WEBVTT NOTE Sentence-level transcript of https://www.youtube.com/watch?v=pMggiOb18tc NOTE One cue per sentence. Cue ids are the line anchors on /transcripts/pMggiOb18tc.html. A cue ends where the next begins, or 2 s after its last word. s1 00:00:13.280 --> 00:00:15.880 Good morning, everyone. s2 00:00:15.920 --> 00:00:17.000 I'm Romain. s3 00:00:17.000 --> 00:00:18.960 Hey everyone, I'm Alexander. s4 00:00:18.960 --> 00:00:20.520 Wow, this room is incredible. s5 00:00:20.520 --> 00:00:25.520 There's over There's over 7,000 AI engineers here today with us. s6 00:00:25.520 --> 00:00:29.680 And you know, it's not just about who's talking about this technology. s7 00:00:29.680 --> 00:00:33.880 It's also about who's actually using it and pushing the frontier every day. s8 00:00:33.880 --> 00:00:37.480 So, we couldn't be more proud to be here with all of you today. s9 00:00:37.480 --> 00:00:42.480 And when we were thinking about this event with Alex, we kept coming back to the World's Fair. s10 00:00:42.480 --> 00:00:48.000 And the World's Fair made actually the future visible to everyone by building it in public, you know? s11 00:00:48.000 --> 00:00:52.560 Ideas that previously sounded impossible were actually suddenly there. s12 00:00:52.560 --> 00:00:53.800 People could see them. s13 00:00:53.800 --> 00:00:55.320 They could walk into them. s14 00:00:55.320 --> 00:00:57.600 And they could even start to believe in them. s15 00:00:57.600 --> 00:01:00.560 And honestly, this event has the exact same energy. s16 00:01:00.560 --> 00:01:03.880 The future of engineering is not arriving from somewhere else. s17 00:01:03.880 --> 00:01:09.320 It's really being built here by the people in this room and much faster than most expected. s18 00:01:09.320 --> 00:01:13.560 And that's why it's a little surprising that people keep saying that engineers are going away. s19 00:01:13.560 --> 00:01:18.800 The argument is that coding is, you know, abstracted away and therefore eventually we won't need engineers. s20 00:01:18.800 --> 00:01:21.160 Well, in fact, we think it's quite the opposite. s21 00:01:21.160 --> 00:01:23.920 You know, software ate the world. s22 00:01:23.920 --> 00:01:27.600 And then AI ate software. s23 00:01:27.600 --> 00:01:31.840 But now, what we're here to say is that the AI engineers are eating the world. s24 00:01:31.840 --> 00:01:34.800 AI engineers are the people here pushing the frontier. s25 00:01:34.800 --> 00:01:36.800 Yes. s26 00:01:37.680 --> 00:01:42.280 And you all are figuring out how this new capability can reach everyone. s27 00:01:42.280 --> 00:01:47.720 And there has never been a better time to be an engineer, in fact, because engineering was never about writing code. s28 00:01:47.720 --> 00:01:53.240 Engineering has always been about solving problems for yourself and for other people as well. s29 00:01:53.240 --> 00:02:02.800 It's about taking the latest science and combining it with design, with taste, with judgement, and most of all, imagination to make something that people can actually use. s30 00:02:02.800 --> 00:02:05.400 And in that sense, it's not the end of engineering. s31 00:02:05.400 --> 00:02:08.399 We think it's a return to the roots of engineering. s32 00:02:08.399 --> 00:02:14.240 And the technology we're building on is accelerating, getting faster and faster. s33 00:02:14.240 --> 00:02:21.080 For example, we used to ship a new model every 15 months or so, and now it's about roughly every 6 weeks. s34 00:02:21.080 --> 00:02:28.880 And in case you missed it, last week we launched a preview of the 5.6 series, and we're super excited to get into all of your hands. s35 00:02:28.880 --> 00:02:34.160 Now, building on top of all these models, the rate of product progress is relentless. s36 00:02:34.160 --> 00:02:39.959 And as a result, I don't have to tell you and the engineering feels completely different. s37 00:02:39.959 --> 00:02:45.440 So, just to go over a couple years of what for me were successive like mind-blowing experiences. s38 00:02:45.440 --> 00:02:53.959 You know, obviously for a long time we've had, you know, completion, and then we went to inline prediction, and then finally we had then we had command K where you could ask model to make a change, s39 00:02:53.959 --> 00:02:55.200 but they wouldn't test the work. s40 00:02:55.200 --> 00:03:00.880 Then models started testing the work, and now we have models taking on long, hard goals until they're done. s41 00:03:00.880 --> 00:03:09.000 And for me, each of these phases, I remember the first time was just mind-blowing, and then obviously afterwards you just get used to it and you're trying to get your work done. s42 00:03:09.000 --> 00:03:14.440 Yeah, in fact like I can't believe that build and test loop was not even part of the models just 2 years ago. s43 00:03:14.440 --> 00:03:22.480 This was a picture of me at Dev Day 2024, and I used 01 in preview at the time to build a mini drone in drone interface from scratch. s44 00:03:22.480 --> 00:03:31.360 And the slightly insane part is the model could not actually run the code or verify its own work, and I knew the demo would work most of the time, but surely not all of the time. s45 00:03:31.360 --> 00:03:32.840 So, I had to cross my fingers. s46 00:03:32.840 --> 00:03:36.200 You can kind of see here that I was pretty nervous, but hey, that's me. s47 00:03:36.200 --> 00:03:40.480 I only do live demos, so I never know what's actually going to happen each time. s48 00:03:40.480 --> 00:03:50.040 Luckily, it did work, and by Dev Day of last year in 2025, I was confident enough now the models could test their own work to kind of control an entire camera system and lighting system live. s49 00:03:50.040 --> 00:03:51.840 But yeah, we've come a long way. s50 00:03:51.840 --> 00:03:58.000 Yeah, so we refer to him as the demo god and you know, before the demo like I'll ask him so hey, how often does demo work? s51 00:03:58.000 --> 00:04:01.160 He'll be like, you know, three times out of four and we're all right, good luck. s52 00:04:01.160 --> 00:04:05.840 Um, obviously we've come miles since then and this year alone has been crazy. s53 00:04:05.840 --> 00:04:08.680 So, what we're putting up here are all the things that we've shipped so far. s54 00:04:08.680 --> 00:04:12.360 Not actually not even all, a selection of the things that we've shipped so far this year. s55 00:04:12.360 --> 00:04:17.600 Uh, my favorite things that we've shipped are like Codex app, goal mode, remote. s56 00:04:17.600 --> 00:04:20.600 These are things that really changes how it feels to do work. s57 00:04:20.600 --> 00:04:31.480 You know, obviously we couldn't do these things if we didn't use Codex to build Codex, but I think to me what is most interesting is that now Codex can do and agents can do any task that you can do on your own computer. s58 00:04:31.480 --> 00:04:39.120 And so, that means they're not just helping you with the coding, but they're helping you with what happens before the coding and they're helping you with what happens after the coding. s59 00:04:39.120 --> 00:04:40.240 And this is really key, right? s60 00:04:40.240 --> 00:04:50.640 I think there's been a lot of talk, there will be a lot of talk today about loops and if you can connect the agent to not only the work that you have to do, but why it has to be done, that's how you can get the agent to start to begin much more work. s61 00:04:50.640 --> 00:04:55.680 And then if you can connect it to what you do afterwards, review and deploy, that's how you help it land much more work. s62 00:04:55.680 --> 00:05:04.520 So, with all of this, of course we can move much faster, but to me as a product person, the most exciting thing is actually that we make better decisions around what to build. s63 00:05:04.520 --> 00:05:10.480 For instance, we try, we prototype many more ideas and we spend much more time with users. s64 00:05:10.480 --> 00:05:12.880 So, yes, that's all of you. s65 00:05:12.880 --> 00:05:18.600 So, wanted to pause and just give you all a big thank you both for the love and the constructive feedback. s66 00:05:18.600 --> 00:05:25.120 I would say it's safe to say that we, Codex and actually the entire industry wouldn't be here without you. s67 00:05:25.120 --> 00:05:26.520 Yeah, thank you so much for all the feedback. s68 00:05:26.520 --> 00:05:28.160 We're constantly listening to all of you. s69 00:05:28.160 --> 00:05:29.468 Thank you. s70 00:05:29.468 --> 00:05:30.360 [applause] s71 00:05:30.360 --> 00:05:31.320 Okay. s72 00:05:31.320 --> 00:05:34.760 So, the models are getting really good. s73 00:05:34.760 --> 00:05:44.120 I would say, if you pick like a medium length computer task and you give me and the model the same amount of time to get that task done, probably s74 00:05:44.120 --> 00:05:48.000 at least in my case, the model will do a better job than me for the average task. s75 00:05:48.000 --> 00:05:51.280 And so, okay, we're we're getting these models, you know, in some ways they're smarter than us. s76 00:05:51.280 --> 00:05:53.600 They can do almost anything. s77 00:05:53.600 --> 00:05:54.720 How should we shape that? s78 00:05:54.720 --> 00:05:57.000 What should the products that we use feel like? s79 00:05:57.000 --> 00:06:00.080 And so, to answer that, we look to our mission. s80 00:06:00.080 --> 00:06:04.120 The part of it here that I've got up is, you know, AGI that benefits all of humanity. s81 00:06:04.120 --> 00:06:08.920 And I think in order to do this, there are two main questions that I think about now. s82 00:06:08.920 --> 00:06:12.840 One is, how do we set up the agents to actually do things in the world? s83 00:06:12.840 --> 00:06:14.160 So, what can they do? s84 00:06:14.160 --> 00:06:16.680 You know, gradually agents are getting connected to more and more things. s85 00:06:16.680 --> 00:06:17.520 Then where do they run? s86 00:06:17.520 --> 00:06:19.240 More on that later. s87 00:06:19.240 --> 00:06:22.360 And then the other question is, how do we use these agents? s88 00:06:22.360 --> 00:06:24.880 For us, you know, what what should the product feel like around them? s89 00:06:24.880 --> 00:06:29.000 And for us, the goal is squarely not to automate engineers. s90 00:06:29.000 --> 00:06:34.200 Instead, the the product shape that we want is one that maximally empowers engineers. s91 00:06:34.200 --> 00:06:38.760 So, you know, if we think about what that product shape is, we actually think it's pretty simple. s92 00:06:38.760 --> 00:06:41.800 I I read a lot of sci-fi, and you know, watch superhero movies. s93 00:06:41.800 --> 00:06:45.400 And I actually think that the the simple ideas in there are approximately right. s94 00:06:45.400 --> 00:06:48.080 So, there are two modalities, roughly. s95 00:06:48.080 --> 00:06:48.919 Chat. s96 00:06:48.919 --> 00:06:50.800 I actually think I know some people think chat is dead. s97 00:06:50.800 --> 00:06:52.480 I think chat is underrated. s98 00:06:52.480 --> 00:06:56.080 Uh and some kind of hands-on experience. s99 00:06:56.080 --> 00:07:02.000 So, what you want is a single entity that you can ask for help with anything, anywhere. s100 00:07:02.000 --> 00:07:09.640 And then you want a a powerful collaborative UI that you can use when you want to inspect, steer, or shape things yourself. s101 00:07:09.640 --> 00:07:16.400 And so, I had Codex image gen me an illustration of this to help understand when you might want to use these things. s102 00:07:16.400 --> 00:07:19.000 And so, my analogy for you, yeah, I hope you like the image gen. s103 00:07:19.000 --> 00:07:22.840 Uh my analogy for you would be it's just like working with a team. s104 00:07:22.840 --> 00:07:27.320 Most of the time, you're just talking about stuff, and your team is just doing stuff. s105 00:07:27.320 --> 00:07:33.840 You don't actually want to watch over the shoulder, or like have to walk over to the workbench of your teammate for every single unit of work. s106 00:07:33.840 --> 00:07:35.760 Mostly, you just want to talk and let them cook. s107 00:07:35.760 --> 00:07:41.720 And then every now and then, you want to dig in and really dig in all the way to the weeds of things, and dig digging that problem together. s108 00:07:41.720 --> 00:07:49.720 And for us, as we build product, we have this idea that we want to make it so that you can retain this feeling of mastery of the work that we're doing, s109 00:07:49.720 --> 00:07:50.800 because that's really powerful. s110 00:07:50.800 --> 00:07:57.800 We don't want to make it feel like actually it's really hard to like get to the details and like, you know, disassemble the hardware in this case. s111 00:07:57.800 --> 00:08:03.640 So, the way that we're bringing to this to life is just the beginning, but this is why we built the Codex app. s112 00:08:03.640 --> 00:08:08.120 You get a very simple chat interface that you can use for coding and for anything else. s113 00:08:08.120 --> 00:08:10.760 And you know, you can have a conversation and then go deep as you want. s114 00:08:10.760 --> 00:08:15.840 So, in the case here, we have Homan's predicted score of this upcoming World Cup match. s115 00:08:15.840 --> 00:08:17.280 I hope I'm right. s116 00:08:17.280 --> 00:08:17.880 I hope I'm right. s117 00:08:17.880 --> 00:08:19.115 We'll see. s118 00:08:19.115 --> 00:08:19.560 [laughter] s119 00:08:19.560 --> 00:08:20.320 Okay. s120 00:08:20.320 --> 00:08:25.800 Um and what you can do here is you could go in and you can point at a very specific thing and say, "Hey, I want you to make this change." s121 00:08:25.800 --> 00:08:27.720 Or you can make this change yourself. s122 00:08:27.720 --> 00:08:38.000 And a fun story here is that actually I remember pitching some of you who I know are in the audience uh this idea before we started and I was told squarely like, "We I will never use such a tool. s123 00:08:38.000 --> 00:08:39.919 I will never leave my terminal." s124 00:08:39.919 --> 00:08:41.760 Uh or Vim or Emacs. s125 00:08:41.760 --> 00:08:44.360 Um but actually those people are now using it. s126 00:08:44.360 --> 00:08:49.600 And even internally like within our team, there were a lot of questions like, "Why should we build this? s127 00:08:49.600 --> 00:08:51.960 People love the CLI, they love the IDE." s128 00:08:51.960 --> 00:08:58.800 And it's a little subtle, but our take is that you can't really build that collaborative interface for any kind of work in a CLI. s129 00:08:58.800 --> 00:09:00.000 It's mostly chat. s130 00:09:00.000 --> 00:09:09.440 And then in an IDE, the order is wrong and so you're starting with the code, but now it's time to transition to like working with teammates where you chat first and you dig in when you need it. s131 00:09:09.440 --> 00:09:10.440 Totally. s132 00:09:10.440 --> 00:09:17.440 And we're moving really fast on this, on the product surface and the model layer, of course, but we're also trying to keep pace with all of you, right? s133 00:09:17.440 --> 00:09:24.400 Honestly, half the time I open X, I see someone in this room doing something that I had not realized Codex could actually do. s134 00:09:24.400 --> 00:09:26.160 And honestly, this is what pioneers do. s135 00:09:26.160 --> 00:09:30.960 You guys experiment, you set up tools for yourself, for your team, and in turn, we get inspired. s136 00:09:30.960 --> 00:09:34.240 We learn from you and eventually everyone benefits. s137 00:09:34.240 --> 00:09:42.120 And so we are helping uh uh you are helping us figure out like what to uh what to build next and also what the future of engineering should look like. s138 00:09:42.120 --> 00:09:49.360 But for that to work, one thing that we we really care about is that Codex cannot be a closed product that only OpenAI can improve. s139 00:09:49.360 --> 00:09:54.120 So we've intentionally designed Codex as a set of layers that anyone can build on. s140 00:09:54.120 --> 00:09:59.520 And we want to show you a little bit of that stack today and how it it manifests. s141 00:09:59.640 --> 00:10:04.160 First, it starts with the model and Alexander showed how quickly we're progressing on models. s142 00:10:04.160 --> 00:10:07.360 And you guys use these models through the responses API and guess what? s143 00:10:07.360 --> 00:10:09.520 This is how we built the Codex app, right? s144 00:10:09.520 --> 00:10:12.600 We use the same models and through the same API. s145 00:10:12.600 --> 00:10:16.080 We actually are building on the same thing that we give to developers. s146 00:10:16.080 --> 00:10:22.640 And when Codex needs something new, we always try to bake it into the API first so you can benefit as well. s147 00:10:22.640 --> 00:10:24.480 One example recently was compaction. s148 00:10:24.480 --> 00:10:30.200 Codex needed a way to compact long contexts for long-running tasks and so we'll build that into the API. s149 00:10:30.200 --> 00:10:35.600 So that means your agents can use the same primitives that we built for ourselves. s150 00:10:35.640 --> 00:10:42.440 Moving on to the next layer, the Codex harness is also open source so you can inspect it, you can fork it, you can adapt it. s151 00:10:42.440 --> 00:10:45.040 And we also took the same approach with Agent MD. s152 00:10:45.040 --> 00:10:54.280 Instead of reinventing a new file format for Codex for to follow instructions, we thought let's pick a name that other agents can actually use as well. s153 00:10:54.280 --> 00:10:59.280 The models are hard to default in the in the harness the models from OpenAI but they're not hard coded in there. s154 00:10:59.280 --> 00:11:03.720 So if you want to use an open model and keep the same agent loop, you can. s155 00:11:03.720 --> 00:11:08.120 And we also bring this Codex harness into the post-training process of our models. s156 00:11:08.120 --> 00:11:15.800 So that means the models can learn to call tools and navigate an environment that's actually something that's open source. s157 00:11:15.800 --> 00:11:17.720 Now take the open code team for instance. s158 00:11:17.720 --> 00:11:28.680 They actually were able to inspect how we have this like reference implementation and they could reuse the parts that make sense to them or change entirely all the rest and make different choices. s159 00:11:28.680 --> 00:11:35.640 I know for instance they were trying to see how we did like signing with ChatGPT and so they could look at the code and and and learn from it. s160 00:11:35.640 --> 00:11:43.480 And we think it's better than having developers reverse engineering how it builds what and and how we launch. s161 00:11:43.480 --> 00:11:46.920 But now let's say, speaking of subscriptions, that we want to go a level higher. s162 00:11:46.920 --> 00:11:49.800 And how you bring this harness into an app? s163 00:11:49.800 --> 00:11:53.200 And how do you let people sign in with their existing Codex subscription, for instance? s164 00:11:53.200 --> 00:11:59.240 Well, it turns out we had the same problem ourselves, cuz we wanted to build a VS Code extension and the Codex app. s165 00:11:59.240 --> 00:12:02.800 And we wanted to have a unified way to like actually control this harness. s166 00:12:02.800 --> 00:12:06.480 So we built app server, and we also made that open source. s167 00:12:06.480 --> 00:12:12.120 And the app server is not kind of a community adapter, it's really the path that we use for our own products. s168 00:12:12.120 --> 00:12:13.720 And you can use it, too. s169 00:12:13.720 --> 00:12:22.880 Toma, for instance, here, aka Demilyan on X, he built his own native app for Codex, called Codex Monitor, before we even launched the Codex app. s170 00:12:22.880 --> 00:12:25.800 Cuz he could build that using the app server. s171 00:12:25.800 --> 00:12:31.520 And now he works on our team, and he actually built Codex for iOS. s172 00:12:31.840 --> 00:12:37.320 And moving up the stack, at the app layer, we also want to make sure that innovation is not blocked on our own ideas. s173 00:12:37.320 --> 00:12:44.360 And so we built extensible primitives here, like the in-app browser that we showed on the screen, and plugins. s174 00:12:44.360 --> 00:12:54.520 So if you take, for instance, browser use and computer use, these were built as plugins using the exact same extension points that we have available for for all of you. s175 00:12:54.520 --> 00:13:04.080 And lastly, we also recently built role-specific plugins for Codex, say, to make it easier to to customize for people who work in data science or design, for instance. s176 00:13:04.080 --> 00:13:05.960 And these plugins are also open source. s177 00:13:05.960 --> 00:13:09.360 You can see under the hood and get inspired from them if that's useful. s178 00:13:09.360 --> 00:13:12.880 Our goal is really to keep making this as open and flexible as we can. s179 00:13:12.880 --> 00:13:24.240 And the best part is people can use their existing subscription in more and more places, from Open Code, Pi, Droids, Open Claw, to even Xcode and JetBrains as IDEs. s180 00:13:24.240 --> 00:13:32.320 And you can see how they're becoming quite a meaningful part of how people use these tools and that's really why we want to care about building this open ecosystem s181 00:13:32.320 --> 00:13:33.240 with all of you. s182 00:13:33.240 --> 00:13:37.600 So really if there's one thing I want you to take away from this section and this talk, it's this. s183 00:13:37.600 --> 00:13:43.520 We're not building one system for Open AI and a second system that's simplified for developers. s184 00:13:43.520 --> 00:13:46.880 At every layer, we actually use the thing that we give to you. s185 00:13:46.880 --> 00:13:55.160 And we want to thank all of you cuz every time you fork the harness, every time you find the edge of capabilities of the models, it means we get to learn and improve. s186 00:13:55.160 --> 00:14:07.480 And honestly, with 7,000 of the finest AI engineers in this room today, I'm confident that all of you will define a lot of how we uh will experience AI and how the world will experience AI in the future. s187 00:14:07.480 --> 00:14:09.840 So thank you. s188 00:14:10.199 --> 00:14:12.199 [applause] s189 00:14:14.120 --> 00:14:16.920 I want to give a shout-out to whoever over there is injecting energy. s190 00:14:16.920 --> 00:14:18.880 That's you, okay, thank you so much. s191 00:14:18.880 --> 00:14:24.760 Uh so, with all of your help, we are making agents explosively useful. s192 00:14:24.760 --> 00:14:28.960 Uh and so now the question is how do we get value out of them? s193 00:14:28.960 --> 00:14:30.920 And you know, that's not token maxing. s194 00:14:30.920 --> 00:14:34.240 We have a term for this that maybe you use as well. s195 00:14:34.240 --> 00:14:35.160 I don't know, is it on screen? s196 00:14:35.160 --> 00:14:36.640 Value maxing. s197 00:14:36.640 --> 00:14:43.680 So, you know, when we talk to engineering leaders, most of the conversation is about some themes relating to the idea of value maxing. s198 00:14:43.680 --> 00:14:51.320 So, we're going to walk you through a few common topics that come up, some things where we've already made a lot of progress, and some things where actually there's a lot more progress to still be made. s199 00:14:51.320 --> 00:14:54.640 So, the first one of these is cost efficiency. s200 00:14:54.640 --> 00:14:57.080 Everyone wants frontier intelligence. s201 00:14:57.080 --> 00:15:00.040 Pick your favorite eval, you want the best model. s202 00:15:00.040 --> 00:15:06.880 So, with terminal bench here for instance, that's GPT 5.6 Sol, and like I said, we can't wait for you to have it. s203 00:15:06.880 --> 00:15:13.200 But okay, you also want as much intelligence as you can get, and that's where efficiency comes in. s204 00:15:13.200 --> 00:15:18.640 Cost efficiency has been a focus for us for quite some time, and the results are continuing to pay off. s205 00:15:18.640 --> 00:15:23.200 So, for example, GPT 5.6 Terra, I think it's in like dark blue in there. s206 00:15:23.200 --> 00:15:27.080 Brings GPT 5.5 level intelligence, but at half the cost. s207 00:15:27.080 --> 00:15:31.600 And Luna there beats some pretty notable models in this eval. s208 00:15:31.600 --> 00:15:37.280 But at only $1 per million input tokens and $6 per million output tokens. s209 00:15:37.280 --> 00:15:41.080 I'll leave it up to you to compare those costs, but that is insane value. s210 00:15:41.080 --> 00:15:47.240 Yeah, we we we really can't wait to uh, to see all of you build uh, with GPT 5.6 and this new family of models. s211 00:15:47.240 --> 00:15:49.640 Now, the next thing I want to touch on is speed, right? s212 00:15:49.640 --> 00:15:56.760 GPT 5.3 could spark showed you what speed can unlock, but we also know that you all want frontier intelligence. s213 00:15:56.760 --> 00:16:01.760 You don't want to have a model that's like not as great as what you can operate at the very best. s214 00:16:01.760 --> 00:16:09.360 Well, this is GPT 5.6 Soul running on Cerebras, the frontier intelligence at now 750 tokens a second. s215 00:16:09.360 --> 00:16:12.360 We can't wait to see what you can build with this next month. s216 00:16:12.360 --> 00:16:19.440 And honestly, to put that in perspective, this is kind of like having a pretty substantial PR written in like 10 seconds. s217 00:16:19.440 --> 00:16:22.800 And it's not just about getting one answer faster, right? s218 00:16:22.800 --> 00:16:25.360 It's about what can you do with that speed? s219 00:16:25.360 --> 00:16:36.320 You can think about an agent taking different approaches, maybe like five or six in parallel, maybe like, you know, coming back and picking the best one in the time you would have taken to not even generate just one. s220 00:16:36.320 --> 00:16:42.640 So, we really can't wait to see what that can unlock when you have frontier intelligence, the very best at that speed. s221 00:16:42.640 --> 00:16:50.839 It really starts to feel less like waiting for an AI to respond and much more like a coworker that's like already showing you the results as it goes. s222 00:16:50.839 --> 00:16:54.200 Speaking of working with coworkers, um, can I get a show of hands? s223 00:16:54.200 --> 00:16:58.800 Who who is familiar with this kind of sight in offices? s224 00:16:59.280 --> 00:17:00.120 Okay. s225 00:17:00.120 --> 00:17:01.760 Okay, wow, a lot of you are very well behaved. s226 00:17:01.760 --> 00:17:03.440 I really see some people up front. s227 00:17:03.440 --> 00:17:09.280 Um, so yeah, a lot of people are keeping their laptops open so that agents can keep working. s228 00:17:09.280 --> 00:17:20.520 And this is funny, but you know, what we really want is to be able to shut our computers um, and we want to be able to run many tasks in parallel isolated on their own box. s229 00:17:20.520 --> 00:17:23.400 Now, we've been actually aiming at this from the start. s230 00:17:23.400 --> 00:17:29.720 Our first major launch was Codex Cloud, and it is due for some major upgrades coming soon. s231 00:17:29.720 --> 00:17:40.120 But, better yet, as we think about this, the future shouldn't have this awkward distinction between like a local task and a cloud task, and you have to decide where to run everything. s232 00:17:40.120 --> 00:17:43.320 Really, what you should have is kind of going back to what I was saying earlier. s233 00:17:43.320 --> 00:17:53.360 You should just have an agent, you talk to it wherever, whenever, about anything, and it should figure out, okay, what do I need to do, which environment is right for my work, and use whatever is available. s234 00:17:53.360 --> 00:17:57.920 In fact, Theo made this prediction over the weekend on this very topic. s235 00:17:57.920 --> 00:17:59.560 And it's a pretty acute tweet. s236 00:17:59.560 --> 00:18:00.880 Like, Alex, what do you think? s237 00:18:00.880 --> 00:18:02.320 Sooner or later than 6 months? s238 00:18:02.320 --> 00:18:07.320 I think the not in maybe not exact details, but the vibe of this tweet much sooner than 6 months. s239 00:18:07.320 --> 00:18:11.720 Yeah, I mean, at at the pace at which everything is going is going, I would not be surprised if it's sooner indeed. s240 00:18:11.720 --> 00:18:17.680 Um, Well, so now, you might be wondering, where's the live demo today? s241 00:18:17.680 --> 00:18:23.120 Uh, well, for this AI engineer, we wanted to do something a little different this time around. s242 00:18:23.120 --> 00:18:29.360 And we think it's a very unique moment for all of us to kind of reimagine how we work and how we build. s243 00:18:29.360 --> 00:18:38.520 And so, we wanted to bring a special guest who has bent what's possible with agents and really has pushed us to be more AGI pilled uh at OpenAI. s244 00:18:38.520 --> 00:18:43.992 So, with that, please welcome to the stage the claw father, Peter Steinberger. s245 00:18:43.992 --> 00:18:45.992 [applause] s246 00:18:48.577 --> 00:18:50.577 [applause] s247 00:18:50.680 --> 00:18:51.840 Peter, take it away. s248 00:18:51.840 --> 00:18:53.120 Thank you all. s249 00:18:53.120 --> 00:18:55.280 Thanks, Al. s250 00:18:56.920 --> 00:18:58.840 Good morning, everyone. s251 00:18:58.840 --> 00:19:06.800 You know, I love this picture because it reminds me just how much has changed in a few months. s252 00:19:06.800 --> 00:19:16.640 I was juggling 10 or more terminal windows, always waiting for one of them to finish, so I could steal the agent and queue new work. s253 00:19:16.640 --> 00:19:21.200 In In January, that felt like peak productivity. s254 00:19:21.200 --> 00:19:23.960 Today, it feels a little bit silly. s255 00:19:23.960 --> 00:19:25.880 I thought I was orchestrating. s256 00:19:25.880 --> 00:19:27.480 Really, I was polling. s257 00:19:27.480 --> 00:19:32.640 I was the scheduler, the router, and the memory. s258 00:19:33.040 --> 00:19:36.440 You know, at first, I paired with one agent. s259 00:19:36.440 --> 00:19:39.520 With 10 terminals, I was no longer pairing. s260 00:19:39.520 --> 00:19:43.280 I was managing 10 direct reports. s261 00:19:43.280 --> 00:19:50.520 Now, I mostly talk to a long-running manager, which delegates work to a team. s262 00:19:50.520 --> 00:19:57.000 For tricky work, I can still drop down and pair directly with a worker, but my default changed. s263 00:19:57.000 --> 00:20:02.600 I manage the manager of a small company of agents. s264 00:20:02.720 --> 00:20:05.720 Three changes made it possible. s265 00:20:05.720 --> 00:20:15.880 Number one, server-side compaction made long-running tasks reliable enough that I stopped optimizing around first sessions. s266 00:20:15.880 --> 00:20:21.560 Coordination lets one thread create and steer the right projects. s267 00:20:21.560 --> 00:20:26.840 And third, automation can wake the same manager when something happens. s268 00:20:26.840 --> 00:20:32.040 So, we have persistent context, delegation, and triggers. s269 00:20:32.040 --> 00:20:34.400 There's your loop. s270 00:20:34.600 --> 00:20:42.280 And once the loop starts working, you discover the next problem, the bottleneck keeps moving. s271 00:20:42.280 --> 00:20:49.120 You know, last year, I was primarily constrained by tokens. s272 00:20:49.160 --> 00:20:53.040 Now, I fixed that by joining OpenAI. s273 00:20:53.760 --> 00:20:57.360 I know I know the strategy does not scale. s274 00:20:57.800 --> 00:21:04.120 Then, my constraint shifted to token uh compute. s275 00:21:04.120 --> 00:21:10.320 All these threads run at the same time, and my MacBook starts sounding like a jet engine. s276 00:21:10.320 --> 00:21:13.480 That's mostly fixed by using test boxes. s277 00:21:13.480 --> 00:21:18.040 So, agents can run tests on a separate machine. s278 00:21:19.560 --> 00:21:24.240 Now, I'm primarily constrained by attention. s279 00:21:24.240 --> 00:21:28.360 And unlike tokens or compute, I can't simply add more of it. s280 00:21:28.360 --> 00:21:36.080 So, the most important skill is today is deciding where to spend it. s281 00:21:36.080 --> 00:21:42.080 Are you still staring at the agent while the code flies by? s282 00:21:42.840 --> 00:21:49.840 I know I know it's it's it feels cool, but with the earlier models, this was necessary. s283 00:21:49.840 --> 00:21:57.400 You know, you you you you see the agent go in a direction you don't like, you hit escape, you steer it, you steer it back. s284 00:21:57.400 --> 00:22:07.400 But, the latest generation of models is so good at understanding intent that it's a little bit of a waste of time to watch the agent generate code. s285 00:22:08.760 --> 00:22:14.440 Imagine someone files an issue on one of my open source projects. s286 00:22:14.440 --> 00:22:23.440 The manager wakes up, reads it against the project's goals, notes, and vision, and decides whether it might be a fit. s287 00:22:23.440 --> 00:22:26.600 If it does, it creates a worker. s288 00:22:26.600 --> 00:22:35.000 That worker investigates, implements the change, runs the tests, and another agent can review the result. s289 00:22:35.000 --> 00:22:43.560 I don't need to watch those agents work or consume every intermediary message. s290 00:22:43.560 --> 00:22:54.680 When the manager needs me, it returns a PR, the original issue, the proposed diff, maybe a video, or even a running build I can VNC into. s291 00:22:54.680 --> 00:23:04.560 I review once, I leave a note, I maybe approve, the loop continues, and it can land after the checks pass. s292 00:23:04.560 --> 00:23:08.640 The agent runs the inner execution loop. s293 00:23:08.640 --> 00:23:14.320 I set the direction and I make decisions in the outer loop. s294 00:23:14.520 --> 00:23:18.800 You know, Paul Paul thought he's already running a version of this. s295 00:23:18.800 --> 00:23:21.000 He pinned his chief of staff. s296 00:23:21.000 --> 00:23:25.840 It wakes up every 10 minutes and it coordinates his GitHub work. s297 00:23:25.840 --> 00:23:33.920 The agent creates threads in the sidebar so Paul can jump in whenever the work needs additional steering. s298 00:23:34.640 --> 00:23:42.000 You know, once the manager is long-lived tying it to a laptop just feels wrong. s299 00:23:42.000 --> 00:23:45.960 Codex can already move work between hosts. s300 00:23:45.960 --> 00:23:49.640 Open claw has a gateway and nodes. s301 00:23:49.640 --> 00:23:54.680 But neither feels like the final form. s302 00:23:54.680 --> 00:23:59.080 I don't even want to think where I work. s303 00:23:59.080 --> 00:24:03.680 My agent should be able to connect to any of my machines. s304 00:24:03.680 --> 00:24:12.640 They should know which work can be done in the cloud or which work requires my local machine. s305 00:24:13.040 --> 00:24:17.760 The manager shouldn't be a session trapped inside your app. s306 00:24:17.760 --> 00:24:24.320 It should be an agent that I can text, steer from Slack, or hear from wherever I am. s307 00:24:24.320 --> 00:24:32.200 Really, why can't I talk to my agent and have it design the whole loop for me? s308 00:24:32.560 --> 00:24:34.760 We haven't solved that yet. s309 00:24:34.760 --> 00:24:41.680 Models are advancing faster than the harnesses and organizations around them. s310 00:24:41.680 --> 00:24:46.320 Designing those things is the next engineering problem. s311 00:24:46.320 --> 00:24:49.720 That's where all of you come in. s312 00:24:49.720 --> 00:24:52.480 The future is not 20 terminals. s313 00:24:52.480 --> 00:24:54.080 It's better loops. s314 00:24:54.080 --> 00:24:55.640 Let's build them. s315 00:24:55.640 --> 00:24:57.720 Thank you. s316 00:24:57.720 --> 00:24:59.880 Thank you.