Andreessen, Horowitz Warn China May Lap US on AI Robotics Hardware

安德森与霍洛维茨警告：中国或在AI机器人硬件领域反超美国 尽管目前在软件开发方面保有微弱的六个月领先优势，但由于本土工业供应链遭到侵蚀，美国面临在长期人工智能竞赛中输给中国的风险。 本次讨论重构了针对人工智能的常见批评，提出真正的人类创新极为罕见，并将语言模型视为重混与拓展创造力的强大工具，如同早期的嘻哈音乐采样。发言者质疑了纯粹智力必然决定领导力或权力的观点，指出人类管理具有复杂、因情境而异的本质，且缺乏实体的AI大脑存在全身性局限。展望未来，他们反驳了AI市场泡沫论，建议创业者依循第一性原理进行构建而非套用陈旧的技术手册，并概述了一场高风险的美中地缘政治竞争——西方在这场角逐中有可能被中国占据主导地位的机器人实体制造生态所超越。 AI用户界面的长期演进依然高度不可预测，这与早期PC的历史如出一辙——当时从文本提示符过渡到图形用户界面（GUI）耗时17年，再到网络浏览器又历时5年。 (0:00) 针对大语言模型缺乏真正原创性的批评忽视了真正的人类创新极其罕见这一现实；即便是贝多芬这样的历史名家，也大量重混了前人的作品。 (0:00) 人类在分布外推理和横向迁移学习方面的能力被严重夸大了，这意味着AI无需达到完美的原创思维，便能创造巨大的社会价值。 (3:44) 传奇嘻哈艺术家将AI视为拓展其创作能力的先进工具，并将大语言模型的训练机制与嘻哈音乐中采样的基础实践进行了类比。 (5:28) 认为顶尖的原始智力自然会带来领导力与权力的假设与现实的人际动态相矛盾，在实际中，高智商的专家通常服务于社交能力更为出色的中等智商通才。 (8:52)

Marc Andreessen and Ben Horowitz argue that artificial intelligence must be evaluated against realistic human cognitive limits rather than romanticized notions of genius or infallible leadership. They contend that current enterprise adoption reflects genuine demand rather than a financial bubble, though the ultimate product interfaces remain unformed and Western manufacturing decline poses severe hardware risks in embodied AI. Human Baselines and the Creative Remixing Debate (0:00) The speakers dispute criticisms that Large Language Models (LLMs) fall short because they merely remix data rather than generate isolated scientific or creative breakthroughs. In their view, true conceptual leaps and cross-domain transfer learning are extraordinarily rare even among humans, where only a tiny fraction of individuals can consistently apply lateral insights across distinct fields. In social science, fluid intelligence (IQ) carries a correlation factor of approximately 0.4 with positive life outcomes, leaving 60 percent of variance unexplained, while major historical achievements routinely build on decades of accumulated precedent. Horowitz illustrates this dynamic through hip-hop, an artistic movement founded entirely on sampling and recontextualizing existing recordings. Over 50 years of genre history, only a handful of pioneering figures—such as Rakim, Dr. Dre, and George Clinton—introduced foundational conceptual leaps. Because human creators contribute localized, lived experiences, human artists retain distinct storytelling advantages, while AI functions as an expansive new instrument for creative remixing. The Limits of Raw Intelligence in Leadership (8:52) Raw intelligence does not dictate societal hierarchy or organizational effectiveness. The speakers observe that high-IQ technical specialists routinely work for generalists who excel at interpersonal navigation, theory of mind, and constructive confrontation. Drawing on US military data utilizing the Armed Services Vocational Aptitude Battery (ASVAB), they note that leadership cohesion breaks down when a leader's IQ diverges by more than one standard deviation (roughly 15 points) from subordinates, because extreme intellectual gaps impair a leader's ability to model follower cognition. Consequently, a theoretical ultra-intelligent machine would struggle to govern human systems effectively. Furthermore, the speakers emphasize that human cognition is an embodied phenomenon shaped by sensory inputs, hormones, and the gut biome, presenting a contrast with disembodied software models. While modern LLMs exhibit capable theory of mind—enabling applications like simulated voter focus groups—they demonstrate an artificial default bias toward consensus and conflict avoidance unless explicitly prompted to sustain adversarial debate. Market Fundamentals and Unformed AI Interfaces (22:50) Addressing concerns of an artificial intelligence bubble, the speakers argue that true financial bubbles require total market capitulation, whereas current AI investment—with infrastructure capex at roughly 1 percent of GDP—is grounded in authentic enterprise demand. They urge observers to evaluate the sector using two ground truths: whether the technology functions as promised and whether paying customers support it. While localized physical bottlenecks like power and cooling exist, the speakers see no structural demand deficit on a five-year horizon. Additionally, the speakers reject the premise that chatbots and search engines represent the final form of AI software. Computing history shows interface transitions take significant time: the personal computer operated under text prompt interfaces for 17 years (1975–1992) before pivoting to graphical user interfaces, followed by web browsers five years later. Because long-term interface designs remain nascent, incumbent tech monopolies face classic disruption vulnerabilities if they cannot sustain execution speed across successive technological shifts. Infrastructure Glut Cycles and Hardware Geopolitics (31:53) In accordance with standard supply economics, the speakers project that intense shortages in AI infrastructure and talent will eventually invert into supply gluts. While elite talent and compute capacity remain constrained, non-traditional researchers are already training competitive frontier architectures, citing Chinese developments like DeepSeek, Qwen, and Kimi alongside Elon Musk's xAI. As tooling commoditizes, the premium on traditional brand-name research teams will diminish. Geopolitically, the speakers estimate that the United States holds only a narrow lead of approximately six months over China in AI software. While Western firms lead in core conceptual software advances, 40 years of Western de-industrialization have left the physical supply chain for embodied robotics heavily concentrated in China. If the AI paradigm transitions from disembodied software to physical machines, China's manufacturing ecosystem creates a substantial risk that it will outpace the West in hardware deployment. While Andreessen and Horowitz anticipate sustained demand and the normalization of current infrastructure constraints, they consider the long-term competitive landscape unresolved due to unformed product interfaces and Western supply-chain deficits in physical robotics.

Marc Andreessen与Ben Horowitz指出，当前对大语言模型创造力的批评多基于对人类普遍心智的不切实际预期，而AI在真实商业与组织中的应用正揭示出智力并非权力和领导力的唯一支柱。针对行业动态，他们认为人工智能市场并无泡沫，但随着技术从纯软件演进至具身机器人，全球竞争的瓶颈将迅速由芯片转向更广泛的硬件与物理制造生态。 心智神话：重组式创新与智力边际 (0:00) 对于大语言模型（LLM）仅能“重新打包或重组现有数据”、缺乏真正原创发明的批评，Andreessen与Horowitz认为这严重夸大了人类自身的普遍能力。在人类文明史上，贝多芬等艺术巨匠也高度借鉴了莫扎特和海顿等前人的成果，几乎所有的重大突破都是数十年累积工作的终点，技术发展通常遵循至少40年的前置研发周期。至于跨领域推理与“分布外”迁移学习，现实中一万人里可能仅有大约三人能稳定做到；在社会科学中，流体智力（IQ）与人生积极结果的相关性仅为0.4左右，剩余60%的方差完全由其他因素决定。因此，即便AI仅达到99.99%人类的平均认知水平，也足以带来巨大的社会价值。 管理与权力：高智力并不等同于统治力 (8:52) 针对AI圈内“更高智力必然统治较低智力”的技术决定论假设，对谈指出真实世界运转逻辑截然相反。在企业与社会组织中，高智商的技术专家（即所谓“空间旋转者”）往往受雇于通才或MBA管理者（即擅长语言与人际动态的“词汇细胞”）。管理的核心挑战不在于单纯解算最优策略，而在于凭借同理心、勇气、心智理论（Theory of Mind）以及对情境动态的敏感度，促使员工执行并不受欢迎但正确的决策。群体聚集时展现的集体智力甚至常常远低于其个体均值。 美国军方通过ASVAB职业素质测试发现了一条关键规律：若领导者的智商偏离下属超过一个标准差（约15分），沟通与领导力就会出现断层。极高智商的人往往难以有效模拟中等智力人群的思考逻辑。Andreessen与Horowitz据此推论，一个拥有“1000智商”的机器或超级智能，由于其认知视角与人类过于脱节，根本无法直接理解并统领复杂的人类现实。此外，人类心智还是融合了神经、肠道菌群与激素的全身性具身体验，脱离肉身的纯算法模型在天然情境理解上仍有根本局限。 市场周期：需求驱动与资本过热的伪命题 (22:50) 在探讨当前AI市场是否存在投机泡沫时，对谈者辩称真正的资本泡沫往往伴随着市场彻底投降、所有怀疑者全数做多，而当下业内对泡沫的密集讨论本身就证伪了全面亢奋状态。虽然目前AI资本支出（Capex）已占到美国GDP的大约1%，且短期面临电力、冷却等物理设施的瓶颈，但底层业务并未出现需求不足的迹象。传统科技巨头在平台跃迁中可能行动迟缓，但新旧力量的胜负根本上取决于技术是否真正解决问题以及客户是否愿意真金白银买单。 就供需经济学而言，Andreessen与Horowitz预测当前的结构性短缺终将诱发供应爆发，进而转向供给过剩。无论是英伟达主导的高利润芯片短缺，还是顶尖AI研究人才的稀缺，都不会无限期持续。中国企业（如DeepSeek、阿里通义千问Qwen、月之暗面Kimi）以及xAI展现出的工程实力表明，行业已开始在著名学术权威之外批量培养青年工程人才，而芯片短缺一旦遭遇电力供应受限等其他瓶颈，很快就可能转变为局部过剩。 产品形态演进与硬件代差的地缘博弈 (28:20) 对谈强调，将当前AI生态狭隘定义为“聊天机器人与传统搜索的较量”忽视了人机界面的长期演进律。从1975年个人电脑问世到1992年图形用户界面（GUI）兴起经历了整整17年，其后仅用5年又转向了Web浏览器。当前的文本对话框绝非终极形态，未来二十年必将诞生截然不同的产品交互架构，创业者必须基于第一性原理进行组织和战略重构，而非照搬上一代软件公司的打法。 在更广泛的地缘博弈层面，中美之间的AI竞赛被形容为领先优势仅有数月的“寸土必争”。美国在核心概念与软件创新上仍具前沿优势，但竞争重心一旦随着具身智能转向物理机器人，西方过去四十年去工业化的结构性缺陷将暴露无遗。机器人制造严重依赖涵盖数千家机械、电子与半导体部件供应商的深厚供应链生态，这一生态默认坐落在中国。即便美国能维持底层算法的领先，也存在被中国凭借强大的制造业基础在物理硬件量产端全面反超的切实风险。 本场对话的核心结论在于，AI的发展不应被单纯简化为算力规模和基准IQ的单向突破，其未来的重塑力量取决于深层人机交互形态的演变与实体工程落地。尽管美国在纯软件模型与初创算法方面维持了微弱领先，但若缺乏对组织情境、真实商业模式以及本土实体制造供应链的协同构建，纯算法优势在地缘竞争与物理机器人时代将面临严峻的硬件承接挑战。

[8:52] not said by anybody who owns a cat. [8:52] a supreme "shape rotator" (a highly technical specialist) can only rotate shapes, but a supreme "word cell" (a master of language and social dynamics) can rotate the shape rotators. [13:58] which ones can you afford [13:58] Mark Zuckerberg: intelligence is not life [13:58] people seldom do what's in their best interest. [18:05] it wants everybody to be happy. [18:05] reputational destruction [22:50] You can be furious about that for 30 years in our business [26:31] Google got their head out of their ass [28:20] tremendous headroom for invention [31:53] the best position probably anybody's ever had in chips [34:51] maybe a little bit of help... maybe a little USB stick in the middle of the night. [34:51] lap us in hardware

[0:00] I think we don't yet know the shape and [0:02] form of the ultimate products. Just one [0:03] just obvious historical analogy is you [0:05] know the personal computer from sort of [0:07] invention in 1975 through to you know [0:09] basically 1992 was a text prompt system [0:13] 17 years in you know the whole industry [0:14] took a left turn into gooies and never [0:16] looked back and then by the way you know [0:17] 5 years after that the industry took a [0:18] left turn into web browsers and never [0:20] looked back right and you know look I'm [0:21] I'm sure there will be chat bots 20 [0:23] years from now but I'm pretty confident [0:24] that both the current chatbot companies [0:26] and many new companies are going to [0:28] figure out many kinds of user [0:29] experiences that are radically different [0:30] that we don't we don't even know [0:34] Please join me in welcoming Mark Andre [0:36] and Ben Horowitz with general partner [0:38] Eric Torberg. [0:41] Follow me into our solo. Get in the flow [0:43] and you can pitch it like a photo. Music [0:45] makes me maintain and make melodies for [0:47] MC's motivation breaks. I'm never [0:50] >> Thank you for the rock Kim. Who did [0:52] that? [0:53] >> Ben picks the Ben picked the music. [0:57] Mark, there's been a lot of talk lately [0:59] about the limitations of LLMs that they [1:02] they can't do true invention of say new [1:05] science that they can't do true creative [1:07] genius that it's just combining or [1:09] packaging. [1:11] You have thoughts here. What say you? [1:13] >> Yeah. So, so for me, yes, you get all [1:14] these questions and yeah, they usually [1:16] come in either, you know, sort of are [1:18] are language models intelligent in the [1:20] sense of like can they actually um you [1:22] know, can they actually process [1:23] information and have sort of conceptual [1:24] breakthroughs? the way that people can. [1:26] And then there's are language models or [1:28] or video models creative, you know, can [1:30] they can they create new art? Um [1:32] actually have genuine creative [1:33] breakthroughs. And of course my my my [1:34] answer to both of those is well can [1:35] people do those things? Um and um I [1:38] think there's two two two questions [1:39] there which is like okay even if some [1:41] people are quote unquote intelligent as [1:43] in having original uh conceptual [1:45] breakthroughs and not just let's just [1:46] say regurgitating the training set um uh [1:49] or following scripts um uh how you know [1:51] what percentage of people can actually [1:52] do that and I say I've only met a few [1:55] some of them are here in the room um but [1:57] uh you know not that many most people [1:59] never do and then creativity I mean how [2:00] many people are actually genuinely [2:01] creative right and so you you kind of [2:03] point to a Beethoven or you know Van [2:05] Gogh or something like that you're like [2:06] Okay, that's creativity and yeah, that's [2:07] creativity and then how many Beethovenas [2:08] and Van Go are there? Obviously, not [2:10] very many. So, so one is just like okay, [2:12] like you know, if it's if it if if if [2:14] these things clear the bar of, you know, [2:16] 99.99% of humanity, you know, then [2:18] that's pretty interesting just in and of [2:20] itself. But then I you dig into it [2:22] further and you're like, okay, like how [2:24] many actual real conceptual [2:26] breakthroughs have there ever been [2:27] actually ever in human history as [2:29] compared to sort of remixing remixing [2:31] ideas. Um and you know like if you look [2:34] at the history of technology it's almost [2:36] always the case that the big [2:37] breakthroughs are the result of you know [2:39] usually at least 40 years of sort of [2:41] work ahead of time you know four decades [2:43] right in fact language models themselves [2:44] are the culmination of eight decades [2:46] right of previous work and so there's [2:48] remixing and then in the arts it's the [2:50] exact same thing you know novels and [2:51] music and everything like you know there [2:52] are clearly creative leaps but you know [2:54] there's just tremendous amounts of [2:56] influence that came in from from people [2:57] who came before and even if you think [2:59] about like somebody with the creativity [3:00] of a Beethoven like there There's a lot [3:02] of Beethoven in Mozart and Heiden and in [3:04] the composers that came before and so [3:06] there's just tremendous amounts of of of [3:07] remixing and combination. And so it's [3:09] it's it's a little bit of an angel's [3:11] dancing on the head of a pin question [3:12] which is like if you can get if you can [3:14] get you know within you know I don't [3:15] know one you know 001% of of kind of [3:19] worldbeating uh you know generational [3:21] creativity intelligence like you're [3:22] you're you're probably all the way [3:23] there. So I so emotionally I want to [3:26] like hold out hope that there is you [3:28] know still something special about human [3:30] creativity and I I certainly believe [3:31] that and I and I and I very much want to [3:33] believe that but um I don't know when I [3:35] use these things I'm like wow they seem [3:37] to be awfully smart and awfully [3:38] creative. So I'm I'm I'm I'm pretty [3:40] convinced that they're going to clear [3:41] the bar. [3:42] >> Yeah. [3:44] the I think that seems to be a common [3:45] theme in your analysis when when people [3:47] talk about the limitations of LMS, you [3:49] know, can they do transfer learning or [3:50] or just learning in general? You seem to [3:52] ask, can people do this? [3:54] >> Yes. Can people do these things? Well, [3:55] it's like lateral thinking, right? So, [3:57] yeah, so it's like reasoning in or out [3:58] of distribution, right? And so it's [3:59] like, okay, I know a lot of people who [4:00] are very good at reasoning inside [4:01] distribution. How many people do I [4:02] actually know who are good at reasoning [4:03] outside of distribution and doing [4:05] transfer learning? And and the answer is [4:06] like I know a handful. Like I I know a [4:08] few I know a few people um where [4:10] whenever you ask them a question, you [4:12] get an extremely original answer. And [4:13] usually that answer involves bringing in [4:15] some idea from some adjacent space and [4:17] basically being able to bridge domains. [4:19] Um and so you know you'll ask them a [4:20] question about I don't know you know [4:21] finance and they'll they'll bring you an [4:23] answer from psychology or you ask them a [4:24] question about psychology and they'll [4:25] bring you an answer from biology right [4:27] or whatever it is. And so I I know you [4:29] know I know I don't know sitting here [4:31] today probably three I probably know [4:33] three people who can do that reliably [4:35] out of the you know I I've got you know [4:37] I've got 10,000 in my address book. Um [4:39] and so three out of 10,000 [4:41] >> Yeah. is not is not that high a [4:42] percentage. By by the way, I find this [4:45] very encouraging. Like I I yeah, [4:46] immediately the mood in the room has [4:47] gone completely to hell. Um I find this [4:50] very encouraging. I find this very [4:52] encouraging because look at what [4:53] humanity has been able to build, right? [4:55] Despite all of our limitations, right? [4:57] And and look at all the creativity that [4:58] we've been able to exhibit and all the [5:00] amazing art and all the amazing movies [5:01] and all the amazing novels and all the [5:02] amazing technical inventions and [5:04] scientific breakthroughs. And so we, you [5:06] know, we've been able to do, you know, [5:07] everything we've been able to do with [5:08] the limitations that we have. Um, and [5:10] so, you know, I think that, you know, [5:12] like, you know, do you need to get to [5:13] the thing where you are 100% positive [5:15] that's actually doing, you know, [5:16] original thinking? I don't think so. Uh, [5:19] I I think it'd be great if you did, and [5:20] I think ultimately we'll probably [5:21] conclude that that's what's happening. [5:23] Um, but like I it's not necessary for [5:25] like just tremendous amounts of [5:26] improvement. [5:28] >> Ben, we were just celebrating some some [5:30] hip-hop legends at your paid in full uh [5:32] event last week and so you think a lot [5:34] about creative genius. How do you think [5:36] about this question? [5:38] Yeah, I mean I think that uh I agree [5:41] with Mark that it's [5:44] whatever it is, it's very useful. um [5:46] even if it isn't all the way that level [5:48] uh I think that [5:51] you know there's something about the [5:54] actual like real time human experience [5:57] that humans are very into um at least in [6:01] art where [6:04] you know with the current state of the [6:06] technology kind of the the pre-training [6:08] doesn't have quite the the right data to [6:11] to to get to um what you really want to [6:15] Uh but it, you know, it's [6:19] pretty good. [6:20] >> It is pretty good. [6:21] >> How many So how many true conceptual So [6:24] Ben Ben Ben Ben's Ben, one of Ben's [6:25] nonprofit activities is something called [6:27] the Pain and Fall Foundation, which is [6:28] honoring and actually providing [6:30] essentially a pension a pension for uh [6:31] you know, for sort of, you know, the [6:33] great innovators in in in rap and [6:34] hip-hop. Um and and so he has he knows [6:38] and has many of we were just at the [6:39] event and he you know has many of the [6:40] kind of leading lights of that field for [6:42] the last 50 years you know perform and [6:43] and it's really fun to meet them and [6:44] talk to them. Um but like how many [6:47] people in that entire field over the [6:48] course of the last 50 years would you [6:50] classify as like a true conceptual [6:51] innovator? [6:53] >> Yeah. Well, you know it's interesting. [6:56] Well, it depends how broadly you define [6:58] it but [7:01] >> you know there were several of them [7:02] there last you know on Saturday. Rock, I [7:06] think. Yeah, Rockm, you'd certainly put [7:07] in that category. Dr. Dre, you'd [7:09] certainly put in that category. George [7:10] Clinton, you'd certainly put in that [7:12] category. Um, [7:14] >> you know, in a narrower sense, like Cool [7:18] G Rap certainly had a new idea. Um, [7:22] but you know, it depends like a [7:26] fundamental kind of musical [7:28] breakthrough, you probably just say like [7:31] Rock Kim and George Clinton. Um, are [7:33] they excited? [7:34] >> So, so two out of [7:37] >> Well, I mean those of the guys who were [7:38] there. [7:39] >> Oh, yeah. Yeah. [7:40] >> Yeah. But yeah, it's a tiny percentage. [7:42] Tiny tiny tiny tiny tiny. [7:44] >> We had Jared at the fireside last night [7:46] with Jared Leto. He was talking about [7:47] how many people in Hollywood are are [7:49] really scared or against this um what's [7:52] happening here is is what do you see in [7:54] you know when you talk to the Dr. JS the [7:55] nas the Kanye's are they excited? Are [7:57] they using it? Are they [7:59] >> Yeah. No, I the so everybody uh who I [8:02] speak to there are definitely people who [8:05] are scared in music but like there are a [8:07] lot of people who are very very [8:08] interested in it and particularly the [8:09] hip-hop guys are interested because um [8:13] it's almost like a replay of what they [8:16] did right that they just took other [8:18] music and they kind of built new music [8:20] out of it and I think that you know AI [8:24] is uh a fantastic creative tool for [8:27] them. like way opens up the pallet and [8:29] then for you know a lot of what hip-hop [8:32] is is it's um kind of telling a very [8:35] specific story of a specific time and [8:38] place which um having intimate knowledge [8:42] and being trained just on that thing is [8:44] is actually an advantage as opposed [8:46] being like a generally smart uh music [8:49] model. Um, people also use the same [8:52] logic of, hey, whatever is more [8:54] intelligent will rule whatever is less [8:56] intelligent. And and Mark, you recently [8:58] uh [8:59] not not said by anybody who owns a cat. [9:02] >> Yeah, exactly. [9:04] Mark, you recently tweeted, "A supreme [9:06] shape rotator can only rotate shapes, [9:08] but a supreme word cell can rotate shape [9:10] rotators." And and also, [9:13] >> someone's clapping here. And also, high [9:15] IQ experts work for midiq generalists. [9:19] What means? [9:20] >> Yeah. What what means? Uh yeah. So it's [9:22] the PhDs all work for MBAs, right? So [9:24] it's like, you know, okay. So yeah, like [9:28] so yeah. Well, I just, you know, just [9:30] take it up a level. It's just like when [9:31] you look at the world today, do you [9:32] think we're being ruled by the smart [9:34] ones, [9:36] right? Like is that is that your big [9:38] conclusion from like current events, [9:39] current affairs, [9:41] right? Like okay, we put the geniuses in [9:42] charge. Like [9:43] >> you mean Kamla and Trump aren't the [9:44] best? [9:45] >> Well, now let's not even be specific [9:46] towards the US. let's just look all over [9:48] the world. [9:49] >> Um, you know, yeah. And so like it's [9:51] just like [9:53] there's this thing. So I think two two [9:56] things are true. One is we we probably [9:58] all kind of underwrite the importance of [9:59] intelligence. Um, and actually there's a [10:01] whole kind of backstory here of like [10:03] intelligence actually turns out to be [10:04] this like incredibly inflammatory, you [10:06] know, kind of topic for lots of reasons [10:08] over the last hundred years. um uh which [10:10] which we could we could talk about in [10:11] great detail but like it it and you know [10:13] it's and the even the just the very idea [10:15] that like some people are smarter than [10:16] other people you just like really freaks [10:18] people out and people people don't like [10:19] to talk about it we really struggle with [10:20] that as a society and so like and then [10:22] it is true that intelligence is like in [10:24] humans intelligence is correlated to [10:26] almost every kind of positive life [10:27] outcome right and so u intelligence [10:30] generally in the social sciences what [10:31] they'll tell you is what they call fluid [10:33] intelligence or or the G factor or IQ is [10:35] sort of it's sort of 0.4 four correlated [10:37] to basically everything. Um, and so it's [10:40] 0.4 correlation to like educational [10:42] outcomes and like you know professional [10:44] outcomes and and you know income and by [10:46] the way also like life satisfaction and [10:48] by the way non-violence you know being [10:50] able to solve problems without physical [10:51] violence and so forth. And so like on [10:53] the one hand like we probably all [10:54] underrate intelligence. Um on the other [10:56] hand the people who are in the fields [10:58] that involve intelligence probably [10:59] overrate intelligence. Um, and you might [11:02] even you might even coin a term like [11:03] maybe like intelligence supremacist or [11:05] something like that where it's just like [11:06] oh like intelligence is very important [11:08] and so therefore maybe it's like the [11:10] most important thing or the only thing [11:12] but but then you look at reality and [11:13] you're like okay that's clearly not the [11:14] case. [11:15] >> Yeah. It's still zero only 0.4 right. [11:17] Yeah. [11:17] >> Well so to start with it's only 0.4 and [11:19] you know in the social sciences 0.4 is a [11:21] giant correlation factor right like most [11:23] most things that where you can correlate [11:24] whether it's you know genes or observed [11:26] behavior whatever to anything in the [11:27] social sciences the correlations are [11:28] much smaller than that. So 0.4 is is [11:30] tiny, but it's still only 0.4. So, even [11:32] if you're like a fullon if you even if [11:34] you're like a full-on genetic [11:36] determinist and you're just like, you [11:37] know, genetic IQ just like drives all [11:39] these outcomes, like it still doesn't [11:41] explain, you know, 6 uh of the [11:44] correlation and so that leaves it but [11:45] but that's just on the individual level. [11:47] Then you just look at the collective [11:48] level. Well, you just look at the [11:49] collective level and it's like a famous [11:51] famous observation is you take a b you [11:52] take a bunch of you take any group of [11:54] people, you put them in a mob and the [11:55] mob is dumber, right, than the average [11:58] and and and you put a bunch of smart [11:59] people in a mob and they definitely turn [12:01] dumber like and you see that all the [12:02] time, right? Um uh and so you put people [12:05] in groups and they they they behave very [12:07] differently and then you you create and [12:09] then you create these you know questions [12:11] around like who's in charge whether [12:12] who's in charge at a at a company or [12:14] who's in charge of a of a country and [12:15] like it it's whatever the filtration [12:17] process it's clearly not it's not it's [12:19] it's not it's certainly not only on IQ [12:21] and it may not even be primarily on IQ [12:23] and so so therefore it's just like this [12:25] assumption that you kind of hear in some [12:26] of the AI circles uh which is like [12:27] inevitably the smart you know kind of [12:29] thing is going to govern the dumb thing [12:31] like I I just think that's like very [12:32] easily uh it's just sort of very easily [12:34] and obviously falsified like [12:35] intelligence isn't sufficient and then [12:37] you just you just you just convey it. [12:39] You know, we're all in this room lucky [12:40] enough to know a lot of smart people and [12:41] you you just kind of observe smart [12:42] people and like some smart people, you [12:44] know, really figure out how to have [12:45] their stuff together and become very [12:46] successful and a lot of smart people [12:48] never do. Um and so there there's there [12:51] there must be there obviously are and [12:53] there and there in fact must be many [12:54] other factors that have to do with [12:56] success um and have to do with like [12:58] who's in charge than just raw [12:59] intelligence. It it begs the the [13:01] follow-up question of what are what are [13:03] some examples of what that might be you [13:05] know skills sort of outside of [13:06] intelligence and more particularly [13:07] specifically why couldn't AI systems you [13:10] know learn them? [13:11] >> Yeah. So Ben like what what what other [13:12] than intelligence what what in your [13:14] experience determines for example [13:15] success in leadership or in [13:16] entrepreneurship or in in solving [13:18] complex problems or organizing people? [13:21] >> Yeah. [13:23] There there are many things um you know [13:25] like a lot of it is uh being able to [13:28] have a confrontation in the correct way [13:30] and like there's some intelligence in [13:32] that but a lot of it is just under [13:34] really understanding who you're talking [13:36] to [13:37] you know being able to interpret [13:40] everything about how they're thinking [13:42] about it and just kind of generally [13:45] seeing decisions through the eyes of the [13:47] people working in the company not [13:49] through your eyes is the [13:52] skill that you you know you develop by [13:54] talking to people all the time, [13:55] understanding what they're saying, so [13:57] forth, these kinds of things. And it's [13:58] just um [14:01] you know, it's certainly not an IQ [14:03] thing. And not that like I I could [14:05] imagine an AI training on any individual [14:08] and like figuring it all out and knowing [14:11] what to say and so forth. Um but then [14:14] you also need that integrated with uh [14:17] you know like whatever the business [14:19] ought to be doing. So you're not you're [14:21] not trying to do what's popular. You're [14:23] trying to get people to do what's [14:26] correct even if they don't like it. And [14:28] you know that's a lot of management. So [14:30] uh [14:32] it's not a problem anybody's working on [14:34] currently but maybe they will. [14:36] >> It's some right some combination of like [14:38] courage some combination of motivation [14:41] some combination of um of emotional uh [14:43] understanding theory of mind. [14:45] >> Yeah. what you know what do people want [14:47] like you know married to you know what [14:51] needs to be done and then like how [14:53] talented are they like which ones can [14:56] you afford like if they jump out the [14:57] window it's fine you know which one's [14:59] not fine you know this kind of thing [15:01] it's a there's a lot of like weird [15:04] subtleties to it uh and it's very [15:06] situational I think the hardest thing [15:08] about it uh and why management books are [15:10] so bad is because it's situational um [15:14] you know like your company your product, [15:15] your people, your org chart is very very [15:18] different than you know, yeah, here are [15:21] the five steps to building a strategy. [15:23] It's like, well, that's the most useless [15:24] thing I ever read because it has [15:27] nothing to do with you. [15:28] >> So, one of the interesting things on [15:30] this like on this is right the the [15:32] concept of theory of mind is really [15:33] important, right? So, theory of mind is [15:34] can you in your head model what's [15:36] happening in the other person's head, [15:37] right? And and and you would think that [15:38] maybe that you know maybe obviously [15:40] people who are smarter should be better [15:41] at that. It turns out that that may not [15:42] be true and I'll the reason to believe [15:44] that that's not true which which is as [15:45] follows. So the so the the US military [15:48] is was the early adopter and has [15:49] continued to be sort of the leading [15:51] adopter in in US society of of of [15:53] actually IQ testing and they uh they [15:55] they they basically launder it through [15:56] something called the ASVAB which is [15:58] their they call vocational aptitude [15:59] battery test but it's basically an I [16:01] it's essentially an IQ test. Um and uh [16:04] so they they they they still use [16:06] basically explicit IQ tests and they [16:07] they slot people uh into different [16:09] specialties and roles uh you know in [16:11] part with according IQ um including into [16:14] leadership roles. Um and and so they [16:16] they know what everybody's IQ is and [16:17] they they kind of organize organize [16:18] around that. And one of the things that [16:19] they found over the years is if the [16:21] leader is more than one standard [16:23] deviation of IQ away from the followers, [16:24] it's a real problem. Um and and and [16:27] that's true in both directions, right? [16:29] Um if if the leader is not smart enough [16:31] to be able to right manage the you know [16:34] is to be able to mo you know for [16:36] somebody who is less smart to model the [16:38] mental behavior of somebody who's more [16:39] smart is inher of course inherently very [16:41] challenging and maybe impossible. But it [16:42] turns out the reverse is also true which [16:44] is if the leader is two standard [16:46] deviations above the norm of the [16:47] organization that he's running. He also [16:49] loses theory of mind. Right? it it's [16:51] it's actually very hard for very smart [16:53] people to um to to model uh the internal [16:56] thought processes of even moderately [16:57] smart people. Um and so there there's [17:00] actually there's actually a real there's [17:01] actually a real need to have a level of [17:03] connection there that's not just right [17:05] and therefore by inference if you had a [17:07] person or a or a machine that had you [17:09] know a thousand IQ or something like it [17:10] it may just be it would be so alien its [17:13] understanding of reality would be so [17:14] alien to the people or the things that [17:16] it was managing that it wouldn't it [17:18] wouldn't even be able to connect in any [17:19] sort any sort of realistic way. So again [17:20] this is a very good argument that like [17:22] it yeah this is the world is going to be [17:23] far from organized by IQ for yeah for [17:26] centuries to come. [17:27] >> Yeah and Zuckerberg had a great line [17:29] which is intelligence is not life and [17:32] life life has a lot of dimensionality to [17:34] it that is independent of intelligence. [17:36] I think that [17:39] you know if you spend all your time [17:40] working on intelligence you lose track [17:42] of that. We sometimes say about uh some [17:45] specific people that they're too smart [17:47] to properly model or or you know um too [17:51] they s sort of assume too much [17:52] rationality on other people or they just [17:54] o overthink things or overrationalize [17:56] them. Um yeah just to your point that [17:58] it's it's on everything. [17:59] >> Yeah. Yeah. Yeah. People often people [18:02] seldom do what's in their best interest [18:04] I should say. [18:05] >> You know I also suspect this kind of [18:06] gets more into the biology side of [18:07] things. I you know there's more and more [18:09] scientific evidence that basically also [18:11] that like human human cognition human [18:14] cognition or human I don't whatever you [18:16] want to call it self-awareness [18:17] information processing decision-m sort [18:19] of experience um uh is is not purely a [18:22] brain like the basically the d the sort [18:24] of mind famous mindbody dualism is just [18:26] not correct like and again this is an [18:28] argument against sort of IQ supremacism [18:29] or intelligent supremacism is it's not [18:32] you know we we human beings didn't [18:33] experience existence just through the [18:35] rational thought u and and and [18:36] specifically not through just the [18:38] rational thought of the brain, but [18:39] rather it's a whole body experience, [18:40] right? And there's there's there's [18:42] aspects of our nervous system and [18:43] there's aspects of everything from our [18:44] gut biome to, you know, to to you know, [18:46] to to smells, you know, to alactory [18:48] senses and and um you know, and hormones [18:50] and like all all kinds of like [18:52] biochemical kind of aspects to life. Um [18:55] I sus if you just kind of track the [18:57] research, I suspect we're going to find [18:58] is human cognition is a full body [18:59] experience. Uh much much more than much [19:01] more than people thought. Um and and so [19:04] therefore to actually and this you know [19:06] this is like a and this is you know one [19:07] of the kind of big fundamental [19:08] challenges in the AI AI field right now [19:10] which is you know the form of AI that we [19:12] have working is is the is the fully mind [19:14] body dual version of it which is it's [19:16] just a disembodi you know like a [19:17] disembodied brain you know the robotics [19:20] revolution for sure is coming when that [19:21] happens when we put AI in in physical [19:23] objects that move around the world you [19:25] know you're going to be able to get [19:26] closer to having that kind of you know [19:28] inte integrated intellectual physical [19:30] you know experience you're going to have [19:31] sensors in the robots are going to be [19:33] a lot more data. [19:45] But it is just to me at least reading [19:47] the research like that all those ideas [19:48] feel very nent and we have a lot of work [19:50] to do to try to figure that out. [19:51] >> Do you have a sense for how they are at [19:54] theory of mind today? Um or do you have [19:57] a where the limitations are? you like to [19:58] talk to them a lot. Are there any [20:00] particular things that are particularly [20:01] surprising to you as you do? [20:02] >> Yeah, I would say generally they're [20:04] really good. Um, yeah. And so like one [20:05] of the one one of the more I find one of [20:07] the more fascinating ways, you know, to [20:08] to work with language models is actually [20:10] have them create personas. Um, and uh [20:12] and then you know, basically have well [20:14] actually so I like I like I like [20:16] basically I like Socratic dialogues. I [20:18] like when things are argued out and like [20:19] a Socratic dialogue. And so you you know [20:20] tell a tell any advanced LLM today to [20:23] create a Socratic dialogue and it'll [20:24] either make up the personas, you can [20:25] tell what it is, it does a good job. it [20:27] has this very very annoying property [20:29] which is it wants everybody to be happy. [20:31] Um and so it wants all of its personas [20:33] to agree. Um and so by default it will [20:35] have a uh it will have a briefly [20:37] interesting discussion and then it will [20:39] sort of figure out you know basically it [20:41] like you're watching I don't know PBS [20:42] special or something. It'll it'll kind [20:43] of figure out how to bring everybody [20:44] into agreement. Everybody's happy at the [20:45] end of the discussion. And of course I [20:46] hate that. Like it drives me [20:48] nuts. I don't want that. So instead I I [20:50] tell it I'm like make the conversation [20:52] more tense, right? And like fraught with [20:55] like anger and like you know people, you [20:58] know, going to get like increasingly [20:59] upset throughout the conversation. And [21:01] then it starts to get really [21:02] interesting. Um and then I and then I [21:04] tell it, you know, bring it, you know, [21:05] use introduce a lot more cursing. Um you [21:08] know, really have them go at it like all [21:09] the gloves come off, they're going for [21:10] pull full, you know, you know, [21:12] reputational destruction of each other. [21:13] >> You do a lot of these skits. [21:14] >> Yeah, these skits. And then I get [21:15] carried away and then I'm like it turns [21:16] out they're all like secret ninjas and [21:18] then they all start fighting and you've [21:19] got Einstein you know you know you know [21:21] hitting you know Neils Bore with [21:22] nunchucks and it and by the way it's [21:24] happy to do that too. Um so you do have [21:26] to you have to you have to control [21:27] yourself but it it is very good a theory [21:29] of mind and then I'll give you another [21:31] example. There's a there's a startup [21:32] actually in the UK uh in uh in in the [21:34] world of politics and and what what they [21:36] found is that um they found that [21:38] language models now are good enough so [21:40] specifically for for politics which is [21:42] sort of a sub subcategory where this [21:44] this idea matters. Um so you know in [21:46] politics people do focus group you do [21:48] focus groups of voters all the time and [21:50] and by the way many businesses also do [21:51] that. Um, you know, so you get a bunch [21:53] of people together from different [21:54] backgrounds in a room and you kind of [21:55] guide them through a discussion and try [21:56] to get their their points of view on [21:58] things. And and focus groups are often [21:59] surprising. Like politicians who if you [22:01] talk to politicians who do focus groups, [22:02] they're often surprising. They're often [22:04] surprised by the things that they [22:05] thought voters cared about is actually [22:06] not the things that voters care about. [22:07] And so you can actually learn a lot by [22:09] doing this. But focus groups are very [22:10] expensive to run. And then there's a [22:12] long lag time because they have to be [22:13] actually physically organized and you [22:14] have to recruit people and vet people [22:16] and and so forth. Um, and so it turns [22:18] out that the the the the [22:19] state-of-the-art models now are are good [22:21] enough at this so they can actually they [22:23] can they can correctly um accurately [22:25] reproduce a focus group of real people [22:27] um inside the model. Um so so so they're [22:29] good enough to clear that bar. In other [22:31] words, you you can basically have a [22:32] focus group actually happening in the [22:33] model where you create personas in the [22:35] model and then it actually accurately [22:36] represents, you know, a college student [22:38] from, you know, Kentucky is contrasted [22:40] to a housewife from Tennessee is [22:41] contrasted to a, you know, whatever [22:42] whatever you you just like specify this. [22:44] And so, you know, they're good enough to [22:46] clear they're good enough to clear that [22:47] bar and, you know, we'll we'll see how [22:48] far they get. [22:50] >> I want to segue to the bubble [22:52] conversation. Uh, Amin and G2, Jensen [22:56] and Matt spoke about the enormous scale [22:58] of physical infrastructure being built [23:00] out. AI capex is 1% of GDP. How should [23:04] we understand and think about this [23:06] bubble question? [23:07] >> Well, I think the fact that it's a [23:09] question means we're not in a bubble. [23:10] That's the first thing to understand. I [23:12] mean, a bubble is a psychological [23:14] phenomenon um as much as anything. And [23:17] in order to get to a bubble, everybody [23:19] has to believe it's not a bubble. That [23:21] that's sort of the the core mechanic of [23:23] it. And that, you know, we call that [23:26] capitulation. Everybody just gives up. [23:28] Like, okay, I'm not going to short these [23:29] stocks anymore. I'm tired of losing all [23:32] my money. I'm going to go long. Uh and [23:34] we saw that actually in, [23:37] you know, and a little bit of question [23:38] like really what was the tech bubble? Um [23:41] but in the kind of.com era right as the [23:45] prices went through the roof Warren [23:47] Buffett started inviting investing in [23:49] tech. So like and he swore he would [23:51] never invest in tech because you didn't [23:52] understand it. And so if he capitulated [23:55] nobody was saying it was a bubble when [23:57] it became like a quote unquote bubble. [23:59] Now if you look at that phenomenon [24:03] um the internet clearly was not a [24:05] bubble. Uh you know it was a real thing. [24:07] It was in the short term there was a [24:09] kind of price dislocation that happened [24:12] because uh the the market um you know [24:17] there were just not enough people on the [24:18] network to make those products go at the [24:21] time uh and then the prices kind of [24:23] outran the market. [24:26] you know, in AI, it's much harder to see [24:28] that because there's so much demand in [24:29] the short term, right? Like we don't [24:32] have a demand problem right now. And [24:34] like the idea that we're going to have a [24:36] demand problem five years from now to me [24:38] seems quite absurd. Uh, you know, could [24:41] there be like weird bottlenecks that [24:44] that appear, you know, like we just at [24:46] some point we just don't have enough [24:48] cooling or something like that? Yeah, [24:50] you know, maybe. But like like right now [24:53] if you look at demand and supply and [24:57] what's going on and multiples uh against [25:00] growth, [25:02] it doesn't look like a bubble at all to [25:04] me. Um but [25:06] I don't know. Do you think it's a [25:07] bubble, Mark? [25:08] >> Yeah. Look, I I would just say this. [25:10] Yeah. Like nobody know So nobody knows [25:12] in the sense of like the experts like if [25:13] you're talking to anybody at like a [25:14] hedge fun or a bank or whatever, like [25:15] they definitely don't know. Um uh [25:18] generally the CEOs don't know. So it the [25:20] >> by the way a lot of VCs don't know that [25:21] they just get upset like VCs get like [25:24] emotionally upset when you guys have [25:26] higher valuations like it it makes them [25:29] like like angry. Uh and you know and I I [25:32] get it all the time and I'm like what [25:34] are you mad about? Like the is [25:37] working man be happy. Come on. But so so [25:40] like there's a lot of emotion around [25:43] like people wanting it to be be a [25:45] bubble. [25:45] >> Yeah. No, nothing's worse than passing [25:47] on a deal and then having the company [25:48] become a great success. Like it's just I [25:49] was just just put [25:50] >> that that valuation is outrageous. [25:52] >> You can be furious about that for 30 [25:53] years in our business. It's it's [25:55] amazing. Um and you can find Yeah. You [25:57] come up with all kinds of reasons to [25:58] cope and and explain why it wasn't your [26:00] mistake. But it's you know it's the [26:02] world it's the world that's wrong, not [26:03] me. Right. So there there's a lot of [26:06] that. Yeah. Uh yeah. So I I just I would [26:08] just I would just say like I would [26:09] always say bring the conversation back [26:10] to ground truth fundamentals. And the [26:12] the two big ground truth fundamentals [26:13] are number one, does the technology [26:15] actually work? Um, you know, can it [26:17] deliver on its promise? And then number [26:18] two, is are customers paying for it? Um, [26:20] and if those if those two things are [26:22] true, then it's very hard to it's very [26:23] hard to uh like as long as those two [26:25] things stay grounded. Um, you know, gen [26:27] generally generally things are going to [26:29] I think are going to be on track. [26:31] >> Yeah. [26:32] When Gavin was up here with DG, he said [26:34] chatbt was a Pearl Harbor moment for [26:36] Google, the moment when the giant wakes [26:38] up. when when we look at history and and [26:40] platform shifts, what determine whether [26:42] the incumbent actually wins the next [26:43] wave versus versus new entrance or how [26:45] should we think about that in in AMP? [26:48] >> Well, you know, reacting to it is [26:51] important. Um, but that doesn't mean [26:55] like I it's a Pearl Harbor moment. I [26:58] think [27:00] Google got their head out of their ass. [27:02] It was the sound of it. [27:06] Uh so you know they're not going to get [27:08] completely run over but [27:11] nonetheless like I don't think open AI [27:13] is going away. So like they they [27:15] definitely let that happened. Um yeah [27:18] some of it to speed and then just look [27:20] it's execution over a long period of [27:22] time and uh you know some of these very [27:25] large companies to varying degrees have [27:28] lost their ability to execute. And so if [27:32] you're talking about a brand new [27:33] platform and you're talking about, you [27:35] know, kind of building for a long time, [27:37] it's like, you know, Microsoft got [27:39] caught with their pants down on Google. [27:41] Um, Microsoft's still like very strong, [27:44] but they missed that whole opportunity. [27:47] They also missed the opportunity. You [27:49] know, Apple was nothing and Microsoft [27:52] fully believed that they were going to [27:53] own mobile computing. They completely [27:55] missed that one. But they were still so [27:57] big from their Windows monopoly they [27:58] could build into other things. So you [28:02] know I think generally the new companies [28:04] have won the new markets. Uh [28:07] and that doesn't mean the big company [28:09] the biggest companies the biggest [28:11] monopolies from the prior generation [28:13] just last a long time is the way I would [28:15] look at it. [28:16] >> Yeah. I I also think we don't quite know [28:18] like it's all happened so fast we we [28:20] actually don't I think we don't yet know [28:22] the shape and form of the ultimate [28:23] products. [28:24] >> Yeah. [28:24] >> Um Right. And so and so like because [28:27] it's it's tempting and this is kind of [28:28] what what always happens. It's kind of [28:30] it's kind of tempting to look I'm not [28:31] saying what that's what these guys did [28:32] on stage but it's kind of tempting to [28:33] look you sometimes you hear the kind of [28:35] reductive version of this which is [28:36] basically it's like oh there's either [28:37] going to be a chatbot or a search engine [28:39] right the competition is between a [28:40] chatbot and a search engine and the [28:42] problem Google has is the classic [28:44] problem of dis you know disruption. are [28:45] you going to disrupt the 10 blue links [28:47] model and swap in you know at you know [28:49] sort of uh AI answers and you know [28:50] potentially disrupt the advertising [28:52] model and then the problem OpenAI has is [28:53] they have the the full you know the full [28:55] chat product but you know they don't [28:56] have the advertising yet and they don't [28:57] have the distribution Google scale [28:59] distribution and so you know you kind of [29:00] say okay that's a fairly that's a fairly [29:02] cl like that'd be straight out of a like [29:04] you know the innovator's dilemma you [29:06] know business textbook like this is just [29:08] a very clear you know one one versus one [29:10] you know kind of dynamic but that [29:12] assumes that you know the mistake that [29:14] you could make in thinking way is that [29:15] assumes that the forms of the product in [29:17] 5 10 15 20 years that that are going to [29:19] be the main things that people use are [29:21] going to be either a search engine or a [29:22] chatbot, right? Um and and you know the [29:25] just there's you know there's just [29:27] obvious historical analogies. One just [29:28] obvious historical analogy is, you know, [29:30] the the personal computer from sort of [29:32] invention in 1975 through to, you know, [29:35] basically 1992, [29:37] you know, was was a was a text prompt [29:39] system, right? Um, you know, and at the [29:40] time, by the way, an interactive text [29:42] prompt was a big advance over the [29:43] previous generation of like punch card [29:45] systems, time sharing systems. Uh, and [29:47] then, you know, it was, you know, 1992, [29:49] so was what, seven, 17 years in, you [29:51] know, the whole industry took a left [29:52] turn into GUIEs and never looked back, [29:54] you know. And then by the way, you know, [29:55] 5 years after that, the industry took a [29:57] left turn in web browsers and never [29:58] looked back, right? And so the very [30:00] shape and form and nature of the user [30:02] experience and how it and how it fits [30:04] into our lives, uh, you know, is is is I [30:06] think still unformed. And so like and [30:08] you know, look, I'm I'm sure there will [30:09] be chat bots 20 years from now, but I [30:10] I'm I'm pretty confident that, um, you [30:12] know, both the current chatbot companies [30:14] and many new companies are going to [30:16] figure out many kinds of user [30:17] experiences that are radically different [30:19] that we don't we don't even know yet. [30:21] Yeah. And by and by the way, that's one [30:22] of the things of course that keeps the [30:23] tech industry fun, which is it, you [30:24] know, especially on the especially on [30:26] the software side, you know, is it's not [30:27] it's not it's not obvious what the shape [30:29] and form of the products are. And [30:30] there's just I think there's just [30:31] tremendous headroom for invention. [30:32] >> As as you're coaching entrepreneurs and [30:34] the entrepreneurs in this room, what [30:35] what else feels different about this era [30:38] or or or other advice that you find [30:40] yourself whether it's around uh sort of [30:42] the talent wars that are going on or [30:44] other aspects that feel unique to this [30:46] era? What what other advice do you want [30:47] to be leaving our entrepreneurs with? [30:50] that's unique to this era. Well, like I [30:52] I I actually think you said the right [30:54] thing, which is this is a unique era. [30:57] And so [30:59] trying to [31:01] learn the organizational design lessons [31:05] of the past or trying to learn um kind [31:08] of too much from the last generation is [31:12] can be deceptive [31:14] because things really are different. [31:15] like the way these you know the way your [31:17] companies are getting built is is quite [31:19] different in in many aspects. Uh and you [31:23] know the types of [31:26] you know what the just like our observ [31:29] observation on like PhD AI researchers [31:33] is just very different than like a [31:36] traditional um engineer full stack [31:39] engineer or something like that. So you [31:42] know I I think you do have to think [31:44] through a lot of things from first [31:46] principles uh because it is different [31:49] and like you know observing from the [31:50] outside it's really different. [31:52] >> Yeah. [31:53] >> Yeah. And I would just offer like I I do [31:54] think things are going to change. So I [31:55] already talked about I think the shape [31:56] and form of products is going to change. [31:58] Um uh and so like I think there's still [32:00] a lot of creativity there. I also think [32:02] and I I I let's say I think that um like [32:06] in a in a world of supply and demand the [32:08] thing that creates gluts is shortages. [32:09] Um right so like when something becomes [32:11] too scarce there becomes a massive [32:13] economic incentive to figure out how to [32:14] unlock new supply and so the the the [32:16] current generation of AI companies are [32:18] really struggling with uh particular [32:20] shortages of of the really talented AI [32:22] researchers and engineers and then [32:23] they're really you know challenged with [32:25] shortage of of infrastructure capacity [32:27] chips and and data centers and power. Um [32:30] I I don't want to call timing on this. [32:31] There will come a time when both of [32:32] those things become gluts. Um and so you [32:35] I don't know I don't know that we can [32:37] plan for that. Um although I I would [32:39] just say the following. Number one, um [32:41] the the the researcher engineer side of [32:43] things, it is striking. It is striking [32:45] to the degree to which um there are [32:47] excellent, you know, outstanding models [32:48] coming out of China now. Um you know, [32:51] and and in a m from multiple companies [32:52] and you know, specifically, you know, uh [32:54] Deepseek and and Quinn and Kimmy. Um it [32:57] is striking how the teams that are [32:58] making those are not, you know, the name [33:00] brand, you know, for the most part, [33:02] these are not like the name brand people [33:03] with their names on all the papers. Um [33:05] and and so like China is successfully [33:07] decoding how to like basically take [33:09] young people and train them up in the [33:10] field. [33:10] >> Well, and XAI to a large extent too. [33:13] >> Yeah. Yeah. And so I I think that I [33:15] think there's going to be and look it [33:17] makes sense up until it it makes sense [33:19] that for a while it's going to be this [33:20] super esoteric skill set and people are [33:21] going to pay through the nose for it. [33:23] But like you know there's no question [33:24] the information is right being [33:26] transferred into the environment. People [33:27] are learning how to do this. Um you know [33:29] college kids are figuring it out. Um, [33:31] and so, um, you know, there's there's [33:33] and I don't know that there's ever going [33:34] to be a talent glut per se, but like I [33:35] think for sure there's gonna there's [33:36] gonna be a lot more people in the future [33:37] who of course know know how to build [33:38] these things. Um, and then and then by [33:40] the way also of course you know AI [33:42] building AI, right? So that the the the [33:43] the tools themselves are going to be [33:45] better better at at contributing to [33:46] that. And so and and I think that I [33:49] think this is good because I think that [33:50] you know the current level of of [33:51] shortage of of engineers and researchers [33:53] is is is too constraining. And then and [33:54] then on the chip side I don't I don't [33:55] want to I'm not a chip guy and I don't [33:56] want to call call it specifically but [33:58] like it it's never been the case. It's [33:59] never been the case in the ship industry [34:01] that there's ever, you know, every every [34:02] shortage in the ship industry has always [34:04] resulted in a glut uh because the the [34:06] profit the profit pool of a shortage, [34:08] the margins get too big, the incentive [34:10] uh for other people to come in and [34:11] figure out how to commoditize the [34:12] function get too big. And so, you know, [34:14] Nvidia has like, you know, the best [34:15] position probably anybody's ever had in [34:16] chips. But notwithstanding that, I I [34:18] find it hard to believe that there's [34:20] going to be this level of pressure on [34:21] infrastructure in 5 years. [34:22] >> Yeah. And even if the bottleneck within [34:24] the infrastructure moves, so if if it [34:26] becomes power, if it becomes cooling or [34:29] or or anything else, then you'll have a [34:31] chip glut for sure. Yeah. [34:33] >> So So I I think over the I I would just [34:34] say this, it's likely the challenges [34:36] that we that we all have in five years [34:37] from now are going to be different [34:38] challenges. [34:39] >> Yeah. Yeah. Yeah. Like don't [34:42] don't definitely this industry of all [34:44] industries don't look at us as static [34:46] like you know the positions uh could [34:49] change very very fast. [34:51] Let's actually close on a more of this [34:53] macro note. Mark, you mentioned China. [34:55] Last month, we were in DC and what one [34:57] of the big questions the senator has is [34:59] how should we make sense of sort of the [35:01] state of the AI race visa v China. Do [35:03] you you want to share just the the high [35:05] level um summary of what what you shared [35:06] with them? [35:07] >> Yeah. So my sense of things and I and I [35:09] think the current I think the current if [35:11] you just observe currently specifically [35:12] like deep sea quan and and these models [35:14] coming out of China I my my sense [35:16] basically is like I would say the US [35:20] specifically in the west generally but [35:21] you know more and more specifically the [35:23] US is like the conceptual innovations [35:24] are you know have been you know coming [35:26] out coming out of coming out of the US [35:27] coming out of the west you know kind of [35:29] the the big kind of conceptual [35:30] breakthroughs um uh China is extremely [35:33] good at picking up ideas and [35:35] implementing them and scaling them and [35:36] commoditizing them and and you know [35:38] that's that they do that obviously [35:39] throughout the manufacturing world. Um [35:41] and and they're doing it now very I [35:43] think successfully uh sort of in AI. Um [35:46] and so I would say that they're running [35:47] they're running the catch-up game like [35:48] really well. Um you know and then [35:50] there's there's sort of always this [35:52] question of like how much of that is [35:53] like being done let's just say like [35:55] authentically um uh you know through [35:57] hard work um and smart people and then [35:59] how much is being done with maybe a [36:00] little bit of help um maybe a little USB [36:03] stick uh in the middle of the night uh [36:05] you know kind of help um Okay. [36:08] >> So, uh, you know, there's always a [36:09] little bit of a question, but like [36:10] either either way, uh, you know, they're [36:11] they're doing a great job. Uh, obviously [36:13] they they aspire to, you know, more than [36:15] that. Um, and there are many very smart [36:17] and creative people in China. And so, [36:18] you know, it will be interesting now to [36:20] see, you know, the level to which the [36:21] conceptual breakthroughs start to come [36:22] from there and whether they whether they [36:24] pull ahead. Um, and so, but like I would [36:26] say like what we tell people in [36:28] Washington is like look this this is a [36:29] foot this is now this is a full-on race. [36:31] It's a foot race. It's a game of inches. [36:33] Like we're not going to have a 5year [36:35] lead. we're going to have like maybe a [36:36] six-month lead. Like we have to run [36:37] fast. We have to win. Like we we have to [36:39] we have to do this. We we can't and then [36:41] we can't put constraints on our [36:42] companies that the the the Chinese [36:44] government isn't putting on their own [36:45] companies. And so um you know we'll just [36:47] lose and you know do do you really want [36:49] do you really want to wake up in the [36:50] morning and live in a world you know [36:51] really controlled and run by Chinese AI? [36:53] Most of us would say no we don't want to [36:55] live in that world. Um and so um so so [36:59] that so there's that and I would say I [37:01] feel moderately good about that just [37:02] because I think that I think we're we're [37:03] really good at software. um you know the [37:05] minute this goes into you know embodied [37:07] AI in the form of robotics I think [37:09] things get a lot scarier and you know [37:11] this is the thing I'm now spending time [37:12] in DC trying to really educate people on [37:14] which is you know the ch because the US [37:16] and the west have chosen to [37:18] de-industrialize to the extent that we [37:19] have over the last 40 years um you know [37:22] China China specifically now has this [37:24] giant industrial ecosystem for building [37:26] you know sort of mechanical electrical [37:28] and um uh and semiconductor and now [37:31] software you know devices of all kinds [37:33] including phones and drones phones and [37:34] cars um and robots. Um and so uh you [37:39] know there's going to be a phase two to [37:40] the AI revolution. It's going to be [37:41] robotics. It's going to happen you know [37:43] pretty quickly here I think. Um and when [37:46] it does like even if the US stays ahead [37:47] in software like the robots's got to get [37:49] built and that's not an easy thing. And [37:51] it's not just like a company that does [37:53] that. It's got to be an entire [37:54] ecosystem. Um and it's you know it's [37:56] going to be you know it's going to you [37:57] know like I mean you know the car [37:59] industry was not three car companies. It [38:00] was thousands and thousands of of of [38:01] component suppliers building all the [38:03] parts. And it's been the same thing for [38:04] airplanes and the same thing for [38:05] computers and everything else. Um it's [38:07] going to be the same thing for robotics. [38:08] Um and you know by default sitting here [38:11] today that's all going to happen in [38:12] China. And so even if they never quite [38:14] catch us in software they might just [38:15] like lap us in hardware and and that'll [38:17] be that. Um you know the good news is I [38:20] I think there's a growing awareness in [38:22] there's a growing awareness I would say [38:23] across the political political spectrum [38:25] in the US that like de-industrialization [38:26] went too far. Uh and there's a growing [38:28] desire to kind of figure out how to [38:29] reverse that. Um, and um, you know, I [38:32] say I'm guardedly optimistic that we'll [38:33] be making progress on that, but I think [38:35] there's a lot of work to be done [38:37] >> on that call to arms. Uh, let's wrap. [38:39] Uh, thank you Mark and Ben. To to wrap [38:42] up, I'd like to welcome back. [38:44] >> Thank you, everybody. [38:46] [Applause] [39:04] [Music]