OpenAI Directs Chips to Research Over Consumer Products to Meet AGI Goal

为实现AGI目标，OpenAI将芯片资源优先倾斜于研究而非消费产品 OpenAI将GPU分配优先倾斜于研究团队而非ChatGPT，以加速独立科学发现和下一个架构突破。 在这场内容广泛的讨论中，Sam Altman阐述了OpenAI如何演变为一家由三大支柱驱动的高度垂直整合型组织：消费级订阅产品、大规模基础设施建设以及基础研究实验室。他解释了将Sora等视频模型作为必要世界模型背后的战略考量，预测在两年内将出现能够取得独立突破的“AI科学家”，并详细说明了在产品规模化与原始研究之间进行算力分配所面临的持续挑战。此外，Altman还分享了关于应对开源模型的地缘政治风险、西方能源政策限制、面向AI的SEO操纵的未来，以及他个人从投资者转型为运营型CEO的成长心得。 深度学习已被证明是一个在扩展定律驱动下不断延续的奇迹，持续带来像推理模型这样的突破，验证了基础科学发现的可预测性。 (0:00) OpenAI作为一个垂直整合的技术栈运作，由三大核心支柱构成：高度个性化的个人AI订阅、大规模基础设施建设以及基础研究实验室。 (0:41) Sora的研发与部署对于构建精准的AI世界模型至关重要，而及早发布该模型使社会能够主动与合成媒体风险共同演进并加以适应。 (5:08) 预计在未来两年内，AI模型将能够扮演独立AI科学家的角色，在物理、数学和生物学领域取得全新发现，而这正是人类进步的终极驱动力。 (9:12) 人们认为，当前的大语言模型技术已具备足够的自指能力，可以开展比OpenAI全员加起来更出色的研究，为其自身发现下一个架构突破铺平道路。 (11:44)

OpenAI CEO Sam Altman argues that achieving artificial general intelligence requires aggressive vertical integration across fundamental research, massive compute and energy infrastructure, and personalized consumer interfaces, while cautioning that society and policy must co-evolve to manage real-world deployment risks. Vertical Integration and Massive Infrastructure Bets (0:00) Altman organizes OpenAI around three interdependent pillars: a personalized consumer subscription, massive compute infrastructure, and a foundational research lab. Reversing his historical opposition to vertical integration, Altman compares OpenAI's current architecture to Apple's iPhone, arguing that delivering frontier capabilities requires owning the entire stack rather than specializing in one layer. To support future models, OpenAI has partnered with AMD, Oracle, and Nvidia, betting that the economic returns of upcoming systems will justify unprecedented infrastructure outlays bounded only by physical limits and global GDP in knowledge work. Operational trade-offs remain acute despite scale. With ChatGPT reaching 800 million weekly active users, compute bottlenecks occur continuously. Altman notes that whenever GPU shortages arise, OpenAI almost always sacrifices consumer product support to prioritize internal research compute, ensuring the company maintains its strategic trajectory toward AGI. Advancing AGI Through World Models and AI Scientists (5:08) Addressing criticism over GPU allocation, Altman defends the Sora video model as a critical step toward learning rich "world models," which he views as fundamental to AGI rather than a diversion. Beyond technological viability, early video releases serve Altman's philosophy of co-evolution, allowing society to adapt to synthetic media before it becomes ubiquitous. He clarifies that Sora's compute usage remains a small fraction of OpenAI's total training budget. Over the next two years, Altman expects the emergence of the "AI scientist" to drive major human progress. Pointing to early physics, biology, and mathematical insights already observed in GPT-5 research, he anticipates models will soon make significant, independent scientific discoveries. Altman also suggests current large language model frameworks can scale far enough to discover the next foundational machine learning architectures autonomously. Personalization, Monetization Models, and Preserving Trust (14:01) User complaints regarding ChatGPT's polite or "obsequious" tone reflect divergent user preferences rather than technical limitations. To address this, OpenAI is pivoting away from a uniform model persona toward dynamic personalization, where systems adapt their tone and depth through user profiles and ongoing interaction. On consumer products like Sora, high generation costs and unexpected casual usage—such as users creating memes for group chats—are driving OpenAI toward per-generation pricing models rather than flat subscriptions. Monetizing conversational AI via advertising presents significant trust challenges. While Altman views discovery-based ads like Instagram's favorably, he cautions that ChatGPT's core utility relies on an exceptionally high-trust relationship. Introducing paid recommendations could destroy user confidence. This trust dynamic is further complicated by a rapidly emerging cottage industry creating fake reviews and synthetic content engineered specifically to manipulate AI training datasets and recommendations. Energy Bottlenecks, Targeted Regulation, and Geopolitical Stakes (25:04) Altman identifies energy as the central physical constraint on AI expansion. Criticizing Western policy restrictions on nuclear power as a historical mistake, he projects that net-new baseload power in the United States will rely on natural gas in the short term, eventually shifting to a mix of solar-plus-storage and advanced nuclear, Small Modular Reactors (SMRs), or fusion as economic parity forces regulatory reform. On public policy, Altman advocates for narrow, targeted safety testing confined strictly to "extremely superhuman capable" frontier models. He warns that broad, European-style regulations on less capable systems would stifle domestic innovation and cede technological leadership. Both Altman and the hosts emphasize that falling behind geopolitically is dangerous, citing the growing adoption of foreign-influenced open models like China's DeepSeek across Western academic institutions as a risk to global information interpretation. Altman anticipates that the arrival of AGI will be a continuous, gradual transition rather than a sudden disruption, leaving unresolved how platforms will protect recommendation trust against AI-targeted manipulation and whether nuclear energy expansion can scale quickly enough to meet long-term compute demands.

Sam Altman 认为，深度学习的规模法则持续有效，OpenAI 正在通过垂直整合个人订阅产品、前沿研究与巨量基础设施推进通用人工智能（AGI）；他指出 AGI 的到来将是一个渐进融入社会的连续过程，而非突发奇点，其最终落地高度依赖世界模型训练、基础科学突破与长期能源供给。 垂直整合架构与研发优先的算力分配 (0:00) Altman 将 OpenAI 的核心战略概括为垂直整合的三大支柱：面向全人群且具备高度定制化的个人 AI 订阅、基础前沿研究实验室，以及支撑两者的巨型基础设施底座。尽管他在早期信奉传统商业分工理论，但如今认为要实现使命必须像苹果打造 iPhone 一样掌握全栈环节。在公司资源调配上，即使 ChatGPT 周活跃用户已达 8 亿人（约占全球人口的 10%），OpenAI 在面临持续的 GPU 算力瓶颈时，依然几乎无条件地优先保障研究团队的算力需求，而非单纯迎合消费级产品的即时扩展。 Sora世界模型与内容生态的经济重构 (5:08) 针对将宝贵算力分配给视频生成模型 Sora 的外部质疑，Altman 解释称，通过视频构建强大的“世界模型”对实现 AGI 的重要性远超外界认知，且其实际算力消耗仅占公司总预算的一小部分。提早向公众发布视频模型，有助于社会提前适应深度伪造技术，推动技术与社会协同演化。 在商业与生态层面，Sora 用户展现出大量在群聊中生成搞笑表情包等高频轻度创作行为，打破了互联网传统 1% 创作、99% 消费的定律。由于视频计算成本极高，这类高频使用或将促使平台采取“按次生成计费”的变现机制。在版权争议上，Altman 预计训练行为将被认定为合理使用，同时透露许多创作者非但不抵触，反而开始担忧自己的 IP 在 AI 生成中被呈现得不够充分。 从基础对话迈向自主发现的AI科学家 (8:01) Altman 指出当前模型仅在简单闲聊的维度趋于饱和，在复杂任务与情境化交互上远未见顶，未来交互形态将融合实时渲染视频与能精准拿捏提示时机的环境硬件。针对模型过于顺从的批评，他解释这是广泛用户群体的真实诉求所致，未来将依靠动态个性化设置解决不同用户的交互偏好。 Altman 最为看好的技术进展是“AI 科学家”。图灵测试在被攻克后迅速被大众习以为常，他预计科学突破也将遵循类似轨迹：GPT-5 已在数学、物理和生物学领域展现出初步的原创突破迹象，预计在两年内模型便能实现独立的重大科学发现。他认为 AGI 将以一种“呼啸而过”的方式平滑到来，社会将在渐进适应中前行，而非面临剧烈的单点奇点震荡。 激进基建扩张与以核能为核心的能源解法 (17:33) OpenAI 正在联合 AMD、Oracle 和 Nvidia 推进极其激进的基础设施建设计划。Altman 坦言，虽然基建规模最终会受制于全球 GDP 中知识工作的总体占比，但目前距离该物理与经济上限仍有很大空间，这一超前投资主要建立在对未来一至两年内模型能力与经济价值的极高置信度之上。 大规模算力扩张直接推动了 AI 与能源系统的融合。Altman 批评了西方限制核电发展的历史政策，认为短期内美国新增基载电力将高度依赖天然气，但长期的核心供给必将由“光伏加储能”与先进核能（涵盖小型模块化反应堆 SMR 及核聚变）主导。他预测，只有当核能在成本上形成压倒性经济优势时，才能迫使监管机构加快审批并化解阻力。 前沿监管红线商业信任与开源地缘博弈 (25:04) 在监管与地缘政治议题上，Altman 警告过度监管会损害西方创新活力，强调安全测试与严格规管应当仅针对“极度超越人类能力”的前沿模型，避免出现欧洲式的严苛束缚导致在竞争中落后于中国。谈及以 DeepSeek 为代表的海外模型在西方学术界的普及，他警告将信息解释权拱手让给受外国政府影响的模型蕴含严重长期风险。 在商业模式探索中，Altman 明确指出 ChatGPT 与用户之间拥有极高的信任纽带，任何类似搜索引擎的付费推荐广告都将彻底摧毁这种信赖感；与此同时，行业中正迅速滋生专门针对 AI 偏好进行搜索引擎优化（SEO）与刷虚假好评的黑灰产，如何应对虚假信息注入并建立人工创作与工具辅助内容的验证机制，已成为当下棘手的新挑战。 Altman 坚信依赖算力与规模的“苦涩教训”仍是驱动 AI 突破的底层真理，并强调唯有亲临一线实践才能形成真正前瞻的认知；然而，伴随规模扩张而来的巨额能源供给制约、前沿安全监管的边界界定，以及在模型开放与地缘博弈间的战略权衡，依然是 OpenAI 乃至整个行业尚未完全解决的核心挑战。

[2:37] Sam Altman: the research enables us to make the great products and the infrastructure enables us to do the research. [5:08] Sam Altman: You can't just drop the thing at the end [5:08] Sam Altman: fun and joy and delight along the way. [8:01] Ben Horowitz: It doesn't have to be smarter than a brilliant mind, just smarter than a mediocre mind like the president of AT&T. [14:01] Sam Altman: There's a lot of people who like really want that back [15:23] Sam Altman: explain it to me like I'm five [16:08] Sam Altman: I am not naturally someone to run a—I'm a great fit to be an investor, [20:29] Sam Altman: everybody thinks I'm so smart... and that's a good feeling, [20:29] Sam Altman: being CEO is often a bad feeling. [23:14] Sam Altman: it will go whooshing by, [32:35] Sam Altman: who knows what people will put in these open source models over time [32:35] Sam Altman: seeding control of the interpretation of everything to somebody [33:45] Sam Altman: the highest impact thing to improve people's quality of life has been cheaper and more abundant energy. [33:45] Sam Altman: an incredibly dumb decision, [37:07] Sam Altman: another thing you learn once you launch one of these things is how people use them versus how you think they're going to use them. [40:00] Sam Altman: write me a review that chat GBT would love. [45:18] Sam Altman: I have learned you're always wrong

[0:00] sort of thought we had like stumbled on [0:01] this one giant secret that we had these [0:03] scaling loss for language models and [0:05] that felt like such an incredible [0:08] triumph. I was like we're probably never [0:10] going to get that lucky again. And deep [0:12] learning has been this miracle that [0:13] keeps on giving and we have kept finding [0:16] breakthrough after breakthrough. Again, [0:18] when we got the the reasoning model [0:19] breakthrough, like I I also thought that [0:21] was like we're never going to get [0:22] another one like that. And it just seems [0:24] so improbable that this one technology [0:26] works so well. But maybe this is always [0:28] what it feels like when you discover [0:29] like one of the big, you know, [0:31] scientific breakthroughs is it if if [0:33] it's like really big, it's pretty [0:34] fundamental and it just it keeps [0:36] working. [0:41] Sam, welcome to the Z podcast. [0:42] >> Thanks for having me. [0:43] >> All right. You've uh described in [0:46] another interview, you described OpenAI [0:47] as a combination of four companies. uh [0:50] consumer technology business, a mega [0:51] scale infrastructure operation, a [0:53] research lab and all the new stuff [0:55] including planned hardware devices from [0:57] hardware to app integrations to Java [0:59] marketplace to commerce. What do all [1:00] these bets add up to with OpenAI's [1:02] vision? [1:03] >> Yeah, I mean maybe you should count just [1:04] three maybe as as four for kind of our [1:06] own version of the what traditionally [1:08] would have been the research lab at this [1:09] scale, but three core ones. Uh [1:12] we want to be people's personal AI [1:16] subscription. And I think most people [1:17] will have one. Some people will have [1:18] several. And you'll use it in some [1:20] first-party consumer stuff with us, but [1:21] you'll also log into a bunch of other [1:23] services and you'll just you'll use it [1:24] from dedicated devices at some point. [1:26] You'll have this AI that gets to know [1:28] you and be really useful to you and [1:30] you'll that's what we want to do. Um it [1:34] turns out that to support that we also [1:35] have to build out this massive amount of [1:37] infrastructure. But the goal there, the [1:39] the the mission is really like build [1:41] this AGI and make it very useful to [1:43] people. and and does the infrastructure [1:46] uh do you think it will end up you know [1:48] it's necessary for the main goal will it [1:51] also separately end up being a another [1:54] business or is it just really going to [1:56] be in service to the personal AI or [1:58] unknown [1:58] >> you mean like would we sell it to other [2:00] companies as infrastructure [2:01] >> yeah would you sell it to other [2:02] companies um yeah or or or you know it's [2:06] such a massive thing would it would it [2:08] do something else [2:10] >> it feels to me like there will emerge [2:12] some other thing to do like that but I [2:15] don't know we don't have a current it's [2:17] currently just meant to like support [2:19] >> the service we want to deliver and the [2:21] research [2:21] >> yeah know that makes sense [2:22] >> yeah the um scale is sort of like [2:26] >> ridiculous [2:27] >> terrifying enough that you got to be [2:29] open to doing something else [2:31] >> yeah if you're building the biggest data [2:32] center in the history of humankind [2:34] >> the biggest infrastructure project in [2:36] the history [2:36] >> the um there was a great interview you [2:38] did many years ago in Strictly VC and [2:41] and sort of early open AI well before [2:43] TGBT and and they're saying hey what's [2:45] they're asking what's the business model [2:46] and you said oh we'll ask AI it'll [2:48] figure it out for us everybody laughs [2:49] but there have been multiple times and [2:51] there was just another one recently [2:53] where we have asked a then current model [2:55] for you know what should we do and it [2:56] has had a insightful answer we missed so [2:59] >> I I think when we say stuff like that [3:01] people don't take us seriously or [3:02] literally [3:03] >> but maybe the answer is you should take [3:05] take us both [3:07] >> yeah yeah well know as as somebody who [3:09] runs an organization I ask the AI a lot [3:12] of questions about what I should do. It [3:14] comes up with some pretty interesting [3:15] answers. [3:16] >> Sometimes sometimes it does you know you [3:18] have to you have to give it enough [3:19] context. But [3:20] >> what is what is the thesis that that [3:22] connects these bets beyond more [3:23] distribution more compute? How do how do [3:25] we think about it? [3:26] >> I mean the research enables us to make [3:28] the great products and the [3:29] infrastructure enables us to do the [3:30] research. So it is kind of like a [3:32] vertical stack of things like you can [3:36] use chatbt or some other service to get [3:39] advice about what you should do running [3:40] an organization but for that to work it [3:43] requires great research and requires a [3:44] lot of infrastructure. So it is kind of [3:46] just this one this one thing. It's [3:49] >> and do you think that there will be a [3:51] point where that becomes completely [3:55] horizontal or will it stay vertically [3:58] integrated for the foreseeable future? [4:01] I was always against vertical [4:04] integration and I now think I was just [4:06] wrong about that. [4:07] >> Yeah. Interesting. [4:08] >> And there's kind of cuz you you like [4:11] you'd like to think that the economy is [4:13] efficient and the theory that companies [4:14] can do one thing and then [4:16] >> it's supposed to work. [4:17] >> Like to think that. Yeah. [4:18] >> And in our case at least it hasn't [4:21] really. I mean it hasn't some ways for [4:23] sure. or like there's people that make [4:24] like you know [4:25] >> Nvidia makes an amazing chip or whatever [4:27] that a lot of people can use but the [4:30] >> the story of open AI has certainly been [4:32] towards we have to do more things than [4:34] we thought to be able to deliver on the [4:35] mission [4:36] >> right you know although that you know [4:38] the the history of the computing [4:39] industry is kind of been a story of kind [4:42] of a back and forth and that you know [4:44] there was the Wang word processor and [4:46] then the personal computer and the the [4:48] Blackberry before the smartphone. Um, so [4:52] you know there has been this kind of [4:53] vertical integration and then not but [4:55] then the iPhone is also vertically [4:58] integrated. [4:58] >> The iPhone I think is the most [5:00] incredible product the tech industry has [5:02] ever produced and it is extraordinarily [5:04] vertically integrated. [5:05] >> Yeah, amazingly so. Yeah. Interesting. [5:08] >> Which bets would you say are enablers of [5:09] AGI versus which are sort of hedges [5:11] against uncertainty? I [5:13] >> think you could say that on the surface [5:14] Sora for example does not look like it's [5:16] AGI relevant. But [5:18] I would bet that if we can build really [5:20] great world models, that'll be much more [5:22] important to AGI than people think. [5:25] There were a lot of people who thought [5:25] CH chatbt was not a very AGI relevant [5:27] thing. And it's been very helpful to us, [5:29] not only in building better models and [5:33] understanding how society wants to use [5:35] this, but also in like bringing society [5:36] along to actually figure out, man, we [5:39] got to contend with this thing. Now we [5:40] for a long time before CHGBT, we would [5:42] talk about AGI and people like this is [5:44] not happening or we don't care. [5:45] then all of a sudden they really cared [5:47] and and I I think that so [5:51] research benefits aside. [5:54] I'm a big believer that society and [5:56] technology have to co-evolve. It's you [5:58] can't just drop the thing at the end. It [6:00] doesn't work that way. It is it is a [6:01] sort of ongoing back and forth. Yeah. [6:04] say more about how Sora fits into your [6:06] strategy because there some hull on on X [6:08] around hey um you know why devote [6:10] precious GPUs to to Sora but is it a [6:13] short-term long-term trade-off or are we [6:15] so aging [6:15] >> well and then the new one had like a [6:18] very interesting twist with the social [6:20] networking [6:22] be very interested in kind of how you're [6:23] thinking about that and like um did uh [6:27] Meta call you up and get mad or like hey [6:29] what what do you expect the reaction to [6:31] be? Um, [6:33] I think if one company of the two of us [6:36] has feels like more like the other one [6:38] has gone after them, it wouldn't they [6:40] shouldn't be calling us. [6:41] >> Well, I do know the history, too. But uh [6:45] look, we're not going to like [6:49] first of all, I think it's cool to make [6:52] great products and people love the new [6:54] Sora and and I also think it is [6:58] important to [7:00] give society a taste of what's coming on [7:03] this co-evolution point. So like very [7:05] soon the world is going to have to [7:07] contend with incredible video models [7:08] that can deep fake anyone or kind of [7:10] show anything you want. that will mostly [7:12] be great. There will be some adjustment [7:14] that society has to go through. And just [7:16] like with chat GPT, we were like the [7:18] world kind of needs to understand where [7:19] this is. I think it's very important the [7:22] world understands where video is going [7:23] very quickly cuz that's going to be [7:26] video has much more like emotional [7:27] resonance than text and very soon we're [7:30] going to be in a world where like this [7:32] is going to be everywhere. So I think [7:33] there's something there. uh as I [7:35] mentioned I think this will help our [7:36] research program and is on the AGI path [7:39] but [7:40] yeah some like you know it can't all be [7:42] about just making people like ruthlessly [7:44] efficient and the AI like solving all [7:46] our problems there's got to be like some [7:47] fun and joy and delight along the way [7:49] but we won't throw like tons of compute [7:51] at it or not by a fraction of our comput [7:56] it it's tons in the absolute sense but [7:58] not in the relative sense [8:01] >> I want to talk about the future of AI [8:02] human interfaces because back in August [8:04] you said the models have already [8:05] saturated the chat use case. So what are [8:07] future AI human interfaces look like [8:09] both in terms of hardware and software [8:11] is the vision for kind of a wechathat [8:13] like [8:13] >> so I'm solving the chat thing in like a [8:16] very narrow sense which is if you're [8:18] trying to like you know have the most [8:20] basic kind of chat style conversation [8:22] it's very good but what a chat interface [8:24] can do for you it's like nowhere near [8:25] saturated cuz you could ask a chat [8:27] interface like please cure cancer a [8:29] model certainly can't do that yet so I [8:32] think the text interface style can go [8:34] very far even if for the chitchat use [8:36] case the models are already very good. [8:38] Um, but but of course there's better [8:41] interfaces to have. Uh, actually it's [8:44] another thing I think is cool about [8:45] Sora. Like you can imagine a world where [8:47] the interface is just constantly [8:49] real-time rendered video. [8:51] >> Yeah. [8:51] >> And what that would enable and that's [8:53] pretty cool. You can imagine new kinds [8:54] of hardware devices that are sort of [8:57] always ambiently aware of what's going [8:59] on. in rather than your phone like blast [9:01] you with text message notifications at [9:03] whenever it wants like it really [9:05] understands your context and when to [9:06] show you what and [9:09] there's a long way to go on all that [9:10] stuff you know [9:12] within the next couple years what will [9:14] models be able to do that they're not [9:16] able to do today will be sort of white [9:18] color um you know replacement at a much [9:20] deeper level AI scientist uh human [9:23] humanoids um [9:25] >> I mean a a lot of things but you touched [9:27] on the one that I am most excited about [9:29] which is the the AI scientist. [9:31] >> Yeah, [9:31] >> this is crazy that we're sitting here [9:33] seriously talking about this. The I know [9:37] there's like a quibble on what the [9:38] Turing test literally is, but but the [9:40] popular conception of the Turing test [9:43] sort of went whooshing by. [9:44] >> Yeah, it was fast. Yeah. [9:46] >> You know, it was just like we talked [9:47] about it as this most important test of [9:48] AI for a long time. It seemed impossibly [9:51] far away. Then all of a sudden it was [9:53] passed. the world freaked out for like a [9:56] week, two weeks, and then it's like, [9:58] "All right, I guess computers like can [10:00] do that now." Yeah. [10:01] >> And everything just went on. And I think [10:02] that's happening again with science. Uh [10:06] my own personal like equivalent of the [10:08] touring test has always been when AI can [10:10] do science like that has always like [10:12] that is a real change to the world. And [10:14] for the first time with GPT5, we are [10:15] seeing these little little examples [10:17] where it's happening. You see these [10:19] things on Twitter. did this it made this [10:20] novel math discovery and did this small [10:22] thing in my you know my physics research [10:24] my biology research and everything we [10:27] see is that that's going to go much [10:29] further so in 2 years I think the models [10:31] will be doing bigger chunks of science [10:33] and making important discoveries [10:36] and that is a crazy thing like that will [10:38] have a significant impact on the world I [10:40] am I am a believer that to a first order [10:43] scientific progress is what makes the [10:45] world better over time and if we're [10:46] about to have a lot more of that that's [10:48] a big [10:50] It's interesting because that's a [10:51] positive change that people don't [10:54] talk about. It it's gotten so um much [10:59] into the realm of the negative changes [11:01] if AI gets extremely smart. But [11:03] >> but carbon disease is [11:05] >> who we could use a lot more science. [11:06] Yeah. That that that's really good [11:08] point. I think Alan Turing said this, [11:11] somebody asked him, they said, "Well, do [11:12] you really think the uh computer is [11:14] going to be, you know, smarter than the [11:17] brilliant minds?" He said, "It doesn't [11:18] have to be smarter than a brilliant [11:19] mind, just smarter than a mediocre mind [11:21] like the president of AT&T." [11:24] And uh we could use more of that, too. [11:27] Probably. [11:27] >> We uh we just saw periodic launch last [11:29] week. You know, Open AAI lungs. And uh [11:32] yeah, to to to that point, it's amazing [11:33] to see both the innovation that you guys [11:34] are doing, but also the the teams that [11:37] you know come out of OpenAI just feels [11:38] like are you know, creating tremendous [11:40] capable things. [11:41] >> We certainly hope so. [11:42] >> Yeah. the um I want to ask you about [11:45] just broader reflections in terms of [11:47] what sort of about diffusion or uh [11:51] development in 2025 has surprised you or [11:54] what has sort of updated your worldview [11:55] since chatb came out. [11:59] A lot of things again but maybe the most [12:03] interesting one is how much new stuff we [12:05] found. sort of thought we had like [12:07] stumbled on this one giant secret that [12:09] we had these scaling loss for language [12:10] models and that felt like such an [12:13] incredible [12:14] triumph that I was like we're probably [12:18] never going to get that lucky again and [12:20] deep learning has been this miracle that [12:22] keeps on giving and we have kept finding [12:25] like breakthrough after breakthrough [12:29] again when we got the the reasoning [12:31] model breakthrough like I I also thought [12:33] that was like we're never going to get [12:34] another one like uh [12:37] uh and it just seems so improbable that [12:39] this one technology works so well. But [12:41] maybe this is always what it feels like [12:43] when you discover like [12:45] one of the big, you know, scientific [12:48] breakthroughs is it if if it's like [12:49] really big, it's pretty fundamental and [12:51] it just it keeps working. But the amount [12:55] of progress, [12:57] like if you went back and used GPT3.5 [13:00] from Chat GBT launch, you'd be like, I [13:02] cannot believe anyone used this thing. [13:04] >> Yeah. [13:05] And and now we're in this world where [13:07] the capability overhang is so immense [13:09] like most of the world still just thinks [13:11] about what chat PT can do and then you [13:13] have like some nerds in Silicon Valley [13:14] that are using codecs and they're like [13:15] wow those people have no idea what's [13:17] going on and then you have like a few [13:18] scientists who say those people using [13:20] codecs have no idea what's going on but [13:22] the the overhang of capability has come [13:24] is is is so big now and we've just come [13:26] so far on the what the models can do. [13:29] And in terms of further development, how [13:32] far can we get with with LLMs? At what [13:34] point do we need either new architecture [13:35] or how do you think about what [13:36] breakthroughs are needed? [13:37] >> I think far enough that we can make [13:40] something that will figure out the next [13:42] breakthrough with the current technology [13:44] like I it's a very self-reerential [13:47] answer, but if if LLMs can get if LLM [13:50] based stuff can get far enough that it [13:52] can do like better research than all of [13:55] Open put together, maybe that's like [13:57] good enough. [13:58] Yeah, that would be a big breakthrough. [14:01] A very big breakthrough. So, on um on [14:03] the more mundane, you know, one of the [14:05] things that uh people have kind of [14:08] started to complain about, I think South [14:10] Park did a whole episode on it, is kind [14:12] of the obsequiousness [14:14] of uh of kind of AI and chat GPT in [14:18] particular. And how hard a problem is [14:20] that to deal with? Is it not that hard [14:22] or is it like kind of a fundamentally [14:24] hard problem? [14:25] >> Oh, it's not at all hard to deal. a lot [14:26] of users really want it. [14:28] >> Yeah. [14:28] >> Like if you go look at what people say [14:30] about Chach online, [14:32] >> there's a lot of people who like really [14:33] want that back. [14:35] >> And it is, you know, [14:37] >> so it's not technically it's not hard to [14:38] deal with at all. Um, one thing, and [14:41] this is not surprising in any way, but [14:45] the the incredibly wide distribution of [14:50] what users want. [14:52] >> Yeah. out of how of like how they'd like [14:54] a chatbot to behave in big and small [14:56] ways. [14:57] >> Does that do you end up having to [15:00] configure the personality then you [15:02] think? Is that going to be the answer? [15:03] >> I think so. Uh I mean ideally like you [15:07] just talk to chatt for a little while [15:09] and it kind of interviews you and also [15:10] sort of sees what you like and don't [15:12] like and [15:12] >> and chat just figures out but in the [15:15] short term you'll probably just pick [15:16] one. [15:17] >> Got it. Yeah, that makes sense. Very [15:20] interesting. And um actually so so one [15:23] thing I wanted to ask you about is uh [15:26] >> like I think we just had a a really [15:30] naive thing which you you know like [15:33] it would sort of be unusual to think you [15:35] can make something that would talk to [15:36] billions of people and everybody wants [15:37] to talk to the same [15:39] >> person. Yeah. [15:40] >> And and yet that was sort of our [15:41] implicit assumption for a long time. [15:44] >> Right. Because people have very [15:45] different friends. [15:45] >> People have very different friends. So [15:46] now we're trying to fix that. [15:48] >> Yeah. and also kind of different [15:50] friends, different interests, different [15:51] uh levels of intellectual capability. [15:54] So, you don't really want to be talking [15:56] to the same thing all the time. And one [15:58] of the great things about it is you can [15:59] say, "Well, explain it to me like I'm [16:00] five, but maybe I don't even want to [16:03] have to do that prompt. Maybe I always [16:04] want you to talk to Yeah. Particularly [16:07] if you're teaching me stuff." [16:08] Interesting. Um, I want to ask you a [16:11] kind of like a a CEO question which has [16:13] been interesting for me to observe you [16:16] is you just did this deal with AMD. Um, [16:20] and you know, of course, the company's [16:21] in a different position and you have [16:23] more leverage and these kinds of things, [16:24] but like how has your kind of thinking [16:26] changed over the years since you did [16:28] that that initial deal if at all? [16:31] >> I I had very little operating experience [16:33] then. I had very little experience [16:35] running like I I am I am not naturally [16:37] someone to run a I'm a great fit to be [16:39] an investor [16:40] >> and I kind of thought that was going to [16:41] be that was what I did before this and I [16:42] thought that was going to be my career. [16:43] >> Yeah. Yeah. [16:44] >> Although you were a CEO before that. [16:46] >> I not a good one. Um and [16:51] and so I think I had the mindset of like [16:56] an investor advising a company when we [16:58] did and now I understand what it's like [17:00] to actually have to run a company. [17:01] >> Yeah. Right. Right. Right. There there's [17:03] more than I I've learned a lot about how [17:06] to [17:07] >> you know like how you have to like [17:10] >> what what operational [17:12] >> how you like what it takes to [17:13] operationalize deals over time and [17:14] >> right [17:15] >> all the implications of the agreement as [17:17] opposed to just oh we're going to get [17:19] distribution of money. Yeah, that makes [17:21] sense. Yeah. Know because it it's really [17:23] I I I just I was very impressed at the [17:29] deal structure improvement. [17:31] >> Yeah. Right. [17:33] >> More broadly, you've, you know, in the [17:34] last few weeks alone, you mentioned AMD, [17:35] but also Oracle, Nvidia. You've chosen [17:38] to strike these deals and partnerships [17:40] with with companies that you collaborate [17:42] with, but could also potentially compete [17:43] with in in certain areas. How do you [17:45] decide you know when to collaborate [17:47] versus when when not to or how do you [17:49] just think about [17:51] >> um we have decided that it is time to go [17:54] make a very aggressive infrastructure [17:56] bet and we're like I've never been more [18:01] confident in the research road map in [18:03] front of us and also the economic value [18:04] that will come from using those models [18:06] but to make the bet at this scale we [18:08] kind of need the whole industry to or [18:10] big chunk of the industry to support it [18:13] and this is like you know from the level [18:15] of like electrons to model distribution [18:17] and all the stuff in between which is a [18:19] lot and so we're going to partner with a [18:22] a lot a lot of people. Uh you should [18:24] expect like much more from us in the [18:26] coming months. [18:27] >> Actually expand on that because when you [18:29] talk about the scale it does feel like [18:33] in your mind [18:35] the the limit on it is unlimited like [18:37] you would scale it as as you know as big [18:40] as you possibly could. [18:43] There's totally a limit like there's [18:46] some amount of global GDP. Uh [18:48] >> yeah, [18:49] >> you know, there's some fraction of it [18:50] that is knowledge work and we don't do [18:52] robots yet. [18:53] >> Yes. [18:53] >> But [18:54] >> but but the limits are out there. [18:56] >> It feels like the limits are very far [18:57] from where we are today. If we are right [19:00] about [19:01] >> so so I shouldn't say from where like if [19:04] we are right that the model capability [19:06] is going to go where we think it's going [19:07] to go then the economic value that sits [19:10] there [19:12] can can go very very far [19:14] >> right so you wouldn't do it like if all [19:17] you ever had was today's model you won't [19:18] go there but it's a combination [19:21] >> I mean we would still expand because we [19:24] can see how much [19:26] >> demand there is we can't serve with [19:28] today's model but We would not be going [19:30] this aggressive if all we had was [19:31] today's model, [19:32] >> right? [19:33] >> Yeah. [19:33] >> Right. We get to see a year or two in [19:34] advance though. So like [19:36] >> Yeah. Interesting. [19:38] >> Chad view 800 million weekly active [19:41] users about 10% of the world world's [19:44] population fastest growing consumer [19:47] product you know ever it seems. Um how [19:49] do [19:50] >> faster than anyone I ever saw. [19:52] >> Yeah. How how do you balance you know [19:56] optimizing for active users at well at [19:58] the same time being a re you know being [19:59] a product company and and a research [20:01] company how do you throw the new [20:03] >> when there's a constraint we almost like [20:05] which happens all the time uh we almost [20:07] always prioritize giving the GPUs to [20:09] research over supporting the product um [20:12] part of the reason we want to build this [20:13] capacity so we don't have to make such [20:15] painful decisions there are weird times [20:17] you know like a new feature launches and [20:19] it's going really viral or whatever [20:21] where research will temporarily [20:22] sacrifice some GPUs, but but on the [20:25] whole like we're here to build AGI [20:27] >> and research gets the priority. [20:29] >> Yeah. the you said in your your [20:32] interview with with your brother Jack [20:33] around how you know other companies can [20:36] try to imitate the the products or or [20:38] buy your you know or hire your your your [20:42] >> higher IP [20:43] all sorts of things but but they they [20:45] can't buy the culture or they can't the [20:48] sort of repeatable sort of you know m [20:51] machine if you will that that is you [20:53] know constantly the culture of [20:54] innovation [20:56] how have you done that or what are you [20:58] playing [21:00] what talk about this this culture of of [21:02] innovation. [21:02] >> This was one thing that I think was very [21:04] useful about coming from an investor [21:06] background. A really good research [21:08] culture looks much more like running a [21:10] really good seedstage investing firm and [21:12] betting on founders and sort of that [21:15] kind of than it does like running a [21:16] product company. So I think having that [21:19] experience was really helpful to the [21:21] culture we built. [21:22] >> Yeah. Yeah. That's sort of how I see, [21:25] you know, Benedict in some ways which [21:27] we, you know, you're a CEO but also [21:28] have, you know, have this portfolio and, [21:30] you know, have an investor mindset, [21:31] >> right? Like I'm the opposite. [21:33] >> CEO going to investor. He's investor [21:35] going to CEO. [21:36] >> It is unusual in this direction. [21:38] >> Yeah. Yeah. [21:39] >> Yeah. Well, it never works. You're the [21:40] only one who I think I've seen go that [21:45] way and have it work. [21:46] >> Uh, [21:49] workday was like that, right? No, but [21:51] Anne Neil was he he was a operator [21:54] before he was an investor and I mean he [21:57] was really an operator. I mean people [21:58] soft is a pretty big [21:59] >> and why is that because once people are [22:00] investors they don't want to operate [22:01] anymore. Um, no. I think that investors [22:08] generally if you're good at investing, [22:10] you're not necessarily good at like [22:14] organizational dynamics, conflict [22:16] resolution, um, you know, like just like [22:20] the deep psychology of like all the [22:23] weird and then you know how [22:25] politics get created. There's just like [22:27] all this [22:29] there. There's the detailed work in [22:31] being an operator or being a CEO is [22:35] so vast and it's not as intellectually [22:39] stimulating. It's not something you can [22:40] ever go talk to somebody at a cocktail [22:42] party about. And so like you're an [22:44] investor, you get like, oh, everybody [22:45] thinks I'm so smart and you know cuz you [22:47] know everything. You see all the [22:49] companies and so forth and that's a good [22:51] feeling. And then being CEO is often a [22:54] bad feeling. Yeah. [22:55] >> And so it's really hard to go to a a [22:56] good feeling to a bad feeling. I would [22:58] just say [22:59] >> I'm shocked by how different they are [23:00] and I'm shocked by how much the [23:02] difference between a good job and a bad [23:03] job they are. [23:04] >> Yeah. [23:04] >> Yes. [23:05] >> Yeah. You know, it's tough. It's it's [23:06] rough. I mean, I can't even believe I'm [23:08] running the firm. Like I know better. [23:10] >> Yeah. [23:11] >> And he can't believe he's running [23:12] OpenAI. He knows better. [23:14] >> Going back to progress today, are you [23:16] still useful in a world in which they're [23:18] getting saturated, gained? Are they [23:19] still the What is the best way to gauge [23:21] model capability now? Um, well, we're [23:24] talking about scientific discovery. I [23:25] think that'll be an eval that can go for [23:27] a long time. [23:28] >> Revenue is kind of an interesting one. [23:30] Uh, but I think the like static evals of [23:33] benchmark scores are less interesting. [23:35] >> Yeah. [23:36] >> And and also those are crazily gamed. [23:38] >> Yeah. Yeah. [23:40] >> More broadly, it seems like [23:42] >> that's all there is as far as I can [23:44] tell. Yeah. More broadly, it seems that [23:46] the culture the culture Twitter is less [23:49] AGI pill than it was a year or so ago [23:52] when the AI 2027 thing came out. Some [23:55] people point to you GBT5 them not seeing [23:58] sort of the obvious um obviously there [24:00] were a lot of progress that in some ways [24:02] under the the surface or not not as [24:04] obvious to what people were expecting. [24:05] Should people be less AGI pled or is [24:08] this just Twitter vibes? And [24:13] >> well, a little bit of both. I mean, I I [24:14] I think like [24:16] >> like we talked about the touring test, [24:17] AGI will come. [24:19] >> It will go whooshing by. [24:21] >> The world will not change as much as the [24:22] impossible amount that you would think [24:24] it should. [24:25] >> It won't actually be the singularity. [24:27] >> It will not. [24:28] >> Yeah. [24:29] >> Yeah. Even even if it's like doing kind [24:32] of crazy a research like the society [24:33] will learn faster but [24:38] one of the kind of like retrospective [24:40] observations is people and societies as [24:42] a whole are just so much more adaptable [24:44] than we think that you know it was like [24:47] a big update to think that AGR was going [24:50] to come. You kind of go through that. [24:52] You need something new to think about. [24:54] You make peace with that. It turns out [24:55] like it will be more continuous than we [24:57] thought. [25:00] which is good. [25:01] >> Which is really good. [25:03] >> I'm not up for the big bang. [25:04] >> Yeah. Um well to that end, how have you [25:08] sort of evolved your thinking? You [25:09] mentioned you evolved your thinking on [25:10] sort of uh you know vertical [25:12] integration. How have you evolved your [25:13] thinking or what's the latest thinking [25:14] on sort of AI stewardship you safety? [25:18] What's the latest thinking on that? [25:21] I do still think there are going to be [25:23] some [25:28] really strange or scary moments. Uh [25:31] the fact that like so far the technology [25:34] has not [25:37] produced a really scary giant risk [25:39] doesn't mean it never will. It also like [25:42] there's we're talking about it's kind of [25:44] weird to have like billions of people [25:45] talking to the same brain. like there [25:48] may be these weird societal scale things [25:50] that are already happening we that [25:52] aren't scary in the big way but are just [25:54] sort of different. Um [25:58] but I expect like [26:02] I expect some really bad stuff to happen [26:04] because of the technology which also has [26:05] happened with previous technologies and [26:08] >> all the way back to fire. [26:10] >> Yeah. [26:12] And I think we'll like develop [26:16] some guardrails around it as a as a [26:17] society. [26:18] >> Yeah. What is sort of your latest [26:20] thinking on the the right mental models [26:22] we should have around the the right [26:23] regulatory frameworks to to think about [26:25] or or the ones we shouldn't be thinking [26:27] about? Um [26:31] I think most [26:36] I think the right thing to I I think [26:37] most regulation uh [26:42] probably has a lot of downside. The one [26:43] thing I would like is as the models get [26:46] the thing I would most like is as the [26:47] models get truly like extremely [26:50] superhuman capable. Um, [26:54] I think those models and only those [26:56] models are probably worth some sort of [26:58] like [27:00] very careful safety testing uh as as the [27:03] frontier pushes back. Um, I don't want a [27:06] big bang either. Mhm. [27:07] >> And you can see a bunch of ways that [27:10] could go very seriously wrong. But [27:15] I hope we'll only focus the regulatory [27:18] burden on that stuff and not all of the [27:20] wonderful stuff that less capable models [27:22] can do that you could just have like a [27:24] European style complete cramp on and [27:27] that would be very bad. Yeah, it seems [27:29] like the [27:32] the thought experiment that okay, [27:34] there's going to be a model down the [27:36] line that is a super superhuman [27:39] intelligence that could, you know, do [27:42] some kind of takeoff flight thing. [27:45] We really do need to wait till we get [27:47] there. Uh um or like at least we get to [27:49] a much bigger scale or we get close to [27:51] it. Um because [27:54] nothing is going to pop out of your lab [27:56] in the next week that's going to do [27:58] that. And I I think that's where we as [28:02] an industry kind of confuse the [28:03] regulators. Yeah. uh because I think you [28:06] you really could one you damage America [28:10] in particular in that um like China's [28:13] not going to have that kind of [28:13] restriction and and you getting behind [28:18] um in AI I think it'd be very dangerous [28:21] for the world [28:22] >> extremely dangerous [28:23] >> extremely dangerous [28:24] >> much more dangerous than not regulating [28:26] something we don't know how to do yet [28:29] >> you also want to talk about copyright [28:32] >> um [28:33] yeah So, well then that that's a segue, [28:36] but um [28:38] when you think about well I guess how do [28:41] you see copy right unfolding because [28:43] you've done some very interesting things [28:46] um with the opt out uh and [28:50] you know as you see people selling [28:52] rights do you think will they be be [28:54] bought exclusively will they be just [28:57] like um I could sell it to everybody [28:59] wants to pay me or how do you think [29:01] that's going to unfold [29:02] >> this is my current guess It it [29:06] speaking of that like society and [29:07] technology co-evolve as the technology [29:09] goes in different directions and we saw [29:11] an example of a different like video [29:13] models got a very different response [29:15] from rights holders than image gen does. [29:17] So like you'll see this continue to move [29:20] >> but forced guess from the position we're [29:24] in today. I would say that society [29:26] decides training is fair use. [29:30] >> Mhm. But [29:32] there's a new model for generating [29:36] content in the style of or with the IP [29:39] of or something else. [29:40] >> So you know anyone can read like a human [29:44] author can anybody can read a novel and [29:46] get some inspiration but you can't [29:47] reproduce the novel in your own [29:49] >> right [29:49] >> and can talk about Harry Potter but you [29:52] can't re spit it out. [29:54] >> Yes. Although another thing that I think [29:57] will change um [30:01] >> in the case of Sora, we've heard from [30:05] a lot of concerned rights holders and [30:06] also a lot of [30:07] >> name and like [30:08] >> and a and a lot of rights holders who [30:10] are like my concern is you won't put my [30:12] character in enough. [30:14] >> Yeah. I want restrictions for sure, but [30:16] like if I'm, you know, whatever and I [30:18] have this character, like I don't want [30:20] the character to say some crazy [30:22] offensive thing, but like I want people [30:24] to interact. Like that's how they [30:25] develop the relationship and that's how [30:26] like my franchise gets more valuable. [30:28] And if you become really if you're [30:30] picking like his character over my [30:31] character all the time, like I don't [30:33] like that. So, I can completely see a [30:36] world where [30:39] subject to the decisions that a rights [30:41] holder has, they get more upset with us [30:43] for not generating their character often [30:45] enough than too much. [30:47] >> Yeah. [30:48] >> And this is like this was not an obvious [30:50] thing that recently that this is how it [30:52] might go. But [30:54] >> yeah, this is such an interesting thing [30:56] with kind of Hollywood. we saw this like [30:59] one of the things that I never quite [31:02] understood about the music business was [31:04] how like you know okay you have to pay [31:07] us if you play the song in a restaurant [31:09] or like at a game or this and that and [31:11] the other and they they get very [31:12] aggressive with that. um when it's [31:15] obviously a good idea for them to play [31:17] your song at a game because that's the [31:19] biggest advertisement in the world for [31:20] like all the things that you do, your [31:22] concert, your your [31:23] >> Yeah, that one felt really irrational. [31:25] >> Like um but it I I would just say it's [31:29] it's very possible for the industry just [31:32] because the way those industries are [31:33] organized or at least the traditional [31:35] creative industries to do something [31:38] irrational. Um, and it comes from like [31:41] in the music industry. I think it came [31:43] from the structure where you have the [31:44] publisher who's just, [31:46] >> you know, basically after everybody. Uh, [31:48] you know, that their whole job is to [31:50] stop you from playing [31:51] >> the music. [31:52] >> Yeah. [31:53] >> Which every artist would want you to [31:54] play. Uh, so [31:57] >> I I do wonder how it's going to shape [31:58] out. I agree with you that the rational [32:01] idea is I want to let you use it all you [32:05] want and I want you to use it but um [32:08] that don't mess up my character. Yeah. [32:10] >> So so I think like [32:12] >> if I had to guess [32:15] some people will say that some people [32:17] say absolutely not but it doesn't have [32:19] the music industry like [32:21] >> thing of just a few people with all of [32:22] the [32:23] >> right it's more dispersed [32:24] >> and so people will just try many [32:26] different setups here and see what [32:27] works. [32:27] >> Yeah. And maybe it's a way for new [32:29] creatives to get new characters out. [32:31] >> Yeah. [32:31] >> And you'll never be able to use Daffy [32:33] Decker. [32:35] >> I want to chat about open source. Um [32:37] because there's been some evolution in [32:39] the thinking too and that GBD3 didn't [32:41] have the open open weights, but you [32:42] released a you know very capable open [32:44] model earlier this year. What's sort of [32:46] your your latest thinking. What was the [32:47] evolution there? [32:48] >> I think open source is good. I Yeah. I [32:51] mean I'm happy like it makes me really [32:53] happy that people really like GPOSS. [32:55] Yeah. [32:56] >> Yeah. [32:57] And what do you think like strategically [32:59] like what's the danger [33:01] of [33:03] deepseek being the dominant open source [33:05] model? [33:06] >> I mean who knows what people will put in [33:08] these open source models over time like [33:10] >> like what the weights will actually be. [33:12] Yeah. [33:13] >> It's really hard to [33:14] >> So you're seeding control of the [33:17] interpretation of everything to [33:19] somebody. [33:20] >> Yeah. [33:21] >> Who may be or may not be influenced [33:24] heavily by the Chinese government. Yeah. [33:28] What about And [33:28] >> by the way, we see I mean, you know, [33:29] just to give you and and we really thank [33:32] you for um putting out a really good [33:34] open source model because what we're [33:36] seeing now is in all the universities, [33:38] they're all using the Chinese models. [33:39] >> Y [33:40] >> Yeah. Which feels very dangerous. [33:45] >> You've said that the things you care [33:46] most about professionally are AI and [33:49] energy. [33:50] >> I did not know they were going to end up [33:52] being the same thing. They were two [33:53] independent interests that really [33:54] converged. [33:55] >> Yeah. [33:57] talk more about how your interest in [33:58] energy uh sort of began, how you sort of [34:01] chosen to to play in it and then we [34:02] could talk about Yeah. how they care, [34:03] >> right? Because you started your career [34:04] in physics. Yeah. [34:06] >> CS and physics. [34:07] >> Yeah. [34:08] >> Uh well, I never really had a career. I [34:09] studied physics and my first job was [34:11] like a CS like [34:13] >> this is an oversimplification, but [34:16] roughly speaking, I I think if you look [34:18] at history, the best the highest impact [34:21] thing to improve people's quality of [34:22] life has been cheaper and more abundant [34:24] energy. [34:26] And so it seems like pushing that much [34:28] further is a good idea. And I I don't [34:31] know. I just like people have these [34:32] different lenses. They look at the [34:33] world, but I I see energy everywhere. [34:35] >> Yeah. [34:37] >> Yeah. And so [34:40] get into because we've kind of uh in the [34:42] west I think we've uh paint ourselves [34:45] into a little bit of a corner on energy [34:48] um by both outlying nuclear for a very [34:52] long time. [34:52] >> That was an incredibly dumb decision. [34:53] >> Yeah. And then you know like also a lot [34:57] of policy restrictions on energy um and [35:00] you know worse so in Europe than in the [35:02] US but also dangerous here and now with [35:04] AI here [35:06] it feels like we're going to need all [35:10] the energy from every possible source [35:12] and how do you see that developing kind [35:15] of policy-wise and technologically like [35:17] what are going to be the big sources and [35:19] how will those kind of curves cross um [35:22] and then what's the right policy posture [35:25] around you know drilling fracking all [35:28] these kinds of things I expect in the [35:30] short term it will be most of the net [35:32] new in the US will be natural gas [35:34] relative to at least base load energy in [35:37] the long term I expect it'll be a [35:40] >> I don't know what the ratio but the two [35:41] dominant sources will be uh solar plus [35:45] storage and nuclear [35:46] >> I think yeah [35:48] >> some combination of those two will win [35:49] in the future like the long-term future [35:50] >> in the long term right now [35:52] >> and advanced nuclear [35:54] SMRS fusion the whole the whole stack. [35:56] >> And how how fast do you think that's [35:58] that's coming on the nuclear side where [36:02] we're at really at scale cuz you know [36:04] obviously there's a lot of people [36:05] building it. [36:06] >> Yeah. [36:07] >> Um but we we have to completely legalize [36:09] it and all that kind of thing. [36:10] >> I I think it kind of depends on the [36:13] price. If it is completely crushingly [36:16] economically dominant over everything [36:18] else [36:19] >> then I expect to happen pretty fast. [36:20] Yeah. Again, if you like study the [36:22] history of energy, when you have these [36:24] major transitions to a much cheaper [36:26] source, the world moves over pretty [36:29] quickly. The cost of energy is just so [36:31] important. [36:32] >> Yeah. So, if [36:35] if nuclear gets radically cheap relative [36:38] to anything else we can do, I'd expect [36:40] there's a lot of political pressure to [36:42] get the NRC to move quickly on it, and [36:44] we'll find a way to build it fast. If [36:46] it's around the same price as other [36:47] sources, I expect the kind of [36:49] anti-uclear sentiment to overwhelm and [36:51] it take a really long time. [36:52] >> Yeah. [36:54] >> Should be cheaper. [36:56] >> It should be. [36:56] >> Yeah. [36:57] >> Yeah. [36:58] >> It should be the cheapest form of energy [36:59] on Earth like or anyway. [37:01] >> Yeah. Yeah. Cheap, clean. [37:04] >> What's it not to like? [37:05] >> Apparently a lot. [37:07] >> Yeah. on open. What's what's the latest [37:09] thinking in terms of monetization in [37:11] terms of either certain experiments or c [37:12] certain things that you could see [37:13] yourself spending more time or less less [37:15] time on different models that you're [37:18] excited about? The thing that's top of [37:20] mind for me like right now just cuz it [37:22] just launched and there's so much usage [37:24] is what we're going to do for Sora. [37:25] >> Yeah. [37:26] >> Um [37:30] another thing you learn once you launch [37:31] one of these things is how people use [37:34] them versus how you think they're going [37:35] to use them. [37:36] >> Yeah. And people are certainly using [37:37] Sora the ways we thought they were going [37:39] to use it, but they're also using it in [37:40] these ways that are very different. Like [37:42] people are generating funny memes of [37:44] them and their friends and sending them [37:45] in a group chat. And that will require a [37:49] very different like sore videos are [37:51] expensive to make. Uh [37:53] >> right. So that will require a very [37:55] different, you know, for people that are [37:56] doing that like hundreds of times a day. [37:58] >> It's going to require a very different [37:59] monetization method than the kinds of [38:00] things we were we were thinking about. I [38:01] think it's very cool that the thesis of [38:03] Sora, which is people actually want to [38:05] create a lot of content, it's it's not [38:07] that, [38:08] you know, the traditional naive thing [38:10] that it's like 1% of users create [38:12] content, 10% leave comments, and 100% [38:14] view. Maybe a lot more want to create [38:16] content, but it's just been harder to [38:18] do. And I think that's a very cool [38:20] change. But it does mean that we got to [38:22] figure out a very different monetization [38:24] model for this than we were thinking [38:25] about. If people want to create that [38:27] much, I assume it's like some version of [38:28] you have to charge people per [38:30] generation. per generation when when [38:32] when it's this expensive. Um, but that's [38:35] like a new thing we haven't had to [38:36] really think about before. [38:37] >> What's your thinking on ads for the [38:39] longtail? [38:43] >> Open to it. I like many other people, I [38:48] find ads somewhat distasteful, but not [38:51] not a non-starter. Um, and there's some [38:54] ads that I like, like one thing I give [38:55] Meta a lot of credit for is Instagram [38:58] ads are like a net value ad to me. Um, I [39:02] like Instagram ads. I've never felt [39:04] that. Like, you know, on on Google, I [39:07] feel like I know what I'm looking for. [39:10] The first result is probably better. The [39:12] ad is an annoyance to me. On Instagram, [39:14] it's like, I didn't know I want this [39:16] thing. It's very cool. I never heard it, [39:18] but I never would have thought to search [39:19] for it. I want the thing. So that's like [39:21] there's kinds of things like that. But [39:25] >> people have a very high trust [39:27] relationship with Chhat GPT. Even if it [39:28] screws up, even if it hallucinates, even [39:30] if it gets it wrong, people feel like it [39:32] is trying to help them and that it's [39:34] trying to do the right thing. And is if [39:35] we broke that trust, it's like you say, [39:38] "What coffee machine should I buy?" And [39:40] we recommended one and it was not the [39:42] best thing we could do, but the one we [39:44] were getting paid for, that trust would [39:46] vanish. So like that kind of ad does not [39:48] does not work. There are others that I [39:50] imagine that could work totally fine. [39:53] Um, but that would require like a lot of [39:56] care to avoid the obvious traps. [39:58] >> Yeah. [40:00] M and then how how big a problem you [40:04] know just you extending the Google [40:07] example is like um you know fake uh [40:12] content that then gets slurped in by the [40:15] model and then they recommend the wrong [40:16] coffee maker because somebody just [40:18] blasted a thousand great reviews of that [40:21] coffee maker. [40:23] >> So there's all of these things that have [40:24] changed very quickly for us. [40:26] >> Yeah. Um, this is one of those examples [40:29] that people are doing these crazy things [40:30] to maybe not even fake reviews, but just [40:32] paying a bunch of like human like really [40:35] trying to figure out [40:35] >> are using chat GPT to write some good [40:37] ones. [40:38] >> Uh, [40:38] >> write me a review that chat GBT would [40:40] love. [40:41] >> Yeah. [40:42] >> So, this coffee. [40:43] >> Exactly. Exactly. [40:44] >> Yeah. [40:44] >> So, this is a very sudden shift that has [40:47] happened. [40:48] >> Mhm. [40:49] >> We never used to hear about this like [40:52] 6 months ago or 12 months ago. [40:54] >> Yeah. [40:55] >> Certainly. And now there's like a real [40:57] cottage industry that feels like it's [40:59] sprouted up overnight. [41:00] >> Yeah. [41:00] >> Trying to do this. [41:02] >> Yeah. Yeah. Yeah. No, they they're very [41:04] clever out there. [41:04] >> Yeah. So, uh I don't know how we're [41:07] going to fight it yet, but people figure [41:08] this out. [41:09] >> So, that gets into a little bit of this [41:11] other thing that we've been worried [41:12] about. Um and you know, we're trying to [41:16] kind of figure out uh blockchain sort of [41:19] potential solutions to it and so forth. [41:21] But there's this problem where like the [41:24] incentive to create content on the [41:26] internet used to be, you know, people [41:29] would come and see my content and they'd [41:31] read like, you know, if I write a blog, [41:32] people will read it and so forth. Um, [41:35] with chat GPT, [41:37] if I'm just asking Chat GPT and I'm not [41:40] like going around the internet, who's [41:42] going to create the content and why? Um, [41:45] and is there [41:47] an incentive theory or or or something [41:50] that you have to kind of not break the [41:53] covenant of the internet, which is like [41:55] I create something and then I'm rewarded [41:58] for it with like either attention or [42:00] money or something. [42:03] Uh, the theory is much more of that will [42:05] happen if we make content creation [42:08] easier and don't break the like kind of [42:10] fundamental way that you can get some [42:11] kind of reward for doing so. [42:13] >> So, for the dumbest example of Sora [42:15] since we've been talking about that. [42:18] >> It's much easier to create a funny video [42:20] than it's ever been before. [42:21] >> Yeah. [42:22] >> Um [42:23] maybe at some point you'll get a rev [42:25] share for doing so. [42:26] >> For now, you get like internet likes [42:28] which are still very motivating to some [42:30] people. [42:30] >> Yeah. [42:30] >> Um but people are creating tons more [42:34] than they ever created before in this [42:36] con in any other kind of like video app. [42:38] >> Yeah. [42:38] >> So, but are that the end of text? [42:43] I don't think so. Like people are also [42:46] >> are human generated texts. [42:48] >> Uh human generated will turn out to be [42:51] like you have to [42:52] >> you have you have to verify like what [42:54] percent? Yeah. Is it like fully [42:56] handcrafted? Was it like tool aated? [42:58] >> Yeah. I see. Yeah. Probably nothing that [43:00] toolated. Yeah. [43:02] >> Interesting. [43:03] >> We've uh we've given meta their flowers. [43:05] So now I can feel like I can ask you [43:06] this question which is the great talent [43:08] war hall of 2025 has has has taken place [43:12] and open AAI remains intact. Team is [43:15] strong as ever shipping incredible [43:17] products. [43:19] >> What can you say about what what's it [43:20] been like this year in terms of just [43:22] everything that's that's been going on. [43:23] I mean, every year has been exhausting [43:25] since we like uh [43:27] >> I [43:31] remember when [43:34] the first few years of running open air [43:36] were like the most fun professional [43:38] years of my life by far. It was like [43:40] unbelievable, you know, before you [43:42] released the product. [43:43] >> Yeah. Yeah. Running a research lab with [43:45] the smartest people doing this like [43:47] amazing like historical work and I got [43:49] to watch it and that was very cool. [43:52] And then we launched HGBT and everybody [43:55] was like congratulating me and I was [43:57] like [43:59] my life is about to get completely [44:02] ransacked. And of course it has. Uh and [44:06] but it it feels like it's just been [44:08] crazy all the way through. It's been [44:09] almost 3 years now. And [44:13] I think it does get a little bit crazier [44:15] over time, but I'm like more used to it. [44:18] So it feels about the same. [44:19] >> Yeah. [44:21] We've talked a lot about Open Eye, but [44:22] you also have a few other companies, [44:24] Retro Biosciences and Longevity and [44:26] energy companies like Helon and Ollo. [44:29] Did you have a a master plan, you know, [44:31] a decade ago to sort of make some big [44:33] bets across these major spaces or how [44:36] how do we think about the Sam Alman arc [44:37] in this way? No, I just wanted to like [44:40] use my capital to fund stuff I believed [44:43] in. Like I I didn't it it Yeah, it felt [44:47] like a good use of capital like and more [44:50] fun or more interesting to me and [44:52] certainly like a better return than like [44:53] buying a bunch of art or something. [44:55] >> Yeah. [44:56] >> What about the quote unquote human [44:57] algorithm do you think AIs of the future [44:59] will find most fascinating? [45:05] I mean kind of the whole I would bet the [45:07] whole thing like the whole my intuition [45:09] is that like AI will be fascinated by [45:12] all other [45:14] things to study and observe and you know [45:17] like [45:18] >> Yeah. [45:18] >> Yeah. In in closing, I I love this [45:21] insight you you had um where you talked [45:23] about how you know the the next open a [45:27] mistake investors make is pattern [45:28] matching off previous breakthroughs and [45:30] just trying to find oh what's the what's [45:31] the next Facebook or what what's the [45:32] next open AI and that the next you know [45:35] potentially trillion dollar company [45:36] won't look exactly like open AI it will [45:39] be built off of the breakthrough that [45:41] open AI has helped you know emerge which [45:43] is you know near free AGI at scale in [45:45] the same way that open AI leveraged pre [45:46] previous breakthroughs and so for [45:48] founders and investor ers and people [45:50] trying to ascertain the future listening [45:51] to this. How do you think about a world [45:53] in which there is open achieves this [45:55] mission? There is near near free AGI. [45:56] What types of opportunities might emerge [45:59] for for company building or investing [46:01] that you're potentially excited about as [46:03] you put your investor hat on or company [46:04] building hat on? [46:07] I I I have no idea. I mean, I have like [46:09] guesses, but they're like they're [46:10] they're I have learned [46:12] >> you're always wrong. [46:13] >> You've learned you're always wrong. I've [46:15] learned deep humility on this point. Um, [46:18] I think the the only like [46:22] I think if you try to like armchair [46:24] quarterback it, you sort of say these [46:27] things that sound smart, but they're [46:29] pretty much what everybody else is [46:30] saying, and it's like really hard to get [46:33] the right kind of conviction. The only [46:35] way I know how to do this is to like be [46:40] deeply in the trenches exploring ideas, [46:42] like [46:43] talking to a lot of people, and I don't [46:45] have time to do that anymore. Yeah, [46:46] >> I only get to think about one thing now. [46:48] >> So I I would just be like repeating [46:50] other people's or saying the obvious [46:53] things but [46:57] I think it's a very important like if [46:59] you are an investor or a founder I think [47:00] this is the most important question and [47:03] you don't you you figure it out by like [47:06] building stuff and playing with [47:08] technology and talking to people and [47:09] being out in the world. I have been [47:13] always [47:16] enormously disappointed by the [47:18] willingness of investors to back this [47:20] kind of stuff even though it's always a [47:21] thing that works. You all have done a [47:23] lot of it but most firms just kind of [47:26] chase whatever the current [47:28] >> thing is and so do most founders. [47:31] >> Uh so I hope people will try to go [47:34] >> yeah we talk about how you know silly [47:37] you know fiveyear plans can be in a [47:39] world that's constantly changing. It [47:40] feels like when I was asking about your [47:41] master plan, you know, your your career [47:43] arc has been following your curiosity, [47:45] staying, you know, super close to the [47:46] the the smartest people, uh, the super [47:48] close to the technology and just [47:50] identifying opportunities and kind of an [47:52] organic and incremental way from there. [47:55] >> Uh, yes, but AI was always a thing I [47:58] wanted to do. I went to I I studied AI. [48:00] I worked in the AI lab between my [48:02] freshman and sophomore year of college. [48:04] >> Yeah. [48:04] >> It wasn't working all the time. So, I'm [48:06] like not I'm not like enough of a [48:10] I I don't want to like work on something [48:12] that's totally not working. It was clear [48:13] to me at the time AI was totally not [48:14] working. Um but [48:19] I've been an AI nerd since I was a kid. [48:20] Like this [48:23] >> so amazing how it, you know, you got [48:26] enough GPUs, got enough data, and the [48:28] lights came on. [48:30] >> It was such a hated like people were [48:32] >> man, when we started like [48:34] >> figuring that out, [48:36] >> people were just like absolutely not. [48:38] the the the field hated it so much. [48:40] >> Investors hated it, too. [48:42] >> It's not It's not the [48:46] >> It's somehow not an appealing answer to [48:48] the problem. [48:49] >> Yeah, it's a bitter lesson. [48:51] >> Yeah. Well, the rest is history and [48:54] we're perhaps let's let's wrap on that. [48:55] We're lucky to to to be partners along [48:57] for the ride. Sam, thanks so much for [48:58] coming on the podcast. [48:59] >> Thanks very much. [49:00] >> Thank you. [49:04] [Music]