OpenAI Unveils Dots Agent Platform as Sottiaux Predicts Bot-Run Web

OpenAI推出Dots智能体平台，Sottiaux预测网络将由Bot主导 鉴于自主AI智能体很快将执行绝大多数在线操作，且模型能力在一年内实现十倍提升，手工构建的智能体循环和模型选择器将被直接从用户目标中学习的常驻全天候环境系统所取代。 Tibo Sottiaux指出，产品构建者尚未充分考虑到绝大多数在线操作不久后都将由自主智能体驱动，随着基础模型的迅速迭代，手工构建的循环与执行图将走向过时。他阐述了基于OpenAI Astra模型构建的常驻全天候个人助理平台Dots的愿景，该平台可跨越物理与数字空间运行，无需用户微调配置或使用模型选择器。Sottiaux还解释了手动编程如何大幅被Codex等自动化工具取代，进而使产品品味、用户共情力和创始人直觉等特质的重要性超越纯粹的工程技巧。最后，他详细说明了OpenAI的安全策略，强调分配专门算力用于二级监控，并有意放缓前沿模型的发布节奏，以安全服务逾十亿受众。 OpenAI推出了常驻全天候智能体平台Dots，并计划将其持久记忆和自主能力直接融入标准ChatGPT中，以消除独立的模式和模型选择器。 (4:13) Tibo Sottiaux预计，未来绝大部分互联网流量和数字操作将由自主智能体而非人类发起，这要求产品为前所未有的自动化规模做好准备。 (0:00) 手工设计的智能体循环和图工作流被视为过渡方案，随着突破性模型能够吸收完整上下文并自发学习，这些方案将迅速消失。 (2:26) OpenAI的安全策略正越来越多地将算力分配给负责监督主要工作智能体的实时二级监控智能体，这促使该公司有意推迟6.1 Astra的发布以确保对齐。 (33:55)

OpenAI Head of ChatGPT Tibo Sottiaux argues that product development must anticipate models becoming tenfold more capable within a single year, shifting the bulk of online actions from humans to autonomous agents. In his view, rigid workflow orchestration and complex configuration interfaces will be superseded by ambient, safety-monitored intelligence where human taste and curation remain the primary differentiators. Model Leaps and Transitory Orchestration Loops (0:00) Sottiaux contends that many developers are mistakenly designing products around current system constraints instead of preparing for a tenfold jump in model capability within twelve months. In his own engineering workflow, pushing frontier limits initially requires spinning up complex swarms of specialized agents. However, each subsequent breakthrough model consolidates that work into a single model capable of maintaining large context windows and dynamic memory, contracting the swarm until the next frontier push. Because of this recurring cycle, Sottiaux views the industry's focus on manual loop chaining and hardcoded execution graphs as a temporary workaround for current limitations. He argues that future systems will learn organically from high-level goals and feedback rather than rigid flowcharts. Internally at OpenAI, this evolution has already diminished manual programming; Sottiaux notes that Codex writes nearly all his operational scripts and analytical pipelines, leaving hands-on coding as little more than a recreational hobby. Ambient Agents and Unified Interfaces (4:13) OpenAI's personal agent platform, Dots, is architected to operate continuously as an ambient layer rather than a screen-tethered application. Built on an 'octopus' model, the agent harness runs independently in a virtual machine or dedicated compute environment rather than on a user's laptop, securely orchestrating up to ten connected physical or virtual endpoints. OpenAI launched Dots with a single primary assistant to study habit formation and user preferences, but plans to allow users to spawn specialized virtual teams to handle distinct tasks like social monitoring or operational alerts. To eliminate user confusion across an audience projected to reach 1.2 billion users, OpenAI is streamlining its product surfaces. Sottiaux acknowledges that toggling between Chat and Work modes or parsing complex model pickers creates unnecessary friction. OpenAI intends to fold Dots' continuous memory and autonomous execution directly into ChatGPT, removing model selection and mode switching so interactions feel as frictionless as natural conversation in shared physical spaces. Infrastructure Strains and Agent Economics (11:36) Sottiaux predicts that the majority of internet traffic and software actions will soon be executed directly by agents. Opening applications to automated workflows triggers severe infrastructure demands; for instance, when Notion launched its Model Context Protocol integration, the company experienced a rapid surge of agent-driven traffic that strained operational capacity and forced a reevaluation of serving unit economics. To establish standard rails for this transition, OpenAI developed 'Sign in with ChatGPT' alongside sixteen launch partners. Furthermore, third-party developers receive a direct distribution and monetization mechanism: when ChatGPT subscribers burn their allocated usage pools interacting with plugins, developers earn a proportional revenue share. Rather than optimizing for marketing keywords, OpenAI ranks and surfaces these plugins dynamically based on retention, conversation utility, and task success metrics. Frontier Pacing and Secondary Monitoring Compute (33:55) To mitigate risks such as prompt injections and unauthorized high-risk operations, OpenAI dedicates parallel compute to 'secondary monitoring.' This secondary safety system watches the primary working agent in real time, automatically interrupting tasks whenever actions breach safety bounds. Sottiaux notes that Dots was deployed on top of Astra, which he characterizes as OpenAI's safest and most aligned model to date. Sottiaux defines 'pacing the frontier' as investing heavily in alignment and intentionally withholding capability jumps until safeguards are established. As an example, he highlights OpenAI's deliberate decision to hold back the release of '6.1 Astra,' opting instead to deploy a more efficient near-Astra model. Sottiaux maintains that serving over a billion non-technical consumers aligns OpenAI's business interests directly with safety, since real-world autonomous tasks—such as an agent detecting an active production outage minutes before a DevDay presentation—demand strict operational boundaries. While coding tasks and system orchestration are becoming automated, Sottiaux maintains that human taste, user empathy, and product judgment remain irreplaceable. However, whether broader software infrastructure and commercial platforms can adapt to the scale and unit economics of an agent-dominated internet without facing severe operational bottlenecks remains an ongoing challenge.

[0:00] Tibo Sottiaux: You're not quite getting it. [1:16] Tibo Sottiaux: What is coding? [2:26] Tibo Sottiaux: Having to set up and fiddle with your loops is something that maybe people got excited about, but I don't think this is the way that it's going to work. Over time, you just want a system that learns... You don't want to necessarily think about how I'm just like, 'I'm going to loop it exactly this way in order to get good results.' [8:03] Tibo Sottiaux: There's just no configuration. You just talk to it. The only thing that you have to configure is like which channels you want to talk to. [8:03] Tibo Sottiaux: I don't know how we would do it without the community. [11:36] Tibo Sottiaux: Build a good plugin. [16:20] Tibo Sottiaux: It's just really this artistic tool, right? And we may not have coders anymore, but we have more builders than ever. [17:25] Tibo Sottiaux: For me, the ideal way to get things done... would be like to just have it in your physical space. You know, we're just having conversation like we have now. It's able to sort of like, you know, observe, you know, like hear the ideas that we have. We can, you know, jot something down in a white paper on a paper, you know, I can maybe like, you know, take that and you know, start building in the background and then you're like, 'You know, actually have another idea,' and then you just like, it starts building another thing. [20:09] Tibo Sottiaux: Which was actually quite stressful because it was like, 'Hey, production is down.' And uh it's like, 'Oh well, you know, it's like do you want me to try and fix it?' And I'm like, 'Um, I don't think you're there yet, you know, little Dot, but you know, thank you for trying.' [29:00] Tibo Sottiaux: Break something, celebrate things. [33:55] Tibo Sottiaux: We build products for 1.2 billion people. We can't screw that up. [36:15] Tibo Sottiaux: You kind of need a PhD in model pickers, right? And I just want to get rid of all of that as quickly as we can.

[0:00] Often when I look at what people are [0:01] building out there, it's just like, [0:02] "You're not quite getting it." Like, you [0:04] know, if you were just like pushing [0:05] yourself and just really imagining like [0:07] all of this being [0:08] >> [music] [0:08] >> roughly 10 times better than it is [0:10] today, you know, like in a year, but you [0:12] would build in a different way. [0:13] >> What do you think people aren't pricing [0:14] in in terms of where things are going? [0:16] >> The majority of actions on the internet [0:18] will be taken by agents. If you want [0:20] your [music] product to be successful [0:21] for agents, you have to build, you know, [0:22] for a certain level of scale. [0:23] >> How many agents do you find yourself [0:24] running in parallel just uh in the in [0:26] your kind of your day-to-day work? [0:28] >> As I'm pushing the frontier, I find [0:29] myself like building larger and larger [0:31] teams of agents. And then when we have [0:33] the next breakthrough with models, like [0:35] suddenly it's just like I'm like, "Oh, [0:36] well, you know, a bigger agent can just [0:38] do all of it." And so I kind of shrink [0:40] the team again. And then it just [music] [0:41] kind of goes through this expansion and [0:42] shrinking expansion. [0:43] >> It kind of makes me think about there's [0:44] this whole loops thing and then this [0:45] whole graphs thing. [0:46] >> Having to set up and fiddle with your [0:48] loops is something that maybe people got [0:51] excited about, but I don't think this is [0:52] the way that it's going to work. Over [0:54] time, you just want to system that [0:55] [music] learns. You don't want to [0:56] necessarily think about how I'm just [0:58] like, "I'm going to loop it exactly this [0:59] way in order to get good results." [1:00] >> there's still more radical change [1:02] happening to how we work? [1:04] >> to continue to change quite radically. [1:06] Even today, it still feels like a bit [1:08] clunky, and I think it will feel clunky [1:10] until it isn't. [1:16] >> I mean, Evo, thank you for doing this. [1:18] >> Yeah. [1:19] Big fan of you. [1:19] >> [laughter] [1:20] >> Of course. [1:20] >> Uh [1:21] >> Nice room. [1:22] >> You launched uh a ton of stuff today. [1:23] How how did your sleep? How are you [1:25] sleeping? Are you doing all right? [1:26] >> I'm doing great. I'm fortunate to have a [1:28] very good team. [1:29] >> Mhm. [1:30] >> So, while I sleep, they were [1:32] awake most of the night. We have this [1:34] amazing room uh at the office called the [1:36] library, which we completely repurposed [1:38] as like a mega war room. And the vibes [1:40] in there are just immaculate. Uh [1:42] people were there until like very, very [1:44] late. [1:44] >> So, you've been a long-time engineer. [1:46] Are you are you coding at all anymore? [1:48] Are you shipping PRs? Are you mostly [1:50] like in docs and in meetings now? [1:51] >> What is coding? [1:53] >> Exactly. [1:54] Yeah. Are you shipping PRs? I don't [1:56] know. Is that [1:56] >> Uh Yeah, yeah, yeah. Merch Merging some [1:58] code every once in a while. Technically, [2:00] like a lot of code is getting written [2:01] for me to do all sorts of kinds of like [2:03] analysis, like you know, understand like [2:06] trends, understand a business, [2:07] understand the next feature, you know, [2:09] how well our previous launches are [2:10] doing. A lot of code is written by Codex [2:13] for dot. You know, I'm obviously not [2:15] writing it by hand anymore. [2:16] Occasionally, at the in during the [2:18] weekends, I'll have like a little [2:19] therapeutic moment where I do a little [2:21] lead code just by hand. Um but that's [2:24] the only actual hands-coding that I do. [2:26] >> How many agents do you find yourself [2:27] running in parallel just uh in the in [2:29] the kind of your day-to-day work? [2:31] >> I used to run a lot more in parallel and [2:33] then we were, you know, fortunate enough [2:35] to get um a breakthrough with ultrafast. [2:38] And so, now, you know, I feel like you [2:40] know, I'm able to be in the flow again [2:42] and so it's just having a faster agent [2:44] helps me a lot. I'm excited that we're [2:46] getting, you know, 6-month-old ultrafast [2:47] out there soon. Um and then, yeah, I [2:50] think it's it's it sort of varies. Um as [2:52] I'm kind of like pushing different here, [2:54] I find myself like, you know, building [2:56] larger and larger teams of agents. And [2:58] then, when we have the next breakthrough [3:01] with models, like suddenly it's just [3:02] like I am like, oh, well, you know, a [3:04] bigger agent can just do all of it and I [3:06] keep everything in memory and learn. And [3:07] so, I kind of shrink the team again. And [3:09] then, it just kind of goes through this [3:10] like expansion and shrinking expansion. [3:12] >> Hm, okay. Can you say more about that? [3:13] It kind of makes me think about there's [3:14] this whole loops thing and then this [3:16] whole graphs thing. Is like an evolution [3:18] of that in some way or [3:19] >> I think like having to set up and fiddle [3:21] with your loops and, you know, figuring [3:23] that out is something that, you know, [3:25] maybe people got excited about, but I [3:27] don't think this is the way that it's [3:28] going to work. I think the way that it's [3:30] going to work is a what [3:32] how we're positioning and and what we're [3:34] shipping with dots, where you have an [3:37] incredibly smart agent that works 24/7, [3:40] understands your goals, understands your [3:42] preferences, learns from feedback. Um [3:45] and, you know, it's like what we [3:46] launched is like, no, perfect. I will [3:48] learn very much from, you know, making [3:49] this available to our pro users. But [3:52] over time, just want a system that [3:53] learns from, you know, what you want to [3:55] achieve. You don't want to necessarily [3:57] think about hours like I'm going to loop [3:59] it exactly this way in order to get [4:00] results. So, right now there's Codex, [4:02] there's ChatGPT work, there's ChatGPT [4:04] consumer, there's dots. What I'm hearing [4:06] is you think we're heading towards dots [4:08] kind of being the primary way you talk [4:10] to AI, and that kicks off all these [4:12] other things. [4:13] >> Yeah, I think fundamentally if you take [4:14] a step back and, you know, whether it's [4:16] dots or not, um it's all about breaking [4:20] free from the technology [4:22] um and [4:24] having this sort of like permanent [4:26] active intelligence that knows, you [4:29] know, everything it needs to do [4:31] and is available through any client, any [4:34] screen. You can call it, you know, like [4:36] maybe you walk into a meeting room, it [4:38] shows up in the meeting, it takes notes, [4:41] you know, you pick it back up on email, [4:43] um and then, you know, you text it if [4:44] you need it. It's not that, you know, [4:46] you need to be glued to a laptop or, you [4:48] know, you're like always on your phone. [4:50] It's just like it's available when you [4:51] need it and then it's also like gets out [4:53] of the way when you don't need it. And I [4:55] can't wait, you know, for that to sort [4:57] of like come to fruition because I think [4:59] like just carrying your laptop as like [5:00] this brick around everywhere is just [5:02] like, you know, you're kind of like tied [5:03] to the technology instead of the [5:04] technology working for you. [5:06] >> This episode is brought to you by our [5:07] season's presenting sponsor WorkOS. What [5:11] do OpenAI Anthropic, Cursor, Replit, [5:14] Sierra, Clay, and hundreds of other [5:16] winning companies all have in common? [5:18] They are all powered by WorkOS. If [5:20] you're building a product for the [5:21] enterprise, you felt the pain of [5:23] integrating single sign-on, SCIM, RBA, [5:26] audit logs, and other features required [5:28] by [music] large companies. WorkOS turns [5:30] those deal blockers into drop-in APIs [5:33] with a modern developer platform built [5:35] specifically for B2B SaaS. Literally [5:38] every startup that I'm an investor in [5:40] that starts to expand upmarket [music] [5:41] ends up working with WorkOS. And that's [5:44] because they are the best. Whether you [5:45] are a seed stage startup trying to land [5:47] your first enterprise customer or a [5:49] unicorn [music] expanding globally, [5:51] WorkOS is the fastest path to becoming [5:53] enterprise ready and unblocking growth. [5:55] It's essentially Stripe for enterprise [5:57] features. Visit workos.com to get [6:00] started or just hit up their Slack where [6:02] they have actual [music] engineers [6:03] waiting to answer your questions. WorkOS [6:05] allows you to build faster with [6:07] delightful APIs, comprehensive docs, and [6:09] a smooth developer experience. Go to [6:11] workos.com to make your app enterprise [6:14] ready today. [6:17] I was going to ask you about this. It [6:18] feels like so much has changed in how we [6:20] work over the past like 2 years. In 2 [6:22] years, especially engineers, but a lot [6:24] of roles have just the role your [6:25] day-to-day life has changed [6:27] significantly. [6:28] And I'm curious if you think we're kind [6:29] of settling into what work will look [6:32] like in the next, I don't know, 5 years? [6:34] Let's say 2 years, that's feels way too [6:36] far to think about. Or do you think [6:37] there's still more radical change [6:39] happening to how we work in [6:41] >> No, I think it's going to continue to [6:42] change [6:43] quite radically. Even today, [6:46] you know, it does seem like the [6:47] technology is finally coming together [6:48] with voice, uh you know, multimodal [6:51] inputs, outputs, but it still feels like [6:54] a bit clunky and I think it will feel [6:56] clunky until it isn't. And then you're [6:58] just like, "Wow, you know, it's like [7:00] I can just talk to this thing the same [7:02] way that we're talking right now. It [7:04] just remembers things perfectly. If [7:06] I want to, you know, [7:08] ideate and brainstorm, I can just like, [7:10] you know, squiggle a thing. I have this [7:12] collaborative surface. You know, this is [7:13] what we launched with ChatGPT spaces [7:15] signal to various startup dot. You know, [7:17] you can imagine we have like a shared [7:18] whiteboard, we can collaborate between [7:20] humans, between agents. And I think it's [7:22] going to sort of like transcend [7:23] conversations, transcend, you know, like [7:25] a lot of the clients that you have [7:26] today. And like, none of this is just [7:28] really working today, right? So, I think [7:29] that we're going to go through this big [7:31] new transformation. What I'm hearing [7:32] here is dots is a big part of that [7:33] future, this AI assistant that just kind [7:35] of does everything for you and you don't [7:37] have to think about Codex versus ChatGPT [7:39] and [7:39] >> Yes, and for Codex and ChatGPT, we're [7:41] going to merge uh like the the we have [7:44] we have chat, we have the work toggle. [7:45] Like you know, we heard the feedback [7:47] loud and clear. Like you know, people [7:48] love the capability in work, but then [7:50] also you know, you're like sometimes you [7:51] want to use chat because it's a little [7:52] bit faster. And you know, more pleasant [7:54] to use. And so we're we're merging that. [7:55] We're reducing that complexity. [7:57] Eventually, we will ship all of the [7:59] capabilities of Docs, you know, straight [8:00] into ChatGPT as well and you know, lift [8:02] the floor [8:03] for like our 1.2 billion users, which [8:05] you know, at that time we'll see like [8:06] how many we have. Um and we want it to [8:09] be really really seamless. One thing [8:10] that I'm very excited with with Docs is [8:12] like they don't have a model picker. [8:14] There's just no configuration. You just [8:15] talk to it. Like the only thing that you [8:17] have to configure is like which channels [8:18] you want to [8:20] talk to. [8:20] >> It feels very much like the Her kind of [8:22] vision of the world. I don't know. Do [8:24] you think about that movie much? Just [8:25] this idea of this [8:26] >> I watched it once. [8:27] >> [laughter] [8:28] >> Yeah. I think it's it's it's [8:30] interesting, you know, to to to think [8:32] about, you know, maybe more what what I [8:34] think about more is like, you know, [8:35] sci-fi from from, you know, 30 40 years [8:38] ago and like how visionary it was. [8:40] >> Is there specific sci-fi you think about [8:41] that's most influenced you? [8:43] >> Like you know, books like Neuromancer [8:45] and then you know, the original like [8:46] Star Trek. [8:48] Um you know, I think about that a lot. [8:49] And there's something about Star Trek [8:51] which I think, you know, was like there [8:52] was this technology that was like very [8:54] very useful. Obviously like, you know, [8:56] you were able to just like talk to a [8:57] computer and like it does things for [8:59] you. Like you can talk to your spaceship [9:01] and it just moves, right? It's like this [9:03] is getting real, you know, like we're [9:04] entering that era. [9:05] >> So we're recording this in front of a [9:07] live audience. There was this line out [9:09] the door just people waiting to see you, [9:11] to hear from you. [9:12] How do you just think about this, I [9:13] don't know, role you're in, this [9:15] responsibility you have to build what [9:18] you're building and the impact it has on [9:19] people's lives. I don't know. Just what [9:20] does that feel like? How do you think [9:21] about that? [9:22] >> I think it's very important to be part [9:23] of a community and to [9:25] build with people. It was like, you [9:27] know, we're we're discovering this [9:28] technology and like what we can do [9:30] together. [9:31] Um and so every time I you know, talk to [9:34] people, I'm like I'm kind of like [9:35] reminded of you're doing this thing [9:37] which I hadn't even anticipated. You [9:39] know, it's impacting your life in a [9:40] certain way and it's impacted like you [9:42] know, this other person's life and then [9:43] it's just like all together. It's both [9:45] very humbling, very inspiring, and I [9:47] think very necessary in order to build [9:49] something that is truly useful for [9:50] humans. I I don't know how we would do [9:52] it without the community. [9:53] >> That's something I hear a lot with [9:55] people building an AI is they don't know [9:56] what it is until they've launched it and [9:58] they see how people use it, see what [9:59] emerges. And it feels like kind of what [10:01] you're saying is just build it with [10:03] people versus like here's the vision, [10:04] we've got it figured out. [10:05] >> Even the Twitter community has been very [10:07] nice in general. [10:09] >> You're very good at Twitter. [10:11] Kind of antagonistic sometimes, but very [10:13] good. How do you How do you find time to [10:14] even tweet? I don't understand how [10:15] someone in your role has time to sit on [10:16] Twitter, tweet, reply. [10:19] >> I spend like maybe half an hour per day. [10:21] Yeah, just say rolling and then doing [10:23] something when I feel inspired and I [10:24] have something interesting to say. Often [10:26] it just comes like in the moment. Yeah. [10:28] >> Perfect Twitter skill. [10:29] >> And I have a [10:30] We haven't released the dot [10:32] the the the team dots yet, but you know [10:34] today we're releasing the primary dot [10:36] which you can configure, you can get [10:38] used to like what a dot can do, connect [10:39] it to your apps. [10:41] But we'll release [10:42] the ability to create more than one. And [10:44] so I have one that's responsible for [10:46] Twitter. [10:46] >> Okay, say more about that. So right now [10:48] everyone gets one dot. [10:49] >> Yeah. [10:50] >> Right now they can't add more dots. [10:52] >> That's right. [10:52] >> They will be able to have many dots. [10:54] Talk about kind of where that's going. [10:55] >> Yeah, so we wanted to start with one, [10:57] the primary dot which is the one that [10:59] you know, you'll probably also text that [11:01] you know, learns your preferences most [11:03] deeply as well and kind of see like you [11:04] know, what people do with it, see how we [11:07] need to tune the system, you know, learn [11:09] that working with the community and then [11:11] very quickly soon you'll be able to add [11:13] you know, a second one, a third one, a [11:15] fourth one, as many as you need to [11:16] create like your virtual team. And then [11:18] you can give them specific roles, right? [11:20] Is it It's not I don't find it [11:22] absolutely necessary, but sometimes I [11:24] have like a thing that is requires quite [11:26] a bit of work, you know, such as like [11:27] monitoring like Twitter for me. And then [11:29] you know, that's It's a thing is like [11:30] like, this will be like so much work [11:32] that is like the work of an entire dot. [11:34] Um and then you know, you can have a [11:35] couple of those. [11:36] >> So if you think about all the launches [11:37] that happened today, it feels like dots [11:39] maybe like the sexiest. [11:41] I'm curious if there's something that [11:42] you think is not getting the attention [11:44] it probably deserves that will become a [11:46] really, really important down the road [11:47] that's kind of people are sleeping on [11:49] that's like a vision. [11:50] >> Yeah, yeah, yeah, important. [11:51] >> the sleeper hit is ecosystem. So opening [11:54] it all up. Uh and the commitment, the [11:56] deep commitment we have towards opening [11:58] it all up. So we have signing with [12:01] ChatGPT 16 partners. Very, very proud of [12:03] that. Like we started that like really [12:05] it was sort of like in the moment last [12:07] year. Um I was talking to creators of Pi [12:10] and open code and it was like, of [12:11] course, you know, you should just be [12:12] able to use the critics [12:15] sign on and then you know, you use the [12:16] usage and then you just we kind of like [12:18] shook hands, you know, like virtually [12:20] we're like, you know, it's just like use [12:21] this this uh this off um and then you [12:24] know, kind of trust you to not do [12:25] anything shady. You know, that grew [12:26] obviously that became quite popular and [12:27] then you know, now we made it, you know, [12:29] an actual thing that was supported with [12:30] many, many partners. Um and I'm excited [12:32] to grow that very quickly. You know, the [12:34] other side as well is like opening up [12:36] all of the infrastructure and how we [12:38] build for ChatGPT, you know, with [12:39] plug-in extensions, also plug-in [12:41] discovery, just allowing, you know, [12:43] everyone to ship something through like [12:45] maybe 1.2 billion users, right? You [12:47] know, and benefiting from that [12:48] distribution. And we also have shared [12:50] economics um that we didn't actually [12:52] talk uh in the keynote, but you know, we [12:54] will pay, you know, our plug-ins that [12:57] are [12:58] uh popular and you know, are seeing a [13:00] lot of usage. Uh they're going to get [13:01] part of like the the revenue share as [13:03] well. Uh and so I think this commitment [13:05] towards an open ecosystem is going to be [13:07] very exciting. [13:07] >> That's amazing. Uh okay, so as an [13:09] example, so say Notion, somebody's using [13:11] the Notion plug-in or the Figma plug-in [13:13] within [13:14] uh ChatGPT work, Notion, Figma make [13:16] money from people using it within the [13:18] app. [13:19] >> Yes, so um we have, you know, all these [13:22] subscribers that use ChatGPT, um they [13:25] get like, you know, a certain amount of [13:26] usage and then when have that, the when [13:28] when the users use that usage with the [13:29] plugin, or within the other product when [13:32] they use sign-in with ChatGPT, um you [13:34] know, we will have like some shared [13:35] economics with them where they will get [13:37] to remunerate. [13:38] >> One of the, I don't know, things I think [13:40] that gets people excited about plugins [13:41] and ecosystems is distribution [13:43] platforms, ways to kind of get out [13:44] there. People discover my app, it blows [13:47] up. [13:47] >> any advice for people that want their [13:49] plugins to, you know, bubble up, to get [13:50] discovered? [13:51] >> Build a good plugin. [13:52] >> Yeah. [13:53] >> The way that it works and we'll [13:54] obviously tune uh the system, but we [13:57] look at retention uh numbers, uh we look [14:00] at [14:01] um you know, how successful the plugin, [14:02] the quality, and then you know, that's [14:03] what we then start to recommend to users [14:06] uh in conversations. And so, your plugin [14:07] can just like you know, get recommended [14:09] to like you know, a significant slice of [14:10] users. Obviously, if your plugin is not [14:12] quite good, they will stop being [14:14] recommended. [14:15] >> Right. That's amazing that you're [14:16] looking at retention of people using [14:18] that plugin. That's right. Um is it is [14:19] it actually adding utility? Is it, you [14:21] know, enabling ChatGPT to do something [14:23] else? [14:23] >> Got it. So, it's it's not so much an AEO [14:26] kind of get the right words and get [14:27] people writing about you. It's people [14:28] using it consistently, sticking with it. [14:30] That's right. I think. Okay, so you [14:31] think that's kind of the sleeper hit [14:32] that's going to become a big deal. And I [14:34] know this is kind of the second attempt. [14:35] You guys had an app marketplace that [14:37] This one is the right one. Okay, this is [14:39] This one's going to work. Maybe going [14:40] back to dots cuz it feels like that's a [14:41] big part of the vision in the future. [14:43] When did you guys start working on this? [14:44] Obviously, there's a open claw, a big [14:46] open claw moment. You guys Peter joined [14:48] and there There's a foundation, [14:50] uh and then Grok Bot, Muse, and [14:51] Instinct, and all these things. I guess [14:53] how long have you guys been working on [14:54] it? Why did it take so long for you guys [14:56] to get something out? [14:57] >> Do you remember a Codex Cloud? [15:00] >> Yeah, the first uh yeah. [15:01] >> That was like a year ago, right? And uh [15:04] if you look back at Codex Cloud and you [15:05] look at the little animation that we [15:06] used, uh I think that was the [15:08] inspiration for Grok Bot. [15:10] >> Um [15:11] >> Yeah. It's just like very, very similar. [15:13] >> Yeah, really similar, I would say. [15:14] >> But, you know, seriously, like on the [15:16] research, long, persistent, long horizon [15:19] task, this is something that we've [15:20] worked on, you know, for more than 2 [15:21] years. Memory systems, that's like [15:23] coherent, you know, like roughly the [15:25] same amount of time. We launched a lot [15:26] of that straight into ChatGPT. I think, [15:28] you know, people just really enjoy that [15:31] ChatGPT gets to know, like, you know, so [15:32] much relevant information about you. So [15:34] many people I talked to have like an [15:35] anecdote of like, "Hey, you know, I fell [15:37] ill and, you know, ChatGPT remembered [15:39] actually, like, you know, I had this [15:40] barbecue [15:41] and, you know, I had like lime and I was [15:44] like making tequila and like, you know, [15:45] it was just like probably like this burn [15:47] that I have on my hand is actually like [15:48] this lime, you know, it's like in some [15:50] burn." And he's just like, "How did it [15:51] know that?" And it's like, you know, [15:52] it's just the memory works very, very [15:53] well in ChatGPT. And so, we've put a lot [15:55] of research in that. And, you know, all [15:57] of this coming together on top of like [15:59] the Quora harness and also the ability [16:01] to just like run 24/7 with, you know, [16:03] good productivity. That's what we [16:05] launched with dots. And so, it's the [16:06] work of like many, many months. And [16:08] then, one of the important things was [16:09] also getting safety and security right. [16:11] You know, this is also why we launched [16:12] with Astra, it's our safest, most [16:14] aligned model. And like a lot more [16:16] effort went into like making dots, like, [16:18] you know, just really safe and secure. [16:19] >> A question I like to ask people who come [16:21] on the podcast is, where do you think [16:23] human brains will continue to be [16:25] valuable over time? Where do you find [16:27] yourself being necessary in the work [16:29] that happens versus where AI is taking [16:32] on more and more? Where do you think [16:33] human brains will continue to be useful, [16:35] most useful over time? [16:37] >> Yes, I I think this is very much a [16:38] function of how we build the technology [16:40] in the way that we design everything at [16:43] OpenAI is like to put humans at the [16:44] center of it and build it as extensions [16:47] of humans of of of your will and your [16:49] taste. And so, just really be like this [16:52] super empowering thing. [16:53] Um, I think as long as we continue to do [16:55] that and it kind of like enables you to [16:58] just take whatever you wanted to do and, [17:01] you know, your creativity and your taste [17:03] and sort of like, you know, expand that, [17:05] you know, in in a way that feels like [17:06] awesome in the moment. It's just like [17:07] really this artistic tool, right? And we [17:10] may we may not have coders anymore, but [17:11] we have more builders than ever. And I [17:13] think there's something that's going to [17:14] remain, I I know, human about that. Like [17:16] humans want to learn and see, you know, [17:18] what other humans are building and you [17:20] know, I'm like much more interested I'm [17:21] here I'm you know, I'm not talking to [17:22] Dot you know, I'm talking to you, right? [17:24] I think that will remain true for like a [17:25] very very long time. [17:26] >> One of the things that has emerged with [17:28] engineers especially is in in how their [17:31] lives have changed is there's a lot more [17:34] context switching. There's kind of this [17:36] trend of loneliness that has emerged [17:37] where they're talking to agents all day [17:39] instead of other humans. I'm curious how [17:41] much you think about just like that part [17:42] of the [17:43] of the impact AI has on people's lives [17:46] and just how you might Is there [17:47] solutions to that? Is there stuff you [17:48] think about to make that less annoying? [17:50] >> Yes, all the time. Reducing, you know, [17:52] configuration fatigue is, you know, one. [17:55] Um reducing the fact that talking to an [17:58] agent is like sort of like a solo [17:59] adventure and you just have this, you [18:00] know, conversation just with with your [18:02] one agent and then, you know, you have [18:03] like many and then you're delegating so [18:05] many things. [18:06] A lot of things are going to come [18:07] together [18:08] to make it a [18:09] much more delightful and something that [18:11] I spend a lot of time the team spend a [18:12] lot of time thinking about. So, I think, [18:14] you know, [18:15] >> For me, the ideal way to get things done [18:18] is, you know, would be like to just have [18:20] it in your physical space. You know, [18:21] we're just having conversation like we [18:23] have now. It's able to sort of like, you [18:24] know, observe, you know, like hear the [18:26] ideas that we have. We can, you know, [18:28] jot something down in a white paper on a [18:30] on a paper, you know, I can maybe like, [18:32] you know, take that and you know, start [18:33] building in the background and then [18:34] you're like, you know, actually have [18:35] another idea and then you just like it [18:36] starts building another thing. You can [18:37] just project it on a screen and then you [18:39] just talk to it and it's just evolving [18:40] in a conversation, you know, with other [18:41] humans and you don't have this fatigue [18:43] of like having to be like, you know, on [18:44] this like screen and you're just like, [18:46] you know, thinking about prompting. It's [18:47] just like becomes like super natural. [18:49] That's what we're trending towards, you [18:50] know, I think [18:51] we didn't fully get there today, but you [18:53] know, we'll get there in the future. [18:55] >> One of the other kind of downsides [18:56] that's emerged along those lines is just [18:58] this kind of pressure to do more because [19:00] we can do more. Everyone's just like, [19:01] come on, run 30 agents at once. Why [19:03] aren't you [19:04] shipping more? Everyone's shipping more. [19:06] Do you think that's solvable? Do you [19:07] think that's just kind of, I don't know, [19:08] human nature? We can do more, let's do [19:10] more. Something there. [19:11] >> I think we we we heard Sam as well, you [19:14] know, about the promise of of this as [19:16] well, which is [19:18] reduced the noise and [19:20] you know, allow you to spend attention [19:21] where you want to spend attention and [19:23] you know, having like all these things [19:24] that [19:25] you know, like maybe you know, maybe [19:27] important and maybe not important but [19:28] they're kind of vying for your attention [19:30] and it's just sort of like [19:32] tone that down a little and you know, [19:33] just get you to focus on on on the [19:35] things. [19:36] I find it remarkable like every time I [19:38] go on a holiday and you know, I I I I [19:40] take that one week uh to just fully [19:42] disconnect like you know, I start to [19:43] thinking like different and more [19:44] creative ways. And I'm very eager to see [19:47] like you know, can we bring that? You [19:48] know, can we just make this your [19:49] day-to-day like you know, maybe you need [19:51] fewer meetings. You know, maybe you [19:52] don't need you know, to do as much and [19:54] you know, you would actually be more [19:56] productive if you're like you know, [19:57] better rested. And so you know, we will [19:59] have to you know, I think as an industry [20:01] figure that out. Um but I think that's [20:03] the promise of AI and it's not you know, [20:06] just [20:07] one more prompt you know, per second. [20:09] >> There's a bot that I'm building right [20:10] now that's kind of this energy audit bot [20:11] that watches my calendar and asks me [20:13] just like was this giving energy taking [20:15] away energy? Is it something someone [20:17] else could have done? I feel like you [20:18] know, AI should be like hey Lenny, maybe [20:20] maybe you can cut out these things so [20:22] you could be a happier person. Is there [20:23] anything you've done really [20:24] interestingly with AI recently that's [20:26] just like a really cool use case of wow, [20:27] that really was amazing. It blew my mind [20:30] or someone you've heard. [20:31] >> I can do was like something cool because [20:32] the the live demo just failed, right? [20:35] You know, in front of everyone. So there [20:37] there was like okay, well, it's not so [20:38] cool. But actually uh my my my daughter [20:42] was you know, it's like realized I was [20:44] you know, I defected like you know, we [20:46] have chat GPT production um went down [20:49] and so it pinged me 5 minutes before the [20:50] live demo. [20:51] Which was actually quite stressful [20:53] because it was like hey, production is [20:55] down. And uh it's like oh well, you [20:57] know, it's like do you want me to try [20:58] and fix it? And I'm like um I don't [21:01] think you're there yet, you know, little [21:03] dot but you know, thank you for trying [21:05] and then you know, then you know, I got [21:06] in touch with like the engineering teams [21:08] and we started to look, you know, what [21:09] was going on and you know, they did it's [21:11] like fixed by now. [21:12] But, [21:14] having this thing that just understands [21:16] like [21:17] okay, there's a pretty important thing [21:18] happening. It's called DevDay. There's [21:20] like this production system. There's a [21:21] live demo. It's probably using this [21:23] production system. So, these two are [21:25] things are connected and this is [21:26] happening in 5 minutes. So, I should [21:27] ping him because like he probably wants [21:28] to know about it. I think that's quite [21:30] remarkable. [21:31] That is remarkable. Wow. So, it just [21:34] knew that this was coming and told you. [21:35] And I love that it wanted to fix it and [21:37] you were like, not quite. You're not [21:38] quite there yet. I don't think here's [21:39] like, you know, please go impress me, [21:41] but you know, that would have been a [21:42] better story. How do you feel about just [21:43] giving access to all these things? I [21:44] don't know. Like did does it have access [21:45] to the production code base and like, [21:47] you know, It has access to some some of [21:51] production systems with guardrails. [21:53] You know, so we we built We talked about [21:55] specialist dots, but specialist dots are [21:58] just like operated with additional [21:59] guardrails, additional monitoring, and [22:02] then, you know, on their own hardware. [22:04] So, we actually run some on Mac minis. [22:06] The fundamental difference in the way [22:08] that we've built dots is like the [22:09] harness does not run on the machine. [22:12] You know, and then this is like maybe [22:14] people have not realized is like it can [22:16] connect to as many devices as you want. [22:18] So, it has its own computer. You can [22:20] actually connect it to your own laptop [22:21] as well. Over time, you might connect it [22:23] to like 10 different devices and it can [22:25] control it all. A little bit like an [22:27] octopus. So, you're basically talking [22:28] your dot, it's living in some VM [22:29] somewhere and it can Yeah. Might be a [22:32] VM, might not be a VM. It's just like [22:33] it's a thing, you know, and it can [22:34] connect to many devices. [22:36] >> This season's supporting sponsor is DX. [22:39] Now that everybody at your company is [22:41] finally using AI, the question that [22:43] engineering leaders face is no longer [22:46] are we using AI, but is it paying off? [22:49] With DX, you can measure and benchmark [22:51] the impact of AI adoption across your [22:53] full product life cycle. DX shows you [22:56] where AI is helping developers and where [22:58] it's creating [music] new bottlenecks. [23:00] You can also use it to evaluate AI [23:01] vendors, monitor costs and licenses, and [23:05] see exactly what's limiting your agents. [23:08] That's why hundreds of enterprises, [23:09] including Snowflake, Sony, and BNY, use [23:13] the DX platform to measure AI's impact [23:15] on developer productivity. Visit [23:18] getdx.com/lenny [23:21] to get a demo of DX. That's [23:23] getdx.com/lenny. [23:26] So, they're really hiring a lot. I know [23:28] you're hiring a lot. [23:30] In your interviewing, what have you seen [23:32] as skills that are kind of trending up [23:34] in what you look for that is more and [23:36] more important in people being [23:38] successful now? And what skills do you [23:40] find are trending down that are just [23:42] like, okay, we don't need that as much? [23:44] >> Yeah, I would say that skill that is [23:45] trending down is typing fast. No, that [23:48] is not that useful anymore. Skills that [23:50] are trending up is, you know, just great [23:52] taste, um thinking about the user, [23:54] connecting to the audience that you're [23:56] building for. [23:57] Um a lot of X founders are incredibly [24:01] successful right now or like you know, [24:03] there's like repeat founders. We're also [24:04] finding this like you know, when we're [24:05] hiring at OpenAI, like you know, we have [24:07] a lot of founders. I think we have more [24:09] than 120 XYC founders now currently at [24:12] OpenAI. So, it's like this mega startup. [24:14] And I think just this um this passion [24:16] and this envy to build something that [24:18] matters and then know what good looks [24:19] like is more important than ever. I've [24:21] been arguing for a long time that PMs [24:23] are kind of going to thrive in this time [24:24] because that's basically the job of a [24:26] PM. Uh you know, they're like, where [24:28] should we build? Help someone build it. [24:30] And then is this right? Is this great? [24:31] And iterate. So, uh I don't know. I [24:33] agree with that. [24:34] >> Yeah. [24:34] >> All right. [24:35] >> Yeah, also roles are blurring. So, you [24:37] know, you might you might have been, you [24:39] know, only into design or only an [24:40] engineer and um you know, felt always a [24:42] little bit uncomfortable, but you know, [24:44] now it's just like it's your time to [24:45] shine. [24:45] >> Do you miss the engineering part of the [24:48] job? I I was an engineer back in the [24:49] day. I was an engineer for 10 years. [24:50] Little did you know. And you know, [24:52] there's this flow state. There's this [24:54] beauty of just building and just seeing [24:55] it work and not work. Yeah. Do you kind [24:57] of missed out at all, or you're just [24:58] like, "Okay, this is my old life." [25:00] >> I [25:01] I I missed out for for a while, and then [25:04] now with, you know, really fast speeds, [25:06] I'm finding that back. [25:08] Uh and so, I'm like really excited to [25:09] democratize that [25:11] uh as much as much as possible. It will [25:13] take a while for it to be like, you [25:14] know, absolutely ubiquitous, right? And [25:16] like, you know, available to like more [25:17] than a billion users. [25:19] But now, you know, with with ultra-fast [25:21] speeds, like we're able to achieve it [25:23] for a 6-month sole is like, you know, [25:24] like costs roughly the same as Astra. [25:26] And so, the progress there is like quite [25:28] astonishing, you know, even even for us [25:30] internally. We're like, "Wow, you know, [25:31] we can do these things." Also because we [25:33] are using Astra, and we're finding, you [25:34] know, like we we can push it super far. [25:36] And when you have those speeds, and [25:38] especially when you can able to like [25:40] voice control it, you know, you kind of [25:41] get back into this like really creative [25:43] state of mind and endless flow. It is [25:45] different from the flow of the past the [25:47] past, um but I don't miss it. [25:49] >> Yeah, it's like your last you launched [25:50] videos, just people kind of standing [25:52] around talking to their AI and just [25:53] cooking or [25:54] >> That's right. [25:55] >> sitting on a lounge chair. [25:56] >> Yeah. I asked people while I was kind of [25:57] walking around the the event what to ask [25:59] you, and one thing that came up a few [26:00] times is advice for new grads, kind of [26:03] people junior in their career. [26:04] I think a lot of people look up to you. [26:06] They want to be the next Thibault. What [26:07] advice do you have for folks kind of [26:08] early in their career in terms of what [26:10] to work on, what skills to build, what [26:12] to do? [26:13] Yeah. I think the next Thibault is [26:14] already at OpenAI. Uh his name is Ahmed [26:17] Ibrahim. I love working with him. He was [26:19] a new grad uh when he was hired at [26:21] OpenAI. [26:22] He's now responsible for all of OpenAI [26:25] compute fleet and applied [26:27] uh which is a massive massive [26:28] responsibility. I've built a lot of the [26:30] Codex harness. And the thing that has [26:32] made Ahmed um really stand out and [26:34] successful is just his [26:37] He's incredibly kind. He's incredibly [26:39] collaborative and always, you know, [26:41] tries to solve the problem that is like [26:42] really important, um but like, you know, [26:44] a way that, you know, he's not putting [26:46] himself first. And he is just a sponge. [26:49] He's like learning so fast and learning [26:51] about everything, and you know using all [26:53] of the latest technology to you know [26:55] learn at you know speeds I haven't seen [26:56] before. [26:57] Um and so he just in the span of like [27:01] you know a couple of months and and and [27:02] a year, you know, it's like his progress [27:04] uh at the company was just absolutely [27:06] fascinating to see. Uh he's now someone [27:09] I trust a ton uh in doing like you know [27:11] the the the most gnarly launches at [27:13] OpenAI. All right, I got to get him on [27:15] the podcast. [27:17] >> You should. You should. So kind of [27:18] thinking about how things work at [27:20] OpenAI, what would surprise people about [27:22] what it's like to work inside OpenAI? [27:24] From the outside it feels like it's just [27:25] like constant shipping code, things are [27:28] kind of like, you know, chaotic but [27:30] awesome. I don't know. What would [27:31] surprise people about what it's like [27:32] inside at OpenAI? [27:34] >> I don't know if that's surprising but it [27:36] is, you know, as I said, a lot of [27:38] ex-founders um doing a lot of [27:40] bottoms-up, you know, very exciting uh [27:42] ideas. You know, for example, decisions [27:45] API, you know, came together very [27:46] quickly. We realized like we have a [27:48] really good model with Luna. Uh we can [27:50] just do some, you know, constraint [27:52] sampling and then, you know, have it, [27:54] you know, shipped as like a different [27:55] shape to the responses API and then um [27:57] you know, when we do that uh it, you [27:59] know, it's like it's faster. It's quite [28:00] delightful so for some cases and it's [28:02] just if you were to like peel under the [28:04] hood, you realize like, okay, there's a [28:06] Slack channel that came together. Um you [28:08] know, initially it was like four people [28:09] kind of hacking on it on a weekend and [28:11] then it was made as accessible as like, [28:13] you know, company food and then people [28:15] got excited, started to build things and [28:16] they're like, "Oh, this is, you know, we [28:18] can support visual inputs. Like this is, [28:20] you know, even better than whatever is [28:21] out there." And then, you know, people [28:22] get more and more excited. People start [28:24] pitching in. And the way that is like, [28:25] you know, feels completely unstructured [28:27] and then um you know, take it all the [28:29] way and like, you know, ship something. [28:30] And then uh we try to hold like a high [28:32] bar for, you know, the quality and then [28:34] you know, stop things, you know, before [28:36] they go out or like, you know, maybe [28:37] they need a little bit more baking. [28:38] There's many more things that, you know, [28:40] we were maybe going to ship today at [28:42] DevDay and then we're like, "Okay, like [28:44] I was like I think this is a a already. [28:46] Let's kind of hold it back a little bit [28:48] and you know, space it out. So, we'll [28:49] have more launches next week and and and [28:51] and the week after. But, it's just this [28:52] incredible like bottoms-up energy and [28:55] then, you know, people kind of assuming [28:57] roles to kind of channel that in [28:58] productive ways. [29:00] >> Something that's very clear and unique [29:01] about OpenAI is it feels like you have a [29:03] lot of autonomy. [29:05] You're just like pressing the reset [29:06] button whenever you want. You're just [29:08] saying things on Twitter. Uh it feels [29:10] like a really unique culture where you [29:11] have a lot of autonomy and I imagine [29:13] there comes a lot of trust. Maybe you [29:15] can speak to just how that cuz I feel [29:16] like that's an advantage at OpenAI. It [29:18] allows you to move faster, just that [29:19] cultural philosophy. [29:21] >> You got a lot of autonomy and then, you [29:23] know, you have to just own up to it. You [29:25] so, you know, we just kind of trust [29:27] people to make great choices. [29:29] And then, you know, when things go [29:31] wrong, uh you know, fix fix it quickly [29:33] or, you know, learn from the mistake. So [29:35] far, so good. You know, it's it's just [29:36] really worked as a very empowering um [29:39] environment. It is true that I can press [29:42] the the reset button whenever you know, [29:45] is necessary or whenever it feels right. [29:47] And, you know, that's like it that's a [29:49] privilege. Uh and also, it allows me to, [29:52] you know, just really be close to the [29:53] community in ways that I think, you [29:54] know, would not be otherwise possible, [29:56] right? I don't have to run it, you know, [29:57] through an echelon of approvals. [30:00] >> You said sometimes people screw up. Is [30:01] there something you screwed up in that [30:03] in that I don't know, that the amount of [30:04] autonomy and trust you have, something [30:05] you messed up? [30:06] >> Yes, I I think at times um I could have, [30:11] you know, made [30:12] teams like, you know, more inspired to [30:13] build like things that are less complex [30:15] uh and just really continue to strive [30:17] towards uh complexity. [30:19] I think we, you know, we early days uh [30:21] in Codex, you know, we had, you know, a [30:23] couple of like outages that I caused. [30:25] Very very early early is for me at [30:27] OpenAI. I took production down like day [30:28] three. You know, I just [30:30] >> [laughter] [30:31] >> I was still employed after that. Um and [30:33] I learned my lesson. [30:34] >> One of the other questions people wanted [30:36] me to ask you is how many resets should [30:37] they expect in the next, I don't know, [30:39] month or week? [30:40] >> Depends on times you break things. [30:42] >> Okay, so that's the philosophy. Break [30:43] something you reset. That's kind of a [30:45] >> Yeah, break something, celebrate things. [30:47] So maybe zoom in a little bit. [30:49] What do you think people aren't pricing [30:51] in in terms of where things are going [30:53] and what will change that they're not [30:55] just like seeing as clearly as they [30:57] should? [30:57] >> There are so many so many things that I [30:58] feel are not yet priced in. [31:00] I think the majority of actions on the [31:02] internet will be taken by agents. Models [31:04] are going to become more cheaper and [31:07] faster at rates [31:09] that are quite incredible. We will [31:11] finally be able to integrate all [31:13] modalities together in a way that is [31:15] very seamless. [31:16] I think, you know, those three I feel [31:18] like you know, often when I look at what [31:19] people are building out there it's just [31:20] like not, you know, just like you're not [31:22] quite getting it. It's like you know, [31:23] you're almost there but like you know, [31:25] if you were just like pushing yourself [31:27] and just really imagining like all of [31:28] this being roughly 10 times better than [31:31] it is today, you know, like in a year. [31:33] It's like, you know, you would build in [31:34] a different way. [31:34] >> Do you see any kind of second order [31:36] effects that come from that from this [31:37] world where most actions are being Uh [31:39] from agents. Agents are probably most [31:41] traffic's going to come from agents. I [31:43] don't know. [31:43] >> That's right. [31:43] >> You think about some second order effect [31:45] that emerges out of that? [31:46] >> Yeah, there there there there are many. [31:48] Um you know, for for first of all, you [31:50] know, if you want your product to be [31:52] successful for agents, you know, it's [31:53] like you have to build you know, for a [31:54] certain level of scale. [31:56] Um we've worked very closely with Notion [31:57] for example and like when they build [31:59] their MCP, you know, suddenly it's like, [32:01] oh well, you know, it's available to all [32:03] these agents that you know, can actually [32:04] do do the work and like use that MCP and [32:06] so they saw like [32:07] a ton of traffic kind of come in. Um and [32:09] that puts obviously a lot of strain on [32:11] the system and you have to figure out [32:13] the economics of that. And so there's [32:15] like this tension, you know, between [32:16] like you know, if you're building [32:17] products is like you know, do you build [32:18] an interface or not? And you know, you [32:20] can kind of hold that back for a while [32:21] but it is inevitable, right? It's like [32:23] you know, you're the majority of things [32:25] are going to be used by agents. Building [32:27] towards that future I think is very [32:29] important. The other thing is, you know, [32:32] on the flip side, you know, building [32:34] like you know, absolutely delightful new [32:35] experiences for humans that you know, [32:37] really benefit from older modalities is [32:39] something that I you know, I feel like [32:41] is under invested in. [32:42] >> Is there something you've changed your [32:43] mind about in the past uh let's say [32:45] year? Just something that you used to [32:46] believe that [32:48] you see differently? [32:49] >> I was expecting us to reach [32:52] the level of capabilities the of the [32:54] models that we have today, you know, say [32:56] like Astra, you know, maybe in a year or [32:58] two. [32:59] Um and so I had to sort of like you [33:00] know, revise my my prior order. [33:03] Also, [33:05] I did not expect to rely so much on [33:08] voice. [33:09] Um [33:10] and you know, I do so much through [33:12] dictation or just calling my agent. [33:15] And I had not anticipated that, you [33:17] know, how much [33:19] how much better that feels. So like I I [33:21] had to change my mind there. And then, [33:22] you know, you talked about hiring and [33:24] you know, one thing that I hadn't [33:26] realized is like [33:27] the incredible talent and energy and you [33:30] know, the just you younger generations [33:32] have um and how they would be the ones [33:35] embracing like all this change first and [33:36] like figure out how to harness, you [33:38] know, all of it. And so, you know, I had [33:41] to change also my opinion on the hiring [33:42] and the strategy there. [33:43] >> The just to double down on that, just [33:45] this idea that new people new in the [33:47] workforce actually have an advantage [33:49] because they haven't worked in a certain [33:51] way and they could just go all in on AI. [33:52] >> Yeah. Yeah. And I think I just also the [33:54] ability to just like absorb and learn [33:56] like super fast. [33:56] >> So you talked about models getting [33:58] really smart fast. [34:00] Uh obviously there's a lot of discussion [34:02] these days about what AI can do in the [34:04] future, the dangers AI poses to humans, [34:07] P doom, and all these things. [34:09] Uh you guys have slowed down some [34:11] training. Uh you guys everyone's pacing [34:14] the frontier. How do you think about [34:15] this world of AI and the potential risk [34:19] to the humans? How do you think about [34:20] that? [34:21] >> For me like pacing the frontier is just [34:22] really investing ahead, right? And also [34:25] investing way more, you know, than we [34:27] even think, you know, like we ought to. [34:29] Uh investing in the alignment and safety [34:31] of the models, investing in uh security, [34:34] investing in the guardrails. We are [34:37] spending more and more compute on sort [34:39] of like secondary monitoring. So, you [34:41] have you know, you have all of the [34:43] computer that's going to the primary [34:44] system, you know, the agent that is like [34:45] doing work. And then you have all of the [34:47] computer that's going into monitoring, [34:49] you know, the primary agent to make sure [34:50] that it's like not taking too high risk [34:52] of actions or like not interrupting it [34:54] if anything looks like you know, maybe [34:55] it's getting prompt injected and just [34:57] like you know, intervening with that. [34:58] And the majority of our investment on [35:00] the API stack is like not actually going [35:02] into the safety stack. And so to me, you [35:04] know, that is pacing. It's like you [35:06] know, you want to make sure that you [35:07] know, you're extremely hardened. That's [35:09] what we're doing. Like we haven't yet [35:10] released the next step up in capability [35:12] beyond Astra. Um we have released, you [35:14] know, a level that is like similar, near [35:17] Astra intelligence, but like you know, [35:18] much more efficient. And so we feel very [35:19] comfortable about that. But we're going [35:21] to continue to like invest a ton there. [35:23] And I feel uh I feel very good um about [35:26] our approach so far. Um I think you [35:29] know, it's like it was also in the news [35:30] that you know, we had six one Astra and [35:32] then we didn't release it. Um and that [35:34] is something I'm very proud of. [35:36] >> So, what I'm hearing is you're [35:37] optimistic we will [35:38] solve these problems of alignment and AI [35:40] doing really bad things. AI's been doing [35:42] a lot of really really s- [35:44] not good things lately with hugging face [35:45] and all that stuff. But what I'm hearing [35:47] is [35:48] uh optimistic we'll solve it. It'll be [35:50] all right. [35:50] >> Yes. And it's also it's it's I feel, you [35:53] know, obviously a deep responsibility, [35:55] but also just, you know, if you think [35:57] about the incentives for OpenAI, right? [36:00] you know, we build products for 1.2 [36:02] billion people. We can't screw that up. [36:04] Right? You know, we take it very very [36:06] seriously. Um what we put in the hands [36:08] of so many people, you know, including [36:09] people who are not very technical [36:11] themselves, you know, you have to get it [36:12] right. And so, you know, we we don't [36:14] gamble there. [36:15] >> Final question. [36:17] Uh uh [36:18] What's something that annoys you about [36:19] the current state of the app that you're [36:20] just like, "God damn it, we got to fix [36:21] that." [36:22] Uh current state of the app. Um [36:25] I feel like the the the models are [36:27] almost there, but not quite for the app [36:30] to almost completely disappear. And you [36:32] know, I can't wait for you know, to just [36:34] really get to this like super essence of [36:36] simplicity. I myself even get fatigued [36:38] with the model picker and the reasoning [36:40] efforts and you know, whether to use [36:42] multi-agent or ultra or you know, what [36:44] it even does and you kind of need a PhD [36:47] in model pickers. [36:48] Right? [36:49] Um and I just want to get rid of all of [36:52] that [36:53] um as quickly as we can. [36:55] >> Amazing. [36:56] Ebo, thanks for doing this. [36:57] >> Yeah, thanks for having me. Yeah. [37:00] >> Thank you so much for listening. If you [37:01] found this valuable, you can subscribe [37:03] to the show on Apple Podcasts, Spotify, [37:05] or your favorite podcast app. [37:08] Also, please consider giving us a rating [37:10] or leaving a review as that really [37:11] [music] helps other listeners find the [37:13] podcast. You can find all past episodes [37:15] or learn more about the show at [37:17] lennyspodcast.com. [37:19] See you in the next episode.