WEBVTT 00:00:00.080 --> 00:00:01.837 You know, Greg, this is our fourth time 00:00:01.920 --> 00:00:04.717 speaking and we've spoken every time 00:00:04.800 --> 00:00:08.637 about OpenAI's product direction and I 00:00:08.720 --> 00:00:11.757 think I'm starting to get it. Um, you 00:00:11.840 --> 00:00:14.317 know, there was this conversation that a 00:00:14.400 --> 00:00:17.517 super app was the wrong term for what 00:00:17.600 --> 00:00:20.797 you were doing with um the app that 00:00:20.880 --> 00:00:23.277 you're building bringing codeex which is 00:00:23.360 --> 00:00:26.477 coding side of OpenAI's product browser 00:00:26.560 --> 00:00:31.037 and chat GPT together. And when you use the word super app, people would be 00:00:31.120 --> 00:00:33.997 like, "No, a super app is actually something that you can just use every 00:00:34.080 --> 00:00:37.917 other app within." Um, and now as we've 00:00:38.000 --> 00:00:41.997 seen these products come together, actually super app might be the correct 00:00:42.080 --> 00:00:45.757 term. You know, at least for us on the outside, we're starting to see it that 00:00:45.840 --> 00:00:48.477 when you need to do anything, um, you it 00:00:48.560 --> 00:00:51.117 will start with a prompt in chat GPT and 00:00:51.200 --> 00:00:54.317 then OpenAI's technology will use either 00:00:54.400 --> 00:00:58.397 your browser or your computer to get that done for you. Is that the right way 00:00:58.480 --> 00:01:01.037 to think about it? >> I I think that's a pretty good 00:01:01.120 --> 00:01:04.397 perspective, right? And I think to really zoom out, the thing we're 00:01:04.480 --> 00:01:09.037 actually trying to build is in AGI, right? That if you think about what 00:01:09.120 --> 00:01:13.037 people have been using since chat GBT, it's a language model, right? There's a 00:01:13.120 --> 00:01:15.277 big gap between these. 00:01:15.360 --> 00:01:20.157 It's amazing. You can talk to chat talks back to you. Great. Wonderful. But when 00:01:20.240 --> 00:01:23.277 we launched in 2022, there was no memory, right? It's not hooked up to any 00:01:23.360 --> 00:01:25.597 tools. Has no context. And so it really 00:01:25.680 --> 00:01:30.877 is that this conversational intelligence is only one part of what people really 00:01:30.960 --> 00:01:34.317 need to get work done to be able to achieve their goals. And where we're 00:01:34.400 --> 00:01:37.757 going is to have an AI that's really looking out for you, right? That you can 00:01:37.840 --> 00:01:41.757 provide the goals, the directions, that it's constantly thinking about what can 00:01:41.840 --> 00:01:45.757 I do for Alex today? Uh that it's able to go and solve super hard problems, 00:01:45.840 --> 00:01:48.957 very mundane problems. you wake up your inbox is organized but also if there's 00:01:49.040 --> 00:01:53.597 like a health plan that you are thinking about that it can help you help you 00:01:53.680 --> 00:01:56.797 achieve that figure out medical treatments uh or you know sort of back 00:01:56.880 --> 00:02:00.797 and forth provide you with that kind of information at least and I think that 00:02:00.880 --> 00:02:04.157 the question of what's the interface you want what is the product that you want 00:02:04.240 --> 00:02:07.357 is what we spend a lot of time thinking about and the answer is you want almost 00:02:07.440 --> 00:02:10.957 no interface you want no product right you want this to be like what's the 00:02:11.040 --> 00:02:15.197 interface between you and me right just being able to talk to a persistent 00:02:15.280 --> 00:02:19.757 entity of some form firm that's able to go and accomplish goals for you. And so 00:02:19.840 --> 00:02:23.757 building that is hard. It will take time, but we have a lot of the pieces, 00:02:23.840 --> 00:02:26.797 right? We're increasingly bringing together the product layer, trying to 00:02:26.880 --> 00:02:31.037 make the models better, trying to make the whole system just so there's less 00:02:31.120 --> 00:02:35.917 like clicking buttons and toggles and changing modes and all these things. Not 00:02:36.000 --> 00:02:38.957 to say that there won't be some of those along the way, but the long-term 00:02:39.040 --> 00:02:42.557 trajectory is towards simplification, unification. Yeah, it's very interesting 00:02:42.640 --> 00:02:45.117 that you say the interface uh will melt 00:02:45.200 --> 00:02:47.117 away. And so to go a little bit deeper 00:02:47.200 --> 00:02:52.157 with my question, um many of us who use products like chatbt today will see that 00:02:52.240 --> 00:02:55.517 the bot will make a suggestion at the end. You know, you ask it about 00:02:55.600 --> 00:02:59.037 nutrition and it says should I make a health plan for you or make a diet plan 00:02:59.120 --> 00:03:02.717 for you? You ask it, sorry Ron John, we just talked about travel, but you ask it 00:03:02.800 --> 00:03:03.997 about travel and then it will give you 00:03:04.080 --> 00:03:07.757 an agenda for instance. And so am I 00:03:07.840 --> 00:03:12.477 hearing you right that what's going to happen within chat GPT just to give an 00:03:12.560 --> 00:03:16.557 example is you talk to it about your health decisions and it might say you 00:03:16.640 --> 00:03:20.397 know you probably need to you know go to this specialist let me make an 00:03:20.480 --> 00:03:24.637 appointment for you and then it will go and actually take that action on your 00:03:24.720 --> 00:03:27.917 behalf. So it goes from simply a conversation interface to actually 00:03:28.000 --> 00:03:31.837 understanding your intent and then going out and accomplishing that for you. that 00:03:31.920 --> 00:03:35.757 that I that's exactly right and I think that if you've used codecs and by the 00:03:35.840 --> 00:03:39.437 way how many people in the room have used have used codecs 00:03:39.520 --> 00:03:43.037 >> people yeah decent number of people um and that our goal is to really bring the 00:03:43.120 --> 00:03:45.997 power of codecs to everyone right to 00:03:46.080 --> 00:03:50.077 bring agents to everyone you that technology exists right now right you 00:03:50.160 --> 00:03:52.797 can hook up like I hook up my codeex to 00:03:52.880 --> 00:03:57.837 slack to my Gmail to my calendar and there are many people within openai 00:03:57.920 --> 00:04:01.037 nontechnical users you know it's got code in the name but it's not really 00:04:01.120 --> 00:04:04.797 about code. It's really about having this general purpose tool using harness 00:04:04.880 --> 00:04:07.677 an agent. And that the kinds of things, 00:04:07.760 --> 00:04:11.837 for example, someone on our comm's team does is um she was organizing an event 00:04:11.920 --> 00:04:16.557 and it would just ask all of the uh event attendees for their dietary 00:04:16.640 --> 00:04:21.757 preferences, set up a whole seating chart, kind of did all of that work uh 00:04:21.840 --> 00:04:24.797 so that she could focus on the parts that she wanted to and and really 00:04:24.880 --> 00:04:27.677 thinking about the vision of what she wanted to achieve. And I think that 00:04:27.760 --> 00:04:31.437 we're going to see this across the board. So the it's not sci-fi anymore to 00:04:31.520 --> 00:04:36.237 think about an AI that's hooked up to these tools. And I remember with you 00:04:36.320 --> 00:04:37.757 know our very first attempt at tool use 00:04:37.840 --> 00:04:41.517 in chatbt was 2023. I think in like 00:04:41.600 --> 00:04:45.677 March or April or something we released plugins. Do people remember plugins back 00:04:45.760 --> 00:04:48.317 in early chat days? >> That didn't work. 00:04:48.400 --> 00:04:51.437 >> It didn't work at all because the models 00:04:51.520 --> 00:04:54.637 the models weren't ready. Right. The form factor is correct. Obviously you're 00:04:54.720 --> 00:04:58.397 going to have an AI that's able to like talk to your Gmail like no question. But 00:04:58.480 --> 00:05:01.917 we could only like have three different connectors exposed to the model at a 00:05:02.000 --> 00:05:06.157 time or start forgetting. You know, we had like 2K maybe 4K token context. Like 00:05:06.240 --> 00:05:09.357 there's just no memory, right? It's kind of like when you had early computers in 00:05:09.440 --> 00:05:13.117 the 60s or 70s or something, right? You know, that you had tiny little me memory 00:05:13.200 --> 00:05:16.237 banks and today you have your phone that's like better than any 00:05:16.320 --> 00:05:19.117 supercomputer from that era. And I think that's where we're going with these 00:05:19.200 --> 00:05:21.917 models, right? That the rate of improvement has been so steep. So now 00:05:22.000 --> 00:05:25.757 you can have hundreds of different tools accessible that we have the ability to 00:05:25.840 --> 00:05:28.957 hook them up to whole file system. So you can almost have the full power of 00:05:29.040 --> 00:05:32.477 the internet and like almost any application you want at the model's 00:05:32.560 --> 00:05:34.877 fingertips. And it's smart, right? It's 00:05:34.960 --> 00:05:38.717 got 52 million token context, depends 00:05:38.800 --> 00:05:41.677 how you squint on it. And the capability 00:05:41.760 --> 00:05:45.997 level is also getting so so powerful, right? These models are now solving 00:05:46.080 --> 00:05:48.797 unsolved math problems and physics problems, right? And really helping 00:05:48.880 --> 00:05:52.077 people be able to achieve things they couldn't otherwise. like we are on the 00:05:52.160 --> 00:05:57.117 era on the edge of this era of agents really transforming how we all operate 00:05:57.200 --> 00:06:01.357 whether it's in software engineering finance legal sales and in our personal 00:06:01.440 --> 00:06:05.437 lives too. So just to unpack that example that you were giving there one 00:06:05.520 --> 00:06:10.077 of your colleagues is chatting with Chad GPT about an event and then suggest hey 00:06:10.160 --> 00:06:14.557 you know uh how should we you know contact event attendees about something 00:06:14.640 --> 00:06:17.997 and instead of like saying okay I have to do that and going into an event 00:06:18.080 --> 00:06:22.957 program basically what happens is the the interface will take over from there 00:06:23.040 --> 00:06:26.877 once it says it's a good idea and you agree and then hook into whatever tools 00:06:26.960 --> 00:06:28.557 you're using and then do it for you. 00:06:28.640 --> 00:06:33.357 >> Exactly. So it uses its you know Gmail connector searches through your inbox to 00:06:33.440 --> 00:06:35.517 find all the people who are attending 00:06:35.600 --> 00:06:38.637 and then if you're on the like what 00:06:38.720 --> 00:06:42.717 what'sever in dietary restriction sees oh these people I already have their 00:06:42.800 --> 00:06:44.237 dietary restrictions these people I do 00:06:44.320 --> 00:06:47.517 not um drafts an email depending on 00:06:47.600 --> 00:06:49.837 exactly how you have things set up it 00:06:49.920 --> 00:06:54.557 might say hey I drafted these emails can I send them if you have a connector that 00:06:54.640 --> 00:06:59.357 doesn't even let it send emails said I drafted it you need to send them And uh 00:06:59.440 --> 00:07:02.317 in a different world, you could also imagine that it's you've built enough 00:07:02.400 --> 00:07:05.037 trust with the system where it says I drafted the emails and I actually sent 00:07:05.120 --> 00:07:07.517 them. And I think that this actually 00:07:07.600 --> 00:07:12.397 points to a really important aspect of the agentic era, which is trust, right? 00:07:12.480 --> 00:07:15.597 That we need to really learn how to build trust with these systems, where 00:07:15.680 --> 00:07:18.957 they're good, where they're not. Figure out what you want to delegate to them 00:07:19.040 --> 00:07:23.197 and how you want to entrust them with responsibility. And that's something we 00:07:23.280 --> 00:07:25.997 view as earned, right? It's not something that that we can we can grant 00:07:26.080 --> 00:07:28.237 but by providing lots of tools and 00:07:28.320 --> 00:07:31.437 control and oversight and supervision to 00:07:31.520 --> 00:07:35.277 the operator to the person who that this AI is operating on behalf of like we 00:07:35.360 --> 00:07:38.317 think that that is going to be such an important thing and so that's a key 00:07:38.400 --> 00:07:40.237 product feature and differentiator. 00:07:40.320 --> 00:07:41.517 >> Yeah. Yeah. And when you go back to some 00:07:41.600 --> 00:07:43.437 of the early attempts at this, there was 00:07:43.520 --> 00:07:48.557 this like move that OpenAI had to let you call an Uber within Chat GPT. And it 00:07:48.640 --> 00:07:51.757 it followed a long line of companies that have tried to get you to take 00:07:51.840 --> 00:07:54.237 action within chat, but it never really 00:07:54.320 --> 00:07:57.277 took off. And the difference here might 00:07:57.360 --> 00:08:00.557 be that the chatbot can take control of 00:08:00.640 --> 00:08:05.197 your browser or take control of your computer and then you don't necessarily 00:08:05.280 --> 00:08:09.037 have to worry about like is this plug-in going to work? it goes and accomplishes 00:08:09.120 --> 00:08:11.197 that for you by taking over your 00:08:11.280 --> 00:08:14.317 machine. So I wonder you know if you 00:08:14.400 --> 00:08:17.437 expect a fight from the user interfaces 00:08:17.520 --> 00:08:19.837 that we have today aka like all the 00:08:19.920 --> 00:08:22.397 other apps all the software where to be 00:08:22.480 --> 00:08:25.597 truly useful chat GPT will have to not 00:08:25.680 --> 00:08:29.597 be blocked to be able to go out and execute these actions on behalf of a 00:08:29.680 --> 00:08:34.317 user. Well, look f first of all, I'd say that it's this is not theoretical at 00:08:34.400 --> 00:08:37.037 this point, right? That people have been using codec. So, it's a separate 00:08:37.120 --> 00:08:39.917 product, separate app. You have to install it separately. Really starting 00:08:40.000 --> 00:08:44.797 to focus on software engineering. But the amount of non-software work that has 00:08:44.880 --> 00:08:47.677 been happening codeex has been absolutely exploding, right? It's been 00:08:47.760 --> 00:08:51.837 this like incredible exponential curve, exactly the thing that you you would 00:08:51.920 --> 00:08:55.437 expect. And within OpenAI, we basically have the same level of penetration now 00:08:55.520 --> 00:08:58.957 in usage as Slack, right? It's like everyone and OpenI is like an entirely 00:08:59.040 --> 00:09:02.317 slack based company. We do not use email for the most part. It's like really like 00:09:02.400 --> 00:09:04.957 if you're not on Slack, you're not going to do any work. Um, and it's kind of 00:09:05.040 --> 00:09:09.117 feeling that way now with codeex app as well. And that everyone's codex is 00:09:09.200 --> 00:09:14.077 hooked up to all of these tools. Um, how the ecos ecosystem evolves. I think it's 00:09:14.160 --> 00:09:17.437 going to be a very nuanced thing because I think one thing that is very important 00:09:17.520 --> 00:09:22.317 is that we believe that there should be an ecosystem that gets to be vibrant and 00:09:22.400 --> 00:09:26.557 thriving and that people can really build and see the benefits. And so we've 00:09:26.640 --> 00:09:30.557 actually seen this from uh partner companies where uh you know that we I 00:09:30.640 --> 00:09:33.357 remember there's a couple different partners where we said hey we really 00:09:33.440 --> 00:09:36.317 want to train our AI to be really good at using your software and we didn't 00:09:36.400 --> 00:09:38.957 know what they would say and actually 00:09:39.040 --> 00:09:43.437 the response we got is this is the most partner friendly outreach we've ever had 00:09:43.520 --> 00:09:47.757 right that the idea that you will make your AI specifically good at using our 00:09:47.840 --> 00:09:51.837 tool and they just see the opportunity because their tool will be used just so 00:09:51.920 --> 00:09:56.557 much more as a result and that everyone is trying to think about how do they not 00:09:56.640 --> 00:10:00.637 just survive as a company into the AI era but thrive? like how do you really 00:10:00.720 --> 00:10:04.317 get the advantages of the fact there's going to be so much more activity and if 00:10:04.400 --> 00:10:08.317 you don't have AI in there if you shut it out then you're actually going to be 00:10:08.400 --> 00:10:12.077 declining not thriving >> right this this kind of makes open AI 00:10:12.160 --> 00:10:15.597 puts open AAI so first of all you're going to bring you talk about people 00:10:15.680 --> 00:10:19.757 using codeex so one of your colleagues shared and I think you've talked about 00:10:19.840 --> 00:10:22.317 this too that you've brought chat GPT 00:10:22.400 --> 00:10:27.277 into codeex so you can bring codeex into chat GPT which is basically like if 00:10:27.360 --> 00:10:30.797 we're users of chat GPT this experience that we talked about of 00:10:30.880 --> 00:10:34.957 chat GPT not only suggesting what you might want to do next but going to do it 00:10:35.040 --> 00:10:38.957 for you that's going to happen and so it makes you effectively an operating 00:10:39.040 --> 00:10:43.517 system don't you think but not the operating system like an iOS where you 00:10:43.600 --> 00:10:48.237 would like go open up the your phone and then tap different apps it's almost as 00:10:48.320 --> 00:10:52.397 if all interaction with all apps will happen through this interface is that 00:10:52.480 --> 00:10:57.037 the ambition >> I think that you could describe it that 00:10:57.120 --> 00:11:00.397 way but I think of it a little differently. Like the way that I think 00:11:00.480 --> 00:11:02.397 about this is that what is the ideal 00:11:02.480 --> 00:11:05.197 interface to an AGI or we call it kind 00:11:05.280 --> 00:11:10.557 of a personal AGI. And I think that it's again the same interface that you and I 00:11:10.640 --> 00:11:14.397 are using right now. You just want to talk to an assistant, right? You want to 00:11:14.480 --> 00:11:15.997 talk to something that can go and and 00:11:16.080 --> 00:11:18.877 and work and and operate on your behalf. 00:11:18.960 --> 00:11:21.997 And so that yes, like that agent, that 00:11:22.080 --> 00:11:26.717 AGI, that AI will have its own computer, right? It'll have its own access to 00:11:26.800 --> 00:11:30.477 things. maybe can you know like ideal co-orker would be they can come over and 00:11:30.560 --> 00:11:32.557 type things on your computer too. So 00:11:32.640 --> 00:11:35.037 some access some delegated access to 00:11:35.120 --> 00:11:37.917 your own own system and you know maybe 00:11:38.000 --> 00:11:42.557 you delegate access to your inbox sometimes maybe it has its own inbox 00:11:42.640 --> 00:11:46.557 with some sort of uh some sort of you know window into into the things that it 00:11:46.640 --> 00:11:50.957 needs you forward emails to it. Th these are not actually if you if you think 00:11:51.040 --> 00:11:52.637 about this is not unprecedented, right? 00:11:52.720 --> 00:11:55.277 It's like the way that you work with an 00:11:55.360 --> 00:11:59.997 assistant who's a person that we've we've actually or any co-orker really, 00:12:00.080 --> 00:12:02.957 we've spent a lot of time really thinking about how do you build these 00:12:03.040 --> 00:12:06.637 trust boundaries and make sure that you're able to operate together. And so 00:12:06.720 --> 00:12:08.637 I think of it as just a different thing. 00:12:08.720 --> 00:12:11.837 It's not it you could think of it as an operating system, but an operating 00:12:11.920 --> 00:12:15.597 system is almost something from a different time, right? It's a different 00:12:15.680 --> 00:12:19.197 layer of the stack. This is really more about how do you interface with 00:12:19.280 --> 00:12:22.157 technology broadly. And I think that the 00:12:22.240 --> 00:12:24.157 beautiful thing about AI is it's really 00:12:24.240 --> 00:12:27.597 about bringing the machine closer to the 00:12:27.680 --> 00:12:29.597 human rather than us having to contort 00:12:29.680 --> 00:12:34.797 ourselves into like files and folders and like all these details that somehow 00:12:34.880 --> 00:12:37.357 are not natural, right? That are more about how the machine operates rather 00:12:37.440 --> 00:12:39.837 than how we operate. >> Yeah. Talking about a personal 00:12:39.920 --> 00:12:44.077 intelligence, it sort of um I don't did you watch WWDC last week? 00:12:44.160 --> 00:12:46.077 >> Uh no, no, I missed it. I was I was 00:12:46.160 --> 00:12:50.237 banned, but um I watched it on TV. Um 00:12:50.320 --> 00:12:54.877 come on, Apple. Anyway, it does look like you and and and Siri, the new Siri 00:12:54.960 --> 00:12:58.557 are going to come kind of into competition, right? Because they're an 00:12:58.640 --> 00:13:01.677 app that's going to sit or an intelligence that will sit on top of all 00:13:01.760 --> 00:13:04.157 of your apps and let you take action. 00:13:04.240 --> 00:13:07.997 And Chacht will be an app on the iPhone. 00:13:08.080 --> 00:13:12.077 So then talk a little bit about whether that positioning is going to be 00:13:12.160 --> 00:13:15.357 difficult for open AI and how you're thinking about that strategically. 00:13:15.440 --> 00:13:18.717 >> Well, I just think again think of it a little differently. Like I think that 00:13:18.800 --> 00:13:21.037 we're in the beginning of this new 00:13:21.120 --> 00:13:25.757 agentic era and the way that this has always gone in AI is that when you have 00:13:25.840 --> 00:13:29.437 a new level of capability, it means you have an opportunity to rethink 00:13:29.520 --> 00:13:31.677 everything, right? Rethink how 00:13:31.760 --> 00:13:34.237 >> people interface, how like what the 00:13:34.320 --> 00:13:39.277 techn is capable of. And I think that this is no no different, right? In my 00:13:39.360 --> 00:13:41.117 mind, like the kinds of things that I 00:13:41.200 --> 00:13:43.117 see on the horizon, 00:13:43.200 --> 00:13:46.797 for example, AI for solving 00:13:46.880 --> 00:13:50.557 scientific problems, right? And I think we're starting to see the inklings of 00:13:50.640 --> 00:13:54.397 this. Like for example, today we announced 00:13:54.480 --> 00:13:59.997 we have in uh peer-reviewed literature people doctors who are using 03. 00:14:00.080 --> 00:14:03.517 Remember 03? >> Yep. >> That was like forever ago now, right? 00:14:03.600 --> 00:14:04.797 That was like one of our earliest 00:14:04.880 --> 00:14:08.317 reasoning models using that to find 00:14:08.400 --> 00:14:13.757 diagnosis for people who I had no no answers from doctors for many many 00:14:13.840 --> 00:14:17.037 years. You know, there's an example of someone who had spent 20 years with a 00:14:17.120 --> 00:14:20.717 mysterious ailment. Finally, it's been diagnosed through the use of this 00:14:20.800 --> 00:14:24.237 technology. And if you're like, okay, you've got models that can do that, they 00:14:24.320 --> 00:14:27.597 can do that. And then it's really about like, you know, the same like 00:14:27.680 --> 00:14:30.797 distribution and like, you know, can you get access to an app? You know, to me 00:14:30.880 --> 00:14:33.837 it's it doesn't type check. It's like we have something fundamentally new. And so 00:14:33.920 --> 00:14:36.477 that's not to say that there won't be competition. I actually think that there 00:14:36.560 --> 00:14:39.517 will be and it's going to be great for everyone, but I just think that the ways 00:14:39.600 --> 00:14:41.997 in which you're going to use this technology, the things it will be 00:14:42.080 --> 00:14:45.517 capable of and what it'll make you capable of doing are just totally 00:14:45.600 --> 00:14:47.437 different from anything we've seen before. 00:14:47.520 --> 00:14:51.117 >> You know, I was going to ask you um well, does it mean that you'll have to 00:14:51.200 --> 00:14:55.677 um you know, create your own device assuming that like my concept is is you 00:14:55.760 --> 00:14:59.037 know that you're going to have to go through Apple to get to the user. Um 00:14:59.120 --> 00:15:03.277 assuming that's somewhat valid. But the answer is you already are right. You're 00:15:03.360 --> 00:15:05.597 so OpenAI is working on a device right now. 00:15:05.680 --> 00:15:09.517 >> It certainly has been publicly reported. 00:15:09.600 --> 00:15:11.997 >> I I was in your office in December and 00:15:12.080 --> 00:15:16.557 Sam told me that this is happening. It's it's multiple devices. Um so if you 00:15:16.640 --> 00:15:21.037 think about the way that again you're going to interface with um with these 00:15:21.120 --> 00:15:26.397 AIs, how does that device play in or series of devices? Well, look, I I think 00:15:26.480 --> 00:15:27.837 again I I would just step back and say 00:15:27.920 --> 00:15:29.597 that I think this is the beginning of 00:15:29.680 --> 00:15:33.917 something very new and that I think 00:15:34.000 --> 00:15:37.997 about the way that 00:15:38.080 --> 00:15:41.357 I think I want to say like I think the biggest shift that has happened in terms 00:15:41.440 --> 00:15:44.157 of interface again it's not even about 00:15:44.240 --> 00:15:45.917 devices and and and things like that. 00:15:46.000 --> 00:15:49.917 It's really about the shift from conversational intelligence like kind of 00:15:50.000 --> 00:15:53.917 the chat paradigm where it's like kind of you have an AI that's personalized 00:15:54.000 --> 00:15:57.517 enough to you that it's worth reading its output, right? You ask it a 00:15:57.600 --> 00:16:01.677 question, you get an answer, it's something that's useful to you to agents 00:16:01.760 --> 00:16:03.277 where they're capable enough to actually 00:16:03.360 --> 00:16:05.677 do things for you. Like that is a big 00:16:05.760 --> 00:16:09.117 shift and that that 00:16:09.200 --> 00:16:12.077 implies a difference in how you want to 00:16:12.160 --> 00:16:13.757 interact. And so you kind of are just 00:16:13.840 --> 00:16:16.397 going to want a single agent that has 00:16:16.480 --> 00:16:21.517 access to your context. And this will be true in personal life. This will be true 00:16:21.600 --> 00:16:23.357 in a business context, right? You 00:16:23.440 --> 00:16:26.797 imagine, for example, having a uh you 00:16:26.880 --> 00:16:28.557 know, imagine you have a PhD in every 00:16:28.640 --> 00:16:33.837 field co-worker, you know, Nobel prizes, multiple of them, and you hire one of 00:16:33.920 --> 00:16:37.277 these, you hire a hundred of them, and you don't invite them to any meetings. 00:16:37.360 --> 00:16:40.877 They're not going to be very useful. And so there's something about how do you 00:16:40.960 --> 00:16:44.077 get context into the AI and not just 00:16:44.160 --> 00:16:48.077 statically but dynamically right as 00:16:48.160 --> 00:16:52.077 context evolves as your business processes evolve how do you have a 00:16:52.160 --> 00:16:55.997 context layer that is accessible to an AI that lets the AI operate to the 00:16:56.080 --> 00:16:59.037 extent of that raw intelligence and so 00:16:59.120 --> 00:17:02.317 finding ways to make that AI be 00:17:02.400 --> 00:17:06.637 accessible so available in your meetings to make it very ergonomic it's very easy 00:17:06.720 --> 00:17:09.277 to get access to Um, I think all of 00:17:09.360 --> 00:17:11.117 that's going to require a rethink, but I 00:17:11.200 --> 00:17:14.557 think it again it's just the core for me 00:17:14.640 --> 00:17:18.877 starts from thinking about the agentic form factor and then working backwards 00:17:18.960 --> 00:17:23.197 to how do you just make this have the context it needs and again the trust is 00:17:23.280 --> 00:17:26.877 going to be such a core part of making this whole equation work. 00:17:26.960 --> 00:17:30.957 >> So kind of like having this this device with you at all times and being like I 00:17:31.040 --> 00:17:34.637 need to get that done and it goes and does it for you. And I think that that 00:17:34.720 --> 00:17:38.717 will be part of it. But I almost even think if you don't have a device like 00:17:38.800 --> 00:17:43.117 that, it's not like you're going to be out of the game, right? Because it's 00:17:43.200 --> 00:17:45.917 this a this AI. It's not because there's one thing there's one version of it 00:17:46.000 --> 00:17:48.877 where you think of it where it's like the device is the AI and you want your 00:17:48.960 --> 00:17:51.917 phone to be the AI. You want, you know, whatever whatever you know custom device 00:17:52.000 --> 00:17:54.797 you're you're thinking about to be the AI, but it's not going to be like that. 00:17:54.880 --> 00:17:56.077 It's going to be more like an interface. 00:17:56.160 --> 00:17:59.517 Like no more than your phone is you, right? It's an interface to you. It's a 00:17:59.600 --> 00:18:02.717 way that I can sort of, you know, call you up whenever I need you. uh whenever 00:18:02.800 --> 00:18:05.997 I want to ask you a question and there's different ways of accessing right 00:18:06.080 --> 00:18:09.837 there's like synchronous phone call I can text you I can email you and I think 00:18:09.920 --> 00:18:13.517 that we're going to be much the same with how we interact with our agents uh 00:18:13.600 --> 00:18:17.437 there's been some reports that open is working on these like birectional voice 00:18:17.520 --> 00:18:21.117 models I think we've talked about that in the past like the goal is to have 00:18:21.200 --> 00:18:26.157 like an AI that you can you can speak with and we'll be able to process that 00:18:26.240 --> 00:18:29.837 and speak back with you in a much more natural way can you share anything about 00:18:29.920 --> 00:18:35.037 that No, [laughter] >> but no, more seriously, um I mean, look, 00:18:35.120 --> 00:18:38.637 I think that the that the general shape of the of the technology, like the way 00:18:38.720 --> 00:18:41.437 that like we had we had we've had voice 00:18:41.520 --> 00:18:45.757 models, um you know, kind of a really cool voice experience for, you know, 00:18:45.840 --> 00:18:47.757 year and a half, two years now. Um you 00:18:47.840 --> 00:18:50.237 know, we first demoed it back in March, 00:18:50.320 --> 00:18:53.437 April of 2024. Um brought it to market, 00:18:53.520 --> 00:18:58.157 um you know, maybe late that late that year. And the way that it works and the 00:18:58.240 --> 00:19:03.117 way that everyone's models work is that you basically chain together um well the 00:19:03.200 --> 00:19:06.477 original way that these things worked was that you would chain together a text 00:19:06.560 --> 00:19:09.437 to or a speechtoext model then you do a 00:19:09.520 --> 00:19:13.517 texttoext model and then you would do a texttospech model. Horribleness, right? 00:19:13.600 --> 00:19:17.117 Like these three things chained together. Um it still has been the case 00:19:17.200 --> 00:19:21.117 that even if you have one unified model that's able to kind of take in input and 00:19:21.200 --> 00:19:24.557 then you know able to output a response, you still have this problem of turn 00:19:24.640 --> 00:19:28.397 taking, right? Imagine that like we have this like you cannot overlap, you cannot 00:19:28.480 --> 00:19:32.877 interrupt. It's just like once you you you speak to me in a turn and then you 00:19:32.960 --> 00:19:36.077 got to wait for me to finish my whole response. That is not how human 00:19:36.160 --> 00:19:40.637 conversation works. And so that we we basically have like a hack where we have 00:19:40.720 --> 00:19:44.477 these models that determine oh it seems like the turn has ended and oh it seems 00:19:44.560 --> 00:19:47.757 like the turn has started and we're like why are we talking about turns, right? 00:19:47.840 --> 00:19:51.517 Like turns again are so unnatural. This is the humans contorting ourselves to 00:19:51.600 --> 00:19:55.037 the machine and its limitations. And so the obvious thing that you want to 00:19:55.120 --> 00:19:59.357 accomplish is a model in AI that works much more like you and I do, right? 00:19:59.440 --> 00:20:04.397 That's able to process input at the same time it's processing output. And all of 00:20:04.480 --> 00:20:07.197 that is of course something that many people in this field are trying to to 00:20:07.280 --> 00:20:09.037 run towards. Um I think it's going to be 00:20:09.120 --> 00:20:10.957 very very exciting as you move to these 00:20:11.040 --> 00:20:16.077 natural very human fluid like conversational interfaces. No one's seen 00:20:16.160 --> 00:20:19.837 anything like it. Like one thing that that I think about is the the current 00:20:19.920 --> 00:20:23.357 interaction with you know chat GBT voice. many ways it's magical, right? So 00:20:23.440 --> 00:20:27.277 many people use it on their commute, able to ask all these questions, but it 00:20:27.360 --> 00:20:30.317 also is so frustrating, right? Whenever it breaks the magic because it's like 00:20:30.400 --> 00:20:34.317 you realize, oh, I want to like add some followup and it keeps talking over you 00:20:34.400 --> 00:20:39.037 and it didn't it's just like that is just it doesn't make sense. And so I 00:20:39.120 --> 00:20:43.517 think that part of what we need part of like the whole point of this AI is to be 00:20:43.600 --> 00:20:47.117 something that you can interact with interact with fluidly and naturally. And 00:20:47.200 --> 00:20:48.477 by the way, I think it's not just going 00:20:48.560 --> 00:20:52.957 to be about the sort of use case, like 00:20:53.040 --> 00:20:55.917 we kind of think about the the personal use case, but it's also really the work 00:20:56.000 --> 00:20:59.357 use case. And I think some of the most magical experiences that that I've had 00:20:59.440 --> 00:21:03.917 with codecs have been when operating it through voice. Like many people, we have 00:21:04.000 --> 00:21:05.837 a voice built in. Some people use third 00:21:05.920 --> 00:21:10.957 party apps for it. Um, and that you just get a very different experience when you 00:21:11.040 --> 00:21:14.637 start to realize that like typing a quick message to give some feedback, 00:21:14.720 --> 00:21:18.157 easy, but like writing out a whole paragraph and everything you want, 00:21:18.240 --> 00:21:21.037 horrible. No one wants to do that, right? You just want to be like saying 00:21:21.120 --> 00:21:24.397 things and you want the real-time feedback loop and all of that is going 00:21:24.480 --> 00:21:26.557 to happen and it's going to be amazing. 00:21:26.640 --> 00:21:31.277 So, let's talk about model improvement briefly. Um, so there was a discussion a 00:21:31.360 --> 00:21:33.277 couple years ago that large language 00:21:33.360 --> 00:21:35.597 models were about to hit a wall. um that 00:21:35.680 --> 00:21:39.277 was wrong. And um something you know 00:21:39.360 --> 00:21:43.357 that I'm thinking about is I think we're all thinking about it is how much better 00:21:43.440 --> 00:21:45.597 can these models get and when will the 00:21:45.680 --> 00:21:49.997 improvement stop? Any thoughts? Well, I 00:21:50.080 --> 00:21:52.957 think that this is a place where when 00:21:53.040 --> 00:21:57.197 you're kind of building these models, you get kind of a sense and an intuition 00:21:57.280 --> 00:22:01.917 that I think is harder to get from the outside because we see all the data 00:22:02.000 --> 00:22:05.277 points and we see also the work that goes into these improvements. And so 00:22:05.360 --> 00:22:07.037 that there's two parts to the answer. 00:22:07.120 --> 00:22:08.477 One is I think that the fundamental 00:22:08.560 --> 00:22:11.597 science is one of the most mysterious 00:22:11.680 --> 00:22:14.477 and important 00:22:14.560 --> 00:22:18.637 just scientific discoveries and empirical observations that that I can 00:22:18.720 --> 00:22:21.917 that I that I'm aware of that I can imagine right that we are able to 00:22:22.000 --> 00:22:24.397 actually build these models and that the 00:22:24.480 --> 00:22:27.277 scaling laws continue right that it just 00:22:27.360 --> 00:22:29.357 is the case that you can just keep 00:22:29.440 --> 00:22:32.797 training these models more data 00:22:32.880 --> 00:22:36.717 more compute better architectures and there's a lot of improvements that go 00:22:36.800 --> 00:22:39.837 in. But every time we've kind of run into a like, oh, this isn't quite 00:22:39.920 --> 00:22:44.797 scaling the way we expect, it's we have a problem, we have a bug, that our math 00:22:44.880 --> 00:22:48.157 wasn't quite right, that oh, our implementation isn't isn't isn't quite 00:22:48.240 --> 00:22:53.197 matching the math, whatever the thing is. And that is, I think, a very 00:22:53.280 --> 00:22:56.717 important thing to to sort of internalize. And actually if you we've 00:22:56.800 --> 00:22:59.997 done studies where you go back to the beginning of the field right that neural 00:23:00.080 --> 00:23:02.397 nets themselves were designed in like 00:23:02.480 --> 00:23:05.437 1940s right before computers right as a 00:23:05.520 --> 00:23:09.677 model of maybe this is maybe this how the brain processes information first 00:23:09.760 --> 00:23:13.277 hardware implementation was 1959 with the perceptron and if you look at 00:23:13.360 --> 00:23:16.077 landmark results in the field that the 00:23:16.160 --> 00:23:20.557 landmark results follow this incredibly smooth deterministic path of more 00:23:20.640 --> 00:23:24.877 compute being poured into them and so 70 years of people maybe 80 years now of 00:23:24.960 --> 00:23:27.597 people saying this stuff is never going to work, never going to scale, going to 00:23:27.680 --> 00:23:29.997 hit the wall. Hasn't hit the wall yet. 00:23:30.080 --> 00:23:34.317 There's still no wall in sight. And so I think that the fundamentals allow it. 00:23:34.400 --> 00:23:36.797 Now the practicality is hard, right? 00:23:36.880 --> 00:23:40.477 Actually building these massive supercomputers. It's hard. It's 00:23:40.560 --> 00:23:44.477 expensive. It's not easy, right? That we have teams that just like work so hard 00:23:44.560 --> 00:23:48.797 to solve these incredibly hard technical problems. We have our own network 00:23:48.880 --> 00:23:52.637 protocol that we've had to design. um that we have people who look at every 00:23:52.720 --> 00:23:56.157 single layer of the stack that there's weird wiggles in the graph and you the 00:23:56.240 --> 00:23:58.717 way to think about these neural nets is that there's like there's no 00:23:58.800 --> 00:24:02.317 abstractions, right? It's almost like any little piece that's wrong can have a 00:24:02.400 --> 00:24:05.677 ripple effect that only shows up down there. And so you need people to deeply 00:24:05.760 --> 00:24:09.277 understand all of it. And yet if you get 00:24:09.360 --> 00:24:11.037 the right team together, put the right 00:24:11.120 --> 00:24:13.597 mission in front of people and people do 00:24:13.680 --> 00:24:18.637 that grind, the outcome, it's worth it, right? And it's achievable and it's 00:24:18.720 --> 00:24:22.157 possible. And so I think that for those reasons the progress will continue. 00:24:22.240 --> 00:24:25.837 >> So then I'd love to hear your perspective if if models can basically 00:24:25.920 --> 00:24:28.157 progress much further from where they 00:24:28.240 --> 00:24:31.357 are today. Um let's say let's say open 00:24:31.440 --> 00:24:33.997 AI builds the best model and it the 00:24:34.080 --> 00:24:38.397 equivalent of like something with like 15 PhDs with excellent emotional 00:24:38.480 --> 00:24:40.557 intelligence that doesn't complain and 00:24:40.640 --> 00:24:43.517 goes out and does stuff for you. Um, and 00:24:43.600 --> 00:24:47.437 then the the next model maker will build 00:24:47.520 --> 00:24:50.077 a a less good, but it has 13 PhDs and 00:24:50.160 --> 00:24:54.477 it's like pretty good uh, you know, EQ and we'll still go and do things for 00:24:54.560 --> 00:24:58.717 you. So, where does the differentiation come in when you get to that level of 00:24:58.800 --> 00:25:02.957 intelligence? Because we've seen the model makers kind of move in lock step. 00:25:03.040 --> 00:25:06.557 One makes an advance, the next one comes in and makes the advance. So, they all 00:25:06.640 --> 00:25:11.037 become that smart. Do they is it possible to differentiate? Well, I think 00:25:11.120 --> 00:25:15.757 there are several dimensions to the answer. Um, number one is I do think 00:25:15.840 --> 00:25:19.917 there's a bit of an attractor state where just like from a business model 00:25:20.000 --> 00:25:23.837 perspective, every provider sells out all their compute. 00:25:23.920 --> 00:25:26.797 Okay. Right. I think that is just like the world that we're heading towards 00:25:26.880 --> 00:25:29.997 where there just is not going to be enough compute to serve all the demand. 00:25:30.080 --> 00:25:32.797 Right? That we're heading to this compute powered economy that everyone's 00:25:32.880 --> 00:25:36.477 going to be using these models all the time to be able to accomplish tasks of 00:25:36.560 --> 00:25:40.797 interest. And we just see it. It's like right now we're talking about compute 00:25:40.880 --> 00:25:43.197 constraints and like like the number of 00:25:43.280 --> 00:25:45.837 people using these agents is like order 00:25:45.920 --> 00:25:47.997 of 10 million 20 million maybe you know 00:25:48.080 --> 00:25:53.117 it's like we're not at planet scale chatbt is like a billion users right but 00:25:53.200 --> 00:25:55.597 we haven't brought the agentic power there yet so you're just looking at 00:25:55.680 --> 00:25:59.597 these factors and the depth of usage is also tiny compared to where we're going 00:25:59.680 --> 00:26:03.277 and so I think that we're just going to be in a world where even if you have 00:26:03.360 --> 00:26:07.677 different vendors different capability level open source models all these 00:26:07.760 --> 00:26:11.437 things these neoclouds like I think that compute is just going to be the scarce 00:26:11.520 --> 00:26:16.317 resource and I think that it's going to go go to use. Um so to some extent I 00:26:16.400 --> 00:26:20.077 think that the the like is this a good business to be in and for new entrance 00:26:20.160 --> 00:26:22.877 to come into and things like that. My answer is actually yes. I think that 00:26:22.960 --> 00:26:27.437 there is like a huge market um that we are just not going to be able to address 00:26:27.520 --> 00:26:31.917 and we need much more energy momentum there. Um but a second thing is that it 00:26:32.000 --> 00:26:33.677 also misses the fact that intelligence 00:26:33.760 --> 00:26:36.477 is not a undimensional thing right that 00:26:36.560 --> 00:26:41.837 that If you really zoom in, being good at different domains is something where 00:26:41.920 --> 00:26:45.357 even if you you have a lot of raw intelligence, getting good if you've 00:26:45.440 --> 00:26:49.597 never practiced like you've never actually done a pitch or something like 00:26:49.680 --> 00:26:52.717 you're not going to be good at it your first time, right? And that there's lots 00:26:52.800 --> 00:26:55.517 of different you've never operated a spreadsheet, right? You're not going to 00:26:55.600 --> 00:27:00.317 be able to succeed at doing some complex modeling. And so I think that there is 00:27:00.400 --> 00:27:03.437 something that we have been internalizing which is that we look 00:27:03.520 --> 00:27:05.917 across different industries and different domains and we have to 00:27:06.000 --> 00:27:11.117 prioritize. We can't possibly be great at every single area at once. There is 00:27:11.200 --> 00:27:14.477 definitely a lot of like, hey, you just get the general intelligence up and 00:27:14.560 --> 00:27:18.157 it'll experience a lot of these things, but to really become a domain expert, to 00:27:18.240 --> 00:27:22.237 really be that PhD, and to really be something that can help push forward the 00:27:22.320 --> 00:27:24.957 ambition of a field, like that's hard. 00:27:25.040 --> 00:27:28.077 And and by the way, one thing I also want to say is that I think 00:27:28.160 --> 00:27:31.357 understanding what happens when you successfully do that. I think that 00:27:31.440 --> 00:27:33.677 having a good mental model of that is important, which is you look at 00:27:33.760 --> 00:27:36.157 something like rewind to AlphaGo, right? 00:27:36.240 --> 00:27:38.317 you remember move 37, this move that 00:27:38.400 --> 00:27:43.517 like changed people's understanding of the game and then now more people play 00:27:43.600 --> 00:27:47.037 go than ever, right? And that it actually inspired people to do even 00:27:47.120 --> 00:27:50.397 more. I think we're just going to see that. And so I think that the depth is 00:27:50.480 --> 00:27:54.877 never going to stop, right? How deep can you go on science, right? Like I I think 00:27:54.960 --> 00:27:58.237 that people have thought sometimes that, hey, we found out all the physics, it's 00:27:58.320 --> 00:28:02.157 all good. We're all done. And I don't think that that's the future we're 00:28:02.240 --> 00:28:04.797 signed up for. I think we're signed up for one where we're got to keep find 00:28:04.880 --> 00:28:08.157 every time you unlock one mystery, >> every every time you solve one mystery, 00:28:08.240 --> 00:28:11.037 it unlocks like 10 more, right? So, I think that there's just going to be so 00:28:11.120 --> 00:28:14.237 much more to do and tons of room for differentiation across different 00:28:14.320 --> 00:28:15.757 companies. >> So, I think I'm reading you right and 00:28:15.840 --> 00:28:20.797 that your belief is um maybe there's a way that everybody can scale up these 00:28:20.880 --> 00:28:25.197 models, but ultimately the company with the most compute is going to win. And 00:28:25.280 --> 00:28:29.197 you know, we spoke a couple months ago and you had mentioned that like you were 00:28:29.280 --> 00:28:32.877 asked internally, how much compute should we buy? And you said, "All of 00:28:32.960 --> 00:28:36.237 it." And they said, "No, really, how much should we buy?" And you said, "No, 00:28:36.320 --> 00:28:40.717 buy all of it." And and OpenAI is definitely the leader in in buying 00:28:40.800 --> 00:28:44.637 compute. I mean, we see the the the money going out. Obviously, a lot of 00:28:44.720 --> 00:28:48.557 money coming in through investment and and now you you're you've built a 00:28:48.640 --> 00:28:52.237 business with customers, but there's a lot of money going out. Do you ever 00:28:52.320 --> 00:28:56.557 wonder, hey, are like, do you ever wonder maybe we're 00:28:56.640 --> 00:28:58.957 not going to be able to pay all this money back because it's a brand new 00:28:59.040 --> 00:29:03.437 category? Well, the way that I look at it is on the fundamentals, right? You 00:29:03.520 --> 00:29:05.677 you need to really look at the fact that 00:29:05.760 --> 00:29:10.957 the way that compute goes is that it's multiple years out before compute 00:29:11.040 --> 00:29:13.837 actually arrives, right? Depending on exactly what you're doing. For example, 00:29:13.920 --> 00:29:18.557 we've been investing in our own chip program now for multiple years. And 00:29:18.640 --> 00:29:20.557 super exciting progress like you know, 00:29:20.640 --> 00:29:22.557 we'll we'll we'll have more to announce 00:29:22.640 --> 00:29:26.157 uh actually pretty soon. Um but the fact 00:29:26.240 --> 00:29:29.837 that we're able to do that is something very unique, right? really think about 00:29:29.920 --> 00:29:32.717 the full vertical integration of the supply chain and I think that the world 00:29:32.800 --> 00:29:36.317 we're heading towards is one where again there's just not going to be enough 00:29:36.400 --> 00:29:40.157 compute in the world to satisfy all the demand and we see this very concretely 00:29:40.240 --> 00:29:44.157 like you look at the exponential I mean rewind to the exponential of chatbt um 00:29:44.240 --> 00:29:46.157 look at the exponentials we're on now uh 00:29:46.240 --> 00:29:47.517 you think about the problems that we 00:29:47.600 --> 00:29:51.837 that we are able to solve you know 00:29:51.920 --> 00:29:56.397 it's it's actually kind of interesting that yeah we just yesterday announced um 00:29:56.480 --> 00:29:58.557 it was two days ago announced a new 00:29:58.640 --> 00:30:02.077 result in uh uh basically chemistry and 00:30:02.160 --> 00:30:06.077 being able to synthesize you know new new improved reaction and all of this is 00:30:06.160 --> 00:30:09.357 without much attention you know the thing I just said of if you go deep in a 00:30:09.440 --> 00:30:12.557 domain you can really transform it and we're not even scratching the surface 00:30:12.640 --> 00:30:16.717 yet and so the way to think about is the economy is so massive right and we see 00:30:16.800 --> 00:30:19.837 it very concretely in terms of our own growth in terms of what people are 00:30:19.920 --> 00:30:23.757 willing to pay and kind of the size and growth of this whole industry and so I 00:30:23.840 --> 00:30:27.997 think that the thing that I think about the most is how do we meet the demand 00:30:28.080 --> 00:30:31.517 and how do you actually have something that can help support all of the work 00:30:31.600 --> 00:30:32.957 that people want to do in the economy? 00:30:33.040 --> 00:30:35.837 And I think that is such a vast thing. I don't think any of us have internalized 00:30:35.920 --> 00:30:39.677 it yet. Hi everyone, Alex Canitz here. I want to tell you about a documentary 00:30:39.760 --> 00:30:41.917 I've made with Gravity to explore the 00:30:42.000 --> 00:30:44.717 future of AI agent security. To find out 00:30:44.800 --> 00:30:46.717 if we're truly ready for autonomous 00:30:46.800 --> 00:30:49.597 agents, I sat down with MIT professor 00:30:49.680 --> 00:30:52.077 Romesh Rosker, former White House CIO 00:30:52.160 --> 00:30:54.397 Terresa Payton, Michelin's group chief 00:30:54.480 --> 00:30:57.677 data and AI officer Ambbea Roger Gopal, 00:30:57.760 --> 00:31:01.837 and Sharon Guy, a former executive at Alibaba. [music] 00:31:01.920 --> 00:31:04.237 They each offer unique insights into 00:31:04.320 --> 00:31:06.797 this evolving landscape. We conclude 00:31:06.880 --> 00:31:09.597 with Rory Blendell, CEO of Gravity, to 00:31:09.680 --> 00:31:13.917 discuss the path forward. with gravity leading the way. Join us on this 00:31:14.000 --> 00:31:17.357 journey. You [music] can watch the full documentary at the link in the show 00:31:17.440 --> 00:31:21.890 notes. [music] 00:31:27.120 --> 00:31:31.197 Yeah, but if I may, there is a price war brewing. I mean, at least that's 00:31:31.280 --> 00:31:33.917 according to the reports. It's great to have you here to talk about it. The Wall 00:31:34.000 --> 00:31:35.917 Street Journal recently had a report 00:31:36.000 --> 00:31:38.797 that an upcoming OpenAI model might have 00:31:38.880 --> 00:31:42.077 significant price cuts. And so again, 00:31:42.160 --> 00:31:45.117 like how can you know if it requires so 00:31:45.200 --> 00:31:50.477 much resources to serve this demand and it is growing demand um in an 00:31:50.560 --> 00:31:54.157 environment where there might be price cuts, how do you make that math work? 00:31:54.240 --> 00:31:58.317 >> Well, again, I look at it from a different angle. So if you look at the 00:31:58.400 --> 00:32:02.237 whole history of what we've done, we actually have been increasing the 00:32:02.320 --> 00:32:06.637 intelligence cutting price right for a fixed amount of intelligence and people 00:32:06.720 --> 00:32:09.837 somehow just like like the Japanese paradox just keeps happening. And so I 00:32:09.920 --> 00:32:13.197 think frontier intelligence will always be something that is going to be you 00:32:13.280 --> 00:32:15.837 know it's always going to be the priciest thing but I think that a year 00:32:15.920 --> 00:32:18.957 from now that level of intelligence is going to feel pretty mundane and like 00:32:19.040 --> 00:32:23.437 you know going to be much more available. Um, and I think that that the 00:32:23.520 --> 00:32:26.957 the world that we're in is one where people are starting to really think 00:32:27.040 --> 00:32:29.917 about value. And it's actually been a 00:32:30.000 --> 00:32:34.557 very interesting shift where over the past, you know, first quarter, maybe up 00:32:34.640 --> 00:32:38.557 until now, people have just been like this AI agent stuff, it's all new. We 00:32:38.640 --> 00:32:41.197 need to bring into our enterprise. Like, we don't want to be left behind. How do 00:32:41.280 --> 00:32:45.117 we be part of this this future? And now people are like, okay, like let's make 00:32:45.200 --> 00:32:48.077 sure this actually delivering ROI and value. And I actually think that's a 00:32:48.160 --> 00:32:51.357 great place to be, right? because people are asking the right questions. And I 00:32:51.440 --> 00:32:54.237 hear this I I had some customer meetings today where people were saying exactly 00:32:54.320 --> 00:32:58.077 this. They were like, "How can we have even just like good spend controls? How 00:32:58.160 --> 00:33:02.637 can we have observability?" And I I think we literally today just released 00:33:02.720 --> 00:33:05.117 spend controls. So, you know, it's like >> Okay. 00:33:05.200 --> 00:33:09.117 >> Exactly. We're we are really investing hard in enterprise readiness and the 00:33:09.200 --> 00:33:10.877 tools that our customers are telling us 00:33:10.960 --> 00:33:13.277 that they need. And I think that that 00:33:13.360 --> 00:33:18.317 for me is the shift that we've also been going through as a company is really not 00:33:18.400 --> 00:33:21.117 just thinking about hey we're just going to release models and you know have a 00:33:21.200 --> 00:33:24.077 model really thinking about the endto end of the business. How do we bring 00:33:24.160 --> 00:33:26.397 this into solving real problems for real 00:33:26.480 --> 00:33:29.677 customers and that is happening so 00:33:29.760 --> 00:33:32.397 quickly across every single industry and 00:33:32.480 --> 00:33:37.357 the number of different companies that still feel like they're wrapping their 00:33:37.440 --> 00:33:40.317 mind around how to best make use of these models we're learning at the same 00:33:40.400 --> 00:33:44.877 time. I think it's just so early in this whole game to me the the like the 00:33:44.960 --> 00:33:48.797 absolute size of the market growing so quickly, our own revenue ramp growing so 00:33:48.880 --> 00:33:52.077 quickly. I think it's still just like none of us are anticipating how steep 00:33:52.160 --> 00:33:55.837 that's all going to go. >> Are you going to cut prices? 00:33:55.920 --> 00:34:00.317 >> So again, the answer is always yes, right? But it's about like I think that 00:34:00.400 --> 00:34:03.677 what's going to keep happening is that we're going to have frontier models. Um 00:34:03.760 --> 00:34:06.957 I don't think there's going to be like a massive shift in in the short term. I 00:34:07.040 --> 00:34:09.197 don't think that that is the kind of thing that's going to happen. But I 00:34:09.280 --> 00:34:14.077 think the thing you should anticipate is that over a year-long time horizon to 00:34:14.160 --> 00:34:17.517 get to today's level of intelligence that feels very premier, it's going to 00:34:17.600 --> 00:34:21.197 be much cheaper. But there's going to be a new thing that is going to be so much 00:34:21.280 --> 00:34:23.757 better and you're going to be like, why would I ever use this other one, right? 00:34:23.840 --> 00:34:26.397 It's [clears throat] just it's just how it's always going to be. 00:34:26.480 --> 00:34:28.317 >> So Sadia Nadell has had some interesting 00:34:28.400 --> 00:34:32.557 tweets and interviews recently. He recently said, "The model is becoming a 00:34:32.640 --> 00:34:35.917 commodity and the valuable asset is company," or this might be a paraphrase, 00:34:36.000 --> 00:34:39.597 "The valuable asset is a company specific AI system that continually 00:34:39.680 --> 00:34:42.877 learns from your data." Um, what do you 00:34:42.960 --> 00:34:44.957 think about that? And is it weird to be 00:34:45.040 --> 00:34:46.877 competing with Microsoft now? 00:34:46.960 --> 00:34:49.917 >> Well, look, I don't think that there's 00:34:50.000 --> 00:34:53.917 any layer of the stack here that is going to just kind of be removed from 00:34:54.000 --> 00:34:57.997 the value chain. I think that these things multiplied together. And if you 00:34:58.080 --> 00:35:00.317 think about the most the base layer of 00:35:00.400 --> 00:35:05.597 compute right that is something where it's just like no compute no AI 00:35:05.680 --> 00:35:08.397 and to some extent you could say oh comput is commoditized it's just flops 00:35:08.480 --> 00:35:12.077 who cares about it but in reality like you look at you look at today's chip 00:35:12.160 --> 00:35:15.277 stocks you look at the uh you know people who are selling compute kind of 00:35:15.360 --> 00:35:18.717 what what is what the market is valuing people at and they see that there's a 00:35:18.800 --> 00:35:22.397 fundamental asset here that is just so critical and I think that is because it 00:35:22.480 --> 00:35:27.437 is a revenue center it is something that anyone who's building AI has to rely on 00:35:27.520 --> 00:35:30.797 and that there's a bunch of very interesting dynamics in terms of the 00:35:30.880 --> 00:35:33.597 efficiencies that you can squeeze out and the margins, all these things. But 00:35:33.680 --> 00:35:36.477 fundamentally, even though it's like you can kind of squint it and say it's 00:35:36.560 --> 00:35:40.637 commoditized, it's not. It's it's it's not that that the value goes away. It's 00:35:40.720 --> 00:35:43.757 not that the margins go away. It's like something that the market will reward 00:35:43.840 --> 00:35:47.677 because it has fundamental value and that the importance of it's going to go 00:35:47.760 --> 00:35:50.477 up over time. You can see that with some of the prices that people are paying for 00:35:50.560 --> 00:35:55.357 H100s, right? Hoppers are, you know, kind of, you know, not not obsolete, 00:35:55.440 --> 00:36:00.317 right? They're they're they're a a previous gen chip and uh nor in any 00:36:00.400 --> 00:36:03.357 normal situation where you're not totally supply constrained, no one would 00:36:03.440 --> 00:36:07.197 be buying them. But instead, the market prices are up relative to to where they 00:36:07.280 --> 00:36:10.157 were before. So there's this inversion that's happening and again I think it's 00:36:10.240 --> 00:36:13.517 going to keep happening where because 00:36:13.600 --> 00:36:17.917 everyone has this avalanche of demand that you're going to see prices and 00:36:18.000 --> 00:36:21.837 margins and all of these things AC continuing to increase at various levels 00:36:21.920 --> 00:36:24.717 of the stack. And >> I think the same kind of applies for 00:36:24.800 --> 00:36:28.317 models where the models themselves are 00:36:28.400 --> 00:36:32.717 also again they're not there's there's a lot of competition there and I think 00:36:32.800 --> 00:36:35.357 that's very good. I think it's good for the enterprise. I think it's good for 00:36:35.440 --> 00:36:37.517 customers, consumers. Um but I think 00:36:37.600 --> 00:36:42.397 that there's a lot of areas where for example our models have always been the 00:36:42.480 --> 00:36:45.437 sort of smartest ones, right? The ones that are able to solve these incredibly 00:36:45.520 --> 00:36:49.197 hard problems. I think we're just starting to reach a phase where you're 00:36:49.280 --> 00:36:52.157 going to see the transformative impact from that, right? It's like if we're 00:36:52.240 --> 00:36:56.877 really able to speed up science through models, the smarter the model, the 00:36:56.960 --> 00:37:00.557 faster it's going to go. And it's very very different from a model that has a 00:37:00.640 --> 00:37:03.677 conversational interface that you're able to, you know, is able to book your 00:37:03.760 --> 00:37:07.037 travel, right? Or organize your calendar. So that's also a dimension I 00:37:07.120 --> 00:37:09.837 think we're going to do a very good job in. But I'm just saying it's it's a 00:37:09.920 --> 00:37:14.317 different area. Um and then I think that the question of well how do you actually 00:37:14.400 --> 00:37:19.277 connect the intelligence to your own customers right to real value to you 00:37:19.360 --> 00:37:23.277 have all these enterprises that have built incredible businesses in different 00:37:23.360 --> 00:37:26.797 domains and it's a huge thing and it's not something where if you don't have 00:37:26.880 --> 00:37:29.837 domain expertise that you're just going to be able to do right and part of it is 00:37:29.920 --> 00:37:33.037 that you need you think about regulated industries you think about any area 00:37:33.120 --> 00:37:37.037 where there's like you know think about education where you have a parent you 00:37:37.120 --> 00:37:39.997 have a teacher you have student you have these different parties that need to 00:37:40.080 --> 00:37:44.877 interact in very thoughtful ways for all of these areas, all of these domains 00:37:44.960 --> 00:37:49.757 that there's a lot of value to be built by being in that area and thinking about 00:37:49.840 --> 00:37:53.117 how the workflow should work, how these models should be orchestrated. And so I 00:37:53.200 --> 00:37:57.437 I really think that there's more than enough to go around and I think that we 00:37:57.520 --> 00:38:00.557 have to work together as a whole ecosystem in order to deliver the kind 00:38:00.640 --> 00:38:02.797 of value that I think is possible from these systems. 00:38:02.880 --> 00:38:06.957 >> Okay, just to go back to the SA point one more time, uh he's called models a 00:38:07.040 --> 00:38:10.877 commodity. He's trying to build his own frontier intelligence. He's telling 00:38:10.960 --> 00:38:14.077 potentially your customers, hey, you got to come work with us because we're going 00:38:14.160 --> 00:38:18.317 to help build these loops that will learn from your data. Um, he's got 00:38:18.400 --> 00:38:20.877 access to your IP, I think, till 2032. 00:38:20.960 --> 00:38:24.077 So, how does it make you feel to hear this coming from Saudia? 00:38:24.160 --> 00:38:25.997 >> Look, I think that the most important 00:38:26.080 --> 00:38:31.277 thing that is happening right now is the usage of AI in the economy to really 00:38:31.360 --> 00:38:34.957 transform the economy and to uplift everyone. And so, I think that that is 00:38:35.040 --> 00:38:36.157 something that I'm really focused on. 00:38:36.240 --> 00:38:39.997 And the more that people are trying to make that happen, I think that that's 00:38:40.080 --> 00:38:41.597 better for everyone. 00:38:41.680 --> 00:38:45.277 >> Um, GPT 5.6 is rumored to be on its way. 00:38:45.360 --> 00:38:49.197 Supposed to be this just a Twitter rumor, but I'm going 00:38:49.280 --> 00:38:51.517 to read it to you. Um, always the best 00:38:51.600 --> 00:38:54.157 rumors. Three times cheaper than Fable 00:38:54.240 --> 00:38:57.277 up to 1.5 million token context. Uh, 00:38:57.360 --> 00:39:00.557 stronger agentic coding workflows. Um, 00:39:00.640 --> 00:39:05.497 how much of that is true? What should we expect for GPT 5.6? 00:39:05.580 --> 00:39:08.397 >> [laughter] >> I mean, look, you should always expect 00:39:08.480 --> 00:39:11.837 better, faster, smarter, the whole 00:39:11.920 --> 00:39:15.977 thing. >> So, everything confirmed. 00:39:16.060 --> 00:39:19.197 [laughter] >> Definitely believe everything you read 00:39:19.280 --> 00:39:21.597 on Twitter. Yeah, maybe not. 00:39:21.680 --> 00:39:22.797 >> That has actually been a source of 00:39:22.880 --> 00:39:26.237 problems in my personal life. Um, okay. 00:39:26.320 --> 00:39:29.837 So, uh, I want to end on health. Um, you brought it up a couple times. You 00:39:29.920 --> 00:39:33.597 actually had a question in the audience about it earlier. Um, you know, 00:39:33.680 --> 00:39:35.677 sometimes there's a story and you read 00:39:35.760 --> 00:39:38.957 it and you say to yourself, I know this 00:39:39.040 --> 00:39:41.197 person is speaking to the media and I 00:39:41.280 --> 00:39:43.357 know that what they're saying sounds 00:39:43.440 --> 00:39:46.317 like maybe it's true. Um, but there's 00:39:46.400 --> 00:39:50.477 something wrong with the story and we're not going to see more of it. Um, and 00:39:50.560 --> 00:39:52.477 I've read a couple of those recently. 00:39:52.560 --> 00:39:54.637 Um, one is I think is it your friend the 00:39:54.720 --> 00:39:59.437 GitLab CEO um, Sid Sberage? He had um he 00:39:59.520 --> 00:40:02.877 he got cancer and used uh he got all the 00:40:02.960 --> 00:40:05.677 diagnostic testing uh he could have so 00:40:05.760 --> 00:40:10.477 just went out and tested like crazy and fed that data into chatt with the 00:40:10.560 --> 00:40:12.477 assistance of some people who had built 00:40:12.560 --> 00:40:17.517 purpose-built application for it and was able I don't know if cure is the right 00:40:17.600 --> 00:40:22.077 word but to beat back the cancer to a degree. There was also this dog Rosie 00:40:22.160 --> 00:40:26.797 the dog in Australia. You guys heard of Rosie? like the craziest story where 00:40:26.880 --> 00:40:28.637 this guy I'm gonna get some detail wrong 00:40:28.720 --> 00:40:33.117 but a guy um biopsied his dog which had 00:40:33.200 --> 00:40:36.717 cancer um ran the mutations across of 00:40:36.800 --> 00:40:39.277 across Alphafold and then was able to 00:40:39.360 --> 00:40:41.917 design an mRNA vaccine that he injected 00:40:42.000 --> 00:40:44.877 into the dog um with the assistance of 00:40:44.960 --> 00:40:47.197 chatbots to build this thing which ended 00:40:47.280 --> 00:40:49.757 up being able to jump over tables again 00:40:49.840 --> 00:40:52.157 and the tumor shrunk. Um, when we think 00:40:52.240 --> 00:40:55.757 about the future of AI and health, um, 00:40:55.840 --> 00:40:57.837 help us sort out the truth with this 00:40:57.920 --> 00:41:02.877 question. Are these a couple of outliers that made good headlines, but there was 00:41:02.960 --> 00:41:05.677 something about the story we weren't hearing, or is this going to become 00:41:05.760 --> 00:41:09.037 standard in the future? >> Absolutely going to become standard. 00:41:09.120 --> 00:41:11.277 Absolutely. And it's I I personally have 00:41:11.360 --> 00:41:14.557 a number of friends who have done very 00:41:14.640 --> 00:41:18.477 similar things of get the data right 00:41:18.560 --> 00:41:20.797 your health diagnostics 00:41:20.880 --> 00:41:26.237 and use codecs right use these models to get insights from them and I think that 00:41:26.320 --> 00:41:27.757 there are many people like I think that 00:41:27.840 --> 00:41:31.277 there's about 230 million people each 00:41:31.360 --> 00:41:35.437 week who use chat GBT for health queries 00:41:35.520 --> 00:41:39.597 right and that's been that's this is like a staggering scale Right. And this 00:41:39.680 --> 00:41:44.157 is pe these are people sometimes you upload a scan. Sometimes you have 00:41:44.240 --> 00:41:48.477 doctors who are telling you conflicting information. And I think that we've been 00:41:48.560 --> 00:41:51.997 in a world where patients are not empowered, right? Patients have to be 00:41:52.080 --> 00:41:53.837 the doctor, right? You're the decider. 00:41:53.920 --> 00:41:57.597 You are accountable, right? You know, doctor makes a mistake and you're going 00:41:57.680 --> 00:42:01.597 to be paying the price for the rest of your life. Like it's just it's a very 00:42:01.680 --> 00:42:05.677 different kind of incentive. And this is very personal for me. you know, my wife 00:42:05.760 --> 00:42:08.797 has a number of health conditions and I think that we've just been we've not 00:42:08.880 --> 00:42:11.437 like I don't even know how we'd be able to manage many of her conditions right 00:42:11.520 --> 00:42:15.037 now without the use of of chat. And I think we're just at the beginning of 00:42:15.120 --> 00:42:19.677 this journey, right? That I think that the degree to which 00:42:19.760 --> 00:42:24.237 even if you have the best medical team, the best access, the best best experts, 00:42:24.320 --> 00:42:28.957 there's only so much that can be done, right? That you think about the things 00:42:29.040 --> 00:42:31.757 that are just outside of the reach of humanity or even just sometimes it's 00:42:31.840 --> 00:42:35.677 like someone didn't even read the chart, right? and kind of missed a detail. All 00:42:35.760 --> 00:42:39.517 of that we should be able to improve massively through these tools. And so I 00:42:39.600 --> 00:42:42.317 think that the personalized medicine and 00:42:42.400 --> 00:42:47.357 sometimes it's going to be about drugs and drug discovery that are of you know 00:42:47.440 --> 00:42:51.837 for for mass market but sometimes it'll be even for the kind of NF1 things like 00:42:51.920 --> 00:42:54.317 the the disease um diagnoses that I 00:42:54.400 --> 00:42:56.317 mentioned earlier today. Sometimes it 00:42:56.400 --> 00:42:59.277 will be for just like trying to 00:42:59.360 --> 00:43:03.677 understand conditions and trying to come up with with new potential therapeutics. 00:43:03.760 --> 00:43:07.517 All of that we're seeing it happening right now in front of our eyes. It's not 00:43:07.600 --> 00:43:11.517 theoretical. It's really happening. And so one of the most I think it's like one 00:43:11.600 --> 00:43:15.517 of the most astounding possibilities of AI is how much it can improve our 00:43:15.600 --> 00:43:17.837 health. And you think about the ripple 00:43:17.920 --> 00:43:19.677 effects of the system, right? Where so 00:43:19.760 --> 00:43:23.277 much spending on the health care system 00:43:23.360 --> 00:43:25.197 happens right now. that's a massive part 00:43:25.280 --> 00:43:27.357 of the economy and that if you're 00:43:27.440 --> 00:43:32.637 actually able to help people prevent issues, right, to to get ahead of 00:43:32.720 --> 00:43:36.237 potential, you know, health health problems, that's 00:43:36.320 --> 00:43:39.597 something that actually then alleviates a lot of burden, a lot of strain. And 00:43:39.680 --> 00:43:42.557 we're in a world where doctors are burned out, nurses are burned out, like 00:43:42.640 --> 00:43:46.237 there's like a real crisis that's happening in front of us. And I think AI 00:43:46.320 --> 00:43:48.717 will be able to help with all of that. 00:43:48.800 --> 00:43:50.717 Like we have that potential if we deploy 00:43:50.800 --> 00:43:54.157 it and use it wisely and well. And so I 00:43:54.240 --> 00:43:58.957 think that applying AI to medicine, like that's something that I is really a 00:43:59.040 --> 00:44:02.477 personal motivation for [music] me in in thinking about this whole journey of of 00:44:02.560 --> 00:44:05.357 what we're building, what we're trying to do with Open AI. And I'm hopeful that 00:44:05.440 --> 00:44:08.557 we as a as a world and a community can make the most of that. 00:44:08.640 --> 00:44:12.237 >> Let's hope. I think we will. I'm very very confident. 00:44:12.320 --> 00:44:13.757 >> Greg, thank you so much. 00:44:13.840 --> 00:44:16.317 >> Thank you. [applause] 00:44:16.400 --> 00:44:19.247 Great. Thank you. Thank you so much. 00:44:19.330 --> 00:44:21.350 [applause] 00:44:25.200 --> 00:44:29.077 Oh my god. Thank you everyone. You have a good time today. 00:44:29.160 --> 00:44:31.180 [applause] 00:44:33.440 --> 00:44:36.927 >> Thank you. Should we do it again next year? 00:44:37.010 --> 00:44:39.597 [applause] >> You going to come? Yeah. 00:44:39.680 --> 00:44:41.760 >> All