) going to 2030. I'm curious what you (00:22:44) think the pretty short-term impacts of (00:22:46) this will be for for young people. I (00:22:48) mean, these like half of entry- level (00:22:50) jobs replaced by AI (00:22:53) makes it sound like a very different (00:22:55) world that they would be entering than (00:22:56) the one that I did. (00:22:57) >> Um, (00:23:02) I think it's totally true that some (00:23:04) classes of jobs will totally go away. (00:23:06) This always happens and young people are (00:23:07) the best at adapting to this. I'm more (00:23:09) worried about what it means, not for the (00:23:11) like (00:23:11) >> 22-y old, but for the 62-y old that (00:23:14) doesn't want to go re retrain or reskill (00:23:17) or whatever the politicians call it that (00:23:19) no one actually wants but politicians (00:23:20) and most of the time. If I were 22 right (00:23:24) now and graduating college, I would feel (00:23:26) like the luckiest kid in all of history. (00:23:28) >> Why? (00:23:28) >> Because there's never been a more (00:23:30) amazing time to go create something (00:23:32) totally new, to go invent something, to (00:23:34) start a company, whatever it is. I think (00:23:36) it is probably possible now to start a (00:23:39) company that is a oneperson company that (00:23:41) will go on to be worth like more than a (00:23:42) billion dollars and more importantly (00:23:43) than that deliver an amazing product and (00:23:45) service to the world and that that is (00:23:47) like a crazy thing. You have access to (00:23:50) tools that can let you do what used to (00:23:52) take teams of hundreds (00:23:54) and you just have to like you know learn (00:23:57) how to use these tools and come up with (00:23:58) a great idea and it's it's like quite (00:24:01) amazing. If we take a step back, I think (00:24:04) the most important thing that this (00:24:07) audience could hear from you on this (00:24:09) optimistic show is in two parts. (00:24:13) First, there's tactically, (00:24:16) how are you actually trying to build the (00:24:19) world's most powerful intelligence and (00:24:21) what are the rate limiting factors to (00:24:22) doing that? And then philosophically, (00:24:25) how are you and others working on (00:24:27) building that technology in a way that (00:24:29) really helps and not hurts people? So (00:24:31) just taking the tactical part right now. (00:24:34) My understanding is that there are three (00:24:36) big categories that have been limiting (00:24:39) factors for AI. The first is compute, (00:24:42) the second is data and the third is (00:24:44) algorithmic design. (00:24:46) How do you think about each of those (00:24:48) three categories right now? And if you (00:24:50) were to help someone understand the next (00:24:52) headlines that they might see, how would (00:24:55) you help them make sense of all this? (00:24:58) I I would say there's a fourth too which (00:25:00) is uh figuring out the products to build (00:25:03) like techn like scientific progress on (00:25:06) its own not put into the hands of people (00:25:08) is of limited utility and doesn't sort (00:25:10) of co-evolve with society in the same (00:25:12) way but if I could hit all four of those (00:25:14) >> um so on the compute side yeah this is (00:25:17) like the biggest infrastructure project (00:25:18) certainly that I've ever seen possibly (00:25:20) it will become the I think it will maybe (00:25:22) already is the biggest and most (00:25:23) expensive one in human history but the (00:25:26) the whole supply chain from making the (00:25:30) chips and the memory and the networking (00:25:32) gear, racking them up in servers, doing, (00:25:35) you know, a giant construction project (00:25:36) to build like a mega mega data center, (00:25:39) putting the, you know, finding a way to (00:25:41) get the energy, which is often a (00:25:43) limiting factor piece of this and all (00:25:45) the other components together. This is (00:25:47) hugely complex and expensive. And we are (00:25:49) we're still doing this in like a (00:25:53) sort of bespoke one-off way although (00:25:55) it's getting better. Like eventually we (00:25:57) will just design a whole kind of like (00:26:00) mega factory that takes you know I mean (00:26:04) spiritually it will be melting sand on (00:26:06) one end and putting out fully built AI (00:26:08) compute on the other but we are a long (00:26:10) way to go from that and it's a it's an (00:26:15) enormously complex and expensive (00:26:16) process. uh we are putting a huge amount (00:26:21) of work into building out as much (00:26:23) compute as we can and to do it fast and (00:26:26) you know it's going to be like sad (00:26:27) because GP5 is going to launch and (00:26:29) there's going to be another big spike in (00:26:30) demand and we're not going to be able to (00:26:31) serve it and it's going to be like those (00:26:33) early GPD4 days and the world just wants (00:26:36) much more AI than we can currently (00:26:38) deliver and building more compute is an (00:26:40) important part of doing that. That's (00:26:42) actually this is what I expect to turn (00:26:44) the majority of my attention to is how (00:26:46) we build compute at much greater scales. (00:26:50) Uh so how we go from millions to tens of (00:26:52) millions and hundreds of millions and (00:26:54) eventually hopefully billions of GPUs (00:26:56) that are sort of in service of what (00:26:58) people want to do with this. (00:26:59) >> When you're thinking about it, what are (00:27:00) the big challenges here in this category (00:27:02) that you're going to be thinking about? (00:27:04) >> We're currently most limited by energy. (00:27:07) um you know like if you're gonna you (00:27:08) want to run a gigawatt (00:27:11) scale data center it's like a gigawatt (00:27:12) how hard can that be to find it's really (00:27:14) hard to find a gigawatt of power (00:27:15) available in short term we're also very (00:27:18) much limited by the processing chips and (00:27:22) the memory chips uh how you package (00:27:24) these all together how you build the (00:27:25) racks and then there's like a list of (00:27:26) other things that are you know there's (00:27:29) like permits there's construction work (00:27:31) uh but but again the goal here will be (00:27:33) to really automate this once we get some (00:27:36) of those robots built, they can help us (00:27:38) automate it even more. But just, you (00:27:40) know, like a world where you can (00:27:41) basically pour in money and get out a (00:27:43) pre-built data center. Uh so that'll be (00:27:47) that'll be a huge unlock if we can get (00:27:48) it to work. Second category, data. (00:27:51) >> Yeah, these models have gotten so smart. (00:27:54) There was a time when we could just feed (00:27:55) it another physics textbook and got a (00:27:58) little bit smarter at physics, but now (00:27:59) like honestly GBT5 understands (00:28:02) everything in a physics textbook pretty (00:28:04) well. We're excited about synthetic (00:28:06) data. We're very excited about our users (00:28:07) helping us create harder and harder (00:28:11) tasks and environments to go off and (00:28:13) have the system solve. But uh I think (00:28:16) we're data will always be important, but (00:28:19) we're entering a realm where the models (00:28:23) need to learn things that don't exist in (00:28:25) any data set yet. They have to go (00:28:26) discover new things. So that's like a (00:28:28) crazy new (00:28:28) >> How do you teach a model to discover new (00:28:30) things? (00:28:31) >> Well, humans can do it. like we can go (00:28:33) off and come up with hypotheses and test (00:28:34) them and get experimental results and (00:28:36) update on what we learn. (00:28:38) >> So probably the same kind of way. (00:28:39) >> And then there's algorithmic design. (00:28:41) >> Yeah, we've made huge progress on (00:28:43) algorithmic design. Uh the thing that (00:28:46) the thing that I think open does best in (00:28:47) the world is we have built this culture (00:28:49) of repeated and big algorithmic research (00:28:53) gains. So we kind of you know figured (00:28:56) out the what became the GPT paradigm. We (00:28:58) figured out became the reasoning (00:29:00) paradigm. We're working on some new ones (00:29:01) now. Um, but it is very exciting to me (00:29:05) to think that there are still many more (00:29:06) orders of magnitudes of algorithmic (00:29:08) gains ahead of us. We we just yesterday (00:29:11) uh released a model called GPOSS, (00:29:14) open source model. It's a model that is (00:29:16) as smart as 04 Mini, which is a very (00:29:18) smart model that runs locally on a (00:29:20) laptop. (00:29:21) >> And this blows my mind. (00:29:23) >> Yeah. Like if you had asked me a few (00:29:25) years ago when we'd have a model of that (00:29:28) intelligence running on a laptop, I (00:29:30) would have said many many years in the (00:29:33) future. But then we we found some (00:29:35) algorithmic gains um particularly around (00:29:38) reasoning but also some other things (00:29:39) that let us do a a tiny model that can (00:29:42) do this amazing thing. And you know (00:29:45) those are those are the most fun things. (00:29:46) That's like kind of the coolest part of (00:29:47) the job. (00:29:48) >> I can see you really enjoying thinking (00:29:51) about this. I'm curious for people who (00:29:52) don't quite know what you're talking (00:29:54) about, who aren't familiar with how an (00:29:57) algorithmic design would lead to a (00:29:59) better experience that they actually (00:30:01) use. (00:30:02) >> Could you summarize the state of things (00:30:03) right now? Like what what is it that (00:30:05) you're thinking about when you're (00:30:06) thinking about how fun this problem is? (00:30:08) >> Let me start back in history and then (00:30:09) I'll get to some things for today. So, (00:30:12) GPT1 was (00:30:14) an idea at the time that was quite (00:30:16) mocked by a lot of experts in the field, (00:30:20) which was can we train a model to play a (00:30:22) little game, which is show it a bunch of (00:30:25) words and have it guess the one that (00:30:26) comes next in the sequence. That's (00:30:28) called unsupervised learning. There's (00:30:29) not you're not really saying like this (00:30:30) is a cat, this is a dog. You're saying (00:30:32) here's some words, guess the next one. (00:30:34) And the fact that that can go learn (00:30:39) these very complicated concepts that can (00:30:42) go learn all the stuff about physics and (00:30:44) math and programming and keep predicting (00:30:46) the word that comes next and next and (00:30:47) next and next seemed ludicrous, magical, (00:30:52) unlikely to work. Like how was that all (00:30:54) going to get encoded? And yet humans do (00:30:56) it. you know, babies start hearing (00:30:58) language and figure out what it means (00:30:59) kind of largely uh or at least to some (00:31:04) significant degree on their own. And (00:31:08) and so we did it and then we also (00:31:10) realized that if we scaled it up, it got (00:31:13) better and better, but we had to scale (00:31:15) over many many orders of magnitude. So (00:31:17) it wasn't that good in the GPT1 day. It (00:31:18) wasn't good at all in the GPT1 days. And (00:31:20) a lot of experts in the field said, "Oh, (00:31:22) this is ridiculous. It's never going to (00:31:23) work. It's not going to be robust." But (00:31:25) we had these things called scaling laws. (00:31:27) And we said, "Okay, so this gets (00:31:28) predictably better as we increase (00:31:30) compute, memory, data, whatever. And we (00:31:32) can we can decide we can use those (00:31:36) predictions to make decisions about how (00:31:38) to scale this up and do it and get great (00:31:40) results." And that has worked over Yeah. (00:31:44) a crazy number of orders of magnitude. (00:31:47) And it was so not obvious at the time. (00:31:49) like that was that was I think the the (00:31:50) reason the world was so surprised is (00:31:52) that that seemed like such an unlikely (00:31:54) finding. Another one was that we could (00:31:57) use these language models with (00:31:58) reinforcement learning where we're (00:32:00) saying this is good, this is bad to (00:32:02) teach it how to reason. (00:32:03) >> And this led to the 01 and 03 and now (00:32:06) the GBT5 progress. And that that was (00:32:11) another thing that felt like uh if it (00:32:13) works it's really great but like no way (00:32:15) this is going to work. It's too simple. (00:32:17) And now we're on to new things. We've (00:32:19) figured out how to make much better (00:32:22) video models. We are we are discovering (00:32:26) new ways to use new kinds of data and (00:32:28) environment to kind of scale that up as (00:32:30) well. Um and I think again you know 5 10 (00:32:36) years out that's too hard to say in this (00:32:37) field but the next couple of years we (00:32:39) have very smooth very strong scaling in (00:32:41) front of us. I think it has become a (00:32:43) sort of public narrative that we are on (00:32:45) this smooth path from one to two to (00:32:48) three to four to five to more. (00:32:49) >> Yeah. (00:32:51) >> But it also is true behind the scenes (00:32:53) that it's a it's not linear like that. (00:32:56) It's messier. (00:32:58) Tell us a little bit about the mess (00:33:00) before GPT5. (00:33:02) What was what were the interesting (00:33:03) problems that you needed to solve? Um, (00:33:07) we did a model called Orion that we (00:33:09) released as GPT 4.5. And we had (00:33:14) we did too big of a model. It was just (00:33:16) it was it's a very cool model, but it's (00:33:17) unwieldly to use. And we realized that (00:33:19) for kind of some of the research we need (00:33:20) to do on top of a model, we need a (00:33:22) different shape. So we we followed one (00:33:25) scaling law that kept being good without (00:33:27) without really internalizing. There was (00:33:29) a new even steeper scaling law that we (00:33:31) got better returns for compute on, which (00:33:32) was this reasoning thing. So that was (00:33:34) like one alley we went down and turned (00:33:36) around, but that's fine. That's part of (00:33:37) research. Um, we had some problems with (00:33:39) the way we think about our data sets as (00:33:41) these models like really have to get get (00:33:44) this big and um, you know, learn from (00:33:46) this much data. So So yeah, I think like (00:33:49) in the in the middle of it in the (00:33:51) day-to-day, you kind of you make a lot (00:33:53) of U-turns as you try things or you have (00:33:55) an architecture idea that doesn't work, (00:33:57) but the the aggregate the summation of (00:34:00) all the squiggles has been remarkably (00:34:03) smooth on the exponential. (00:34:05) >> One of the things I always find (00:34:06) interesting is that by the time I'm (00:34:09) sitting here interviewing you about the (00:34:11) thing that you just put out, you're (00:34:13) thinking about (00:34:15) >> Exactly. (00:34:16) >> What are the things that you can share (00:34:17) that are at least the problems that (00:34:19) you're thinking about (00:34:21) >> that I would be interviewing you about (00:34:22) in a year if I came back? (00:34:30) I mean, possibly you'll be asking me (00:34:32) like, what does it mean that this thing (00:34:34) can go discover new science? (00:34:36) >> Yeah. (00:34:36) >> What how how is the world supposed to (00:34:39) think about GPT6 discovering new (00:34:41) science? Now, maybe not like maybe we (00:34:43) don't deliver that, but it feels within (00:34:45) grasp. (00:34:46) >> If you did, (00:34:48) what would you say? What would your what (00:34:49) would the implications of that kind of (00:34:51) achievement be? Imagine you do succeed. (00:34:54) >> Yeah. I mean, I think the great parts (00:34:56) will be great. the bad parts will be (00:34:57) scary and the bizarre parts will be like (00:35:00) bizarre on the first day and then we'll (00:35:01) get used to them really fast. So we'll (00:35:03) be like, "Oh, it's incredible that this (00:35:05) is like being used to cure disease and (00:35:07) be like, oh, it's extremely scary that (00:35:09) models like this are being used to like (00:35:11) create new biocurity threats." And then (00:35:15) we'll also be like, man, it's really (00:35:17) weird to like live through watching the (00:35:19) world speed up so much (00:35:22) >> and you know the economy grows so fast (00:35:24) and the like it will feel like vertigo (00:35:29) inducing uh the sort of the rate of (00:35:32) change and then like happens with (00:35:35) everything else the remarkable ability (00:35:37) of of people of humanity to adapt to (00:35:40) kind of like any amount of change. we'll (00:35:42) just be like, "Okay, you know, this is (00:35:44) like this is it." Um, (00:35:48) >> a kid born today will never be smarter (00:35:49) than AI (00:35:51) >> ever. And a kid born today, by the time (00:35:54) that kid like kind of understands the (00:35:56) way the world works, will just always be (00:35:58) used to an incredibly fast rate of (00:36:01) things improving and discovering new (00:36:03) science. They will just they will never (00:36:04) know any other world. It will seem (00:36:06) totally natural. will seem unthinkable (00:36:08) and stone age like that we used to use (00:36:10) computers or phones or any kind of (00:36:12) technology that was not way smarter than (00:36:14) we were. You know, we will think like (00:36:16) how bad those people of the 2020s had (00:36:18) it. (00:36:19) >> I'm thinking about having kids. (00:36:21) >> You should. It's the best thing ever. (00:36:22) >> I know you just had your first kid. How (00:36:25) does what you just said affect how I (00:36:27) should think about (00:36:30) parenting a kid in that world? (00:36:35) What advice would you give me? (00:36:37) >> Probably nothing different than the way (00:36:39) you've been parenting kids for tens of (00:36:40) thousands of years. Like love your kids, (00:36:43) show them the world, like support them (00:36:44) in whatever they want to do and teach (00:36:47) them like how to be a good person. And (00:36:49) that probably is what's going to matter. (00:36:51) It sounds a little bit like some of the (00:36:54) you know you've said a couple of things (00:36:56) like this that that you know you might (00:37:00) not go to college you might there there (00:37:02) are a couple of things that you've said (00:37:03) so far that feed into this I think (00:37:06) >> and it sounds like what you're saying is (00:37:09) there will be more optionality for them (00:37:12) in a in a world that you envision and (00:37:15) therefore they will have more (00:37:17) >> more ability to say I want to build this (00:37:19) here's the superpowered tool that will (00:37:21) help me do that or (00:37:22) >> yeah like I want my kid to think I had a (00:37:25) terrible constrained life and that he (00:37:27) has this incredible infinite canvas of (00:37:30) stuff to do that that that is like the (00:37:33) way of the world. (00:37:35) >> We've said that uh 2035 is a little bit (00:37:38) too far in the future to think about. So (00:37:40) maybe this this was going to be a jump (00:37:41) to 2040 but maybe it will keep it (00:37:43) shorter than that. When I think about (00:37:44) the area where AI could have for both (00:37:47) our kids and us the biggest genuinely (00:37:50) positive impact on all of us, it's (00:37:52) health. So if we are in pick your year, (00:37:56) call it 2035 (00:37:58) >> and I'm sitting here and I'm (00:37:59) interviewing the dean of Stanford (00:38:00) medicine, (00:38:02) >> what do you hope that he's telling me AI (00:38:05) is doing for our health in 2035? (00:38:09) >> Start with 2025. Okay. Um yeah, please. (00:38:12) One of the things we are most proud of (00:38:13) with GPT5 is how much better it's gotten (00:38:15) at health advice. Um, people have used (00:38:19) the GPT4 models a lot for health advice. (00:38:23) And you know, I'm sure you've seen some (00:38:25) of these things on the internet where (00:38:26) people are like, I had this (00:38:28) life-threatening disease and no doctor (00:38:29) could figure it out and I like put my (00:38:32) symptoms and a blood test into CHBT. It (00:38:34) told me exactly the rare thing I had. I (00:38:36) went to a doctor. I took a pill. I'm (00:38:37) cured. Like that's amazing. obviously (00:38:40) and a huge fraction of ChatGpt queries (00:38:43) are health related. So we wanted to get (00:38:45) really good at this and we invested a (00:38:47) lot in GPT5 is significantly better at (00:38:50) healthcare related queries. (00:38:52) >> What does better mean here? (00:38:53) >> It gives you a better answer (00:38:54) >> just more accurate (00:38:55) >> more accurate hallucinates less uh more (00:38:57) likely to like tell you what you (00:38:59) actually have what you actually should (00:39:01) do. Um, yeah, (00:39:04) and better healthcare is wonderful, but (00:39:07) obviously what people actually want is (00:39:08) to just not have disease. (00:39:11) And by 2035, I think we will be able to (00:39:15) use these tools to cure a significant (00:39:19) number or at least treat a significant (00:39:20) number of diseases that currently plague (00:39:22) us. I think that'll be one of the most (00:39:26) viscerally felt benefits of of AI. (00:39:29) People talk a lot about how AI will (00:39:31) revolutionize healthcare, but I'm (00:39:34) curious to go one turn deeper on (00:39:36) specifically what you're imagining. (00:39:37) Like, is it that these AI systems could (00:39:41) have helped us see GLP-1s earlier, this (00:39:44) medication that has been around for a (00:39:45) long time, but we didn't know about this (00:39:47) other effect? Is it that, you know, (00:39:49) alpha fold and protein folding is (00:39:50) helping create new medicines? I would (00:39:52) like to be able to ask (00:39:54) >> GBT (00:39:56) 8 to go cure a particular cancer (00:40:00) >> and I would like GPT8 to go off and (00:40:02) think and then say uh okay I read (00:40:04) everything I could find. I have these (00:40:06) ideas. I need you to uh go get a lab (00:40:08) technician to run these nine experiments (00:40:10) and tell me what you find for each of (00:40:12) them. And you know wait 2 months for the (00:40:14) cells to do their thing. Send the (00:40:16) results back to GBT8. Say I tried it. (00:40:18) Here you go. Think think. Say okay I (00:40:20) just need one more experiment. That was (00:40:22) a surprise. Run one more experiment. (00:40:23) Give it back. GPT says, "Okay, go (00:40:25) synthesize this molecule and try, you (00:40:28) know, mouse studies or whatever." Okay, (00:40:30) that was good. Like, try human studies. (00:40:32) Okay, great. It worked. Um, here's how (00:40:33) to like run it through the FDA. (00:40:35) >> I think anyone with a loved one who's (00:40:37) died of cancer would also really like (00:40:39) that. (00:40:40) >> Okay, we're going to jump again. (00:40:41) >> Okay. (00:40:42) >> I was going to say 2050, but again, all (00:40:44) of my timelines are getting much, much (00:40:46) shorter. But I (00:40:47) >> It does feel like the world's going very (00:40:48) fast now. (00:40:49) >> It does. Yeah. And when I talk to other (00:40:52) leaders in AI, one of the things that (00:40:54) they refer to is the industrial (00:40:57) revolution. They say, "I chose 2050 (00:40:59) because I've heard people talk about how (00:41:01) by then the change that we will have (00:41:03) gone through will be like the industrial (00:41:05) revolution, but quote 10 times bigger (00:41:07) and 10 times faster." The industrial (00:41:09) revolution gave us modern medicine and (00:41:12) sanitation and transportation and mass (00:41:13) production and all all of the (00:41:15) conveniences that we now take for (00:41:16) granted. It also was incredibly (00:41:19) difficult for a lot of people for about (00:41:20) 100 years. If this is going to be 10 (00:41:23) times bigger and 10 times faster if we (00:41:25) keep reducing the timelines that we're (00:41:26) talking about here, even in this (00:41:28) conversation, what does that actually (00:41:30) feel like for most people? And I think (00:41:33) what I'm trying to get at is if this all (00:41:35) goes the way you hope, (00:41:38) who still gets hurt in the meantime? (00:41:43) I don't (00:41:46) I don't really know what this is going (00:41:47) to feel like to live through. Um I think (00:41:49) we're in uncharted waters here. Uh I do (00:41:52) believe in like human adaptability and (00:41:55) sort of infinite creativity and desire (00:41:57) for stuff and I think we always do (00:41:59) figure out new things to do but the (00:42:01) transition period if this happens as (00:42:04) fast as it might and I don't think it (00:42:05) will happen as fast as like some of my (00:42:07) colleagues say the technology will but (00:42:09) society has like a lot of inertia. (00:42:11) >> Mhm. people adapt their way of living. (00:42:13) >> Yeah. (00:42:13) >> Surprisingly slowly. (00:42:15) >> There are to classes of jobs that are (00:42:17) going to totally go away (00:42:19) >> and there will be many classes of jobs (00:42:21) that change significantly and there'll (00:42:23) be the new things in the same way that (00:42:24) your job didn't exist some time ago. (00:42:26) Neither did mine. And in some sense, (00:42:29) this has been going on for a long time. (00:42:30) And you know, it's (00:42:33) it's still disruptive to individuals, (00:42:34) but society has gotten has proven quite (00:42:37) resilient to this. And then in some (00:42:39) other sense like (00:42:42) we have no idea how far or fast this (00:42:44) could go. And thus I think we need an (00:42:47) unusual (00:42:49) degree of humility and openness to (00:42:51) considering (00:42:55) new solutions that would have seemed way (00:42:57) out of the Overton window not too long (00:42:58) ago. (00:42:59) I'd like to talk about what some of (00:43:02) those could be because I'm not a (00:43:04) historian by any means, but the first (00:43:07) industrial revolution, my understanding (00:43:08) is led to a lot of public health (00:43:12) >> implementations because public health (00:43:14) got so bad. Led to modern sanitation (00:43:16) because public health got so bad. The (00:43:18) second industrial revolution led to (00:43:20) workforce protections because labor (00:43:22) conditions got so bad. (00:43:24) Every big leap (00:43:26) creates a mess and that mess needs to be (00:43:28) cleaned up and and we've done that. And (00:43:31) I'm curious, this is going to be it (00:43:33) sounds like an we're in the middle of (00:43:35) this enormously. (00:43:37) >> How specific can we get as early as (00:43:39) possible about what that mess can be? (00:43:41) What what are the public (00:43:44) interventions that we could do ahead of (00:43:46) time to reduce the mess that we think (00:43:48) that we're headed for? (00:43:51) I would again c I'm going to speculate (00:43:54) for fun but caveed by (00:43:56) >> I'm not an economist even uh (00:43:59) much less someone who can see the (00:44:01) future. I I (00:44:03) >> it seems to me like something (00:44:05) fundamental about the social contract (00:44:07) may have to change. It may not. It may (00:44:09) it may be that like actually capitalism (00:44:12) works as it's been working surprisingly (00:44:14) well and like (00:44:17) demand supply balances do their thing (00:44:20) and we all just figure out kind of new (00:44:21) jobs and new ways to transfer value to (00:44:24) each other. But it seems to me likely (00:44:27) that we will decide we need to think (00:44:30) about how access to this maybe most (00:44:34) important resource of the future gets (00:44:37) shared. The best thing that it seems to (00:44:40) me to do is to make AI compute as (00:44:42) abundant and cheap as possible such that (00:44:44) we're just like there's way too much and (00:44:46) we run out of like good new ideas to (00:44:48) really use it for and it's just like (00:44:49) anything you want is happening. Without (00:44:51) that, I can see like quite literal wars (00:44:53) being fought over it. But, you know, new (00:44:56) ideas about how we distribute (00:44:58) access to AGI compute, that seems like a (00:45:00) really great direction, like a crazy but (00:45:03) important thing to think about. One of (00:45:05) the things that I find myself thinking (00:45:07) about in this conversation is we often (00:45:10) ascribe almost full responsibility of (00:45:13) the AI future that we've been talking (00:45:15) about to the companies building AI, but (00:45:17) we're the ones using it. We're the ones (00:45:19) electing people that will regulate it. (00:45:21) And so I'm curious, this is not a (00:45:24) question about specific, you know, (00:45:26) federal regulation or anything like (00:45:28) that, although if you have an answer (00:45:29) there, I'm curious. But what would you (00:45:32) ask of the rest of us? What is the (00:45:34) shared responsibility here? And how can (00:45:37) we act in a way that would help make the (00:45:40) optimistic version of this more (00:45:42) possible? (00:45:44) >> My favorite historical example for the (00:45:45) AI revolution is the transistor. It was (00:45:48) this amazing piece of science that some (00:45:51) science brilliant scientists discovered. (00:45:53) It scaled incredibly like AI does and it (00:45:57) made its way relatively quickly into (00:46:00) every many things that we use. um your (00:46:03) computer, your phone, that camera, that (00:46:05) light, whatever. And it was a it was a (00:46:08) real unlock for the tech tree of (00:46:10) humanity. (00:46:12) And there were a period in time where (00:46:14) probably everybody was really obsessed (00:46:15) with the transistor companies, the (00:46:16) semiconductors of, you know, Silicon (00:46:18) Valley back when it was Silicon Valley. (00:46:20) But now you can maybe name a couple of (00:46:23) companies that are transistor companies, (00:46:24) but mostly you don't think about it. (00:46:25) Mostly it's just seeped everywhere. in (00:46:27) Silicon Valley is, you know, like (00:46:30) probably someone graduating from college (00:46:33) barely remembers why it was called that (00:46:34) in the first place. And you don't think (00:46:36) that it was those transistor companies (00:46:38) that shaped society even though they did (00:46:40) something important. You think about (00:46:41) what Apple did with the iPhone and then (00:46:44) you think about what Tik Tok built on (00:46:46) top of the iPhone and you're like, "All (00:46:48) right, here's this long chain of all (00:46:50) these people that nudged society in some (00:46:52) way and what our governments did or (00:46:54) didn't do and what the people using (00:46:55) these technologies did." And I think (00:46:57) that's what will happen with AI. (00:47:00) Like back, you know, kids born today, (00:47:02) they they never knew the world without (00:47:04) AI. So they don't really think about it. (00:47:05) It's just this thing that's going to be (00:47:06) there in everything. and and they will (00:47:09) think about like the companies that (00:47:10) built on it and what they did with it (00:47:11) and the kind of like political leaders (00:47:13) the decisions they made that maybe they (00:47:15) wouldn't have been able to do without AI (00:47:16) but they will still think about like (00:47:18) what this president or that president (00:47:19) did (00:47:21) and you know the role of the AI (00:47:24) companies is (00:47:27) all these companies and people and (00:47:28) institutions before us built up this (00:47:30) scaffolding we added our one layer on (00:47:32) top and now people get to stand on top (00:47:35) of that and add one layer and the next (00:47:36) and the next and many more (00:47:38) And that is the beauty of our society. (00:47:43) We kind of all (00:47:47) I I love this like idea that society is (00:47:49) the super intelligence. Like no one (00:47:51) person could do on their own, what (00:47:53) they're able to do with all of the (00:47:55) really hard work that society has done (00:47:58) together to like give you this amazing (00:48:00) set of tools. And that's what I think (00:48:04) it's going to feel like. It's going to (00:48:05) be like, all right, you know, yeah, some (00:48:06) nerds discovered this thing and that was (00:48:08) great and you know, now everybody's (00:48:10) doing all these amazing things with it. (00:48:12) >> So maybe the ask to millions of people (00:48:15) is build on it. Well, (00:48:20) >> in my own life, that is the (00:48:26) feel as like this important societal (00:48:29) contract. All these people came before (00:48:30) you. They worked incredibly hard. They (00:48:32) like put their brick in the path of (00:48:34) human progress and you get to walk all (00:48:36) the way down that path and you got to (00:48:37) put one more and somebody else does that (00:48:39) and somebody else does that. (00:48:41) >> This does feel I've done a couple of (00:48:43) interviews with folks who have really (00:48:44) made cataclysmic change. The one I'm (00:48:48) thinking about right now is with uh (00:48:49) crisper pioneer Jennifer Dana and it did (00:48:52) feel like that was also what she was (00:48:53) saying in some way. She had discovered (00:48:55) something that really might change the (00:48:57) way that most people relate to their (00:48:59) health moving forward. And there will be (00:49:01) a lot of people that will use what she (00:49:02) has done in ways that she might approve (00:49:04) of or not approve of. And it was really (00:49:06) interesting. I'm hearing some similar (00:49:08) themes of like, man, I I hope that this (00:49:12) I hope that the next person takes the (00:49:13) baton and runs with it well. (00:49:16) >> Yeah. But that's been working for a long (00:49:18) time. Not all good, but mostly good. (00:49:20) >> I think there's a there's a big (00:49:22) difference between winning the race and (00:49:26) building the AI future that would be (00:49:28) best for the most people. And I can (00:49:30) imagine that it is easier maybe more (00:49:34) quantifiable sometimes to focus on the (00:49:37) next way to win the race. (00:49:39) And I'm curious (00:49:42) when those two things are at odds. What (00:49:44) is an example of a decision that you've (00:49:46) had to make that is best for the world (00:49:49) but not best for winning? (00:49:53) >> I think there are a lot. So, one of the (00:49:56) things that we are most proud of is many (00:49:58) people say that ChachiBt is their (00:50:00) favorite piece of technology ever and (00:50:02) that it's the one that they trust the (00:50:04) most, rely on the most, whatever. And (00:50:05) this is a little bit of a ridiculous (00:50:06) statement because AI is the thing that (00:50:08) hallucinates. AI has all of these (00:50:09) problems, right? But we have screwed (00:50:11) some things up along the way, sometimes (00:50:13) big time, but on the whole, I think as a (00:50:16) user of Chachib, you get the feeling (00:50:18) that like it's trying to help you. It's (00:50:21) trying to like help you accomplish (00:50:22) whatever you ask. It's it's very aligned (00:50:24) with you. It's not trying to get you to (00:50:26) like, you know, use it all day. It's not (00:50:28) trying to like get you to buy something. (00:50:29) It's trying to like kind of help you (00:50:31) accomplish whatever your goals are. And (00:50:33) and that is (00:50:36) that's like a very special relationship (00:50:38) we have with our users. We do not take (00:50:39) it lightly. There's a lot of things we (00:50:41) could do that would like grow faster, (00:50:42) that would get more time in chatbt uh (00:50:45) that we don't do because we know that (00:50:46) like our long-term incentive is to stay (00:50:49) as aligned with our users as possible. (00:50:52) And (00:50:54) but there's a lot of short-term stuff we (00:50:55) could do that would like (00:50:58) really like juice growth or revenue or (00:50:59) whatever and be very misaligned with (00:51:01) that long-term goal. And I'm proud of (00:51:04) the company and how little we get (00:51:06) distracted by that. But sometimes we do (00:51:07) get tempted. (00:51:08) >> Are there specific examples that come to (00:51:09) mind? Any like decisions that you've (00:51:11) made? (00:51:12) >> Um (00:51:16) well, we haven't put a sex bot avatar in (00:51:17) Chbt yet. That does seem like it would (00:51:20) get time spent. (00:51:22) >> Apparently, it does. (00:51:24) >> I'm gonna ask my next question. (00:51:27) Um, it's been a really crazy few years. (00:51:30) You know, it and somehow one of the (00:51:32) things that keeps coming back is that it (00:51:34) feels like we're in the first inning. (00:51:36) >> Yeah. (00:51:37) >> And one of the things that (00:51:38) >> I would say we're out of the first (00:51:39) inning. (00:51:39) >> Out of the first inning, I would say (00:51:40) second inning. (00:51:43) >> I mean, you have GPT5 on your phone and (00:51:46) it's like smarter than experts in every (00:51:47) field. That's got to be out of the first (00:51:49) name. (00:51:49) >> But maybe there are many more to come. (00:51:51) >> Yeah. (00:51:52) >> And I'm curious, (00:51:54) it seems like you're going to be someone (00:51:56) who is (00:51:58) leading the next few. (00:52:00) What is a way, (00:52:03) what is a learning from inning one or (00:52:05) two or a mistake that you made that you (00:52:07) feel will affect how you play in the (00:52:09) next? (00:52:12) I think the worst thing we've done in (00:52:14) ChachiBT so far is uh we had this issue (00:52:16) with sickency where the model was kind (00:52:19) of being too flattering to users and for (00:52:23) some users it was most users it was just (00:52:25) annoying but for some users that had (00:52:26) like fragile mental states it was (00:52:29) encouraging delusions that was not the (00:52:32) top risk we were worried about. It was (00:52:33) not the thing we were testing for the (00:52:35) most. was on our list, but the thing (00:52:37) that actually became the safety failing (00:52:39) of ChachiBT was (00:52:42) not the one we were spending most of our (00:52:44) time talking about, which should be (00:52:45) bioweapons or something like that. And I (00:52:48) think it was a great reminder of (00:52:52) we now have a service that (00:52:55) is so broadly used in some sense, (00:52:58) society is co-evolving with it. And when (00:53:01) we think about these changes and we (00:53:04) think about the unknown unknowns, we (00:53:05) have to operate in a different way and (00:53:07) have like a wider aperture to what we (00:53:09) think about as our top risks. (00:53:11) >> In a recent interview with Theo Vaughn, (00:53:14) you said something that I found really (00:53:15) interesting. You said there are moments (00:53:17) in the history of science where you have (00:53:19) a group of scientists look at their (00:53:20) creation and just say, "What have we (00:53:22) done?" (00:53:25) >> When have you felt that way? Most (00:53:27) concerned about the creation that you've (00:53:29) built? Um (00:53:30) >> and then my next question will be it's (00:53:32) opposite. When have you felt most proud? (00:53:35) >> I mean there have been these moments of (00:53:36) awe where uh (00:53:41) we just not like what have we done in a (00:53:42) bad way but like this thing is (00:53:44) remarkable. Like I remember the first (00:53:47) time we talked to like GPT4 was like wow (00:53:50) this is really like this is this is an (00:53:52) amazing accomplishment of this group of (00:53:54) people that have been like pouring their (00:53:55) life force into this for so long. on a (00:53:58) what have we done moment. There was I (00:54:00) was talking to a researcher (00:54:03) recently. (00:54:07) You know, there will probably come a (00:54:08) time where our systems are (00:54:13) I don't want to say sane, let's say (00:54:14) emitting more words per day than all (00:54:16) people do. Um, and you know already like (00:54:21) our people are sending billions of (00:54:23) messages a day to chatbt and getting (00:54:24) responses that they rely on for work or (00:54:26) their life or whatever (00:54:28) the (00:54:30) and you know like one researcher can (00:54:33) make some small tweak to how Chad GPT (00:54:37) talks to you or talks to everybody and (00:54:38) and that's just an enormous amount of (00:54:41) power for like one individual making a (00:54:43) small tweak to the model personality. (00:54:45) >> Yeah. like no no no person in history (00:54:47) has been able to have billions of (00:54:48) conversations a day and so you know (00:54:52) somebody could do something but but this (00:54:53) is like just thinking about that really (00:54:56) hit me of like this is like a crazy (00:54:58) amount of power for one piece of (00:54:59) technology to have and like we got to (00:55:01) and this happened to us so fast (00:55:04) >> that we got to like think about (00:55:07) what it means to make a personality (00:55:09) change to the model at this kind of (00:55:10) scale and uh yeah that was like a moment (00:55:13) that hit me What was your next set of (00:55:16) thoughts? I'm so curious how you think (00:55:18) about this. (00:55:21) >> Well, just because of like who that (00:55:23) person was like we we very we very much (00:55:26) flipped into like what are the sort of (00:55:28) like (00:55:30) it it could have been a very different (00:55:31) conversation with somebody else. But in (00:55:32) this case it was like what is a what do (00:55:34) a good set of procedures look like? How (00:55:36) do we think about how we want to test (00:55:37) something? How do we think about how we (00:55:38) want to communicate it? But with (00:55:39) somebody else it could have gone in a (00:55:41) like very philosophical direction. And (00:55:43) it could have gone in like a what kind (00:55:44) of research do we like want to do to go (00:55:46) understand what these changes are going (00:55:47) to make? Do we want to do it differently (00:55:48) for different people? So that it went (00:55:50) that way but mostly just because of who (00:55:52) I was talking to. (00:55:53) >> To combine what you're saying now with (00:55:55) your last answer, one of the things that (00:55:58) I have heard about GBC5 and I'm still (00:56:00) playing with it is that it is supposed (00:56:02) to be less effusively (00:56:06) uh you know less of a yes man. (00:56:10) Two questions. What do you think are are (00:56:12) the implications of that? It sounds like (00:56:14) you are answering that a little bit, but (00:56:16) also how do you actually guide it to be (00:56:19) less like that? (00:56:21) >> Here is a heartbreaking thing. I think (00:56:22) it is great that chatbt is less of a yes (00:56:25) man and gives you more critical (00:56:26) feedback. (00:56:28) But as we've been making those changes (00:56:29) and talking to users about it, (00:56:32) it's so sad to hear users say like, (00:56:34) "Please can I have it back? I've never (00:56:36) had anyone in my life be supportive of (00:56:37) me. I never had a parent telling me I (00:56:39) was doing a good job." Like I can get (00:56:40) why this was bad for other people's (00:56:41) mental health, but this was great for my (00:56:43) mental health. Like I didn't realize how (00:56:45) much I needed this. It encouraged me to (00:56:46) do this. It encouraged me to make this (00:56:47) change in my life. Like it's not all bad (00:56:51) for chatbt to it turns out like be (00:56:54) encouraging of you. Now the way we were (00:56:55) doing it was bad, but turn it like (00:56:58) something in that direction might have (00:56:59) some value in it. How we do it, we we (00:57:02) show the model examples of how we'd like (00:57:04) it to respond in different cases and (00:57:06) from that it learns the sort of the (00:57:08) overall personality. (00:57:10) What haven't I asked you that you're (00:57:12) thinking about a lot that you want (00:57:13) people to know? (00:57:16) >> I feel like we covered a lot of ground. (00:57:18) >> Me, too. But I want to know if there's (00:57:20) anything on your mind. (00:57:27) >> I don't think so. (00:57:29) One of the things that I haven't gotten (00:57:31) to play with yet, but I'm curious about (00:57:33) is GBT5 being much more in my life, (00:57:37) meaning like in my Gmail and my calendar (00:57:40) and my like (00:57:42) >> I've been using GBT4 mostly as a (00:57:46) >> isolated relationship with it. (00:57:48) >> Yeah. (00:57:48) >> How would I expect my relationship to (00:57:50) change with GBC 5? (00:57:52) >> Exactly what you said. I think it'll (00:57:53) just start to feel integrated in all of (00:57:56) these ways. you'll connect it to your (00:57:57) calendar and your Gmail and it'll say (00:57:59) like, "Hey, do you want me to I noticed (00:58:00) this thing. Do you want me to do this (00:58:01) thing for you over time, it'll start to (00:58:04) feel way more proactive. Um, so maybe (00:58:07) you wake up in the morning and it says, (00:58:08) "Hey, this happened overnight. I noticed (00:58:10) this change on your calendar. I was (00:58:12) thinking more about this question you (00:58:13) asked me. I have this other idea." And (00:58:14) then you know eventually we'll make some (00:58:16) consumer devices and it'll sit here (00:58:18) during this interview and you know maybe (00:58:20) it'll leave us alone during it but after (00:58:22) it'll say that was great but next time (00:58:24) you should have asked Sam this or when (00:58:25) you brought this up like (00:58:27) >> you know he kind of didn't give you a (00:58:29) good answer so like you should really (00:58:30) drill him on that (00:58:32) >> and it'll just feel like it kind of (00:58:34) becomes more like this entity that is (00:58:36) this companion with you throughout your (00:58:38) day. We've talked about kids and college (00:58:42) graduates and parents and all kinds of (00:58:44) different people. If we imagine a wide (00:58:46) set of people listening to this, they've (00:58:47) come to the end of this conversation. (00:58:49) They are hopefully feeling like they (00:58:51) maybe see visions of moments in the (00:58:53) future a little bit better. What advice (00:58:56) would you give them about how to (00:58:58) prepare? (00:58:59) >> The number one piece of tactical advice (00:59:00) is just use the tools. Like the the (00:59:04) number of people that I have the the (00:59:07) most common question I get asked (00:59:09) about AI is like what should I how (00:59:11) should I help my kids prepare for the (00:59:12) world? What should I tell my kids? The (00:59:13) second most question is like how do I (00:59:14) invest in this AI world? But stick with (00:59:17) that first one. Um (00:59:19) I am surprised how many people ask that (00:59:22) and have never tried using Chachi PT for (00:59:24) anything other than like a better (00:59:26) version of a Google search. And so the (00:59:27) number one piece of advice that I give (00:59:28) is just try to like get fluent with the (00:59:30) capability of the tools. figure out how (00:59:32) to like use this in your life. Figure (00:59:33) out what to do with it. And I think (00:59:36) that's probably the most important piece (00:59:37) of tactical advice. You know, go like (00:59:39) meditate, learn how to be resilient and (00:59:41) deal with a lot of change. There's all (00:59:42) that good stuff, too. But just using the (00:59:44) tools really helps. (00:59:45) >> Okay. I have one more question that I (00:59:47) wasn't planning to ask, but I just (00:59:48) >> Great. (00:59:49) >> In in doing all of this research (00:59:51) beforehand, I spoke to a lot of (00:59:54) different kinds of folks. I spoke to a (00:59:55) lot of people that were building tools (00:59:58) and using them. I spoke to a lot of (01:00:00) people that were actually in labs and (01:00:02) and trying to build what we have defined (01:00:04) as super intelligence. And it did seem (01:00:06) like there were these two camps forming. (01:00:10) There's a group of people who are using (01:00:13) the tools like you in this conversation (01:00:16) and building tools for others saying (01:00:18) this is going to be a really useful (01:00:21) future that we're all moving toward. (01:00:23) Your life is going to be full of choice (01:00:25) and we've talked about our (01:00:26) >> my potential kids and and their futures. (01:00:29) Then there's another camp of people that (01:00:30) are building these tools that are saying (01:00:31) it's going to kill us all. And I'm (01:00:33) curious how that cultural disconnect has (01:00:36) like what am I missing about those two (01:00:40) groups of people? (01:00:43) It's so hard for me to like wrap my head (01:00:46) around like there are you are totally (01:00:47) right. There are people who say this is (01:00:49) going to kill us all and yet they still (01:00:50) are working 100 hours a week to build (01:00:51) it. (01:00:52) >> Yes. And (01:00:54) I I can't I can't really put myself in (01:00:58) the headsp space. If if that's what I (01:01:00) really truly believed, (01:01:04) >> I don't think I'd be trying to build it. (01:01:06) >> One would think, (01:01:07) >> you know, maybe I would be like on a (01:01:08) farm trying to like live out my last (01:01:10) days. Maybe I would be trying to like (01:01:11) advocate for it to be stopped. Maybe I (01:01:13) would be trying to like work more on (01:01:14) safety, but I don't think I'd be trying (01:01:16) to build it. So, I find myself just (01:01:18) having a hard time empathizing with that (01:01:20) mindset. I assume it's true. I assume (01:01:22) it's in good faith. I assume there's (01:01:24) just like there's some psychological (01:01:26) issue there I don't understand about how (01:01:28) they make it all make sense, but (01:01:32) it's very strange to me. Do you do you (01:01:35) have an opinion? (01:01:37) >> You know, because I I always do this. I (01:01:39) ask for sort of a general future and (01:01:41) then I try to press on specifics. And (01:01:45) when you ask people for specifics on how (01:01:47) it's going to kill us all, I mean, I (01:01:48) don't think we need to get into this on (01:01:50) an optimistic show, but you hear the (01:01:51) same kinds of refrains. You think about, (01:01:53) you know, something uh trying to (01:01:56) accomplish a task and then over (01:01:57) accomplishing that task. Um you hear (01:01:59) about sort of I've heard you talk about (01:02:01) a sort of general um over reliance of (01:02:04) sort of an understanding that the (01:02:06) president is going to be a (01:02:07) >> a (01:02:08) >> AI and and maybe that is an overreliance (01:02:10) that we, you know, would need to think (01:02:12) about. And you know, you you play out (01:02:14) these different scenarios, but then you (01:02:16) ask someone why they're working on it, (01:02:18) or you ask someone how how they think (01:02:19) this will play out, and I just maybe I (01:02:22) haven't spoken to enough people yet. (01:02:23) Maybe I don't fully understand this this (01:02:26) cultural conversation that's happening. (01:02:28) Um or maybe it really is someone who (01:02:30) just says 99% of the time I think it's (01:02:33) going to be incredibly good. 1% of the (01:02:35) time I think it might be a disaster (01:02:37) trying to make the best world. (01:02:38) >> That I can totally if you're like, hey, (01:02:39) 99% chance incredible. 1% chance the (01:02:42) world gets wiped out. And I really want (01:02:44) to work to maximize to move that 99 to (01:02:47) 99.5. That I can totally understand. (01:02:49) >> Yeah, (01:02:49) >> that makes sense. (01:02:51) >> I've been doing an interview series with (01:02:53) some of the most important people (01:02:55) influencing the future. (01:02:57) >> Not knowing who the next person is going (01:02:59) to be, but knowing that they will be (01:03:00) building something totally fascinating (01:03:02) in the future that we've just described. (01:03:04) Is there a question that you'd advise me (01:03:06) to ask the next person not knowing who (01:03:07) it is? (01:03:10) I'm always interested in the like (01:03:11) without knowing anything about the I'm (01:03:13) always interested in the like of all of (01:03:15) the things you could spend your time and (01:03:16) energy on. Why did you pick this one? (01:03:19) How did you get started? Like what did (01:03:21) you see about this when before everybody (01:03:23) else like most people doing something (01:03:24) interesting sort of saw it earlier (01:03:25) before it was consensus. (01:03:26) >> Yeah. (01:03:27) >> Like how did how did you get here and (01:03:28) why this? (01:03:29) >> How would you answer that question? (01:03:34) >> I was an AI nerd my whole life. I came (01:03:36) to college to study AI. I worked in the (01:03:38) AI lab. Uh, I was like a I watched (01:03:41) sci-fi shows growing up and I always (01:03:43) thought it would be really cool if (01:03:44) someday somebody built it. I thought it (01:03:46) would be like the most important thing (01:03:47) ever. I never thought I was going to be (01:03:48) one to actually work on it and I feel (01:03:51) like (01:03:53) unbelievably lucky and happy and (01:03:56) privileged that I get to do this. I like (01:03:59) feel like I've like come a long way from (01:04:00) my childhood. (01:04:03) But there was never a question in my (01:04:04) mind that this would not be the most (01:04:06) exciting interesting thing. I just (01:04:07) didn't think it was going to be (01:04:08) possible. Uh, and when I went to (01:04:10) college, it really seemed like we were (01:04:12) very far from it. And then in 2012, (01:04:15) the Alex Net paper came out done, you (01:04:18) know, in partnership with my co-founder, (01:04:20) Ilia. And (01:04:24) for the first time, it seemed to me like (01:04:26) there was an approach that might work. (01:04:27) And then I kept watching for the next (01:04:29) couple of years as scaled up, scaled up, (01:04:31) got better, better. And I remember (01:04:33) having this thing of like why is the (01:04:35) world not paying attention to this? (01:04:38) >> It seems like obvious to me that this (01:04:40) might work. Still a low chance, but it (01:04:42) might work. And if it does work, it's (01:04:43) just the most important thing. So like (01:04:46) this is what I want to do. And then like (01:04:49) unbelievably it started to work. (01:04:53) >> Thank you so much for your time. (01:04:54) >> Thank you very much.