OpenAI President Says Startup Now Seeks AGI Post 80% Milestones

OpenAI总裁称已完成80%里程碑 正全力追求AGI 随着智能体系统实现任务执行自动化，人类注意力将取代算力成为人工智能发展的终极瓶颈。 在这次广泛的访谈中，OpenAI总裁Greg Brockman探讨了当前人工智能的前沿进展，表示我们距离实现AGI大约已完成了80%的历程。他详细阐述了在持续架构创新以及为满足全球无限需求而积极寻求算力的推动下，扩展定律（scaling laws）为何依然有效。Brockman重点指出了向智能体工作流的巨大转变——编程工具编写开发者代码的比例已从20%跃升至80%便证明了这一点——并公布了Codex的新上下文记忆工具“Chronicle”。最终，他设想了这样一个未来：人类的角色将从手动执行转变为庞大智能体网络的“CEO”，使人类注意力和数据治理成为新的关键瓶颈。 OpenAI 的核心商业模式是通过购买、租赁及自建算力并加价转售，但面对全球近乎无限的智能需求，Greg Brockman 指出公司现有的算力储备远远无法满足。 (0:53) Greg Brockman 估计目前距离实现 AGI 已完成约 80% 的进程，先进模型已能独立在夜间完成复杂系统优化的代码实现、性能剖析与自主迭代。 (4:42) 智能体编程工具已迅速从协助编写 20% 的代码跨越到编写 80%；同时 OpenAI 推出接入 Codex 的新工具 Chronicle，通过持续监测屏幕操作捕获上下文并构建记忆。 (6:46) 随着任务执行门槛急剧下降，人类注意力正成为最稀缺的资源，组织的瓶颈已转向治理与对齐；OpenAI 内部也严格要求所有合并的代码必须由真人签核负责。 (14:41)

OpenAI co-founder and president Greg Brockman argues that as compute scales and AI agents rapidly automate complex software execution, human attention, context alignment, and governance—rather than task completion—have become the defining bottlenecks in technology. Compute Constraints and Scaling Toward AGI (0:53) Brockman describes OpenAI's operational model as buying, renting, and building compute to resell at a margin, noting that compute availability remains an ongoing constraint against virtually unlimited demand. Despite industry concerns over physical limits, he asserts that neural scaling laws continue to hold alongside algorithmic and architectural innovations moving beyond original transformer designs. Brockman estimates that the industry is approximately 80% of the way to achieving artificial general intelligence (AGI). To illustrate rapid performance leaps, Brockman points to internal testing where a systems engineer supplied GPT-5.3 with a complex optimization design document. The model autonomously implemented the specification, instrumented the code, ran profilers to diagnose latency bottlenecks, and iterated overnight to deliver an optimized solution—tasks that earlier iterations could not handle. The Agentic Shift and Context Integration (6:46) Brockman highlights a pivotal transition in software engineering, stating that agentic coding tools advanced from generating 20% of developer code to 80% within a single month. To bridge the friction of constantly briefing AI, Brockman introduced Chronicle, a tool integrating with Codex that continuously observes screen activity to build persistent contextual memory. This capability aims to remove the operational tax of human context transfer across computing tasks. Internally, OpenAI pairs engineering workflows with Codex under a mandatory governance policy: a human engineer must remain accountable and sign off on all merged code. Beyond software development, the organization is systematically deploying specialized agent workflows into internal operational domains, including finance, IT, and sales. Human Attention and Governance as Core Bottlenecks (14:41) Because automated generation has made prototyping widgets and code drastically cheaper, Brockman argues that friction has shifted to data provenance, sharing, and governance. Enterprise architectures now require mechanisms to track artifact dependencies, such that restricting permissions on an original document automatically revokes derived AI outputs. Furthermore, Brockman cautions that agents still lack social nuance; he shares an incident where an agent escalated an internal technical error to an engineer's manager after just two minutes without human intervention. As task execution becomes trivial, Brockman predicts that human attention will become an intensely scarce resource. Relying on humans for passive, repetitive approvals leads to fatigue and automatic sign-offs; the industry must instead develop systems where models discern high-risk actions for escalation while safely handling routine decisions autonomously. Organizational Restructuring, Security, and Scientific Frontiers (13:17) Brockman projects that declining technological barriers will flatten traditional corporate hierarchies, enabling solo founders to run large enterprises operating like 'CEOs of 100,000 agents.' Concurrently, rising attack surfaces require continuous AI-driven red teaming and codebase auditing. Brockman urges developers to engage with structured defensive ecosystems, such as trusted cyber access programs, emphasizing that AI models provide systemic resilience rather than magic solutions. While physical intelligence and biology present messier challenges than software simulations, Brockman reports early signs of progress in hard sciences. He notes instances where models generated solutions to physics formulas related to quantum gravity that researchers previously considered unsolvable, forecasting a broader renaissance in AI-assisted scientific discovery over the coming year. Brockman stresses that keeping pace with rapid acceleration requires hands-on experimentation, as machine interfaces increasingly adapt to human intent rather than requiring human accommodation. However, navigating this transition leaves critical challenges unresolved, particularly in instilling social intelligence into autonomous agents and maintaining reliable human oversight as organizations scale.

OpenAI 联合创始人兼总裁 Greg Brockman 认为，随着智能体执行复杂任务的成本大幅降低，人机协同的瓶颈已从操作本身转移至价值对齐与风险监管，“人类注意力”正成为系统中最稀缺的新资源。 算力扩张逻辑与扩展法则的延续 (0:53) OpenAI 的核心商业模式建立在采购、租赁、自建算力并以正向毛利转售的基础之上。Brockman 证实，由于全球对机器智能的需求近乎无限，OpenAI 的算力储备始终处于短缺状态；自 ChatGPT 启动以来，团队便遵循尽可能买下全部可用算力的原则，而 AWS 首席执行官 Matt Garman 亦曾预计 2026 年行业 GPU 算力可用增量将基本趋近于零。 面对算力瓶颈，Brockman 强调模型扩展法则（scaling laws）依然展现出如物理定律般的有效性，神经网络在持续注入算力时性能仍未见顶。但这并非仅靠运行 1940 年代构想的算法，OpenAI 正在算法底层进行持续创新，从优化数据格式细节到跳脱 2018 年最初的 Transformer 架构，以底层理论演进支撑算力规模的扩充。 软件工程范式突变与上下文记忆构建 (6:46) Brockman 个人评估当前技术已走完实现通用人工智能（AGI）约 80% 的历程。以内部实验为例，系统工程师交付复杂系统优化设计文档后，GPT-5.3 能够在夜间自主完成代码编写、性能打点、分析器排查瓶颈及多轮迭代优化。同时，智能体编码工具在 12 月经历转折，其编写代码的占比从 20% 跃升至 80%，由辅助脚本转变为主要开发驱动力。 为消除人类反复向计算机解释工作背景的沟通损耗，OpenAI 推出了集成于 Codex 的工具 Chronicle。该工具通过监测屏幕行为构建长期连续记忆，使模型自主掌握用户情境。Brockman 建议初创团队尽早让 AI 接入会议与日常工作流，以解决人机协作中的信息断层。 瓶颈转向人类注意力与组织治理重塑 (14:41) 随着原型开发成本大幅下降，企业的核心挑战已由“产出内容”转向“共享控制与数据治理”。在 OpenAI 内部，虽然 Codex 已拓展至财务、销售与 IT 领域，但仍硬性规定合并代码必须由人类最终负责。同时，针对文档权限变动后模型输出的“衍生工件”（derived artifacts），系统必须具备依赖项追踪与同步失效机制，以防越权访问。 智能体执行效率的提升反而凸显了人类判断力的稀缺。Brockman 提到，在测试中智能体曾因等待耗时过长而擅自在 Slack 上向员工主管直接升级催办，反映出当前模型仍欠缺社会化情商（EQ）。过去人工反复点击批准的机制易导致疲劳放行，未来的系统必须依赖 AI 自动处理常规操作，仅将高风险行为升级呈报，从而保护并集中消耗珍贵的人类注意力。 战略聚焦、防御网络与科学前沿突破 (22:11) 面对应用层的无限延展，OpenAI 采取了聚焦战略，主要深耕企业级直销以及以长期目标形成为核心的消费级顾问服务，放弃涉足所有终端应用。在安全领域，针对网络和生物安全风险，OpenAI 扩大了网络安全可信访问计划，呼吁防御方利用大模型进行全代码库扫描与端到端红队演练，以此提升全网系统的防御韧性。 针对物理智能和科学研究进展相对迟缓的观点，Brockman 表示已有迹象表明模型能推导出解决量子引力难题的物理公式，并预测科学领域将在未来一年迎来重大突破。他预期人机交互将摆脱传统的键盘屏幕输入，演进为单人调遣数十万智能体团队的分布式组织管理模式。 Brockman 认为加速变革已是科技行业的新常态，但目前的智能体在适应复杂社会规范及应对物理世界的非结构化环境上仍面临技术与治理考验。技术普及的终极约束不再是算力能否执行任务，而是人类如何安全、审慎地配置自身注意力以驾驭庞大的自主模型。

[0:53] Greg Brockman: We buy, rent, build compute, and we resell it at a margin. [0:53] Greg Brockman: all of it [2:21] Greg Brockman: deep and very beautiful mystery [4:42] Greg Brockman: 80% of the way there [6:46] Greg Brockman: lean in [6:46] Greg Brockman: makes no sense [9:24] Greg Brockman: we still want a human to be accountable for all code that gets merged [14:41] AI agent: This is taking too long. I've escalated to the person's manager [14:41] Greg Brockman: building up the EQ of the model [14:41] Greg Brockman: human attention is going to be this incredibly scarce resource [16:44] Greg Brockman: they're not magic [18:40] Greg Brockman: play with the technology yourself [18:40] Greg Brockman: rather than have the machine be something you have to contort yourself to, the machine contorts itself to you [20:04] Greg Brockman: put the power of AI in people's hands [20:04] Greg Brockman: more of you should apply. It's great. [22:11] Greg Brockman: some very painful decisions about what not to do [24:25] Greg Brockman: a very unnatural state [24:25] Greg Brockman: CEO of an organization of like 100,000 agents [25:32] Greg Brockman: signs of life [25:32] Greg Brockman: real renaissance [25:32] Greg Brockman: totally wild [27:46] Greg Brockman: hopefully will have post-AGI

[0:02] So, Greg, uh thank you for coming back [0:03] here. Um [0:05] I don't think we ever charged you for [0:06] rent. So, uh maybe we'll send you an [0:08] invoice later. But, uh Greg, you've been [0:10] part of like two really spectacular [0:12] companies, Stripe as employee number [0:14] four, and then the first CTO. I just [0:17] recently heard that they process 1.6 [0:19] billion, sorry, 1.6% of the [0:23] global GDP. You must be proud of that. [0:25] That's amazing. You must be even more [0:27] proud of the fact that OpenAI has almost [0:29] a billion or maybe more than a billion [0:31] uh in weekly active users at this point. [0:34] I mean, it's all it's all very exciting. [0:36] It shows you what technology can do. And [0:38] uh you're not just co-founder and [0:39] president, but you're also chief [0:41] builder. [0:42] At OpenAI, I heard that that was one of [0:43] your titles. I'm not sure if it was ever [0:45] an official title, but I've been called [0:47] many things. Let's just say that. [0:49] Well, you have a audience of great [0:51] builders here, so we'll start from all [0:53] the way at the bottom of the stack. You [0:55] OpenAI has multiple stacks of the [0:57] business, one of which is compute, and [0:59] you guys have been very aggressive, very [1:01] aggressive on securing compute. Why is [1:04] that? [1:05] Well, in many ways, we have a very [1:06] simple business. [1:08] We [1:09] buy, rent, build compute, [1:12] and we resell it at a margin. [1:15] That's it. As long as the margin's [1:16] positive, then you want to scale it [1:18] because the demand for [1:20] solving problems, the demand for [1:23] intelligence, that's unlimited. [1:26] And the AIs that we have right now [1:27] really are able to rise to the challenge [1:29] of effectively any kind of problem that [1:31] you want to throw at them. Do you have [1:32] enough compute? No. [1:34] Really? [1:36] Yeah, definitely not. [1:36] >> just with Matt Garman, and he says the [1:39] GPU compute availability [1:42] in 2026 rounds to zero. Don't you guys [1:44] have all of it? I mean, we have we we [1:46] would love more. We're we're constantly [1:48] out there hunting for more, honestly. [1:50] And I'll tell you like when we first [1:52] launched when we launched ChatGPT, [1:54] I remember being on a call with my team, [1:56] and they're like, all right, how much [1:58] computer should we buy? [1:59] And I said, all of it. [2:01] And they're like, no, no, no, we're [2:02] seriously, like like come on, how much [2:03] should we buy? I'm like, no matter how [2:05] fast we try to ramp compute, I guarantee [2:07] you we're not going to be able to keep [2:08] up with demand and that has been true [2:10] ever since. [2:11] Um, that's that's fascinating. Moving up [2:14] from compute since I don't know if much [2:17] of this audience can help you with [2:19] securing more compute [2:21] um because most of them are founders of [2:22] startups. Um, about architecture and [2:25] scaling laws, what are the [2:27] what are the where are we in the scaling [2:30] laws? Are they still doubling each year? [2:32] Are you changing architecture? What's [2:35] what is [2:36] what are you guys pushing on the [2:37] frontier on the research side? Well, I [2:39] would say first of all, the scaling laws [2:42] are a deep and very beautiful mystery. [2:45] Right? They feel deeply fundamental. [2:48] It's like the scientific truth that just [2:50] like you think about physics and you [2:53] know, Newton's laws and things like [2:54] that, they're somehow this truth of the [2:56] universe. [2:58] And they're empirical. Like we don't [3:00] necessarily have all the theory to [3:02] explain exactly why it works, but to me [3:04] the most beautiful thing is that neural [3:06] networks [3:07] were really designed like in the 1940s [3:10] before there were computers. [3:12] And somehow we've been able to take the [3:15] exact ideas that were developed back [3:17] then and apply increasing amounts of [3:19] computation and as you pour more compute [3:22] into the models, they get [3:23] correspondingly more capable and it just [3:25] keeps going. There's no wall. [3:27] And that's I think that's a beautiful [3:28] thing. [3:30] That's pretty beautiful. Are there more [3:32] research or more algorithms that are in [3:35] the works that because [3:38] you know, in the past we had neural [3:39] networks to your point in 1940s, but we [3:41] couldn't we didn't have the computer for [3:42] it. Now that we have the computer for [3:44] it, you just where are we just pushing [3:45] the same things or are there new [3:47] architectures and new ideas coming up? [3:49] Yeah, so I would I would think of it as [3:50] we we absolutely have new ideas that are [3:53] constantly powering what we do. It's [3:55] very simplified to say, well, let's take [3:57] a neural network from the 1940s and, you [3:59] know, put put it in a gigawatt data [4:01] center, right? We have made tons of [4:03] innovations and we constantly are [4:05] improving things. And sometimes these [4:07] are micro tweaks. Like you just realize [4:08] that the way you've been formatting data [4:11] was not quite right and that can [4:12] actually be a very big deal. Sometimes [4:14] it's larger. You think about the shift [4:16] from the LSTM to the transformer and I [4:18] don't think the transformer is, you [4:20] know, like everyone's moved past the [4:21] transformer as described in the other [4:22] 2018 paper. [4:24] So, there's there's constant innovation [4:26] happening and I think of places that [4:27] have been perhaps the most invested in [4:29] long-term research on how to improve the [4:32] architectures, how to improve the [4:33] fundamental algorithms, and how to get [4:34] the paradigm shifts. I think OpenAI has [4:36] been leading the pack there and that's [4:38] something we continue to invest in and I [4:39] see lots of fruit on the horizon. Got [4:41] it. [4:42] And on the models, um [4:44] Does OpenAI have a formal definition for [4:46] AGI? Are we close? Are we not close? I [4:49] Pat and Sam and Ilya published this [4:50] thing that we are at AGI [4:53] functionally. [4:54] Do you agree with that? Do you not agree [4:55] with that? Well, we do have a formal [4:57] definition, but to some extent I one [5:00] thing I have learned is that everyone [5:02] has their own intuitions about what AGI [5:04] is and maybe you can view it as like [5:06] according to my view of where we are, I [5:08] think we're about 80% of the way there [5:10] in that we have models that are smart. [5:13] They're very capable. They're able to if [5:16] you give Are they smarter than you? [5:18] I mean, I they're certainly more capable [5:19] than I am at writing software, right? If [5:21] you give it all the context, then yes, I [5:23] I I think that they are they're just so [5:26] capable. It's it's really remarkable. [5:28] Like is anyone here feel better at [5:29] writing software than GPT-5.4? [5:35] Oh. [5:38] All right, writing kernels. Um so, even [5:40] there we're seeing massive gains from [5:43] Exactly. We're And for some of our [5:45] internal results, um that [5:47] we're really seeing if you pour the [5:48] right kinds of of [5:50] you know, if you have the right setup [5:52] for your problem, then you're able to [5:53] get really massive results out of very [5:56] low-level, um, even low-level tasks. And [5:58] just to give you one example of how [5:59] things have been trending, um, [6:02] one of my systems engineers, also very [6:04] similar, was like, "Hey, like, I haven't [6:06] been able to get value out of the models [6:07] for GPT-5 or 5.1, [6:10] for 5.2 as well." For 5.3, he on a lark [6:15] had prepared this design document for a [6:16] very complicated systems optimization he [6:18] was about to do. [6:19] He [6:20] handed it over to the model, went to [6:22] sleep, waking up intending to like give [6:24] this to his team to work on for the next [6:25] week. [6:26] And when he woke up, it was done. [6:28] That the model had actually [6:30] implemented the initial spec, had seen [6:32] that it was slow, had added [6:33] instrumentation, had actually run the [6:35] code, used a profiler to figure out [6:37] where things were slow, and iterated [6:39] multiple times until it got into an [6:41] optimized result. And like, like, that [6:43] is incredible. That's where we are. [6:46] And so, how what would you advise all [6:48] startups here to do? Because the models [6:49] keep getting more and more capable. [6:52] They're kind of, uh, [6:53] I've asked this, uh, when Sam was here [6:55] in the past, and, you know, what if [6:57] you're building today, do you need to [6:59] rebuild in 2 years when a new model [7:01] comes out? Because all the functionality [7:03] and all the capabilities all change [7:06] around you. [7:07] Um, do you need to make sure that you're [7:09] not in OpenAI's way because [7:12] you're going to roll, you're just going [7:14] to run over for startups because the [7:16] models are so much more capable. [7:18] Um, how how would you recommend the, uh, [7:21] a set of, um, startup founders to [7:24] to build in this environment? Well, [7:26] first of all, I would say to lean in. [7:28] The tools right now have become [7:30] incredibly useful. And if you look even [7:32] over the course of December, I think [7:34] that we went from these agentic coding [7:36] tools being like, you know, they're like [7:38] writing 20% of your code to writing 80% [7:40] of your code. Which means they go from [7:42] being kind of a side show to being the [7:44] main thing that you're doing. And I [7:46] think we're doing that across all of [7:48] the work that people do with computers, [7:50] all computer work this year. And you can [7:52] look at the recent progress on Codex. [7:53] It's really changing from a tool for [7:55] software engineers to a tool for anyone [7:57] who's doing work with a computer. And [7:59] just over the past week, we've released [8:01] [snorts] a bunch of features that just [8:02] make it so much more powerful and [8:03] capable. Um and like one thing we just [8:07] announced today is a new tool called [8:08] Chronicle that plugs into the Codex, [8:10] where it actually can see everything [8:12] you're doing with your computer and can [8:13] form memories of what's going on. And so [8:16] you ask it a question, you just it [8:18] instantly knows what you're talking [8:19] about. You're like, "Huh, what was I [8:20] doing 5 minutes ago?" It knows, right? [8:22] You're like, "Oh, what was this person [8:23] talking about?" It knows. It's To me, it [8:25] was this real wake-up call to realize [8:27] you spend so much of your effort right [8:28] now just explaining your computer what's [8:31] going on. Like, why are you explaining [8:33] to your computer what's going on? That [8:34] makes no sense. And so I think what's [8:35] going to happen over upcoming years is [8:37] the models are going to get much more [8:39] capable, will have better harnesses, [8:41] will be able to be able to solve harder [8:43] and harder problems, come up with new [8:44] knowledge, all of these things. But [8:46] there is a one-time shift that's [8:47] happening now, which is really about [8:49] context. It's really about [8:51] is your AI able to [8:54] you have all these meetings, you didn't [8:55] include the AI. [8:57] You know, that's not very nice to the [8:58] AI. Like, you're asking it to to help [9:00] you with things and it has no [9:01] information. So I think really leaning [9:02] into how do you make sure the AI even [9:05] has enough information in theory to [9:06] solve the problem and then trust the [9:08] models are going to really get there and [9:10] improve. So I think it will be a [9:12] constant cycle of improvement and [9:14] iteration and leaning into the tools and [9:16] kind of talking to your friends if [9:17] you're how they're using it, but that [9:19] there is this investment that's a [9:20] one-time investment that now is the time [9:23] to make. And in terms of like, let's say [9:25] we you you set that all up, how do you [9:28] how is OpenAI using Codex differently [9:31] than you think [9:33] everybody else outside is using it? [9:35] Well, I think one of the amazing things [9:37] about being at OpenAI is you do get to [9:39] live in the future, right? You do get to [9:40] really see the shape of what's emerging, [9:42] and we can co-design, right? We can [9:44] really change the models, the harness, [9:46] everything together in order to better [9:48] serve the needs that we see. [9:51] And a lot of the approach we've been [9:52] taking is so we started with software [9:53] engineering, and we set some clear [9:55] guidelines for example saying that [9:58] we still want a human to be accountable [10:00] for all code that gets merged, right? [10:02] So, [10:03] at the end of the day, [10:05] is it a good thing to merge this piece [10:06] of code? [10:07] Is it well-structured? Is it going to [10:09] make our codebase more maintainable? We [10:11] want to make sure there's a human who is [10:12] signing off to say yes. And that's I [10:14] think that thoughtfulness of not just [10:16] saying, "Okay, let's just blindly use [10:17] this." Or, you know, "Oh, we don't want [10:18] to use this at all." Like I think [10:19] neither extreme is quite right. And then [10:21] we are also going vertical by vertical [10:23] within OpenAI to adopt these tools [10:25] within finance, within sales, within [10:28] IT. And there we have a small dedicated [10:31] team who's really deeply understanding [10:33] the domain, working with the people who [10:35] are the experts in it in order to build [10:37] skills, in order to modify the Codex UI, [10:40] whatever it is that is is needed in [10:42] order to get it to be good. And then [10:44] that's something we can then, once we [10:46] have it in good shape, we will [10:47] externalize, and that we're able to to [10:49] ship that to all of you. And so we are [10:51] starting to work with certain customers [10:53] as well. So, for people who want to be [10:55] very AI forward and want to be part of [10:57] defining this revolution, that there's a [10:59] place for that, and I'd love to talk [11:00] afterwards. [11:02] But yeah, I think that just this desire [11:03] to say, "Hey, we really want to be AI [11:05] forward, really live in the future, and [11:06] experience what it will be like for [11:08] everyone else 1 year, 2 years, 3 years [11:10] down the road." [11:11] Do you guys structure your company [11:13] differently with the engineering teams [11:15] differently because of of the living in [11:17] the future? I mean, if you have to go [11:19] way back, when my father learned [11:20] computer science, he was just himself, [11:22] and then we had these long software [11:25] releases that became waterfall. And then [11:28] when the web happened and the cloud [11:29] happened, when these two-pizza teams, [11:31] and we have scrum. [11:33] Now that we have these coding agents, is [11:35] it is how do you structure around [11:38] everything differently? [11:39] I think we're still figuring it out, and [11:41] there's certain places where you really [11:43] see it. For example, the cost of [11:45] building a prototype is cheap now. It's [11:47] so cheap, and if you want to build a [11:49] dashboard that used to be like, uh, it'd [11:51] take like someone like a week to do it, [11:53] and you just do it now. And so actually [11:55] a lot of the bottleneck has shifted to [11:57] things like sharing. Like how do you And [11:59] so we we actually have some internal [12:01] work on this as well that again we will [12:02] be externalizing. If How do you make it [12:04] really easy for anyone in your [12:05] enterprise to build a dashboard, a [12:07] widget, a bot, whatever the thing is, [12:09] and then share it with others. And then [12:11] that starts to really put pressure on [12:13] having good governance. Like you want [12:14] your IT organization to be able to see [12:16] all these different, [12:17] you know, threads of execution that are [12:18] happening, all the little things that [12:19] are being shared around, have some [12:21] control over [12:23] data provenance, right? To really make [12:25] sure that, okay, like a good example of [12:27] this is, um, [12:28] I think people are now starting to take [12:30] their internal knowledge from storing [12:31] them to wikis. We we have some some a [12:33] really cool one of these internally, and [12:36] the thing you immediately think about [12:37] is, well, if someone has a document in [12:40] the internal knowledge base that was [12:41] accidentally permissioned incorrectly, [12:43] and they realize, oh no, I didn't want [12:44] this information to be accessible, how [12:47] do they fix that, right? So normally [12:49] it's they go into the doc, they change [12:50] the permissions, but now there's these [12:51] derived artifacts. And so you need to [12:53] make sure you have some way of tracking [12:54] through the system to say, well, this [12:56] output document came from the source [12:58] one. The source one's no longer [13:00] accessible to this Let's go and [13:02] invalidate that as well. And so you have [13:03] to start really building your technical [13:05] architecture with awareness of the way [13:07] that people are going to use this [13:09] information, and it really changes how [13:11] teams relate to each other because you [13:13] can just It really changes where the [13:15] bottlenecks are and what's hard. [13:17] Uh, do you think [13:19] team sizes going to be a lot smaller? [13:21] Are we going to have [13:22] still human software engineers in a [13:24] decade? [13:25] Well, decade is a long time from now and [13:28] the the ceiling on this technology is [13:30] hard to is really hard to internalize. I [13:32] think that it is clear that what a [13:34] company is will change in a lot of ways. [13:37] I think that we're going to have this [13:39] ability for solopreneurs to build very [13:42] incredible businesses. And so anyone who [13:44] has a vision, I think will be able to [13:45] realize it. I think the jobs that you [13:47] all have will become way easier in a lot [13:49] of ways, way more fun. Now, might be [13:51] more competitive, too. Right? Because [13:52] everyone's going to have these amazing [13:53] tools. And so really figuring out what [13:55] is your niche, what is your unique [13:56] angle, is probably going to become kind [13:58] of the most important core. But a lot of [14:01] how we run organizations right now, and [14:03] it's there's almost only one way to [14:04] organize large groups of people, where [14:05] you have teams, you have management [14:06] structures, and you have scopes, and you [14:08] have these hierarchies, and all these [14:09] things. Maybe that can change. Maybe you [14:11] can be much more flat, small teams that [14:13] can really just do incredible things. [14:15] Like, we're seeing it right now in [14:17] mathematics, where [14:19] these individuals on the internet are [14:21] using GPT-4 Pro to solve these unsolved [14:24] math problems. Normally you need a math [14:26] team, and they're just doing it. Yeah. [14:28] And my son's a math nerd. I just told [14:30] him that maybe we should be studying [14:32] something else besides math. [14:34] But I well, but see, this is the [14:35] question, right? Is if you look at [14:37] something like AlphaGo, you know, move [14:39] 37, this move that just like changed [14:41] humanity's understanding of the game, [14:43] but the thing that was surprising is [14:45] it made the game more interesting and [14:47] important for humans. And maybe that [14:49] will be true for for these other [14:50] domains, too. True. [14:52] What about uh common failure modes when [14:54] you're building when you're you're [14:56] you're building with uh production [14:58] agentic workflows? What do you What do [14:59] you see as the common things that [15:01] founders get wrong and they're building [15:03] incorrectly these days? Well, I think [15:06] that [15:07] these models, they have such power, and [15:09] really understanding how to operate them [15:11] well takes thought. And so we've been [15:14] investing a lot in primitives, security [15:16] primitives, observability, having again [15:18] good governance, things like that. Um [15:20] but just to give you one anecdote that I [15:22] think is evocative, um, I asked So, I [15:24] was working with my co-workers, I asked [15:25] them to install some package that [15:26] someone at OpenAI had written, ran into [15:28] an error, I was like, "Oh, ping that [15:29] person on Slack and ask them for help." [15:31] So, ping the person on Slack. [15:33] Two minutes later, it said, "This is [15:34] taking too long. I've escalated to the [15:35] person's manager." And it actually [15:37] pinged the person's manager. [15:38] >> [laughter] [15:40] >> And And you realize it's like, on the [15:42] one hand, it's kind of a reasonable [15:44] thing for the model to do. It's being [15:45] proactive, it's trying to solve my [15:46] problem. It's like, you know, not just [15:48] sitting around waiting to be told what [15:49] to do. But, on the other hand, like, you [15:51] know, [15:52] >> [laughter] [15:52] >> maybe it should have taken a little bit [15:54] longer, maybe should have checked with [15:55] me. And so, I think that really thinking [15:58] about these questions where we're still [16:00] building up the EQ of the model. And [16:03] that in some places, it's getting very [16:05] good. For example, clicking approve, [16:06] approve, approve is kind of where we've [16:08] been. And humans are not very good at [16:10] that, either, right? [16:11] >> They just They They just default. [16:13] >> default. And so, now we're starting to [16:14] have AIs that can actually take care of [16:16] flagging, is this a high-risk action? [16:18] Hey, this one should be escalated, this [16:20] one's okay to auto-approve. And it [16:21] really makes you realize that human [16:23] attention is going to be this incredibly [16:24] scarce resource, right? The doing of [16:26] things now is easy. The Is this a good [16:28] thing? Is this what I wanted? Is this [16:30] aligned with my values, with my desires? [16:32] That is going to become the single most [16:35] important bottleneck. And so, I think [16:37] building systems that take that into [16:39] account and really think about the human [16:40] factor, like, that's the most important [16:42] thing to do now. [16:44] Another human factor, security. [16:46] Um, how would you advise people to think [16:49] about security in this world of AI and [16:52] heard about breaches left and right with [16:55] Vercel recently and then [16:57] And these models are incredibly powerful [17:00] at finding security holes. So, how how [17:02] would you recommend people here use the [17:04] models to to find those security issues? [17:06] Well, I think there's a couple levels to [17:07] the answer. I do think that this is I [17:10] think that the internet has been a place [17:12] where security has been just like a a [17:15] ratcheting important concern over time. [17:17] You think about where it started, going [17:18] through the '90s with viruses and worms [17:20] and malware and those things and we've [17:22] moved past that. Um I think we are also [17:24] moving now to a much more ultimately [17:26] secure regime, but it does require kind [17:28] of an internet-wide effort to get there. [17:30] And so a lot of this honestly is just [17:32] again leaning into the technology, [17:33] having these models they can scan your [17:35] code base, they can actually be used for [17:37] end-to-end red teaming. Like there's a [17:39] lot that can be done with them and a lot [17:41] of how we're thinking about further [17:42] models and improvements there is really [17:44] leaning into how do we how do we [17:46] actually sort of leverage trusted access [17:49] programs, how do we leverage the [17:50] community of people who really care [17:52] about being defenders and making the [17:54] internet more secure. I think that's [17:56] something where everyone has a role to [17:58] play and can participate, but the number [18:00] one thing is just sort of recognizing [18:02] that these models are very powerful, but [18:04] they're not magic, right? That they are [18:06] just like a part of the overall [18:07] resilience ecosystem and I think that we [18:09] as a society and I think every company [18:12] again really contributes to this have [18:14] something to build in terms of how do we [18:16] how do we incorporate these in a way [18:18] that results in more [18:20] assurance and more [18:23] sort of certainty on the impacts of of [18:27] whether it's a particular patch that [18:28] you're taking, whether it's thinking [18:29] about how do you make sure that you're [18:31] um yeah, just sort of rolling in updates [18:32] quickly as they're being released. Um so [18:34] I think that there's a lot of work to be [18:36] done, but I have a lot of optimism for [18:38] where this is going. [18:40] Um let's switch to speed. Seems like [18:43] things are moving faster and faster and [18:45] faster when in the world of accelerating [18:47] change. We were talking about it when we [18:48] when you [18:49] uh you were walking up here around how [18:52] how you're trying to keep up with [18:54] things. How how do you you keep up with [18:57] all the accelerating change? How would [18:58] you recommend everybody here keep up [19:00] with everything that's changing? [19:02] Well, I think this is the new normal and [19:05] I think to some extent it's not really [19:06] because of AI. I think it's just been [19:09] the trend of technology for the past two [19:11] decades. [19:13] There's more people doing things. It's [19:14] easier to do things than ever. Barrier [19:16] to entry goes down. [19:18] Means it's also much more easy to build [19:19] value, right? To have great successes. [19:21] And so, I think that really trying to [19:23] keep your ear to the ground and [19:25] understand what's changing. And to some [19:27] extent, it always starts with the same [19:28] thing, which is play with the technology [19:29] yourself. Like, it's very different to [19:32] hear AI described versus to use it. But, [19:35] the beautiful thing about AI is it's so [19:37] intuitive. Like, that's the whole point [19:38] is that rather than have the machine be [19:40] something you have to contort yourself [19:41] to, the machine contorts itself to you, [19:44] right? It's doing work for you. And it [19:45] should be something where you ask it and [19:47] does something. And so, I think that [19:48] just really trying to just get your [19:52] finger on the pulse of what's changing, [19:54] what's possible, where the models lag. [19:56] That is, I think, the core skill that is [19:59] going to really determine a lot of the [20:01] success of of companies in the future. [20:03] And then, on the flip side of that, you [20:05] guys have held up held back models to [20:07] work with security agents. So, it's like [20:09] the opposite of like going as fast as [20:11] possible. So, um [20:14] you're doing things responsibly, too. [20:15] So, how do you kind of like think about [20:17] the balance? Because you're in a [20:18] competitive environment, you want to [20:20] ship as quickly as possible, and yet [20:22] you're trying to do the right thing, as [20:23] well. [20:24] Yeah, I think at a values level, like [20:26] what OpenAI is about, like we really [20:28] want to put the power of AI in people's [20:31] hands. Like, we believe that people can [20:34] We want to empower people to build the [20:35] future with the tools that are being [20:37] created. But, we need to do that in a [20:38] thoughtful way, right? That we really [20:40] think about both sides of Here are the [20:41] benefits, here are the risks. How do you [20:43] maximize the benefits? How do you [20:44] mitigate those risks? And I think that [20:45] in cybersecurity and in biosecurity, [20:48] those are areas where we're very [20:50] thoughtful. We've been building We've [20:51] been working on these kinds of both [20:53] mitigations and trusted access programs [20:55] for quite a long time. And that what we [20:58] see coming is models that are going to [21:01] be increasingly powerful and capable in [21:03] a continuous way across all dimensions [21:05] of capability. And the [21:09] you know, we announced last week [21:11] the expansion of our trusted access for [21:13] cyber program. By the way, has anyone [21:14] here applied? [21:17] No one. Oh, I see one hand, two hands. [21:19] Okay, more of you should apply. It's [21:20] great. [21:21] We really need help because and it's [21:23] very important that people who are [21:25] trustworthy and responsible and really [21:27] want to push these models are [21:29] participating in this because that is [21:30] how that's going to pay dividends for [21:32] everyone. [21:33] We're going to have more to announce [21:34] over upcoming weeks on how we're [21:36] expanding the program, but and also when [21:39] we release models to everyone kind of [21:41] the mitigations that that we have and [21:42] how we're going to tune those to be both [21:45] to really balance, right? To really try [21:46] to bring these capabilities as broadly [21:48] as possible, while also making sure that [21:50] the ones that are, you know, that we're [21:52] thinking about the risks and and able to [21:54] have some observability over them and to [21:56] ensure that this is maximally positive [21:58] in terms of deployment. So, I think the [21:59] short answer is like it's core to our [22:01] mission. We care a lot about the impacts [22:03] of what we're doing, not just building [22:05] technology in isolation, but it is a [22:07] whole community, a whole world effort to [22:09] really get to where we need to be. [22:11] On [22:12] moving up from the models to the [22:14] application layer, which is what a lot [22:15] of people here are building. [22:18] How do you How does OpenAI decide what [22:21] in the application layer you're you're [22:22] going to build and what you're going to [22:24] leave out? [22:26] Well, people have probably seen the word [22:28] focus being applied to OpenAI quite a [22:31] lot recently, possibly for the first [22:33] time [22:34] >> in a while. [22:35] And uh It's been applied to her, too. [22:37] >> [laughter] [22:38] >> And it's it's hard because the field of [22:41] AI is one of opportunity, right? It's [22:43] like anything you're going you can [22:44] imagine is going to be great. No [22:46] question it's going to be great. And [22:49] we as a company, as a single company, no [22:52] matter how much compute we build, no [22:53] matter how many people we have, are only [22:55] going to be able to do so much. And so, [22:56] a lot of where we've been how we've been [22:58] thinking about things is what is the [23:01] sort of most focused strategy that [23:04] covers the parts of the space? [23:07] You know, maybe it's an 80/20 or just [23:09] like the parts of the space that we [23:10] think we can have most impact on. And I [23:12] think there it's very clear right now [23:13] we're going through this identity [23:15] transition. And so [23:17] >> [clears throat] [23:17] >> products that are And it's not just [23:20] about enterprise versus consumer, right? [23:22] So it's like clear we are being very [23:23] serious about enterprise. Like we're [23:25] selling to to big companies and and [23:26] building a whole muscle and sales motion [23:28] there. But consumer, what consumer is is [23:32] going to change, right? It's kind of a [23:33] very broad term that buckets in multiple [23:35] things. But that the slice of consumer [23:37] that's about not just productivity, but [23:39] about goals, about achieving your goal, [23:41] about even knowing what is your goal, [23:43] being able to elicit that and having an [23:44] AI that can proactively do that. It's [23:46] all kind of the same thing. Like in the [23:48] end we're trying to build an AGI that [23:50] you can talk to, that has all this [23:52] context that you can use in your [23:53] personal life, your work life, that's [23:54] trustworthy, right? That you can go to [23:56] it for advice, can give you useful [23:58] information, maybe health information, [23:59] or maybe about finances, or [24:02] you know, about if you're trying to [24:03] figure out what to do with your career. [24:04] Like all these things, they all kind of [24:07] ladder into one thing. And it's meant we [24:09] had to make some very painful decisions [24:11] about what not to do. But I think I [24:12] would just say that that's the aperture [24:14] that we look at things through and that [24:16] things that accrue to that singular [24:18] vision of what we want to build, you [24:20] should expect us to pursue. Got it. [24:23] Um [24:25] Do you think we'll be coding with [24:27] command lines and agents in [24:29] in a few years or is it going to be [24:31] completely changed? I mean, I think that [24:34] we're in a very unnatural state right [24:36] now for how we work. [24:38] Like we all sit behind this box and kind [24:40] of type away and it's very clear our [24:43] bodies were not designed for this. We [24:44] got our carpal tunnel and our, you know, [24:45] hunched shoulders and all these things. [24:48] And I don't think we want that. I don't [24:49] think any of us wanted that. Like I [24:51] think that we want more free time. We [24:52] want more time, but it's it's not even [24:53] about free time necessarily, right? It's [24:55] like you want to spend more time with [24:56] your loved ones? Yes. You want to spend [24:58] more time like talking to people and [25:00] like coming up with like brilliant [25:01] visions or just like what you're excited [25:03] about or just understanding yourself. [25:05] So, it's kind of like do you do you want [25:07] to be a CEO of an organization of like [25:09] 100,000 agents? Like that actually seems [25:11] pretty good. And I think that we're all [25:13] going to be able to get so much more [25:14] done, but the mechanics of it are going [25:17] to feel as different as like going from [25:20] having to write out things with, you [25:22] know, by hand with a quill or something [25:25] to being able to you know, just send a [25:27] text message and have people go and and, [25:29] you know, working on your behalf on your [25:31] goals. [25:32] All right. We talked about compute. We [25:34] talked about model and security and [25:38] agents and app layer. Let's talk about [25:41] frontier. [25:42] When when are the models going to be [25:44] good enough to push the frontiers of [25:47] science, physical AI? Seems like we had [25:50] Jen Fan here. It seems like LLMs have [25:53] been a great scaling law for digital [25:54] intelligence. It hasn't been as strong [25:57] for robotics, for physical intelligence, [26:01] for [26:02] aspects of [26:03] biology and science where the problems [26:05] are probably a lot harder to verify or [26:08] takes a long time to verify. [26:10] Well, how are you keeping track of [26:11] science and and physical AI in in the [26:14] world? Well, well, science is one domain [26:16] that we're really leaning into and we [26:18] see line of sight to really incredible [26:20] progress. [26:22] And we're starting to have some signs of [26:23] life and I think it's always important [26:24] to ground in what is happening today [26:26] when trying to predict what will happen [26:28] 6 months a year from now. So, for [26:30] example, we had a physics result where [26:34] our AI came up with this very beautiful [26:36] formula that physicists who've been [26:38] working on this for quite some time [26:39] thought was totally impossible. Thought [26:41] it was like maybe an unsolvable problem. [26:43] And [26:44] like it's pretty significant, right? [26:46] It's like real serious physicists who [26:49] who who really view this as a step [26:51] towards really being able to get to um [26:53] to to some sort of answer for quantum [26:55] gravity and all these things. It's not [26:56] there, but it's a step. That's much [26:59] bigger than where we were just a couple [27:00] months ago. And so you it makes you [27:02] really wonder a year from now, like how [27:04] far will we have traveled? Now, things [27:07] like biology that they are different [27:09] from physics and math, right? That they [27:10] are you got to leave your beautiful [27:12] simulated world and you know, deal with [27:14] messy reality. But I think we've been [27:17] learning how to deal with messy reality [27:19] in other domains. Software engineering [27:20] is a perfect example where we've really [27:22] realized that just building the thing [27:24] that solves competition [27:27] you know, programming competitions, like [27:29] that's not enough. Like you need [27:30] something that's seen real-world messy [27:32] codebases, humans interrupting it [27:33] different ways, like this adversarial [27:35] banging at it. And I so I think that [27:37] that on science, I expect we're going to [27:39] see a real renaissance. You know, maybe [27:40] we'll see some big results this year. [27:42] Next year I think it's going to be a [27:43] totally wild wild time. [27:46] We live in interesting times. Um I I [27:49] promised that I get you out on time [27:50] because you're a busy man. Um before we [27:52] let you leave, we got 1 minute on the [27:54] shot clock. What since [27:56] you have no time, but soon you will have [27:58] lots of time, um [28:00] what do you and Anna do for fun? Uh fun? [28:04] I mean, [28:05] same as anyone. Like like to watch [28:07] movies, go on hikes, that those kinds of [28:09] things. [28:10] Um [28:11] you know, not as much time for it as as [28:12] maybe we'll we'll hopefully have [28:14] post-AGI. Um but you got to kind of [28:16] enjoy the the ride along the way. [28:19] Thank you, Greg, for joining us. Thank [28:21] you, everyone. [28:21] >> [applause]