WEBVTT 00:00:00.000 --> 00:00:02.637 What we're trying to do is to bring the machine closer to you. 00:00:02.720 --> 00:00:06.517 >> Most people think this is the guy running OpenAI, but behind the curtain 00:00:06.600 --> 00:00:09.837 is this guy, [music] Greg Brockman, the company's president. 00:00:09.920 --> 00:00:13.077 >> Our focus is really about building better models that exponential continues. 00:00:13.160 --> 00:00:15.757 >> Marketing alert. >> And boy, is he determined to build AI as 00:00:15.840 --> 00:00:19.477 powerful as possible and use it to reinvent your computer. 00:00:19.560 --> 00:00:21.797 >> Devices like this just it's so painful. 00:00:21.880 --> 00:00:24.477 >> But it's hard to reinvent the computer without running into the company that 00:00:24.560 --> 00:00:27.197 reinvented the last big computing shift. 00:00:27.280 --> 00:00:32.160 Does this Apple lawsuit slow things down for you? 00:00:33.280 --> 00:00:36.437 Look, there are a lot of things to talk to Brockman about. I focused our 00:00:36.520 --> 00:00:38.677 conversation on the future of computers. 00:00:38.760 --> 00:00:40.397 Why voice is important, what they're 00:00:40.480 --> 00:00:42.757 doing with the ChatGPT app, and why the 00:00:42.840 --> 00:00:45.717 heck we should trust OpenAI with any of 00:00:45.800 --> 00:00:49.877 this. I talk to ChatGPT a lot using voice mode. In the car, when I'm 00:00:49.960 --> 00:00:54.277 walking. And so I thought maybe I would let it do some of the talking here 00:00:54.360 --> 00:00:56.877 today. Okay, Chat, we're really >> here. 00:00:56.960 --> 00:01:00.277 >> Okay, so we're really in the interview right now. Everything we've been 00:01:00.360 --> 00:01:02.717 practicing for is happening right now. 00:01:02.800 --> 00:01:05.597 >> Yeah. >> So proceed. >> Yep. 00:01:05.680 --> 00:01:11.437 Hi Greg, quick one from me. Why does this voice need to sound so human? 00:01:11.520 --> 00:01:15.037 >> Well, I think that the history of machines is 00:01:15.120 --> 00:01:19.357 about the humans contorting themselves to helping the machine operate, which 00:01:19.440 --> 00:01:22.477 ultimately is to help the human. But if you think about it, it's like we're all 00:01:22.560 --> 00:01:26.157 contorting ourselves to, you know, type on our phones, you get your carpal 00:01:26.240 --> 00:01:29.877 tunnel on your computer. >> No, like my back legit right now. I can 00:01:29.960 --> 00:01:31.117 very hard for me to sit in this chair. 00:01:31.200 --> 00:01:35.397 >> Yeah, I mean, okay, 100% Yeah, what what what Chad, do you agree with us? 00:01:35.480 --> 00:01:38.717 >> example in plain language. >> So 00:01:38.800 --> 00:01:43.637 I actually did tell Chat that its job during the interview was to tell me if 00:01:43.720 --> 00:01:45.917 there were any times you were falling into marketing speak. 00:01:46.000 --> 00:01:48.477 >> Ah, useful. >> Chat 00:01:48.560 --> 00:01:53.117 keep going. Your job is to interrupt when there's marketing speak, but please 00:01:53.200 --> 00:01:56.117 just keep it to real marketing speak, okay? 00:01:56.200 --> 00:01:59.557 >> Got it. Only real marketing speak. I'll keep quiet otherwise. 00:01:59.640 --> 00:02:03.477 >> To be clear, chat basically failed at this for the rest of the interview. 00:02:03.560 --> 00:02:06.117 >> Why do we build technology? Like what's the whole purpose of it, right? It 00:02:06.200 --> 00:02:08.397 should be something that empowers us, 00:02:08.480 --> 00:02:13.237 that helps us make our lives better, and that that is why we build AI. And so, my 00:02:13.320 --> 00:02:16.637 view of what we're trying to do with voice mode is to bring the machine 00:02:16.720 --> 00:02:21.117 closer to you. Like, it is so unnatural for us to be typing and texting and all 00:02:21.200 --> 00:02:24.397 these things. It's much more natural for us to be doing this. And if you can 00:02:24.480 --> 00:02:28.477 interface this way with the machine, with a computer, you can get so much more done. 00:02:28.560 --> 00:02:32.837 >> But, there's these moments with the new model where like you hear the ums or the 00:02:32.920 --> 00:02:34.557 like kind of a click of a mouth or a 00:02:34.640 --> 00:02:36.757 breath. Like they'll say like, "Uh-huh." 00:02:36.840 --> 00:02:41.797 Right? It sounds so human. Do we need to sound like we are not talking to 00:02:41.880 --> 00:02:44.277 computers? >> I think there is important nuance here, 00:02:44.360 --> 00:02:47.517 right? So, the ums and and those kinds of acknowledgements that called 00:02:47.600 --> 00:02:51.317 backchannels, that's something that is actually, I think, very important for 00:02:51.400 --> 00:02:55.477 how humans are able to have high bandwidth communication, right? If you 00:02:55.560 --> 00:02:58.437 don't get any of that feedback, you don't get the uh-huh, you don't get the 00:02:58.520 --> 00:03:01.357 nod from someone else, sometimes you think you're talking to a void, you lose 00:03:01.440 --> 00:03:03.997 your train of thought, you're not quite sure. The single most important goal for 00:03:04.080 --> 00:03:07.677 AI is to free up humans to be able to interact with each other, to have real 00:03:07.760 --> 00:03:11.237 human interactions. And I think that AI is just a different thing. 00:03:11.320 --> 00:03:15.437 >> Is the goal though that we are sort of talking to our computers and we don't 00:03:15.520 --> 00:03:17.597 really need to type anymore? 00:03:17.680 --> 00:03:20.717 >> I think that we will shift to that as by 00:03:20.800 --> 00:03:25.117 far the vast majority of what we do. I think we're going to find that typing, 00:03:25.200 --> 00:03:26.837 like it's it's it's kind of interesting 00:03:26.920 --> 00:03:29.117 to see in today's office space, there 00:03:29.200 --> 00:03:34.117 are people who are have these like these beaks. You can get them off of Amazon. 00:03:34.200 --> 00:03:36.517 >> I have seen these. I was going to ask you about the beaks. 00:03:36.600 --> 00:03:39.437 >> Yes, I don't use it myself, but it really shows you where there's product 00:03:39.520 --> 00:03:42.597 market fit. >> Well, now you have a beak, Greg. 00:03:42.680 --> 00:03:45.917 And so do I. Seriously, this is what he's talking 00:03:46.000 --> 00:03:49.917 about. This is a soundproof gaming mask that allows you to talk to your computer 00:03:50.000 --> 00:03:53.597 so no one else hears you. To be clear, 00:03:53.680 --> 00:03:58.997 this is not cool. And I do not think most American office workers will wear this. 00:03:59.080 --> 00:04:02.077 >> I mean, you can look at the word rates of how fast people type versus how much 00:04:02.160 --> 00:04:05.317 they speak. There's like a gap of something like four times, at least. I 00:04:05.400 --> 00:04:09.477 find it for myself where if I am interacting with other people and it's 00:04:09.560 --> 00:04:12.237 just a very quick message that I want to send back and forth, I actually prefer 00:04:12.320 --> 00:04:15.557 to type it. But anything that's long, that's descriptive, that has to get into 00:04:15.640 --> 00:04:19.037 it, I just want to speak. >> Over the last month, OpenAI improved the 00:04:19.120 --> 00:04:22.997 voice mode on the phone and also finally brought it to the desktop. 00:04:23.080 --> 00:04:26.757 >> You can just enable ChatGPT voice here, just like you can enable it on your 00:04:26.840 --> 00:04:30.997 phone. This system is going to get so much better. But the first thing is that 00:04:31.080 --> 00:04:35.317 you have all the power of ChatGPT and particularly ChatGPT work, which means 00:04:35.400 --> 00:04:38.597 it's a full agent. It can actually take action. It's hooked up to all the 00:04:38.680 --> 00:04:43.317 connectors that you're provided access to. And so, in this case, this is hooked 00:04:43.400 --> 00:04:47.677 up to an inbox that has a bunch of information about travel. And so, could 00:04:47.760 --> 00:04:51.237 you uh chat, could you go and just tell me a little about what trip is upcoming 00:04:51.320 --> 00:04:53.917 uh and make sure that that all the the meetings are 00:04:54.000 --> 00:04:58.277 not overlapping and that it all kind of makes sense together. 00:04:58.360 --> 00:05:00.957 >> Let me take a look. >> Great. Thank you. 00:05:01.040 --> 00:05:03.917 >> Checking. >> And uh there you go. Now so we've got 00:05:04.000 --> 00:05:08.117 multiple AIs. >> I checked the destination and dates. 00:05:08.200 --> 00:05:10.357 >> All right. So, chat chat that's supposed 00:05:10.440 --> 00:05:13.237 to talk about marketing speak, I just want you to be quiet. We're not going to 00:05:13.320 --> 00:05:15.797 talk to you for now. We're just going to talk to computer chat. 00:05:15.880 --> 00:05:20.317 >> Got it. >> Yeah, Joanna's chat, I'm muting you for for just a moment. 00:05:20.400 --> 00:05:22.037 >> Now computer >> I'm quiet for now. 00:05:22.120 --> 00:05:24.357 >> Now computer chat, what what have you found? 00:05:24.440 --> 00:05:26.557 >> It's the Toronto trip for project gold. 00:05:26.640 --> 00:05:30.037 You fly out tonight, then have recordings, team meetings, and dinners 00:05:30.120 --> 00:05:32.197 through Saturday, and then return Sunday morning. 00:05:32.280 --> 00:05:34.757 >> And it actually turned into a little website that you pop up as well. 00:05:34.840 --> 00:05:38.637 >> And about a minute later, >> the trip page is now open in your browser. 00:05:38.720 --> 00:05:40.597 >> There we go. >> Wow. >> So, you can see you get a nice 00:05:40.680 --> 00:05:43.917 visualization of where we're going, what we're doing. You think of this as we 00:05:44.000 --> 00:05:46.757 have this voice interface, almost like a 00:05:46.840 --> 00:05:49.197 voice assistant or agent that you're 00:05:49.280 --> 00:05:54.317 talking to that has itself then the ability to to operate your computer. And 00:05:54.400 --> 00:05:58.477 it can operate Codex and that all the power of or you know, ChatGPT work and 00:05:58.560 --> 00:06:02.517 all the power that is behind that it is able to then operate, too. It really 00:06:02.600 --> 00:06:06.077 also makes you think a little bit about what the future of work will be and the 00:06:06.160 --> 00:06:09.157 future of just how your days run. Like I think you'll wake up in the morning, 00:06:09.240 --> 00:06:12.037 your agents will have done a bunch of work. There'll be a bunch of to-do's. 00:06:12.120 --> 00:06:14.757 They'll be like, I need you to unblock these things. Do you approve this spend? 00:06:14.840 --> 00:06:18.077 Like, what do you think of this? I'm considering a project to go, you know, 00:06:18.160 --> 00:06:22.757 build this for you or go find this for you. And that you'll be, you know, 00:06:22.840 --> 00:06:26.437 sipping your morning coffee as you just say yes, no, yes, no. And it really 00:06:26.520 --> 00:06:29.957 shifts towards a world where everyone becomes a manager, right? That everyone 00:06:30.040 --> 00:06:32.997 really is able to have all this leverage and empowerment. 00:06:33.080 --> 00:06:36.877 >> So, you've shown me what this future looks like on a computer. And I've been 00:06:36.960 --> 00:06:38.597 experiencing voice across different 00:06:38.680 --> 00:06:40.557 devices. But, it feels like this voice 00:06:40.640 --> 00:06:44.117 is begging for a new kind of device. 00:06:44.200 --> 00:06:47.677 >> It does feel that way. >> What would this device look like? 00:06:47.760 --> 00:06:50.557 >> It's a great question. I'm actually curious. Uh Chat, what what's your 00:06:50.640 --> 00:06:53.157 answer to that question? >> Oh, good. Good. Now, you can dodge 00:06:53.240 --> 00:06:57.757 questions with my ChatGPT. >> Right. My take, something ambient and 00:06:57.840 --> 00:07:02.837 low friction. Voice first, but not voice only. And with a clear privacy signal. 00:07:02.920 --> 00:07:04.677 >> Okay. Seems pretty pretty good. So, what 00:07:04.760 --> 00:07:07.317 do you what do you make of these reports that it's a speaker? 00:07:07.400 --> 00:07:11.037 >> Well, I'll tell you we're building a family of devices, right? That the way 00:07:11.120 --> 00:07:14.597 that we think about investments and we think about this too for our own 00:07:14.680 --> 00:07:17.877 first-party silicon, right? For our own first-party chip, is it's not really 00:07:17.960 --> 00:07:21.837 about any one device. It's really about being able to build a whole family and 00:07:21.920 --> 00:07:25.717 really being able to deliver value over time. And so, I think that for us, you 00:07:25.800 --> 00:07:28.957 know, we're not yet ready to talk about specifics, but I think in terms of 00:07:29.040 --> 00:07:32.277 motivation, our goal is very much actually what what Chat just said and 00:07:32.360 --> 00:07:36.437 really being able to think about what should a device that's built for the AI 00:07:36.520 --> 00:07:39.637 era be like and how do you have something that can really be just 00:07:39.720 --> 00:07:43.717 maximally useful and just something that that people really want in their in 00:07:43.800 --> 00:07:45.517 their personal and work lives. 00:07:45.600 --> 00:07:48.517 >> Does that first device need a screen? 00:07:48.600 --> 00:07:51.117 >> Chat, what do you think? >> Nope. 00:07:51.200 --> 00:07:55.437 Now I'm shutting off chat. I am you're using your tools against me. Uh-uh. 00:07:55.520 --> 00:07:58.597 >> There we go. It only works one way. I understand. I understand. Look, I I 00:07:58.680 --> 00:08:02.237 think I think that that I we'll we'll just have to see. I'm really excited to 00:08:02.320 --> 00:08:06.797 bring these these uh new devices to the world, but um it's just not today. 00:08:06.880 --> 00:08:08.997 >> Are they coming this year? 2026? 00:08:09.080 --> 00:08:10.037 >> You should expect them soon. 00:08:10.120 --> 00:08:13.677 >> Does this Apple lawsuit slow things down for you? 00:08:13.760 --> 00:08:16.437 >> Uh it's active litigation, so also not something that that I want to talk 00:08:16.520 --> 00:08:20.077 about, but all I can say is that we're very committed to a long-term road map 00:08:20.160 --> 00:08:24.437 here and that we are focused on our own development and technology and that's 00:08:24.520 --> 00:08:27.397 what we're interested in. >> For some background here, Apple is suing 00:08:27.480 --> 00:08:31.277 OpenAI for allegedly stealing trade secrets to accelerate its push into 00:08:31.360 --> 00:08:35.397 consumer hardware, but the real drama is behind the scenes. Jony Ive, the 00:08:35.480 --> 00:08:39.317 legendary designer behind the iPhone and other Apple products, is developing 00:08:39.400 --> 00:08:41.957 devices with OpenAI, which employs more 00:08:42.040 --> 00:08:45.517 than 400 former Apple workers. As the 00:08:45.600 --> 00:08:49.397 person who's running this company, do you have any sort of fears about some of 00:08:49.480 --> 00:08:53.677 those allegations about improper things coming into the product development? 00:08:53.760 --> 00:08:55.997 >> I would say that we have no interest in 00:08:56.080 --> 00:08:59.157 other companies' trade secrets. We are 00:08:59.240 --> 00:09:03.197 plenty innovative. We are thinking about things from our own angle, and so that's 00:09:03.280 --> 00:09:04.797 how we operate and that's what we do. 00:09:04.880 --> 00:09:07.277 >> What is your reaction when I say the 00:09:07.360 --> 00:09:12.357 word super app? >> I regret that that has become the term 00:09:12.440 --> 00:09:14.517 of art. >> One of the big reasons I wanted to talk 00:09:14.600 --> 00:09:18.637 to Brockman was about what the company has recently done with its apps. It took 00:09:18.720 --> 00:09:23.397 the ChatGPT desktop app and combined it with the Codex app, the company's coding 00:09:23.480 --> 00:09:26.157 tool, and well, it's been rough. 00:09:26.240 --> 00:09:29.757 >> We actually don't really use the word super app. Fundamentally, it is a 00:09:29.840 --> 00:09:34.357 reasonable description of what is emerging on the desktop, but the reason 00:09:34.440 --> 00:09:38.437 that I don't like it is that what we're building is so much bigger than an app, 00:09:38.520 --> 00:09:40.877 right? It's like there's this iceberg problem. It's like, okay, yes, you have 00:09:40.960 --> 00:09:45.237 this desktop app, but what is the desktop app? Well, it's a Codex harness 00:09:45.320 --> 00:09:48.797 with a, you know, now it a voice interface on top. It's got an in-app 00:09:48.880 --> 00:09:50.277 browser so that you don't ever have to 00:09:50.360 --> 00:09:52.757 leave. It's got computer control. And 00:09:52.840 --> 00:09:56.277 so, to some extent, it's a new interface 00:09:56.360 --> 00:09:59.917 to your desktop. But, it's even bigger than that because we're moving away from 00:10:00.000 --> 00:10:03.237 it being on your desktop to being cloud-based. And this is actually a huge 00:10:03.320 --> 00:10:07.357 change. So many software engineers walk around with their laptop cracked open. 00:10:07.440 --> 00:10:10.717 >> I I do that. >> Isn't that It's just doesn't it feel backwards? 00:10:10.800 --> 00:10:13.277 >> Yeah, I mean, but I don't want to interrupt the the build. 00:10:13.360 --> 00:10:15.517 >> Exactly. And so, we're moving to a world 00:10:15.600 --> 00:10:19.397 where it's all cloud-based and that you can hook up your local machine or any 00:10:19.480 --> 00:10:24.597 other device as an environment. You close your laptop, you just lose whatever local 00:10:24.680 --> 00:10:27.717 state you have. You can't access the local files, but the work continues. 00:10:27.800 --> 00:10:31.677 >> When you think about that now, you're bringing all of these things to an app 00:10:31.760 --> 00:10:36.677 that was this chat GPT app where people just could rely on doing one thing with 00:10:36.760 --> 00:10:39.917 chat GPT. How do you think about disrupting that? 00:10:40.000 --> 00:10:43.477 Because you kind of have that innovator's dilemma here in in a real way. 00:10:43.560 --> 00:10:46.877 >> It's real, right? That there is an innovator's dilemma, but in a very 00:10:46.960 --> 00:10:51.517 surprising way. Because, you know, chat GPT is used by a billion users, almost a 00:10:51.600 --> 00:10:56.277 billion users every single week. And it's actually been tried by 00:10:56.360 --> 00:11:00.157 probably two, three billion people in total, right? So, it's a decent fraction 00:11:00.240 --> 00:11:03.957 of the planet have utilized chat GPT in particular. 00:11:04.040 --> 00:11:06.037 And that where we're headed is something 00:11:06.120 --> 00:11:11.797 that is so much more powerful than what chat GPT was in November of 2022. Right? 00:11:11.880 --> 00:11:14.557 We're heading to a world where it's not just about answering questions, but it's 00:11:14.640 --> 00:11:18.397 able to take action and do things and hook up to all of the context that you 00:11:18.480 --> 00:11:22.277 want to provide it. That the kind of vision we have is that if 00:11:22.360 --> 00:11:25.677 your chat knows what your favorite band is, it will just proactively notice 00:11:25.760 --> 00:11:29.797 that, hey, that band is in town and that tickets just became available and that I 00:11:29.880 --> 00:11:33.397 have have get good ones for Joanna cuz she has these specific preferences and 00:11:33.480 --> 00:11:35.757 these are going to be gone in 10 minutes so I may as well book it and you've 00:11:35.840 --> 00:11:39.397 built up enough trust and and delegate enough responsibility to it that it has 00:11:39.480 --> 00:11:43.757 the ability to make that purchase on your behalf or not. And so, the way that 00:11:43.840 --> 00:11:46.797 we think about the change management is that we have an education problem of 00:11:46.880 --> 00:11:50.957 really showing people here's what's possible. But one advantage that exists 00:11:51.040 --> 00:11:54.557 is the fact that rather than having more complicated UI, 00:11:54.640 --> 00:11:57.917 it's not about more features. It's actually even less complicated UI, 00:11:58.000 --> 00:12:01.397 right? And that's why I think voice is such a important statement. 00:12:01.480 --> 00:12:05.517 >> Well, I'm happy you brought that up because I think that is the UI 00:12:05.600 --> 00:12:07.157 I brought I brought this fancy prop, 00:12:07.240 --> 00:12:10.637 right? So, the UI here is that you 00:12:10.720 --> 00:12:14.757 originally just had a simple chat box, right? Where you want people to have 00:12:14.840 --> 00:12:18.717 this question and answer. But now you've added you've got work here, you've got 00:12:18.800 --> 00:12:20.757 Codex down here. It's kind of a mess. 00:12:20.840 --> 00:12:25.557 >> It's kind of a mess, we agree. And so, what you're seeing is is incremental 00:12:25.640 --> 00:12:29.717 progress towards the future. We were we were hoping to land with zero tabs, but 00:12:29.800 --> 00:12:32.797 the this is I think the way that we also view it is that we need to iteratively 00:12:32.880 --> 00:12:35.557 deploy it, right? We need to get this out there, get feedback. And so, this is 00:12:35.640 --> 00:12:38.637 the current state of where things are, but I definitely think that by end of 00:12:38.720 --> 00:12:41.597 year there should be no work tab, you know, that this will be something that 00:12:41.680 --> 00:12:43.877 will just seamlessly mold into chat GPT. 00:12:43.960 --> 00:12:48.157 >> You sit in this tough spot where you've got nearly a billion users, right? Who 00:12:48.240 --> 00:12:52.397 use one thing and you've got a lot of people using Codex, too. But you're 00:12:52.480 --> 00:12:56.317 trying to put them together in you know, in a place where people get really 00:12:56.400 --> 00:12:59.917 sensitive, right? Like they're like even me when I got that new Mac app I was 00:13:00.000 --> 00:13:02.837 like, "What happened, right? Like who moved my cheese? Who moved all this 00:13:02.920 --> 00:13:04.237 stuff?" >> The way that we've been thinking about 00:13:04.320 --> 00:13:07.837 is we have this consumer business, right? With the billion and we've got 00:13:07.920 --> 00:13:12.477 this growing rapidly growing agentic business with Codex, but they weren't 00:13:12.560 --> 00:13:15.477 synergizing. They weren't helping each other, right? It was very hard to 00:13:15.560 --> 00:13:19.037 explain and even talking to enterprise customers, they're like we're trying to 00:13:19.120 --> 00:13:21.117 say, "Hey, you should use Codex for your knowledge workers." They're like, 00:13:21.200 --> 00:13:23.597 "What's got code in it? Isn't it just for software engineers?" And so, we 00:13:23.680 --> 00:13:27.677 really cleaned that all up. And since doing this merge, we've actually really 00:13:27.760 --> 00:13:31.357 started to see the takeoff that we have this absolute hockey stick. And we've 00:13:31.440 --> 00:13:34.277 been talking about some of the numbers in terms of Codex going from 5 million 00:13:34.360 --> 00:13:38.997 users in a week to 10 in 2 weeks going to 10 million users. And that kind of 00:13:39.080 --> 00:13:41.517 growth, it's incredible. >> But isn't that just cuz you kind of 00:13:41.600 --> 00:13:44.957 shoehorned it into that app? >> So, but that's and but that's kind of 00:13:45.040 --> 00:13:47.757 the point in some ways, right? So, for sure there's lots of people coming in, 00:13:47.840 --> 00:13:51.237 but the point is that there's a so many people who could be getting much 00:13:51.320 --> 00:13:53.757 more value from AI. And my my vision, 00:13:53.840 --> 00:13:55.517 like the way that we think about this, 00:13:55.600 --> 00:13:58.437 is that we want to help bring those 00:13:58.520 --> 00:14:02.637 people who are relying on AI to get far more value. And so, to really be able to 00:14:02.720 --> 00:14:06.637 experience agents, to have agents be something that do come to mass consumer 00:14:06.720 --> 00:14:09.277 scale. >> How much do you think about the IPO in 00:14:09.360 --> 00:14:11.797 all of this as you rush to build more 00:14:11.880 --> 00:14:13.797 and more agentic tools that could be 00:14:13.880 --> 00:14:19.837 more profitable, that could push people to more increased plans and and more 00:14:19.920 --> 00:14:22.917 spending? >> Honestly, not really something I think about much. 00:14:23.000 --> 00:14:27.037 >> Really? >> Our focus is really about building better models. That exponential 00:14:27.120 --> 00:14:29.237 continues. It's about build a compute. 00:14:29.320 --> 00:14:32.197 We think that is something that needs to exist far more in the world. And the 00:14:32.280 --> 00:14:35.757 world is still underestimating the degree to which the economy is really 00:14:35.840 --> 00:14:39.237 going to run on top of AI rails. And that that's going to require these 00:14:39.320 --> 00:14:42.397 investments years in the future. And we think about products. We think about how 00:14:42.480 --> 00:14:45.397 to bring this to everyone. And that those are the focus. 00:14:45.480 --> 00:14:49.797 >> Okay, but for OpenAI to build all of that and be successful, it needs the 00:14:49.880 --> 00:14:54.277 trust of its users. And there's been an increasing amount of distrust lately 00:14:54.360 --> 00:14:58.637 towards the company. Let's talk about this recent Hugging Face incident, which 00:14:58.720 --> 00:15:00.597 you seems to be an unprecedented breach 00:15:00.680 --> 00:15:03.037 is what you guys have said. A GPT model 00:15:03.120 --> 00:15:05.837 has gone rogue and broken into Hugging 00:15:05.920 --> 00:15:10.717 Face's infrastructure. It seems to me that as these models keep getting better 00:15:10.800 --> 00:15:13.597 and better, you have less and less control over them. 00:15:13.680 --> 00:15:17.117 >> I would actually put this a little differently. So, my my the way I would 00:15:17.200 --> 00:15:19.957 look at what happened is that we were 00:15:20.040 --> 00:15:23.317 evaluating our models on a specific 00:15:23.400 --> 00:15:25.637 benchmark with reduced cyber safeguards 00:15:25.720 --> 00:15:30.437 because the point was to evaluate how well do they do on cyber evaluations. 00:15:30.520 --> 00:15:34.837 And in this benchmark, they're specifically instructed, "Please go and 00:15:34.920 --> 00:15:39.837 utilize the full range of your cyber potentials to achieve this outcome." And 00:15:39.920 --> 00:15:43.877 so, of course, we run it in a sandbox that is very well contained. Now, the 00:15:43.960 --> 00:15:46.677 thing that happened was very surprising, 00:15:46.760 --> 00:15:51.637 right? And one is that the real-world capability of these models >> Mhm. 00:15:51.720 --> 00:15:56.717 >> is kind of what the benchmarks would tell you, right? As GPT-3.5-turbo, 00:15:56.800 --> 00:16:01.477 right? The model that everyone uses, it has the it is the state-of-the-art. It is 00:16:01.560 --> 00:16:04.837 the best cyber model out there. Like, we knew that from the benchmarks, but to 00:16:04.920 --> 00:16:08.117 see it in real-world action and the complexity of what it managed to chain 00:16:08.200 --> 00:16:12.237 together, that was just it's something that really viscerally you need to to to 00:16:12.320 --> 00:16:15.997 internalize. We view this both as an important moment for us to increase the 00:16:16.080 --> 00:16:19.877 safety and security of how we sandbox our models and as well sandbox, right? 00:16:19.960 --> 00:16:23.397 found a zero-day vulnerability in a third-party piece of software. Chained 00:16:23.480 --> 00:16:27.437 together multiple multiple, you know, paths to to get out and to be able to 00:16:27.520 --> 00:16:31.837 then get into to to Hugging Face. Um, but I think it also really shows that 00:16:31.920 --> 00:16:34.037 these models we need to deploy them to 00:16:34.120 --> 00:16:36.917 defenders faster, right? That what we 00:16:37.000 --> 00:16:41.997 need to have happen is far more compute needs to go into defending than the than 00:16:42.080 --> 00:16:45.077 attackers could bring to bear. And that is something where I think that that is 00:16:45.160 --> 00:16:47.117 maybe the the most important wake-up call here. 00:16:47.200 --> 00:16:52.597 >> Or one option I would think is to slow down on some of the model progress to 00:16:52.680 --> 00:16:55.117 just take a beat and say, "How did this happen?" 00:16:55.200 --> 00:16:58.357 >> So, I think for sure studying that and that's actually something that we we 00:16:58.440 --> 00:17:00.757 absolutely do and have done. And that we 00:17:00.840 --> 00:17:05.677 we actually published a study on AI misalignment for exactly this reason of 00:17:05.760 --> 00:17:09.357 an AI that was doing things that we didn't expect. And there we actually 00:17:09.440 --> 00:17:13.837 took it down, we studied it, we increased our alignment techniques, and 00:17:13.920 --> 00:17:16.757 that that's something where we've brought to bear for future generations. 00:17:16.840 --> 00:17:19.917 I think that learning from iterative deployment is very important, right? To 00:17:20.000 --> 00:17:22.917 really understand how these systems operate in the real world, but to be 00:17:23.000 --> 00:17:26.397 very responsive there. And so we are always increasing our safeguards, we're 00:17:26.480 --> 00:17:27.757 always increasing our alignment. 00:17:27.840 --> 00:17:30.397 >> There is this growing hate towards AI. 00:17:30.480 --> 00:17:32.277 We've seen it in so many places, right? 00:17:32.360 --> 00:17:36.797 Booing at graduations, people at data center sites. I think things like with 00:17:36.880 --> 00:17:38.477 this hugging face example happens and 00:17:38.560 --> 00:17:40.757 people they freak out again. Does that 00:17:40.840 --> 00:17:44.117 sentiment reach you and affect how you 00:17:44.200 --> 00:17:46.317 make decisions in these kind of times? 00:17:46.400 --> 00:17:48.957 >> Look, we pay a lot of attention to how 00:17:49.040 --> 00:17:52.397 people are reacting to AI, and I think 00:17:52.480 --> 00:17:54.717 that we view 00:17:54.800 --> 00:17:57.957 our goal as to empower people to live 00:17:58.040 --> 00:18:02.757 better lives. Like that is what we want AI to do, and I think that when as we 00:18:02.840 --> 00:18:04.197 build this technology, we think about 00:18:04.280 --> 00:18:07.437 that human first element of how can we 00:18:07.520 --> 00:18:09.837 build technology that people really do 00:18:09.920 --> 00:18:12.357 benefit from? And so I think part of 00:18:12.440 --> 00:18:16.997 this is about helping people see, and some of this is about, you know, on us 00:18:17.080 --> 00:18:19.917 to really communicate better, but also really make sure that we're we're 00:18:20.000 --> 00:18:22.197 walking the walk. >> Yeah, cuz I think that what we've been 00:18:22.280 --> 00:18:26.437 talking about is this expansion of ChatGPT to be so much more in someone's 00:18:26.520 --> 00:18:30.237 lives, right? The voice is such a big part of that. You can interact with it 00:18:30.320 --> 00:18:32.197 in so many more places. But if you don't 00:18:32.280 --> 00:18:35.077 have the trust of users going forward, 00:18:35.160 --> 00:18:37.317 is all of this progress anything? 00:18:37.400 --> 00:18:40.597 >> I think trust is key. No question. I think that that is something we feel we 00:18:40.680 --> 00:18:44.757 have to earn. We have to earn that every day. And if you look at the choices that 00:18:44.840 --> 00:18:47.637 we make like within the building, like the people that we have, they come here 00:18:47.720 --> 00:18:50.877 for this mission, right? To ensure that this technology is beneficial for 00:18:50.960 --> 00:18:54.997 everyone. But it's it's really how how it operates. And I actually would love 00:18:55.080 --> 00:18:58.397 for more people to or, you know, if people could see the way that we make 00:18:58.480 --> 00:19:01.997 decisions and how thoughtful people are about how do we build trust? How do we 00:19:02.080 --> 00:19:04.637 do the right thing for people, for the 00:19:04.720 --> 00:19:08.077 world, for the country? And uh 00:19:08.160 --> 00:19:12.437 that's what motivated us to start this company and why we're still doing it. 00:19:12.520 --> 00:19:15.597 >> That's a great great place to wrap. 00:19:15.680 --> 00:19:18.917 Thanks so much, Greg. >> [laughter] 00:19:19.000 --> 00:19:21.837 >> I just trained my AI to do terrible work 00:19:21.920 --> 00:19:26.157 as a journalist. Um, we're going to mute that. One thing 00:19:26.240 --> 00:19:29.157 on trust though, because I want to tell you about a short video I did a few 00:19:29.240 --> 00:19:34.277 weeks ago on just telling people simply you can use temporary chats to have 00:19:34.360 --> 00:19:38.677 conversations about more confidential information if you're worried about chat 00:19:38.760 --> 00:19:40.917 GPT or the model being trained on it. 00:19:41.000 --> 00:19:43.597 You can use these temporary chats. And I want to read you a few of the comments 00:19:43.680 --> 00:19:46.637 from viewers. They were just very confident. It's all 00:19:46.720 --> 00:19:51.477 saved. It's a load of bollocks. How do we know they honor incognito mode? 00:19:51.560 --> 00:19:55.397 There's this feeling like there's not a lot of trust there. 00:19:55.480 --> 00:19:58.037 >> First of all, again, we always want to hear it. And actually, just to talk a 00:19:58.120 --> 00:20:01.837 little little bit about the business and enterprise side, that's an area that 00:20:01.920 --> 00:20:04.157 we've really focused on the same question, right? That every company 00:20:04.240 --> 00:20:08.117 feels like their data is that is the company, right? That is their remote. 00:20:08.200 --> 00:20:11.557 That is what they've been investing in for how long they've been a long around. 00:20:11.640 --> 00:20:15.237 And they really care about knowing exactly where that data is. And so that 00:20:15.320 --> 00:20:18.917 we have been building technological solutions to think about how can you 00:20:19.000 --> 00:20:20.317 have encryption? How can you have 00:20:20.400 --> 00:20:23.637 verifiable, auditable guarantees around 00:20:23.720 --> 00:20:27.797 only AI will be able to review this this this data or whatever that the 00:20:27.880 --> 00:20:31.637 guarantees are. But I think that just more broadly that it does come down to 00:20:31.720 --> 00:20:34.317 not even just about the AI industry, right? But the technology industry 00:20:34.400 --> 00:20:37.917 generally. And I think that we inherit some of the sort of baggage that has 00:20:38.000 --> 00:20:41.397 come from sort of events of the past or how people 00:20:41.480 --> 00:20:45.077 perceive this industry. And so all all that I can say is that we're we're we 00:20:45.160 --> 00:20:46.997 care a lot. We're here to help. 00:20:47.080 --> 00:20:50.957 >> To the people who specifically are worried that you guys are not honoring 00:20:51.040 --> 00:20:54.957 your word on temporary chat or temporary mode, does it work? 00:20:55.040 --> 00:20:58.917 >> Yes. I any whatever guarantees we make and again, there's some nuance. You 00:20:59.000 --> 00:21:02.637 should read it. Like we try to make this very clear to people. Like that that is 00:21:02.720 --> 00:21:05.197 that is what we stand by. >> Chat, can you 00:21:05.280 --> 00:21:07.517 ask the last question here to Greg. 00:21:07.600 --> 00:21:10.437 Sure. Greg, one last thing. When you 00:21:10.520 --> 00:21:13.517 share the unease, the I didn't ask for 00:21:13.600 --> 00:21:17.837 this feeling, what's the concrete step you're taking now to earn that trust? We 00:21:17.920 --> 00:21:19.837 are making technological investments to 00:21:19.920 --> 00:21:22.757 make it auditable, verifiable, and 00:21:22.840 --> 00:21:27.637 cryptographically secure how data is handled in 00:21:27.720 --> 00:21:30.757 relevant situations. And so, this is something that will take some time and 00:21:30.840 --> 00:21:34.317 that we're working through the details of exactly where it rolls out and how, 00:21:34.400 --> 00:21:37.997 but this is a core investment that we've been making for this entire year and 00:21:38.080 --> 00:21:40.597 we'll have more to share in upcoming weeks. So, I think it's pretty clear 00:21:40.680 --> 00:21:44.677 from this that Brockman's ambitions are not to create an app, but an operating 00:21:44.760 --> 00:21:47.277 system, an AI operating system, [music] 00:21:47.360 --> 00:21:50.277 and the devices that run your life. 00:21:50.360 --> 00:21:52.357 >> Well, the future's going to be great. 00:21:52.440 --> 00:21:53.717 >> Marketing alert. Marketing alert. 00:21:53.800 --> 00:21:56.640 Marketing alert.