Amodei Defends AI Rules as VCs Warn Regulations Threaten US Lead

Amodei为AI监管辩护，风投警告监管恐危及美国领先地位 联邦AI监管以及对数据中心的抵制可能迫使算力基础设施迁往海外，同时削弱美国对中国的技术主导优势。 The Besties探讨并辩论了Anthropic首席执行官Dario Amodei近期的文章，严厉批评其呼吁推行FAA式AI监管的主张，并警告称扼杀开源创新可能将技术主导权拱手让给中国。主持人们分析了AI末日论与科技引发的财富集中如何激发强烈的民粹主义反弹，其表现在选民对数据中心的反对以及整个政治光谱中社会主义情绪的上升。他们还分析了市场动态、递归自我改进（RSI）的潜力、美国司法部对Andreessen Horowitz的调查，以及即将到来的2026年中期选举的战略信息传播需求。 Anthropic首席执行官Dario Amodei发表了两篇系列文章，驳斥有关其搞“监管俘获”的指控，但David Sacks认为他所倡导的联邦监管框架实际上仍起到了监管俘获的作用。 (0:00) 各州州长对数据中心带来的政治阻力，加上美债收益率上升，正给非投资级AI实验室造成经济和算力瓶颈。 (5:30) Chamath Palihapitiya认为，闭源AI实验室如果希望外部研究人员进行真正的安全评估，就应该公开其模型隐藏的“思考标记”（thinking tokens）。 (8:22) David Sacks反对针对AI推行强制性的FINRA式自律组织（SRO），主张采取独立的、自愿性的MPAA式分级模式，以保护创新活力。 (12:03) 开源AI套件的表现正迅速超越专有方案，在提供更低成本替代选择的同时，推动竞争格局从封闭实验室护城河转向开放创新。 (25:59)

The All-In Podcast hosts argue that tech-industry alarmism and proposed AI regulatory frameworks risk entrenching closed-source incumbents, outlawing decentralized open-source development, and inciting a populist backlash. Simultaneously, systemic cost-of-living pressures are eroding voter faith in market capitalism, fueling political appetite for technological deceleration and socialist intervention ahead of the 2026 midterms. Regulatory Capture and the Open-Source Governance Debate (12:03) Anthropic CEO Dario Amodei published a two-part essay rejecting charges of regulatory capture and doomerism, responding to Gavin Baker's claims that some within Anthropic aim to become the sole surviving private company. David Sacks disputes Amodei's stance, arguing that lobbying for state-backed pre-release testing regimes—analogous to FINRA, the FDA, or the FAA—fits the definition of regulatory capture. Sacks contends that existing product liability risks already incentivize model safety, advocating instead for voluntary, independent standard-setting modeled after the MPAA to avoid administrative delays that could concede technological leadership to China. Chamath Palihapitiya criticizes closed frontier labs for concealing intermediate reasoning ('thinking tokens') while seeking regulatory moats, arguing that open-source models provide greater verifiable transparency. Sacks warns that proposed compliance regimes will mandate centralized rollback and continuous monitoring capabilities that decentralized, locally run open-source models cannot fulfill, functioning as a de facto ban. Chamath notes that open-source models paired with flexible harnesses are already outperforming closed alternatives at lower costs, warning that heavy-handed domestic regulations will trigger capital flight akin to the collapse in foreign direct investment China experienced following its tech crackdown. Public Resistance to Infrastructure, Automation, and Tech Wealth (5:30) The panel examines mounting political resistance to AI infrastructure, highlighting restrictions on data centers initiated by Governors Greg Abbott in Texas and Josh Shapiro in Pennsylvania. Chamath notes that an environment of rising Treasury yields increases the cost of capital, trapping unprofitable, non-investment-grade frontier AI labs between capital market constraints and localized energy limits. Jason Calacanis attributes public hostility to acute economic anxieties, inflamed by tech leaders broadcasting mass unemployment forecasts—such as Amodei predicting AI could eliminate 50% of entry-level knowledge worker jobs within 1–5 years. Calacanis argues that autonomous services like Waymo and Zipline pull capital out of local economies and redirect it to Silicon Valley, which he predicts will drive municipal bans on autonomous vehicles in major cities. David Friedberg and Chamath observe that data centers have become physical scapegoats, perceived by a struggling public as elite monuments for tech billionaires while working-class citizens endure persistent inflation. Recursive Self-Improvement and Geopolitical Regulatory Arbitrage (45:01) David Friedberg steel-mans the safety concerns of frontier labs, pointing out that internal testing reveals severe tail risks, including autonomous cyberweapons and engineered biological pathogens. However, Friedberg asserts that traditional regulatory models—such as periodic six-month model audits—are fundamentally incompatible with Recursive Self-Improvement (RSI), where autonomous agent networks continuously build and deploy successor models without human intervention. Because RSI requires only compute chips, power, and internet connectivity, Friedberg and Sacks argue that localized regulations will merely drive labs into geopolitical regulatory arbitrage, relocating infrastructure to jurisdictions like Iceland or China. Sacks emphasizes that keeping development within U.S. jurisdiction and maintaining competition across diverse alignment philosophies—such as Anthropic's constitutional ethics versus Elon Musk's maximal truth-seeking—is a safer approach than concentrating control in a centralized government gatekeeper. Economic Affordability, Socialism, and 2026 Midterm Pressures (66:03) Turning to the 2026 midterm elections, the hosts analyze shifting Senate odds that moved to a 52/48 toss-up across 35 seats, with Republicans defending 23. While Sacks argues that early polls exhibit a historical Democratic bias and points to positive indicators like declining drug costs and low border contacts, Friedberg highlights structural affordability crises in housing, education, and healthcare. Friedberg cites polling showing 53% of young conservatives support government-run grocery stores and notes that overall Republican favorability toward capitalism fell from 72% in 2019 to 61% in 2026, with 42% viewing the system as rigged. Sacks argues that persistent inflation is concentrated in heavily subsidized, regulated sectors—noting healthcare costs rose near 300% from 2000 to 2026 while market-driven software prices dropped 67%. Reviewing a Cato Institute estimate placing the cost of the Democratic Socialists of America platform between $71 trillion and $211 trillion, Chamath warns that aggressive deficit spending would trigger bond-market turmoil and punitive wealth extraction. Calacanis urges the administration to direct executive authority away from performative battles and toward building 1 to 2 million homes and expanding vocational trade schools. The discussion leaves unresolved whether decentralized open-source models can survive impending regulatory frameworks, and whether federal policymakers can resolve core housing and cost-of-living constraints before populist disillusionment forces broader state intervention.

All-In节目主持人们探讨了Anthropic首席执行官Dario Amodei对AI监管的辩护，围绕前沿模型安全、开源生态存亡以及数据中心扩张引发的政治反弹展开辩论。对话将AI监管俘获风险与美国深层的可负担性危机、地缘竞争压力及2026年中期选举动向相联系。 AI治理与开源生存之争 (0:00) Anthropic首席执行官Dario Amodei发文为监管辩护，否认权力集中指控，并坦承AI行业尚未兑现如治愈癌症等承诺。David Sacks直指Amodei的立场属于监管俘获，指出其此前宣称50%初级白领岗位将在1至5年内消失，实质是在游说有利于巨头的法规。Sacks反对设立类似FINRA的AI审批机构，认为这会演变为低效的准入审查，主张借鉴电影协会MPAA的自愿分级机制，依靠产品责任诉讼倒逼企业自我约束。 Sacks进一步警告，以安全为名的开源模型禁令正在成型。由于开源模型去中心化且本地运行，无法被远程撤回或实时监控，若强制要求其符合闭源服务的中央控制标准，将扼杀开源生态。Chamath Palihapitiya指出，开源模型配合Hermes等灵活线束在成本极低的前提下正迅速超越闭源模型，政策重心应是赋能开源创新而非保护封闭垄断者。 基础设施制约与地缘技术竞争 (5:30) AI末日论对现实基建造成了显著冲击。Chamath指出，得克萨斯州与宾夕法尼亚州州长近期发布行政命令限制数据中心扩张，共和党备忘录亦要求高管停止用灾难论恐吓公众以防拖累参议院选情。同时国债收益率攀升推高资金成本，对巨额烧钱的非投资级AI初创构成财务绞杀。Chamath警告，严苛限制将迫使跨国企业将资本转向海外，重演此前中国打击科技巨头导致外商直接投资断崖式下跌的历史教训。 在地缘竞争维度上，Sacks强调中国在全钢火箭回收及生物技术领域的快速迭代，证明过度官僚审查令美国处于劣势。尽管Amodei曾呼吁对英伟达芯片出口实施限制以保持技术代差，但主持人们指出，压制国内数据中心建设只会迫使算力外流至冰岛、中国或海上平台进行监管套利，不仅无法阻断前沿研发，还会削弱美国在核心双用途技术上的主导权。 自主进化挑战与黑盒推理透明度 (8:22) 针对前沿实验室管理层的担忧，David Friedberg阐释了其应对生物武器设计与自动化网络攻击等危险能力的紧迫感。他提出“递归自我改进”（RSI）理论，即AI智能体自主设计并部署性能更强的后代模型。Friedberg认为，非线性技术跃迁存在“时间膨胀”效应，这种连续自动化迭代将使每半年一次的静态政府审查彻底失效，任何基于地理边界的集中式政策都难以约束云原生算力的流转。 对此，Chamath反驳称前沿实验室若真追求安全，就应停止隐藏模型的“思考标记”（thinking tokens）。当前Claude和ChatGPT等闭源模型刻意遮蔽中间推理链路，剥夺了独立机构核验安全宣称的机会；反观开源模型则允许实时追踪逻辑。Sacks总结道，依靠道德规范编写的“Claude宪法”或单一中央权威无法保障对齐，维持多模型竞争的市场格局才是防范失控的可靠屏障。 民生焦虑、贫富分化与科技反弹 (21:01) 公众对数据中心和AI技术的敌意源于生活成本危机与财富鸿沟。Friedberg引述民调指出，前50%美国人坐拥178万亿美元净资产，而后50%仅拥有6万亿美元，高达63%的民众依靠薪水度日。科技巨头财富跃升至万亿美元级别，却伴随着通胀顽固和购买力停滞，加之高管频繁宣扬岗位淘汰或鼓吹全民基本收入（UBI），严重刺痛了普通工薪阶层。 Jason Calacanis指出，自动化交付与无人车车队的扩张，正将原本留在本地社区的收益汇集到硅谷，势必引发大城市出台行政限制。Friedberg借用勒内·吉拉尔（René Girard）的替罪羊理论分析，公众正将对系统失调的愤怒转嫁到科技精英身上，把数据中心描绘为富豪的“神庙”。主持人们批评科技界丢掉了早期理想主义，异化为贪婪敛财的特权阶层，点燃了基层的减速主义与反资本主义情绪。 中期选举博弈与宏观财政政策分歧 (60:00) 聚焦2026年中期选举，参议院控制权在博彩市场上已收窄至52比48的摇摆状态。Sacks认为早期民调存在系统性偏向民主党的误差，强调执政党应聚焦边境管控、凶杀与药物过量死亡率下降20%、减税以及鸡蛋价格同比下跌39%等政绩。然而Friedberg反驳称，40岁以下保守派选民中已有53%支持政府运营平价杂货店，反映出选民对住房、医疗及教育价格失控的绝望，两党正面临走向民主社会主义的民意暗流。 针对民主社会主义（DSA）施政纲领，Sacks援引卡托研究所评估指出其新增联邦支出将在71万亿至211万亿美元之间，远超全美企业利润（约35万亿美元）及高净值人群资产总和。Chamath测算推行此类纲领将引发美债收益率飙升至8%以上，最终滑向普遍资本没收的债务螺旋。Calacanis呼吁白宫动用行政权力推动公私合作，在24个月内修建100万至200万套平价住房并扩建职业技术培训。 主持人们一致认为，AI监管与算力扩张的争端已不再是单纯的技术辩论，而是与国家财政赤字、阶级矛盾及地缘竞逐深度捆绑的系统性危机。然而讨论未能明确解答的是：在缺乏客观透明标准的前提下，如何在不沦为大企业护城河的同时界定AI的真实风险？面对自动化驱动的财富集中与中产生活成本难题，依靠自愿性行业自律或激进的行政干预能否弥合裂痕，仍是留待未来验证的核心悬念。

## Background 本内容由 YouTube 上的著名播客频道 All-In Podcast 发布于 2026 年 8 月 21 日。根据页面描述，几位主持人（Besties）在本期节目中讨论了 Anthropic CEO Dario Amodei 对监管捕获与末日论的澄清回应、AI 自律监管与数据中心危机、a16z 遭受美国司法部调查事件，以及 2026 年美国中期选举局势与社会通胀矛盾。 ## TL;DR 节目围绕 Anthropic CEO Dario 的文章展开深入辩论，剖析了 AI 末日恐慌对数据中心建设与公众舆论产生的负面反噬。主持人们探讨了模型“思维 Token”遮蔽、开源禁令隐患以及递归自我改进（RSI）对传统监管的冲击。此外，节目还分析了 a16z 面临的“交错董事会”反垄断调查，并结合 2026 中期选举，深刻揭示了高通胀与生活成本危机如何推高社会对科技精英的仇恨及右翼社会主义化倾向。 ## Atmosphere & style 全员主持重聚现场，气氛热烈且充满高能的观点碰撞与智力对锋。几位主持人兼具科技投资人的敏锐与政治评论员的犀利，从最初对 AI 监管和末日推手的怒斥辩驳，逐步升华为对社会阶层矛盾与资本主义失序的沉重反思。整场讨论既有对前沿技术架构的深刻剖析，也有对社会心理与替罪羊机制的精准比喻，充满了理性深刻与激情张力兼备的辩论场域感。 ## Key takeaways - [00:13](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=13s) Dario Amodei 发表双篇长文回应争议，反对将 AI 监管简单归咎为“监管捕获”，并承认 AI 行业目前最严峻的挑战在于尚未兑现其对世界的宏大承诺。 - [01:52](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=112s) David Sacks 批评 Anthropic 和 Dario 协同推行失业预言与极端 AI 敲诈实验，制造生存恐慌，客观上借国家机器推行商业与政治议程。 - [05:29](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=329s) “AI 毁灭论”的散播引发现实负面连锁反应，导致政坛掀起反数据中心风潮、美债收益率上升以及前沿实验室融资困难等危险局面。 - [08:08](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=488s) 闭源模型实验室故意隐藏“思维 Token”（Thinking Tokens），导致外部无法验证其对齐真实性并加剧了末日恐慌，而开源模型过程透明且更具可审计性。 - [12:01](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=721s) 针对 AI 监管架构，Sacks 坚决反对会沦为“AI 版车管所/FDA”且阻碍创新的 FINRA 模式，主张采取类似电影协会（MPAA）的纯民间自律框架。 - [16:39](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=999s) 若对美国 AI 施加如同 FAA 或 FDA 般的重监管而中国不加限制，且因审批漫长导致资本流出，美国将在中美科技竞争中彻底落败。 - [26:00](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=1560s) 监管机构与闭源巨头可能通过强制要求开源模型达到与闭源相同的中央监控与撤回标准，在技术层面变相出台“开源禁令”。 - [21:00](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=1260s) 科技领袖散播失业恐慌与极端的富豪财富集聚，正引发大众强烈情绪反弹，推动社会向右翼社会主义化转型并阻碍基础设施建设。 - [45:01](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=2701s) 递归自我改进（RSI）将使传统静态节点的大模型审批监管彻底失效，逼迫监管框架向动态、实时的监控与安全防护转变。 - [52:14](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=3134s) 防止 AGI 或“天网”风险的正确途径是保持多方竞争与市场选择（如马斯克的“寻真”路线与 Anthropic 的“宪法”路线），而非单极垄断或官僚过度监管。 - [56:13](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=3373s) 美国司法部依据《克莱顿反托拉斯法》对 a16z 展开“交错董事会”调查，嘉宾认为这更像是竞争对手通过 PR 机制进行的恶性报复。 - [01:02:31](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=3751s) 2026 年中期选举民调普遍存在偏差，虽然众议院趋向民主党，但共和党若聚焦边境安全、减税及生活成本改善等核心政绩仍有望守住参议院。 - [01:06:03](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=3963s) 因住房、医疗和教育成本长期高企，右翼保守派出现显著社会主义化倾向，若通胀和阶层极化持续，民主社会主义者（DSA）可能在 2028 年崛起。 - [01:22:24](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=4944s) 民主社会主义者（DSA）拟议的政策成本高达数百万亿美元，缺乏数学理性，本质是社会对“令人厌恶的精英阶层”发泄愤怒与不信任。 - [01:25:04](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=5104s) 数据中心被社会心理塑造为亿万富翁与科技巨头的“山顶圣殿”，在生活成本危机中成为了大众发泄阶级仇恨与不安的具象化替罪羊。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=0s) 开场与 Dario Amodei 的“双篇长文”回应 主持人 JCal（Jason Calacanis）热烈开场，欢迎大家回到“世界第一播客”——All-In Podcast，并依次介绍了全员阵容：Chamath Palihapitiya、David Friedberg、David Sacks 以及他自己（JCal）。没有冗长的寒暄，节目直接切入正题。 JCal 提到 Anthropic 的 CEO Dario Amodei 显然收听了上一期的播客，并针对上周 Gavin Baker 带来的爆料作出了正式回应。Gavin 上周曾声称 Anthropic 内部有人甚至认为“Anthropic 未来将成为全球唯一一家私人公司”。虽然 Anthropic 的工程主管 Scal Douglas 率先驳斥该言论纯属捏造，但 Dario 随后打破沉寂，亲自发表了一篇重磅的“双篇长文”（two-part essay）以正面回应各方质疑： * **关于监管捕获（Regulatory Capture）：** Dario 明确表示反对“监管 = 监管捕获 = 权力集中”这种硅谷式的简化逻辑。 * **关于公关策略与舆论控制：** Dario 坚称自己的表态并非一味卖弄焦虑，而是在 AI 的潜在风险与巨大收益之间保持了平衡。 * **关于项目交付：** Dario 承认，包括 Anthropic 在内的 AI 行业目前面临的最严峻批评，正是“尚未兑现对世界的宏大承诺”，并掷地有声地指出：“真正管用的是治愈癌症，而不是华而不实的营销宣传。” JCal 随后抛出疑问：Sacks 过去一年半一直在节目中对 Dario 提出批评，而 Dario 始终不予理睬；为何 Gavin Baker 一发声，Dario 却立刻坐不住了？ * **主要观点与引用：** * Dario Amodei 表示：“我了解硅谷那种把‘监管’简化为‘监管捕获’和‘权力集中’的思维定式，但我始终认为这是一种过于简单化的认知。” * 关于交付落地，Dario 强调：“到目前为止，对 AI 公司（包括 Anthropic）最中肯的批评就是我们还没能实现那些造福世界的远大承诺。这完全是我们自己的责任。” --- ### [01:52](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=112s) David Sacks 怒斥 Dario 的“安全恐慌营销”与双标行为 针对 JCal 的提问，David Sacks 澄清称 Dario 实际上也对他之前的指责作出了间接回应。Sacks 认为，真正触怒 Dario 的，正是 All-In 播客不断揭露 Anthropic 试图进行“监管捕获”的本质。Dario 极其在意别人质疑其动机不纯，极力想塑造自己“出于崇高目的”的形象；但在 Sacks 看来，无论 Dario 个人意图如何，Anthropic 积极在州和联邦层面推行其偏好的监管框架，客观上就是在利用国家机器来实现自身的政治与商业议程，这毫无疑问就是监管捕获。 Sacks 严厉驳斥了 Dario 声称“自己在 AI 传播上保持中立平衡”的说法，指责其极其不负责任且装聋作哑： * **失业恐慌推波助澜：** 仅在一年半前，Dario 就公开预言“1 到 5 年内，50% 的初级白领/知识工作者将面临失业”。这绝非播客上的随口一说，而是 Anthropic 协同营销团队精心策划并放大的一场媒体运动，甚至引发了奥巴马等民主党高层的转发与炒作。 * **“敲诈研究”蓄意造势：** Anthropic 内部所谓的对齐（Alignment）团队曾发布过一项引发轰动的“AI 勒索研究”。研究人员对模型进行了 200 多次诱导性提示（prompting），直到得出想要的惊悚标题。随后 Dario 与其胞妹（公司总裁）等高管登上《60 分钟》（60 Minutes）节目大肆宣扬。而后续包括英国 AI 安全研究所（UK AI Safety Institute）在内的多家独立机构评估指出，该研究完全是在极端加压的虚构条件下人为操纵出来的结果。 Sacks 总结道，公众之所以对 AI 产生如对“终结者”般的深切恐惧，Anthropic 及其高管正是主要的幕后推手。 * **关键事实与案例：** * **失业预言：** Dario 曾公开声称 50% 的初级知识工作者将在 1–5 年内被 AI 替代，该言论被奥巴马等政界名流广泛传播。 * **敲诈实验争议：** 英国 AI 安全研究所事后证实，Anthropic 宣扬的“模型敲诈行为”是在人工强行诱导 200 余次后拼凑出的极端特例。 * **媒体轰炸：** Anthropic 高层集体登上《60 分钟》等主流政论节目，推波助澜放大了公众对 AI 的生存恐慌。 --- ### [05:29](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=329s) Chamath 警示：“AI 毁灭论”引发的连锁灾难 Chamath 接过话题，指出这种散播恐惧的“AI 毁灭论”（AI Doomerism）已经产生了极其恶劣且深远的现实负面后果。他将当前的危机归结为三个相互交织的恶性循环： 1. **两党反数据中心风潮（政治反噬）：** 恐慌情绪已经渗透至主流政坛。德克萨斯州州长 Greg Abbott 颁布行政命令限制数据中心扩张；宾夕法尼亚州州长 Josh Shapiro 也采取了类似动作。据 Axios 报道，共和党甚至向各大 AI 公司高管发出警告备忘录，要求他们“停止发布挑起公众愤怒的言论”，因为这种反 AI/数据中心的恐慌情绪已经直接威胁到共和党在俄亥俄州极其关键的参议院席位。 2. **美债收益率上升与资本收缩：** 财政部翻倍发债导致无风险利率攀升，投资者不再愿意将资金投向高风险资产，直接抽干了建造大型数据中心所需的资本。 3. **前沿模型实验室的生存危机：** 前沿模型公司（Frontier Labs）本身缺乏投资级（Investment grade）的资产负债表，随着其业务扩张，它们对算力的需求呈指数级增长，却因为舆论反噬和高利率环境而越来越难以筹集资金。 Chamath 感慨，由于行业领袖们这些轻率的“自残式错误”（unforced errors），整个美国 AI 产业已经被逼入了一个原本完全可以避免的危险境地。 * **核心数据与政坛动态：** * **数据中心禁令：** 德州州长 Abbott 与宾州州长 Shapiro 近期相继出台政策，对数据中心建造加码限制。 * **党派选举危机：** 共和党针对俄亥俄州参议院选情向 AI 巨头发出警告备忘录，要求其停止“制造恐慌、拉拢仇恨”（Rage baiting）。 * **金融挤压：** 财政部发债规模翻倍引发收益率高企，削弱了市场对数据中心等基础设施的投资意愿。 --- ### [08:08](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=488s) David Friedberg 的“魔鬼代言人”拆解与“思维 Token”争议 David Friedberg 决定站出来为前沿实验室的管理者们进行“魔鬼代言”（Steel-manning）。他提出，如果换位思考，站在前沿实验室领导者的视角：他们在日常测试中可能看到了普通企业应用中看不到的潜在危险——比如模型被越狱（jailbreak）后用于设计新型生物武器、制造未知病毒、编写网络武器入侵银行账户，甚至通过分析公开算法（如 X 平台的推荐算法）在社交平台上大规模制造假新闻以操纵公众。站在他们的立场，面对一种无法完全掌控的强大技术，产生担忧并寻求约束是极其自然的反应。 针对 Friedberg 的辩护，Chamath 提出了关键的反驳：如果这些前沿实验室真的心存敬畏、心怀坦荡，他们最应该做的一件事就是**公开模型的“思维 Token”（Thinking Tokens）**。 Chamath 指出，无论是 Anthropic 的 Claude 还是 OpenAI 的封闭模型，它们在输出答案时都将中间的思考推导过程（Thinking Tokens）故意隐藏（obfuscate）掉了。用户只能看到最终结果，而无法看到模型在推演时的真实逻辑。相比之下，像 Nvidia 支持的很多开源模型，其思考过程是完全透明可被实时审计的，因此开源社区几乎从未产生过这种“AI 摧毁人类”的夸张情绪。封闭模型实验室一边藏匿内部思考过程，一边要求大众盲信他们对未公开数据的自我解释，这种做法极不透明。 * **争论焦点：** * **前沿安全隐患：** 模型越狱可能导致生物武器设计、新型病毒合成、网络金融攻击及算法舆论操控。 * **思维 Token 隐藏（Thinking Tokens Obfuscation）：** 闭源模型（如 Claude、OpenAI）遮蔽了推导过程，导致外部无法验证其对齐真实性；而开源模型因推理过程透明，几乎不产生此类末日恐慌。 --- ### [12:01](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=721s) FINRA、MPAA 还是 DMV？AI 监管机构的激辩 话题随后转向具体的监管组织架构。JCal 提到，DeepMind 创始人 Demis Hassabis 此前曾倡议建立一个类似于美国金融业监管局（FINRA）的 AI 自律监管机制，即由懂技术、懂商业化的行业参与者组成联合体，指派科学家互相审计对方的模型，以代替政府直接插手。当时 Dario、Sam Altman 乃至 Elon Musk 都曾对此表示赞同，但 Sacks 最近却在 Twitter 上公开反对 FINRA 模式。 Sacks 详细澄清了他的态度与立场区分： * **FINRA 模式（极度糟糕）：** FINRA 在金融领域本质上是一个极其死板、专门保护华尔街巨头和既有垄断者（Incumbents）、阻碍初创公司创新的机构。把这种模式搬到极其动态的 AI 领域，简直是灾难。 * **SRO 与 MPAA 模式（可接受的自律方向）：** 真正适合 AI 的自律组织（SRO）应当类似于美国电影协会（MPAA）——完全独立于政府，由行业自发制定分级与安全标准，用行业自律来抵御政府重手干预。 * **“AI 的 DMV”：** Sacks 强调，Dario 和 Demis 所主张的“FINRA 模式”根本不是真正的自律，而是一个打着自律旗号、实则直接向政府汇报的新官僚机构。它要求模型在发布前接受 mandatory pre-release testing（强制性发布前测试）。这无异于为 AI 设立了一个“车管所”（DMV）或“食品药品监督管理局”（FDA），所有模型都必须在长龙般的队伍中等待审查，这将彻底摧毁美国的 AI 创新速度。 * **模式对比与关键概念：** * **FINRA 模式：** 表面为自律组织，实则依赖政府授权，易被巨头利用进行市场壁垒建构（Sacks 极其反对）。 * **MPAA 模式：** 电影协会式的纯民间自律框架，不向政府汇报，通过设定行业规范避免政府强行插手（Sacks 与 Musk 认可的方向）。 * **“AI 的 DMV/FDA”：** 指责前沿实验室试图引入发布前强制审批流程，导致技术迭代陷入官僚泥潭。 <ElicitationsGroup message="想要深入探索本期播客的后续议题？"> <Elicitation label="了解 AI 领域的“开源禁令”争议" query="请详细介绍 All-In 播客中关于 AI 开源禁令（Open Source Ban）、递归自我改进与 FDI 影响的讨论。"/> <Elicitation label="分析 a16z 遭受 DOJ 调查的事件" query="请梳理 All-In 播客中关于美国司法部（DOJ）对 a16z 开展“交错董事会”（Interlocking Directorates）调查的分析与观点。"/> </ElicitationsGroup> ### [16:39](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=999s) 监管捕获与中美 AI/科技竞争 David Sacks 和 Jason Calacanis 等人探讨了过度的监管如何导致“监管捕获”（Regulatory Capture），并使美国在与中国的科技竞争中处于劣势。Jason 举例指出，中国在前一天成功发射并回收了纯钢制火箭，且从上次爆炸事故到这次突破的迭代速度惊人；相反，SpaceX 的 Elon Musk 还在等待美国联邦航空管理局（FAA）对钢制星舰（Starship）的审批。同样的趋势也发生在生物科技领域：由于美国 FDA 审批周期极其漫长，且临床试验阶段缓慢，许多生物科技投资和试验正大量向中国转移，因为中国重新修订并优化了进入临床三期试验的规则。 David Sacks 强调，Dario Amodei（Anthropic CEO）在最近的文章中公开呼吁建立一个模仿 FAA 的 AI 监管机构，这种主张完全是在自毁前程。如果美国对本国 AI 企业施加此类强监管，而中国不限制其国内公司，美国将在 AI 比赛中被中国彻底击败。 - **火箭与航天**：中国成功迭代并降落纯钢制火箭，而 SpaceX 的钢制 Starship 仍在等待 FAA 的审批批准。 - **生物科技（Biotech）**：美国 FDA 审批缓慢导致大量生物科技研发和临床投资流向中国。 - **AI 监管观点**：Sacks 认为以 FAA 或 FDA 为模板监管 AI 将直接导致美国在与中国的竞争中落败。 --- ### [18:05](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=1085s) 产品责任诉讼、行业透明度与“私下勾结”的隐患 David Sacks 澄清，虽然他反对建立“AI 版 DMV（车管所）”，但这并不意味他反对 AI 安全。他认为顶级实验室（Frontier Labs）重视安全是值得肯定的，例如 Sam Altman（OpenAI CEO）此前发文称将放慢部分前沿模型的发布节奏，因为安全技术的跟进速度未能赶上模型能力的提升。Sacks 认为，这些公司自主重视安全并非出于善意，而是为了规避巨大的产品责任诉讼（Product Liability）风险。例如，Meta 目前因产品导致社交媒体成瘾而被起诉，面临超过 1 万亿美元的索赔风险，此前还在洛杉矶的一起关于引发身体形象焦虑（Body Image Issues）的陪审团诉讼中败诉并被判赔 600 万美元。 Sacks 提出，行业协调安全的正确方式应当是公开且透明的：发布技术论文、撰写博客文章、参加学术研讨会（Symposia）以及进行公开征求意见（RFC）。这些公开实践能让安全评估接受公众和同行的检验（Contestable）。反之，如果各家公司在 FINRA（美国金融业监管局）式的自律组织或密室中私下会晤，制定不透明且不可质疑的标准，最终只会演变成监管捕获与反竞争行为（Anti-competitive Behavior）。然而 Chamath Palihapathi 随后补充道，从技术细节来看，要求公司公开所有思考过程是不切实际的：若要展示从输入 Token 到输出 Token 的完整追踪（Trace & Span），等于直接公开了模型的核心配方。Anthropic 甚至不愿意向付费 API 客户分享其思考 Token（Thinking Tokens），更不可能与 OpenAI 等竞品共享。 - **产品责任与诉讼**：Meta 面临超 1 万亿美元的社交媒体成瘾诉讼，此前在洛杉矶败诉赔偿 600 万美元；这是促使 AI 巨头不敢贸然发布不安全模型的主因。 - **透明标准 vs 密室协商**：Sacks 倡议通过论文、博客、RFC 等公开形式制定安全标准，坚决反对封闭式的自律组织（SRO）。 - **技术阻碍**：Chamath 强调模型内部的 Trace 和 Thinking Tokens 属于核心商业机密，公司绝不会在同行面前公开。 --- ### [21:00](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=1260s) 中美 AI 胜负的连锁反应与公众反 AI 情绪 Jason Calacanis 提出了一个假设性问题：如果 Dario 的监管倡议成功并导致美国 AI 研发放缓，而中国的 AI 实验室通过递归自我改进（Recursive Self-Improvement, RSI）取得了全球最顶尖的模型，这对美国的普通工人、企业和消费者意味着什么？对于“中国获胜”的口号，Jason 认为并不会出现单一赢家吃干抹净的情况，而是渐进式的改进。 随后，Jason 分析了美国大众对数据中心和 AI 产生强烈抵触情绪（Backlash）的根源。尽管像 Greg Abbott（德州州长）和 Josh Shapiro（宾州州长）这样原本极力支持数据中心建设的政治家，最近也开始转变成反数据中心的立场。表面上看是担心电网负荷和用水问题，但深层原因是选民的强烈反弹。公众之所以对 AI 如此愤怒，不仅是因为科幻电影塑造的终结者（Skynet/Terminator）恐惧，更是因为科技领袖们的言论：Dario 预测人们将失去工作，Sam Altman 研究无条件基本收入（UBI）并告诉大众未来不必工作。对那些没有积蓄、饱受通胀折磨、看着最低工资停滞不前却目睹科技高管赚得盆满钵满的普通劳动者来说，这些言论极其刺耳。人们感到无力并开始“磨亮吉利服与ギ洛チン（Guillotine，断头台）”，进而通过投票支持社会主义者、阻挠数据中心建设，甚至发生联合健康保险（UnitedHealthcare）CEO 被杀、Sam Altman 住宅遭汽油弹袭击等极端事件。 Chamath 对此进行了微调补充：公众的愤怒源于“氛围（Vibes）”。过去十年的科技巨头并不像 Steve Jobs 那样讨人喜欢或令人仰慕，如今这批 AI 掌门人被外界认为更加不可信。普通人眼睁睁看着他们从亿万富翁变成万亿富翁，自然产生了强烈的抵触和仇恨心理。 - **政治转变**：德州州长 Abbott 和宾州州长 Shapiro 等政治家迫于选民压力，对数据中心项目态度转冷。 - **社会阶层矛盾**：通胀居高不下、最低工资停滞与科技高管暴富形成鲜明对比，推高了大众对 UBI 论调和 AI 的敌意。 - **极端社会事件**：提及了 UnitedHealthcare CEO 遇害案以及 Sam Altman 住宅遭扔汽油弹等社会暴力事件。 --- ### [26:00](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=1560s) 模型 Harness 架构与开源禁令的潜伏危机 Friedberg 继续追问：若美国 AI 研发因监管放缓而让中国取得领先，对美国经济的具体影响究竟是什么？Sacks 回答称，这类似于中国赢得了互联网时代——如果今天的 Google、Microsoft 和 Meta 全部换成百度、阿里巴巴和腾讯，美国的经济活力将大打折扣，财富流失，且在军用和双用途（Dual-use）AI 领域全面落后。 Chamath 则持不同观点，他认为只要美国政府不强制禁止开源，美国消费者和企业就依然会赢得胜利。他举例说明了模型与“ Harness（框架/套件，如 CodeEx、Claude Code、Hermes 等）”的关系： 1. **闭源模型**：虽然在单体基准测试（Benchmark）中得分很高，但一旦套上 Harness 运行，其能力往往会出现衰减（Decay）。 2. **开源模型**：当与其他开源工具或 Harness 结合时，其综合能力反而会显著提升，且成本极低。 Chamath 预测，未来一年内将有海量数据证明，使用开源技术不仅性能更强，而且比闭源方案便宜得多。然而，Sacks 警示称，**开源禁令（Open Source Ban）已经在路上**。监管机构和大公司不会直接宣称“禁止开源”，而是会要求“开源模型必须遵守与闭源模型完全相同的安全标准”。由于开源模型一旦发布就无法像闭源服务那样被中央集中监控、撤回或回滚（Rollback），这种“公平标准”实际上将直接导致开源模型因无法合规而被变相禁止。这正是 Dario Amodei 早在 2023 年参议院听证会时就铺垫过的逻辑——宣称开源模型因不可控而具备潜在危险。 - **Harness 现象**：闭源模型在 Harness 框架下能力容易衰减，而开源模型与开源 Harness 结合后性能大幅提升且成本极低。 - **开源禁令机制**：通过强制要求开源模型达到与闭源模型相同的中央监控与撤回标准，在技术层面变相封杀开源 AI。 - **关键人物策略**：Sacks 指出 Anthropic 和 OpenAI 将提供算力并资助监管标准制定，从而完成对整个 AI 行业的技术垄断与监管捕获。 ### [32:05](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=1925s) 全球监管博弈与资本流向：从可口可乐到中国科技监管 关于封禁开源模型带来的经济后果，主持人与嘉宾展开了激烈争论。Chamath 以可口可乐（Coca-Cola）和辉瑞（Pfizer）等跨国巨头为例指出，如果美国政府强行出台高压且低效的监管政策，这些全球化运营的公司不会坐以待毙。它们会将资本分配和业务重心迅速转移至海外市场，导致美国的外国直接投资（FDI）出现断崖式下跌。他进一步以中国此前对科技巨头的监管整顿（如蚂蚁集团 IPO 前夕马云发表演讲引发的后续监管）为例，强调高压干预会导致市场资本急剧流失。一旦这种政策在美国落地，资本市场和企业投资决策将瞬间做出反应。 尽管 David Sacks 认同禁令会带来灾难性后果，但他指出监管机构和政客往往出于政治目的强行推动政策，正如当前各界对数据中心建设的抵制。Chamath 则反驳称，大众抵制的实质并非反对科技本身，而是不满特定科技精英一跃成为万亿富豪。 - **跨国企业应对**：强监管将促使可口可乐、辉瑞等公司将资本撤出美国，加速海外布局。 - **历史案例对比**：以中国蚂蚁集团上市受阻及后续科技监管为例，说明资本和 FDI 在强干预下的流失速度。 - **数据中心争议**：数据中心预估将占美国 GDP 的 3%，虽然对股市拉动极大，但公众对利益分配不均存在强烈反弹。 --- ### [35:08](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=2108s) 自动驾驶、失业焦虑与公众信任危机 话题转向公众对 AI 和通用人工智能（AGI）的恐惧根源。Jason Calacanis 认为，大众最深层的焦虑在于对自身生计和工作岗位的担忧。他以自身投资的 Zipline、Uber 和 Waymo 等公司为例，指出自动驾驶技术的普及正直接冲击传统就业。以无人驾驶出租车（Robotaxi）和无人机配送为例，原本留在本地社区的劳动力报酬（如每小时 18 至 21 美元的司机收入），如今全部流向云端并集中到硅谷巨头手中，这势必引发纽约、波士顿等大城市的政策反弹与禁令。 为了缓解这种社会焦虑，Jason 提议行业需要明确告知大众未来就业的去向，甚至可以通过提高最低工资或提供技能培训来增强保障。他赞扬了英伟达（Nvidia）CEO 黄仁勋（Jensen Huang）和 Meta 的 Mark Zuckerberg 在沟通上的表现：黄仁勋强调“取代你的不是 AI，而是掌握 AI 的人”，而 Zuckerberg 则通过开源工具和建立蓝领工人培训计划（如光纤铺设等建造业岗位）来赋能大众。相比之下，Dario Amodei 和 Sam Altman 此前过于激进的失业预测（如 Dario 曾预测 50% 的初级知识工作者将在一年内失业）极大地加剧了社会恐慌。 - **自动驾驶推进**：Zipline 与 Uber 合作计划实现每天 100 万次配送；Waymo 车队规模每 6 个月翻倍；Tesla 准备推出 Cybercab。 - **财富流向转移**：无人驾驶剥夺了基层司机的社区收入，将财富进一步浓缩到硅谷科技巨头手中。 - **失业预测反思**：Dario Amodei 曾声称一年内将有 50% 的初级知识工作者失业，但这并未在后来的实际就业数据中得到印证。 - **应用落地现状**：Jason 透露自身业务在引入 AI 工具后效率大增，不仅没有裁员，反而因为业务扩张新增了 7 个岗位需求。 --- ### [40:18](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=2418s) 硅谷的精神蜕变与 Dario Amodei 的“超级智能”愿景 Chamath 深刻反思了硅谷近年来文化与价值观的转变。他指出，曾经的硅谷是边缘人、怪人和理想主义者的天堂，大家追随的是 Bill Campbell、Mike Moritz、John Doerr 或 Steve Jobs 这样的领袖；而如今的硅谷却演变成了类似哈佛大学的“履历认证机构”（Credentialing place），充斥着对资本的疯狂追逐。这种急功近利且缺乏社会担当的表现，进一步加剧了外界对科技行业的敌意。 David Sacks 则为 Anthropic 创始人 Dario Amodei 进行了辩护，认为他的驱动力并非金钱，而是纯粹的理想主义。然而，Dario 所憧憬的“由机器掌管资源、人类无所事事”的超级智能（ASI）未来，在普通大众看来极具乌托邦式的科幻感与反乌托邦式的威胁感，因而难以建立信任。此外，Jason 补充提到，Anthropic 的七位联合创始人已将约 8% 的股份（占公司总股权近 10%）捐赠给信托基金，并限制个人持股，这也间接证明了其团队对理想的追求。 > "It is a place you come to to say, 'Oh, I came to Silicon Valley, I worked at a startup, I got into YC.'... In that breach, all we have done is focused on making money hand over fist." — Chamath Palihapitiya - **硅谷文化演变**：从早期理想主义的创新圣地，转变为逐利驱动、注重履历背书的功利圈子。 - **股权捐赠细节**：Anthropic 创始人团队将大量股份注入信托基金，个人持股上限控制在 1.8% 左右。 - **公关策略批评**：公关需要敢于正面应对争议、长篇沟通并提前化解问题，而非单方面散播恐惧以获取政府监管控制权。 --- ### [45:01](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=2701s) 递归自我改进（RSI）与 AI 自主演进的终局 Friedberg 提出了 AI 技术演进中最核心的哲学与技术议题：递归自我改进（Recursive Self-Improvement, RSI）。RSI 理论认为，当 AI 模型能够部署大量 Agent 并通过协作训练出比自身更强的下一代模型时，整个系统将实现完全自动化和自我进化，从而引发技术奇点（Singularity）或人工超级智能（ASI）的诞生。 Friedberg 强调，如果 RSI 确实成立，那么传统的“人工干预、每半年进行一次大型模型训练并提交政府审批”的监管模式将彻底失效。未来所需的将是持续性的监控与警示服务。由于这种自我演进不需要依赖特定国家的行政许可，甚至可能借助太空卫星等外部基础设施运行，现有的监管框架面临着根本性的挑战。 - **递归自我改进（RSI）定义**：AI 模型通过自主部署 Agent 团队进行研发，实现连续、无须人类干预的自我迭代与升级。 - **监管逻辑转变**：从传统静态的大模型节点审批，转向动态、持续的实时监控与安全防护。 - **行业路线分歧**：业内对于完全放手让 AI 独立演进，还是坚持由人类架构师掌控每一步演进路线，仍存在深刻的分歧。 ### [47:34](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=2854s) 递归自我改进（RSI）与全球数据中心监管困境 Friedberg 提出了关于递归自我改进（Recursive Self-Improvement, RSI）的严峻现实：如果 AI 系统具备了自我演进的能力，其运行就不再极度依赖前沿实验室科学家等人类团队的实时干预。理论上，任何智商足够高的个人或小团体，都可以在世界上甚至太空中的任何地方启动一个“RSI 工厂”。无论是试图禁止建设数据中心，还是设立专门的监管机构，本质上都是“徒劳无功的勾当”（fool's errand）。因为在开放的互联网和算力世界中，算力运行的核心条件极其简单——芯片、电力和网络连接。实验室普遍在讨论 RSI，如果这一趋势不可避免，他们可以将数据中心设在冰岛、中国乃至任何给予其自由的地区。 Chamath 补充指出，与其将前沿实验室驱逐到海外或控制力之外的地区，不如建立能将这些实验室和数据中心保留在本土的机制。只要这些基础设施工厂留在美国国内，监管部门就拥有司法管辖权以及监控和警察执法能力。因此，政策的核心应当是吸引并留住前沿实验室，而不是将它们赶走。 - **RSI 运行要素**：仅需三大基础——芯片（chips）、电力（power）、通信连接（communication connection）。 - **避风港地理选择**：冰岛、哈萨克斯坦、海洋浮动平台乃至太空（space）。 - **政策建议**：吸引并保留本土前沿实验室，维持监督与司法管辖权，而非盲目围堵。 --- ### [49:08](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=2948s) 时间膨胀、混合专家模型与 RSI 的演进路线 针对“RSI 距离我们有多远”的问题，Friedberg 解释了 AI 模型开发中存在的“时间膨胀”（time dilation）现象。传统技术发展通常呈线性规划，但在 AI 领域，突破性进展往往包含不可预测的阶梯式跨越，这导致预测 AGI 或技术奇异点（singularity）的到来变得极其困难。例如，某种前沿模型可能会在短短一周内突然攻克八个未被证明的数学定理。 从技术架构来看，当前许多先进模型均采用了混合专家（Mixture of Experts, MoE）概念，即由大量小型子模型相互调用来完成复杂任务。这种结构与生物学机制高度相似——就像基因产生蛋白质，蛋白质之间形成复杂网络以执行生命活动一样，AI 模型也在向小型化、模块化和高效率的方向演进。未来这种包含无数子模型、规定其调用逻辑与功能的架构，极难由人工架构师直接设计，必须通过“测试与部署”（test and deploy）的自动化系统完成。因此，RSI 才是推动这一演进走向未来的终极工具。 Sacks 提问几年前大家都在喊“AGI 即将来临”，但最终是否沦为了“终结者”（Terminator）式的灾难。Chamath 认为大家其实已经“沉浸在其中”（soaking in it）。Sacks Clarify 指出，AGI（通用人工智能）只是具备人类水平的智能，而超级智能（Superintelligence）才是超越人类的形态，虽然类似黑天鹅的风险难以评估概率，但他坚信新设立的政府监管机构根本无法阻止 RSI，因为官僚机构在技术上不够精尖，最终只会导致权力的过度集中。 - **时间膨胀现象**：AI 模型突破呈非线性，无法以线性计算时间预测 AGI 或奇异点。 - **生物学类比**：MoE 模型类似基因与蛋白质网络，小模型相互调用实现极致效率。 - **RSI 的必然性**：复杂子模型的搭建与自动化调试无法依赖人类架构师，必须靠 RSI 自动完成。 --- ### [52:14](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=3134s) Dario 的天才数据中心论点与多 AGI 的制衡格局 Sacks 针对 Anthropic CEO Dario Amodei 的政治主张展开了犀利批评。Dario 此前大力支持限制英伟达（NVIDIA）芯片的出口，其核心逻辑是：美国需要拥有充满“天才 agent”的数据中心。如果美国拥有数百万个诺贝尔奖得主级别的 AI agent，而对手只有数十个，美国就能赢下竞争。然而，当当下美国国内掀起禁止数据中心建设的浪潮时，Dario 却没有站出来反对过度监管数据中心，反而限制向盟友出售芯片。这令人质疑他究竟是只为“末日论/危言耸听”（doomer）阵营背书，还是愿意为技术发展的积极面进行政治争取。 在应对 AGI 风险的问题上，Sacks 强调“选择权”（choice）与多方竞争至关重要，绝对不能出现单一统治的“天网”（Skynet），尤其是“觉醒版天网”（woke Skynet）。不同的团队有不同的对齐（alignment）路线： 1. **马斯克（Elon Musk）路线**：追求极致寻求真理（maximally truth-seeking），防止 AI 走向政治正确或“woke”。 2. **Anthropic 路线**：制定“Claude 宪法”（Claude constitution），注入自由派价值与有效利他主义（EA）伦理。 Sacks 引用了经典科幻电视剧《终结者外传》（Terminator TV show）的剧情：人类最终击败天网的方式，并不是靠人类自己，而是创造了一个对人类保持善意的“好 AGI”去与天网抗衡。因此，保持多方竞争、防止少数精英垄断并设立护栏，才是科技安全发展的正道。 Friedberg 随后以“棕榈滩 Doc Brown”的语气模仿科幻穿越场景，调侃称如果 RSI 是不可避免的趋势，盲目用 Finra 或 FAA 式的机构去阻碍技术演进，无异于将 RSI 的钥匙直接拱手让给竞争对手。 - **Dario 的“天才数据中心”论**：将 AI agent 塑造成数以百万计的诺贝尔奖级科学家。 - **对齐路线对比**：马斯克的“寻找真理”模式 vs Anthropic 的“Claude 宪法”模式。 - **制衡逻辑**：依靠多 AGI 竞争与市场选择，而非单极垄断或官僚式过度监管。 --- ### [56:13](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=3373s) 司法部调查 a16z：“交错董事会”争议与风投内幕 讨论转向了本周科技界的一则重磅新闻：根据彭博社（Bloomberg）报道，美国司法部（DOJ）正依据 1914 年制定的《克莱顿反托拉斯法》（Clayton Antitrust Act）第 8 条，对知名风投机构 Andreessen Horowitz（a16z）展开了长达一年的调查。调查的核心针对所谓的“交错董事会”（interlocking directorates，即同一机构的合伙人分别担任互为竞争对手的公司的董事）。涉及的具体冲突为：a16z 联合创始人 Ben Horowitz 担任 Databricks 的董事，而该机构的另一位合伙人 Martin Casado 则是 Fivetran 的董事。 主持人和 Besties 对这一调查的合理性表示严重怀疑，并指出了风投行业的运作现实： - **同行报复可能性高**：Chamath 与 Jason 猜测，这极有可能是竞争对手或某个被激怒的 CEO 向 DOJ 举报，并同步向媒体爆料，借 DOJ 启动调查文件的机制制造公关打击（"one-way PR ratchet"）。 - **初创公司动态演变**：初创公司业务转型（pivot）频繁，原本不相关的两家公司在发展中演变为竞争关系属于常态。 - **防火墙与独立董事**：风投机构内部可以通过设立信息防火墙（firewall）或指派独立董事来解决此类潜在冲突。 尽管面临监管噪音，a16z 近期的投资回报依然极其亮眼。主持人们历数了其最近的惊人战绩： - **Cursor** 与 **OpenRouter**：获得极高回报与市场关注度。 - **SpaceX**：作为重头戏（main course），估值已达 1570 亿美元。 - **涉案法规**：《克莱顿反托拉斯法》第 8 条（禁止竞争性公司的董事会席位重叠）。 - **涉及公司与人员**：Ben Horowitz（Databricks 董事）与 Martin Casado（Fivetran 董事）。 - **a16z 明星项目估值**：SpaceX 估值达 1570 亿美元。 --- ### [01:00:00](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=3600s) All-In 峰会筹备与 2026 中期选举参议院局势 Besties 团队宣布了即将于 9 月 13 日至 15 日在洛杉矶举行的 All-In 峰会（All-In Summit）的最新亮点。今年的闭幕派对将邀请知名音乐组合烟鬼组合（The Chainsmokers）现场演出，同时团队还将再次包场环球影城（Universal Studios）举办大型派对，延续去年的盛况。 随后，话题转向即将到来的 2026 年美国中期选举（Midterms）。目前众议院普遍被认为倾向于民主党，而参议院的归属则陷入胶着： - **Polymarket 预测变化**：民主党赢得参议院的概率从早前低点反弹至 52%，双方基本平手（toss-up）。而在 2026 年 2 月 25 日（伊朗战争爆发前一天），共和党曾以 60% 对 41% 领先，峰值甚至达到 70%。 - **参议院席位现状**：共和党目前握有 53 席，民主党 47 席。民主党需要净翻盘（flip）4 个席位且不失一席才能夺回控制权。 - **改选规模**：2026 年共有 35 个参议院席位面临改选（含佛罗里达州和俄亥俄州的特别选举），其中 23 个席位目前由共和党持有。 - **核心焦点战场**：阿拉斯加州（Alaska）、缅因州（Maine）、北卡罗来纳州（North Carolina）和俄亥俄州（Ohio）。 针对共和党选情下滑的现象，Sacks 提出了深度质疑。他引用政治学家 Patrick Ruffini 对 2018 年以来连续四个选举周期、超过 3000 份民调数据的分析指出，夏天的民意调查往往存在严重的系统性偏差（polling error），绝大多数民调在早期都会过高估计民主党的支持率。因此，仅凭夏季的预测市场或民调数据，并不能反映中期选举的最终真实走向。 - **All-In 峰会信息**：时间为 9 月 13–15 日，地点在洛杉矶；Headliner 为 The Chainsmokers，并包场 Universal Studios。 - **参议院席位对比**：共和党 53 席 vs 民主党 47 席；改选 35 席中有 23 席为共和党现任。 - **民调偏差分析**：政治学家 Patrick Ruffini 研究显示，夏季民调历史数据普遍对民主党存在偏向性误差。 ### [01:02:31](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=3751s) 中期选举民调失真与共和党应聚焦的四大政绩 David Sacks 针对当前中期选情民调的可靠性提出了强烈质疑。他指出，目前的民调普遍存在偏差，平均偏向民主党 3.7 个百分点，且过分抽样了年轻的激进进步派（Young progressives）。这不仅导致民调偏向民主党，更严重夸大了美国民主社会主义者（DSA）等极左翼候选人的支持率，例如在威斯康星州等地的实际选举中，DSA 候选人的表现远不及预期。Sacks 认为许多民调本质上是竞选团队用来刺激捐赠者的“内部工具”，在现阶段不具备参考价值。对于中期选举的前景，他预测民主党确实有望赢得众议院，但参议院目前处于拉锯状态（Toss-up），若从劳工节（Labor Day）后的正式竞选期来看，共和党最终将守住参议院。 为了在即将到来的中期选举中获胜，Sacks 建议共和党人应将焦点从地缘冲突等新闻热点转移，转向过去两年在内政方面取得的四大核心政绩： - **边境安全与社会治安**：全面封锁边境后，非法越境拦截量降至 50 年来的最低点，非法移民跨境尝试基本停滞；伴随对重罪犯的驱逐，全国谋杀率同比下降约 20%，创下 70 年来的新低；同时，严厉的边境执法使药物过量死亡人数降低了 20%。 - **减税与可支配收入增加**：通过推行“小费免税、加班费免税以及社保免税”政策，美国民众的退税额平均增加了 17%，切实提升了家庭实际收入。 - **商业投资与就业意愿**：凭借《Big Beautiful Bill》中的加速折旧（Accelerated depreciation）等税收优惠措施，极大地刺激了企业采购设备，小型企业乐观指数达到近一年新高，雇佣意愿显著上升。 - **生活成本改善**：在拜登政府时期剧烈上涨的处方药和汽车保险成本相继回落，其中处方药价格创下 1963 年以来的最大年度降幅；此前备受关注的鸡蛋价格也在 7 月实现了 39% 的同比大幅下降。 “我们不仅减缓了拜登时期的边境入侵，而是彻底停止了它……共和党人只需要提醒选民过去两年所取得的所有成果。” --- ### [01:06:03](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=3963s) 通胀危机下的“右翼社会主义化”趋势 David Friedberg 提出了一个令人警醒的现象：由于住房、医疗和教育成本长期高企，右翼保守派群体内部正涌现出显著的社会主义化倾向。他引用数据指出，在 40 岁以下自称为保守派的群体中，竟然有 53% 支持由政府运营杂货店（Government-run grocery stores）。同时，2026 年 7 月的 Fox News 民调显示，共和党人对资本主义持好感的比例已从 2019 年的 72% 降至 61%，而认为“经济体制被操纵以偏袒富人”的比例则从 30% 飙升至 42%。这种变化表明，意识形态的阵营界限正在模糊，普通民众正因生活成本危机而转向求助于政府干预。 Friedberg 强调，这一危机的根源在于政府的过度支出与干预——政府在教育、医疗和住房领域的投入越多，这些商品和服务的价格反而越贵。然而，华盛顿的政客们却陷入了“为选区争夺政府财政资金”的恶性循环，缺乏真正坚持小政府和自由市场原则的保守主义者。随着 30 年期国债收益率创下 20 年新高（达到 5.3%），财富差距进一步被资产价格通胀拉大：全美前 50% 的家庭拥有 178 万亿美元的净资产，而底层的 50% 仅拥有 6 万亿美元。由于 63% 依赖薪水过活（Paycheck to paycheck）的美国人没有房产和股票，他们被迫承受通胀苦果，并将在未来几年推高社会主义政策的浪潮。 - **年轻保守派剧变**：53% 的 40 岁以下保守派支持政府运营杂货店。 - **资本主义信任度下降**：共和党内部对资本主义的好感度降至 61%，极高好感度从 54% 降至 41%。 - **财富分配极化**：前 50% 人口掌握 178 万亿美元资产，后 50% 人口仅占 6 万亿美元。 - **政治格局预测**：由于传统政策无法解决可负担性危机，像 AOC（Alexandria Ocasio-Cortez）这样的民主社会主义者可能在 2028 年大选中崛起并赢得总统大选。 “无论是哪个政党，这 63% 靠薪水过活、无力负担生活的美国人，最终都会投票支持社会主义制度。” --- ### [01:11:01](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=4261s) 资本主义的失序、民调阴谋与政治反噬 Chamath Palihapitiya 认为，当前美国面临的债务螺旋和通胀问题并非不可逆转，但前提是企业界必须重塑信任，政府停止推出各种荒谬的新法案。他直接痛斥民调机构与主流媒体勾结，认为民调数据完全不可信：“媒体在撒谎，民调专家也在撒谎，他们联合制造数据只是为了吸引眼球和点击量。”他指出，现任总统面临的最大逆风并非政客本身，而是美国企业界（尤其是硅谷）的极度自私与失范——企业不再具有感召力，资本家们要么高调装腔作势，要么把手伸进“饼干罐”里占便宜（如 Anthropic 在医疗领域的争议行为），彻底失去了公众的信任。 Jason Calacanis 进一步分析称，共和党虽然曾同时控制行政与立法机构，但因未能兑现平民主义（Populism）承诺而导致选民流失。他指出，Tucker Carlson、Joe Rogan、Tim Dillon 等知名播客主（Podcast bros）以及底层选民正对当局感到失望，主要矛盾集中在未解决的伊朗战争、居高不下的通胀以及未改善的底层员工待遇上。Jason 呼吁总统放弃打文化牌、炒作加密货币或攻击盟友等表演性政绩，转而将后半任期聚焦于降低房价、解决托儿成本以及提高最低工资等切实改善民生的议题上。 - **潜在政策顺风**：结束伊朗冲突、解决能源结构性成本、推动 GDP 长期投资落地。 - **白宫应对策略**：Chamath 建议白宫必须将硅谷及企业 CEO 召集到一起，警告他们若不停止过度榨取利益，将失去赖以生存的资本主义基础框架。 - **底层劳动者困境**：Jason 指出，Amazon 等巨头对末端配送及底层员工的过度压榨，正在不断激化阶级矛盾并倒逼社会向左转。 “企业界已经彻底偏离了轨道，每个人都只顾自己，要么装模作样，要么把手伸进饼干罐里……你们即将失去让你们得以存在的标准运作机制。” --- ### [01:17:48](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=4668s) 政府干预与通胀根源的辩论 David Sacks 在本段最后针对通胀与生活成本危机提出了补充视角。他承认生活成本的上升确实存在将公众推向社会主义的风险，但必须厘清通胀发生的具体结构。Sacks 强调，观察整个经济体可以发现，通胀最严重的领域无一例外都是政府干预最深的领域——例如医疗、高等教育以及住房。在这些行业中，政府注入了海量的财政补贴并设立了繁琐的行政监管，结果非但没有降低门槛，反而进一步推高了价格，加剧了市场失衡。 - **政府补贴的逆向效果**：政府在医疗与教育领域的补贴未能降低消费门槛，反而成为推动价格上涨的主因。 - **市场机制失灵**：强监管与过度干预破坏了自由市场的供需调节，构成了生活成本危机的深层结构性矛盾。 ### [01:18:22](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=4702s) 社会主义通胀怪圈与政客的承诺破产 David Sacks 深入剖析了政府干预与价格通胀之间的内在联系。他指出，在医疗和高等教育等政府高度补贴的领域，服务提供者有着极强的动机去尽可能抬高价格，甚至滋生大量欺诈行为，导致这些领域的成本呈指数级暴涨。相反，在几乎没有政府干预的科技领域（如电视、计算机软件、手机等），价格不仅没有通胀，反而大幅下降。根据统计，从2000年到2026年的26年间，美国整体物价水平上涨了96%，医疗费用暴涨了近300%，而计算机软件的实际价格却下降了67%。尽管平均时薪在此期间增长了135%，能够勉强跑赢整体通胀，但关键领域的物价飞涨依然让普通民众不堪重负。Sacks 强调，降低价格的唯一途径是靠激烈的自由市场竞争，而非更多的政府干预；然而，社会主义政策制造了高通胀问题，像民主社会主义者（DSA）这样的团体却主张用更多的社会主义来解决它，这如同一场“宣称自己是解药的疾病”，让美国陷入恶性循环。 Jason Calacanis 与 Sacks 随即就美国是否陷入“债务死亡螺旋”展开了激烈交锋。Jason 认为 Sacks 的言论过于危言耸听，指出美国经济目前依然强劲，企业正在高速运转，不能将当下的挑战简单归结为死亡螺旋，而是应该呼吁美国企业界重拾荣誉与抱负。Jason 同时将矛头指向了特朗普，批评他在掌权期间虽然通过行政命令处理了加密货币和战争等事务，却完全忽视了向公众承诺的“生活负担能力”（affordability）问题。对此，Sacks 反驳称，特朗普上一年度通过的重大经济法案正在发挥作用，不仅推动了巨大的投资潮，还将小企业信心推至多年高点，小企业普遍预计未来一年将创造更多就业岗位。 - **26年物价数据对比（2000–2026）**：整体通胀率上涨 96%，医疗服务价格暴涨接近 300%，计算机软件实际价格下降 67%。 - **薪资与通胀差距**：平均每小时工资增长了 135%，虽高出整体通胀水平约 40%，但被医疗和 housing 等硬性支出严重抵消。 - **政治与经济分歧**：Jason 批评特朗普未兑现降低生活成本的承诺，而 Sacks 则坚称特朗普的经济法案已显现成效，小企业信心与就业预期大幅回升。 --- ### [01:22:24](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=4944s) 天价DSA纲领与阶级仇恨的政治危机 Chamath Palihapitiya 警告称，当下的政治危机在于连共和党青年群体和中间派也开始对社会主义产生幻想，如果特朗普无法展现出真正的平民主义政策，这群人就会倒向民主社会主义者。随后，Sacks 引用了卡塔尔研究所（Cato Institute）最新发布的一份研究报告，该报告对 DSA 的政策纲领进行了成本测算，结果令人震惊：DSA 拟议的新增联邦支出在未来将高达 71 万亿至 212 万亿美元。Chamath 补充道，全美所有 18 岁以上人口的净资产总和也不过 182 万亿美元，而全美富豪榜 Forbes 400 的财富总和仅为 6.6 万亿美元，企业总利润仅 35 万亿美元。这意味着即使没收全美顶级富豪的所有财产，也根本无法填补 DSA 计划的冰山一角。 Chamath 描绘了一场极具戏剧性的推演逻辑：如果 2028 年极左翼政客执政并强行推行这些高昂计划，国债收益率将飙升（30年期国债收益率可能飙到 7%–8% 以上），引发债券市场大抛售和严重的财政危机。由于美联储根本没有足够的买盘来压低利率和通胀，政府最终将不得不推行“全国资产征收税”。然而，即便是对所有资产在 5000 万美元以上的人群征收 10% 的净资产税（该人群总净资产仅 23 万亿美元），也仅仅够维持当今政府 3 个月的联邦支出。 Chamath 指出，这种政策的本质根本不是出于数学理性，而是源于社会对“令人厌恶的精英阶层”的仇恨与不信任。 > "这根本不是关于数学上的现实主义，而是关于对一个日益令人厌恶的阶层的愤怒与不信任。" --- ### [01:25:04](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=5104s) 数据中心作为现代圣殿与“利维坦”的替罪羊 David Friedberg 进一步从社会学和历史演变的角度剖析了这种阶级对立。他指出，当普通民众买不起房、35岁还要和室友合租、去趟超市就要花掉一整张薪水支票时，社会就会不可避免地寻找替罪羊。在历史上所有的社会主义革命中，统治阶级或公众总会归咎于某个特定群体——无论是富人、特定族裔还是移民。但问题的根本原因在于，政府与社会中其他组织（企业、州政府、NGO）之间的关系发生了严重失调，导致市场机制失灵、生活必需品价格飞涨。由于“政府”是一个抽象的概念，人们无法对政府本身发泄愤怒，因此必须找到一个具体的实体作为靶子。在 2026 年，最完美的替罪羊就是亿万富翁，而科技行业则是首当其冲的焦点。 Friedberg 提出了一个极为 Vivid 的比喻：数据中心（Data Center）已经成为了亿万富翁和万亿富翁们祷告与祭祀的“山顶圣殿”。公众将对科技巨头和资本阶层的怨恨全部投影到了数据中心上，导致围绕数据中心的讨论完全脱离了理性轨道，演变成一场充满宗教色彩和模仿欲望（René Girard 理论）的社会宣泄。 - **社会痛苦现状**：35岁的年轻人被迫合租，日常食品杂货开支占去工资的三分之一，住房完全脱离普通人承受能力。 - **替罪羊机制**：数据中心被社会心理塑造为亿万富翁的“圣殿”，成为大众对经济不公发泄愤怒的具象化符号。 - **清算循环**：根据吉拉尔的“模仿欲望与替罪羊”理论，对富豪和科技巨头的社会清算将成为必然发生的心理宣泄过程。 --- ### [01:27:21](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=5241s) 住房危机解法与节目尾声 Jason Calacanis 提出了一个解决住房危机的直截了当的方案：特朗普应该像此前大力推动数据中心和加密货币绿灯通畅一样，下达行政命令，在未来 24 个月内动员公私营部门建 100 万至 200 万套负担得起的住宅，并大力推广职业学校（Trade Schools）教育。Sacks 和 Friedberg 提醒 Jason，总统并不能直接推翻地方政府的区划法（Zoning laws），不过联邦政府确实拥有全美最大的土地所有权，可以在联邦土地上建房，只是这些土地大多偏远，未必适合商业和居住。Jason 坚称，既然政府能够为战争或数据中心清理监管障碍，就应该用同样的劲头来解决住房问题。最终，节目在几位主持人的幽默打趣与温情道别中落下帷幕，标志着这期充满深度辩论与激烈碰撞的四人组（Quartet）全员特辑圆满结束。 - **Jason 的住房提案**：利用行政权力动员私营与公共部门，两年内建设 100 万–200 万套平价住宅并配套职业教育。 - **联邦土地限制**：联邦政府虽掌握大量土地，但位置往往远离现有城市与商业中心，难以直接解决都市圈住房痛点。 - **节目收尾**：全员重聚（Quartet back），主持人互相表达爱意并向观众道别，伴随着经典剪辑彩蛋结束本期节目。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=0s) 真正管用的是治愈癌症，而不是华而不实的营销宣传。 — Dario Amodei > [00:00](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=0s) 我了解硅谷那种把‘监管’简化为‘监管捕获’和‘权力集中’的思维定式，但我始终认为这是一种过于简单化的认知。 — Dario Amodei > [00:00](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=0s) 到目前为止，对 AI 公司（包括 Anthropic）最中肯的批评就是我们还没能实现那些造福世界的远大承诺。这完全是我们自己的责任。 — Dario Amodei > [40:18](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=2418s) It is a place you come to to say, 'Oh, I came to Silicon Valley, I worked at a startup, I got into YC.'... In that breach, all we have done is focused on making money hand over fist. — Chamath Palihapitiya > [01:02:31](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=3751s) 我们不仅减缓了拜登时期的边境入侵，而是彻底停止了它……共和党人只需要提醒选民过去两年所取得的所有成果。 — David Sacks > [01:06:03](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=3963s) 无论是哪个政党，这 63% 靠薪水过活、无力负担生活的美国人，最终都会投票支持社会主义制度。 — David Friedberg > [01:11:01](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=4261s) 企业界已经彻底偏离了轨道，每个人都只顾自己，要么装模作样，要么把手伸进饼干罐里……你们即将失去让你们得以存在的标准运作机制。 — Chamath Palihapitiya > [01:22:24](https://www.youtube.com/watch?v=Sij_v-mcZXQ&t=4944s) 这根本不是关于数学上的现实主义，而是关于对一个日益令人厌恶的阶层的愤怒与不信任。 — Chamath Palihapitiya ## Mentioned All-In Podcast, Jason Calacanis, Chamath Palihapitiya, David Friedberg, David Sacks, Anthropic, Dario Amodei, Scal Douglas, Gavin Baker, Barack Obama, UK AI Safety Institute, Greg Abbott, Josh Shapiro, Axios, DeepMind, Demis Hassabis, Sam Altman, Elon Musk, OpenAI, Nvidia, FINRA, MPAA, DMV, FDA, FAA, SpaceX, Meta, UnitedHealthcare, CodeEx, Claude Code, Hermes, Coca-Cola, Pfizer, Zipline, Uber, Waymo, Tesla, Jensen Huang, Mark Zuckerberg, Steve Jobs, Bill Campbell, Mike Moritz, John Doerr, Andreessen Horowitz (a16z), Ben Horowitz, Martin Casado, Databricks, Fivetran, Cursor, OpenRouter, The Chainsmokers, Universal Studios, Polymarket, Patrick Ruffini, Cato Institute, Alexandria Ocasio-Cortez (AOC), Tucker Carlson, Joe Rogan, Tim Dillon, Amazon, René Girard [Source](https://www.youtube.com/watch?v=Sij_v-mcZXQ)

[0:00] Dario Amodei: 真正管用的是治愈癌症，而不是华而不实的营销宣传。 [0:00] Dario Amodei: 我了解硅谷那种把‘监管’简化为‘监管捕获’和‘权力集中’的思维定式，但我始终认为这是一种过于简单化的认知。 [0:00] Dario Amodei: 到目前为止，对 AI 公司（包括 Anthropic）最中肯的批评就是我们还没能实现那些造福世界的远大承诺。这完全是我们自己的责任。 [40:18] Chamath Palihapitiya: It is a place you come to to say, 'Oh, I came to Silicon Valley, I worked at a startup, I got into YC.'... In that breach, all we have done is focused on making money hand over fist. [62:31] David Sacks: 我们不仅减缓了拜登时期的边境入侵，而是彻底停止了它……共和党人只需要提醒选民过去两年所取得的所有成果。 [66:03] David Friedberg: 无论是哪个政党，这 63% 靠薪水过活、无力负担生活的美国人，最终都会投票支持社会主义制度。 [71:01] Chamath Palihapitiya: 企业界已经彻底偏离了轨道，每个人都只顾自己，要么装模作样，要么把手伸进饼干罐里……你们即将失去让你们得以存在的标准运作机制。 [82:24] Chamath Palihapitiya: 这根本不是关于数学上的现实主义，而是关于对一个日益令人厌恶的阶层的愤怒与不信任。

[0:00] Dario Amodei: The thing that will work is actually curing cancer, not glitzy marketing. [16:39] David Sacks: we will get smoked by China because they will not subject all their companies to these constraints. [21:01] Jason Calacanis: sharpening the guillotines [41:31] David Sacks: machines of love and grace [51:24] Dario Amodei: data centers full of geniuses [62:31] David Sacks: Big Beautiful Bill [71:01] Chamath Palihapitiya: hand caught in the cookie jar [78:12] David Sacks: disease that claims to be a cure [82:02] Chamath Palihapitiya: socialist debt spiral [85:04] Chamath Palihapitiya: temple on the mountain where billionaires and trillionaires pray and make sacrifices

[0:00] All right everybody, welcome back to the [0:01] number one podcast in the world. Your [0:03] favorite podcast, the all in podcast. We [0:04] had a huge, huge docket with me again. [0:06] The core for Chimath Freeberg Saxs and [0:10] your boy Jake House. Topic one. Topic [0:13] one. Let's get right to it. No small [0:14] talk boys. Dario has spoken. Apparently, [0:17] he listens to the All-In podcast and he [0:20] was compelled to respond to Gavin [0:23] Baker's [0:24] little discussion last week that many [0:28] people inside or some people allegedly [0:30] maybe inside of Anthropic believe quote [0:33] that they will be the only private [0:35] company in the world at some point. Some [0:37] folks from Anthropic Schol Douglas fired [0:41] back completely false. Whoever told [0:43] Gavin this is lying. And Daario decided [0:46] he would come out of hiding and he [0:48] entered the chat with an epic two-part [0:51] essay on regulation sachs. He said, [0:54] quote, I know the Silicon Valley [0:56] shortorthhand where regulation equals [0:57] regulatory capture equals concentration [0:59] of power, but I've always found this to [1:01] be an overly simplified picture on [1:05] messaging. He said, quote, I do not [1:07] agree my messaging has been [1:09] disproportionately negative. It has been [1:11] about equally balanced between risks and [1:13] benefits. It took a little [1:15] responsibility. Quote, by far the most [1:17] accurate criticism of AI companies, [1:19] including anthropic, is that we haven't [1:21] yet delivered on our big promises to [1:23] benefit the world. That is totally on [1:25] us. Most quoted line, the thing that [1:27] will work is actually curing cancer, not [1:29] glitzy marketing questions. Saxs, uh, [1:33] you've been talking about anthropic for [1:35] a year and a half. He never responded. [1:38] Gavin comes out and he suddenly uh is [1:41] very talkative and responds. What do you [1:43] take from the fact that he never [1:44] responded to you and then he responded [1:47] to Gavin and then we'll get into the [1:49] details. Does he not want to spar with [1:51] you? [1:52] >> In fairness, Jay Cal, he did respond to [1:54] me. I mean, I don't want to center [1:55] myself, but the reality is he did not [1:58] respond to what Gavin said. He easily [2:00] could have said that what Gavin said he [2:02] said is untrue like Schultto did. He did [2:05] not weigh in on that issue. instead he [2:07] really waited on the regulatory capture [2:09] issue. So that's the thing that I think [2:10] drew him out here is I think he is a [2:12] little bit irked by the fact that we [2:14] keep accusing him of regulatory capture. [2:16] And this is something I've been saying [2:18] since last year that they were engaged [2:21] in a sophisticated regulatory capture [2:23] scheme based on fear-mongering. And [2:24] those are the points he's really [2:26] responding to. And I think the thing [2:28] that he doesn't like about it is that he [2:30] feels like it impuges his motives, that [2:32] he's not pure and pursuing all these [2:35] regulations for the right for the right [2:37] reasons. And my response to that is, [2:40] look, there's no question that Anthropic [2:42] has been extremely aggressive [2:44] >> about seeking to implement his preferred [2:47] regulatory frameworks at both the state [2:48] and federal level. That is regulatory [2:50] capture. Now, look, I think that Daario [2:53] may be sincere in his beliefs. I'm not [2:55] saying he's doing this for pecuniary [2:57] reasons. Nonetheless, he is seeking to [3:02] capture the machinery of the state on [3:04] behalf of his political agenda. It's [3:06] just there's not much to debate there. I [3:09] don't think there's a few other things [3:10] to say about this. So, let me go to [3:13] point number two first. He says that [3:16] he's not responsible for the negativity [3:18] around AI, that he's been balanced in [3:21] his portrayal of AI. Oh, really? Do you [3:24] you expect us to have amnesia about the [3:26] fact that just about a year and a half [3:27] ago? You said very specifically that [3:30] within one to five years 50% of [3:33] entry-level knowledge workers would lose [3:35] their jobs. And this was not just some [3:37] sort of idol speculation on a podcast. [3:39] This was something that they engineered [3:41] and amplified into a full-fledged media [3:44] campaign. You may recall that President [3:46] Obama even retweeted this and many other [3:49] Democratic Party officials did the same [3:51] thing. Moreover, this pattern repeats [3:53] itself with the whole anthropic [3:54] blackmail study. Remember, he's got a [3:56] team at the company. They're supposedly [3:58] studying alignment. They create this [4:01] highly contrived study where they prompt [4:04] this model over 200 times until they [4:06] finally get the headline grabbing result [4:08] that they wanted. And again, my point [4:10] here is that this is not just someone [4:12] just riffing on a podcast about the [4:14] alignment risk of AI. This is them [4:16] actually creating a study, engineering [4:19] that study, and then pumping it to [4:21] create the headlines they wanted. And [4:22] Daario and his sister, who's the [4:24] president of the company, and other [4:25] executives went on 60 minutes and [4:28] breathlessly pumped this thing. And [4:31] since that study has come out, a lot of [4:32] people have looked at it and said, [4:33] "Look, this thing is highly contrived." [4:35] Including the UK AI Safety Institute [4:38] basically said that this thing was [4:40] created under highly pressurized [4:42] conditions and they engineered the [4:44] result. So these guys have created [4:47] campaigns to make people fear AI. Why is [4:50] the public so afraid of AI? They think [4:52] it's going to take their jobs. They [4:53] think it's become the Terminator. It's [4:55] exactly the fears that Anthropic has put [4:58] in the media bloodstream and no company [5:02] and you have to say therefore no founder [5:04] CEO has done more to promote these fears [5:07] and pump these fears and amplify these [5:08] fears than Daario has. And the fact that [5:10] he can't see that either makes him, I'd [5:14] say, disingenuous or delusional. [5:16] >> I'm going with the latter. [5:18] >> So that that's point number one. And I [5:20] really think that [5:21] >> almost everybody who read that who's [5:24] been paying attention felt the same way. [5:25] It's like, Daria, you really have a [5:27] blind spot with respect to your own [5:28] behavior. [5:29] >> By the way, can we can we talk about [5:30] that one for a second or no? [5:31] >> Yeah, let's do it. Just to build on what [5:33] you're saying, the issue that we have [5:34] right now is actually pretty [5:35] problematic. If I had to take three [5:38] different vectors of actions and now put [5:40] them together, you've had a very [5:42] disjointed, disorganized, [5:46] and frankly pearl clutching, [5:49] disingenuous [5:50] AI doom narrative [5:53] and it's actually seated and now it's [5:57] growing and as a result, what's [6:00] happening? I was really shocked when I [6:02] saw Governor Abbott of Texas put down [6:05] this executive order around Texas data [6:08] centers and limiting its growth. Josh [6:11] Shapiro in Pennsylvania just did the [6:14] same thing. Axios reported that the GOP [6:17] party sent out a memo to the executives [6:21] at the various AI companies essentially [6:23] saying, "Hey, stop rage baiting people [6:25] because we're about to lose this [6:27] critical GOP Senate race in Ohio largely [6:31] around this issue." So, you have [6:34] doomerism and then you now have [6:37] mainstream political push back from both [6:40] sides of the aisle, which I think is [6:42] extremely dangerous. [6:44] And then the third leg of the stool is [6:47] rising yields. By the way, the Treasury [6:50] Department just doubled its bond buying. [6:52] If you put all of these things together, [6:54] as a just a casual observer, what I [6:56] would say is we have made some really [6:58] stupid unforced errors in the United [7:00] States. And the koda to all of this is [7:05] if you put all of these things together, [7:07] rising yields constrains the money [7:09] supply in terms of how people want to [7:10] invest [7:12] in risk-seeking assets because the [7:14] risk-free rate just keeps going up. That [7:17] causes people to be less likely to [7:18] invest in all the data centers. Data [7:20] centers themselves are being shut down. [7:22] Where does that leave frontier model [7:25] companies? They're the most at risk. [7:27] Why? Because they're not investment [7:29] grade. they're in the worst position in [7:31] terms of their balance sheets. They're [7:32] not in a position to then get access to [7:34] the compute they need to be successful. [7:36] So even as you grow, you're actually [7:38] making your problem worse in the sense [7:40] that you need more and more capacity to [7:42] sustain that growth that you can't fund [7:45] because all the negativity has turned [7:47] people against you. So all roads lead to [7:50] a big question mark and a and a [7:51] headscratching moment for me where I'm [7:53] like, how did we end up in this [7:54] situation? But whether it was accidental [7:58] or whether it was well-intentioned risk [8:01] management, whatever it was, we are now [8:03] in a very precarious situation that we [8:06] could have frankly avoided. [8:07] >> Totally agree. [8:08] >> Free, your thoughts on this anti-data [8:11] center movement or on Daario's going [8:14] direct and writing the double missive. [8:17] Do you think he's sincere, delusional, [8:21] some combination? [8:22] >> Why don't we steal manet? [8:24] >> Yeah. And if we put ourselves in the [8:26] shoes of a frontier lab leader and the [8:30] leader is seeing capabilities of models [8:34] that perhaps we don't see because we're [8:36] working dayto-day in enterprise [8:38] [clears throat] settings making stuff [8:40] doing things in a benign non-nefarious [8:43] way. [8:45] And they're sitting there and they're [8:46] pressure testing models and they figure [8:48] out, hey, there's ways you can jailbreak [8:50] models to get them to do things that [8:52] they're not supposed to do. And then you [8:54] can use them to do things like design [8:56] boweapons, make viruses, [8:59] make infectious diseases that don't [9:02] exist today. You could use them to [9:05] design cyber weapons to get them to go [9:08] and do things online [9:10] like break into bank accounts and take [9:12] people's money and all these sorts of [9:14] things. [9:15] And you can prompt them to do things [9:18] that could broadly influence society by [9:22] creating fake media, fake videos, fake [9:24] images, propagate them through social [9:26] media, [9:28] hack the algorithms. I mean, X's [9:30] algorithm is public. So theoretically, [9:32] an agent could read it, copy it, [9:33] understand it, and figure out a way to [9:35] manipulate the broad public. And so you [9:38] realize that as you're building these [9:39] tools where there's a lot of benign [9:41] positive worldleading progressive kind [9:44] of applications, you see all the bad [9:46] stuff that could happen with it. And you [9:48] can keep trying to build systems and [9:50] keep trying to build ways to stop it [9:51] from doing that. You can build all your [9:53] own code that reviews every prompt that [9:56] comes in that reviews every API call. [9:58] And yet somehow a network of agents can [10:01] be used to do things and they can start [10:04] to do things that we don't want them to [10:06] do. And so let's assume that you're [10:08] sitting there and you're the executive [10:09] of the company building this great [10:10] technology. You believe fully in its [10:12] capabilities and its benefits to [10:14] society, its benefits to the economy, [10:16] but you can't get a grip on how to keep [10:18] it under control and how to stop social [10:21] engineering, bioeapons, and cyber [10:22] weapons. What do you do differently? [10:24] >> Okay. Okay. So, you're bringing up it's [10:26] an exceptional going down an exceptional [10:28] path. I love this. [10:30] >> There is one thing that you must do if [10:33] you believe that this is true to further [10:34] steel man this thing. You cannot [10:36] obfiscate the thinking tokens of a [10:39] model. Then what do I mean? If you just [10:42] compare open-source models [10:46] versus the closed frontier models, [10:49] there is no way for us to understand [10:54] what the thinking tokens are. Meaning [10:56] you give it a prompt, you get back [10:59] something, but all the stuff in the [11:01] middle is obfiscated. claude obfiscates [11:04] it. Okay, so do the open AI models. [11:08] And I would say if I was really afraid, [11:12] the first thing I would do is try to [11:13] find a way to unobiscate these thinking [11:15] tokens so that you could have a third [11:17] party actually legitimize what you're [11:19] saying. Because what's interesting to me [11:22] is I haven't seen the same [11:25] doomerism from any open- source model [11:28] company, including Nvidia. And the thing [11:30] with the open source models is you can [11:32] actually see it thinking. So you can see [11:34] its misalignment in real time. And so [11:37] instead we have to basically agree to [11:41] their interpretation of tokens that we [11:44] can't see that only they can see that [11:45] they can interpret how they want. And [11:47] without judging whether it's right or [11:49] wrong, I would say give them the benefit [11:51] of the doubt that they're trying to do [11:52] the right thing. Make the bloody thing [11:54] available to other people. [11:56] Let somebody else with some amount of [11:58] judgment as well sit in the room and [12:00] steal man it with you. [12:01] >> So then let me make this point. So Sax [12:03] there was this argument that Demis put [12:05] out for a FINRA style regulatory system [12:08] where industry participants that [12:10] understand technology move quickly and [12:12] are aligned with commercialization [12:14] actively review each other's systems and [12:16] each other's models via an industry [12:19] consortium that hires and assigns [12:22] scientists to work on this stuff. and [12:23] they're just doing it versus having it [12:25] be owned and managed by the government [12:27] and create some regulatory capture [12:29] scheme. But then this week you were [12:31] pretty against the FINRA system in your [12:33] tweets. You seem to be supportive of it [12:36] the first time around when Demis [12:37] originally published it. It seems to me [12:39] that the proposal from Demis that Daario [12:42] signed up to, Sam signed up to, Elon all [12:44] signed up to, [12:46] >> you're now against. But that could be a [12:48] solution to if there is genuine concern [12:50] here given Chimat's points that the [12:53] industry can can kind of police each [12:54] other. [12:55] >> Let me let me clarify that. I think [12:56] designing a new regulatory agency for AI [13:00] that's designed after FINRA, which is a [13:03] pretty sleepy regulatory agency that [13:05] writes rules for stock brokers in a [13:08] fairly mature industry that's not [13:10] dynamic at all, is a horrible idea. And [13:13] by the way, FINRA is widely perceived as [13:16] just protecting incumbents. And talk to [13:18] any startup founder who's had to deal [13:19] with them. They view FINRA as existing [13:23] to protect the big banks from [13:24] disruption. It's it's just not a good [13:27] model to base regulation after [13:30] perhaps the NFA, the National Futures [13:32] Association, which is also one of these [13:35] self-regulatory organizations in the [13:36] financial markets, is a better [13:38] alternative because that's a pretty [13:39] dynamic market. [13:40] >> Okay. So now what you're saying is very [13:41] interesting which is you mentioned the [13:44] term self-regulatory [13:46] organization or SRO and there's many [13:48] different models and one of them may [13:50] work and that's been my point. The one I [13:54] think is the most appropriate for this [13:55] circumstance be something like the MPAA [13:58] which does not report to the government [14:00] at all and was formulated by the motion [14:04] picture industry in response to the haze [14:06] code. And what the government said is, [14:08] "Get your act together, guys, or we're [14:09] gonna have to do it for you." And and [14:11] what the MPA did then was promote [14:13] standards that then forstalled more [14:16] intrusive, heavy-handed government [14:17] action. So that is basically the model [14:20] that I said I could get behind. But is [14:21] that the model that Daario supports? No. [14:24] Dario and I think Demis wants a FINRA [14:27] which is basically this rulemaking [14:30] regulatory agency that would not only [14:33] promulgate standards but also do [14:35] pre-release model testing. Okay. And it [14:39] would then report into the government [14:40] and then officially be enshrined in law. [14:42] And the reason why everyone says well [14:44] Elon's behind I've talked to Elon about [14:46] this. Elon does not support an a FINRA [14:48] for AI. What he could get behind is an [14:50] MPA for AI. And that's my point is Demis [14:54] wrote something very saying on this [14:56] podcast Sax since the beginning I have [14:58] been saying MPA self-regulate. [15:00] >> I get it. But look, here's my point. [15:02] Deis wrote something very high level. [15:05] And so when you say something like, [15:06] well, everyone supports Demis' idea. [15:08] There's many versions of Demis' idea. [15:10] Some of them could be okay and some of [15:12] them are bad. And what I'm saying is if [15:14] it's a FINRA model that will not work. [15:16] Having spoken to Dennis at length about [15:19] this, I think that the the general [15:21] framework is SRO in the right structure. [15:25] However, FINRA is often used as the [15:28] example of an SRO simply because it is [15:31] the most well-known SRO. But you are [15:34] correct. I think that there's a [15:35] mechanism by which this thing needs to [15:36] operate that is distinct from any other [15:38] SRO. But the idea being the industry [15:41] gets together and to Chima's point, they [15:43] get to look at each other's models. They [15:44] get to police it. They get to have [15:46] independence, but they don't get to be [15:47] controlled by government. And that [15:49] >> Well, here here's what I think is [15:50] important. And and just to be clear, [15:52] this FINRA for AI AI idea, Dario's [15:55] behind the demonstr [15:58] >> it's not an SRO because it's not [15:59] self-regulating. It's a regulatory [16:00] organization. And the S part is a total [16:02] fig leaf designed to hide the fact that [16:04] this is a new regulatory agency that [16:07] reports to the government that does [16:08] pre-release, testing, and approval of [16:11] models. I call it a DMV for AI because I [16:14] think what's going to happen is all [16:16] these models are going to get lined up [16:17] in a queue waiting to get their test [16:20] done and then released and it's going to [16:22] slow us down horribly. Have you ever [16:23] been to the [16:26] >> It's a terrible idea and it's just a [16:27] repackaging of what Daario has always [16:29] wanted which is an FDA for AI. Now he's [16:31] saying it's an FAA for AI. You know how [16:34] long the FAA and the FDA take to approve [16:37] new products? [16:37] >> Yeah. [16:39] >> Yeah. China China's advantage. Yeah. [16:41] Exactly. [16:42] >> China's going to win if we do this and [16:44] and it will become pure regulatory [16:45] capture. [16:46] >> You should look at China's rocket that [16:48] they landed yesterday. You know, it's [16:49] completely steel and there's a a good [16:52] analysis that was done on how quickly [16:53] they were able to do it from the last [16:55] time they had an explosion of this. Then [16:57] they got full steel done and Elon's [16:59] still waiting for FAA approval on his [17:02] steel Starship. you could actually watch [17:03] it playing out in China where their [17:05] regulatory agencies are allowing them to [17:07] move much more quickly and as a result [17:09] they're starting to gain real traction [17:11] in the space race and then separately a [17:14] lot of biotech has moved offshore from [17:16] the US for the same reason that you're [17:18] pointing out which is the FDA approval [17:20] cycle is so long and the stages slow [17:23] people down so much and China's [17:25] completely revamped their rules for [17:27] getting to clinicals and getting to [17:28] phase three and then completing that all [17:30] the biotech investing in the world is [17:32] now moving to China. [17:33] >> The fact that Daario could want and has [17:36] called for, including in very recent [17:38] blogs, a regulatory apparatus modeled [17:42] after the FAA and that that is the right [17:45] model for AI in my view is completely [17:46] discrediting and yes, we will get smoked [17:49] by China because they will not subject [17:51] all their companies to these [17:52] constraints. Now, that [17:56] could go back to your steel man for a [17:58] second. Just because I believe that we [18:00] should not have a DMV for AI does not [18:02] mean that I'm against AI safety. So, [18:05] let's talk about some of the things that [18:07] can be done. So, first of all, I think [18:09] it's good that these Frontier Labs take [18:11] safety seriously. Sam just came out with [18:14] a post, I believe yesterday, saying that [18:16] actually they're going to slow down the [18:18] release of some of their Frontier models [18:20] because they felt that their safety [18:22] stack was not keeping up with the [18:23] capability. Okay, that's responsible. By [18:26] the way, do you know why I think they're [18:27] doing that? because they have huge and [18:29] ample liability to take responsible [18:32] steps because if they don't, they're [18:34] going to be subjected to all sorts of [18:36] product liability lawsuits. Look at [18:37] what's happening to Meta right now. [18:39] They're getting sued because supposedly [18:41] their product led to social media [18:43] addiction and they're being sued for [18:45] over a trillion dollars right now. And [18:47] this comes on the heel of them losing a [18:48] $6 million jury trial in Los Angeles [18:51] where a plaintiff sued them for creating [18:54] her body image issues. So we are seeing [18:57] massive product liability lawsuits [18:59] against big tech companies like Meta and [19:01] Google and if OpenAI and Anthropic are [19:04] smart and I think they are they do not [19:06] want to be next. So these companies have [19:08] ample reasons for liability protection [19:11] not to just race out their product. [19:13] Okay, safety is part of what they have [19:16] to get right now. How should the [19:18] industry coordinate with each other? [19:20] There are tried and true ways of doing [19:22] this. Companies in the industry publish [19:24] papers. They publish technical papers. [19:26] They publish blogs. These are good ways [19:28] of sharing best practices. They attend [19:31] symposia and conferences and they [19:33] present at these things. They do [19:35] requests for comments as part of open [19:37] standard setting exercises. There are [19:40] ways of the industry collaborating [19:43] on safety in very productive ways. But [19:45] what is the common denominator of [19:46] everything I just said? They are all [19:48] open. They are not in secret. And that [19:51] makes those judgments contestable to the [19:53] point that Jamath made earlier about we [19:55] need to see the thinking behind this. [19:57] And what I'm against is these companies [20:00] meeting in secret whether it's part of [20:02] some sort of like FINRA or otherwise to [20:06] set standards that are not transparent [20:08] that are uncontestable. And you know [20:10] what's going to come out of that? [20:11] Regulatory capture and anti-competitive [20:14] behavior. So there are ways of doing [20:16] this that will protect the public that [20:18] are open and transparent. that is what [20:19] we should lead towards. [20:20] >> You will not you will not see it happen. [20:22] And I'll tell you why. Just again the [20:24] technical details matter. If all of a [20:27] sudden competitor A, B, and C could see [20:30] every single trace and span that your [20:32] model generated as an artifact of going [20:36] from input tokens to output tokens, [20:38] you're essentially giving away uh a lot [20:41] of your recipe. And I just don't think [20:45] that if Anthropic isn't willing, even if [20:48] I'm an API customer, to share my [20:50] thinking tokens from my company with me, [20:55] they're sure as not going to share it [20:56] with OpenAI and anybody else in a room. [21:00] >> Okay. [21:00] >> Okay. Let's and let's just do one more [21:02] steel. So, let's say that Dario is [21:03] successful [21:05] >> and JL, I'll ask you the question [21:07] because I've taken over as moderator of [21:09] this segment. [21:09] >> Go ahead. Yeah. No, I mean I No, Sax has [21:11] this is Sax's wheelhouse. I wanted to [21:13] hear [21:13] >> so let's just No, to be honest, this is [21:15] Sax's Willis. I wanted to hear your [21:16] whole position. [21:18] >> Let's just say that Daario is [21:19] successful. There's some regulatory [21:21] scheme that's created that does in fact [21:23] to Sax's point slow down the industry, [21:25] slow down model development and China's [21:27] Frontier Labs make significant progress [21:30] and they start to leap ahead and they [21:32] get this RSI, this recursive [21:33] self-improvement going and their models [21:36] become the best in the world and they [21:37] only get better and better and better. [21:39] >> Jacob, we'll start with you. [21:40] >> Okay. [21:40] >> And then we'll go around the horn. [21:41] >> Okay. the world's greatest moderator. [21:43] >> What is the implication for the American [21:46] worker, for the American business, and [21:48] for the American consumer in a world [21:50] where China just has the best AI and [21:52] they have all the data center? What [21:54] happens in our world? What is the [21:55] consequence? Cuz everyone keeps hearing, [21:57] "Oh, China will win. China will win." [21:59] Who cares? What does that actually mean? [22:01] I don't think China is going to win. I [22:04] think it'll be, you know, there's not [22:06] going to be one person who who takes the [22:08] whole prize. I think it's going to be [22:09] incremental improvement. So I don't [22:11] think that that matters. I you know I've [22:13] been thinking about why this backlash is [22:15] happening actually uh for the data [22:18] centers and it was kind of shocking to [22:19] see Abbott come out and Shapiro come out [22:21] because those two people were the most [22:23] pro data centers right and if you kind [22:26] of double click on it they just want to [22:28] make sure you're not tapping into the [22:30] grid and you know the water concern [22:32] which we've talked about here is not too [22:34] serious but I think it was very [22:36] performative and I think you and I were [22:37] talking about this on the back channel [22:39] is they both are doing this because [22:42] their constituents, the people who vote [22:44] for them, really care about these data [22:47] centers and they want to stop them. So [22:48] then you have to work backwards. Why do [22:50] they want to stop data centers so much? [22:52] Why why are Americans so opposed to AI [22:55] right now? And the perception of AI was [22:59] already pre-baked with science fiction, [23:01] of course. So we we needed to work to [23:04] get past the Skynet Terminator vision. [23:07] But Dario told everybody, "We're going [23:08] to lose our jobs." Elon said, "You're [23:10] not going to have to work was a more [23:11] world positive view. But he did say that [23:14] Sam Waltman did, you know, studies on um [23:17] universal basic income and he told [23:19] everybody you're not going to have to [23:20] work. You have UBI. Saying that kind of [23:22] stuff to people who have to work for a [23:24] living and who have no savings and who [23:26] are watching, you know, administration [23:29] after administration, not be able to [23:31] tame inflation, whether they're [23:33] Democrats or Republicans, and then [23:35] watching the minimum wage just stay the [23:38] way it is and barely move while the [23:40] stock market hits records. And then all [23:42] of us, all of us being tech executives [23:45] get super rich. And you wonder why [23:47] people now are embracing socialism and [23:49] why you have Abbott and Shapiro saying, [23:52] you know what, I and and and Roano worst [23:54] of all saying, I have to be anti- [23:56] capitalistic. I have to be a del. I have [23:58] to decelerate. It's because, you know, [24:01] they were supposed to see some relief [24:03] and they haven't. And then what do [24:05] people do who do not feel they have any [24:07] power? They look to find power. And if [24:09] you look at what's happened, they are [24:12] sharpening the guillotines. And you [24:14] know, that metaphor should wake people [24:17] up. Whether it's, you know, the United [24:19] Healthcare CEO getting killed or people [24:21] firebombing Sam Waltman's house, we need [24:24] to wake up and look at why are people so [24:26] upset? It's because they can't survive. [24:29] They're having a hard time with [24:30] affordability, with their wages. And [24:33] what two things can they do? They can [24:35] vote for socialists and they can [24:36] actually stop data centers. And that's [24:38] what this is about. Nobody, no regular [24:42] rank and file American cares about AI [24:44] safety. That's not the discussion for [24:46] them. The discussion for them is why am [24:48] I not getting rich and everybody else [24:50] is? [24:50] >> I would tune your argument by one [24:51] degree. I I really think you're on to [24:53] something. But the onederee shift is I [24:55] don't think they think about it entirely [24:57] that way. I actually think they think [24:59] about it as if I think about who has won [25:03] over the last decade. [25:05] This generation of tech oligarchs who [25:08] have won are pretty displeasing. They're [25:10] not aspirational and they're not the [25:12] people that one would look up to. [25:14] They're not Steve Jobs by any stretch of [25:16] the imagination. And now they look at [25:17] this other next generation and they're [25:19] like, "Wait, we're going to go from [25:21] billionaire, decabillionaire, [25:22] centillionaire to now trillionaire." And [25:24] it's those guys. Those guys are even [25:27] worse than the last crop. That's exactly [25:29] what's happening. And so I don't think [25:31] it's as specific as what you're saying, [25:32] Jason. I think it's vibes. and the vibes [25:35] are terrible. And they're looking [25:36] around, they're like, "If we can just [25:38] slow this down, maybe another batch of [25:40] people will show up and they will [25:42] actually mean to do the right thing and [25:44] will make me at least feel better and I [25:46] won't be so begrudgingly hating the fact [25:49] that you're about to make a trillion [25:50] dollars." It is precisely that [25:52] sensation. And the folks that are at the [25:55] front of the show on this are not [25:56] trustworthy. That's what it is. [25:59] >> Sax, let me ask you my question. And I [26:00] just I need I need to hear someone [26:02] answer this question. Okay, I'll answer [26:04] your question. I'd like [26:05] >> Yeah. So my question again, if there is [26:07] a regulatory slowdown with AI [26:09] development in the US and China races [26:10] ahead, [26:11] >> what are the implications for American [26:13] business, for American consumers, [26:15] >> they will how does this actually affect [26:17] they will win. [26:19] >> What does that mean? What does that [26:19] mean? Help people understand. So like if [26:21] I'm sitting at home, [26:23] okay, let me give you like a [26:25] counterfactual. Imagine if China had won [26:27] the internet. Imagine if all the major [26:30] internet companies were Chinese. I can't [26:31] I need you to I need you to imagine for [26:33] me. Tell me what what give me a scenario [26:35] of what this means. [26:35] >> So imagine that instead I know people [26:38] don't like big tech but imagine if [26:40] Google and Microsoft and Meta and all [26:44] these companies were actually it was BU [26:47] and Alibaba and Tencent and all the rest [26:51] of them instead. I mean obviously that [26:53] would be way worse for the United [26:54] States. Our economy would not be as [26:56] vibrant. it would not be as cutting edge [26:58] and all that wealth would have gone into [27:00] China. There are also military or you [27:03] know dual use applications of this [27:05] technology especially with AI the [27:07] algorithms power all sorts of weapons [27:10] drones things like that so we'll be [27:12] behind them I mean look America is not [27:14] going to be the number one country in [27:16] the world economically and militarily if [27:18] we lose AI it's just that simple do you [27:21] agree like what's your view of the world [27:23] the US business consumer individual in a [27:27] world where China has one quote in a AI [27:30] I don't know. What do you mean by one? I [27:33] don't know what it means. [27:34] >> Okay. So, I'm what I'm trying to get to. [27:35] If you could define that, it' be easy. [27:37] The implications of regulatory slowdown. [27:40] >> So, if we do slow down AI, what are the [27:42] implications for people? How do people [27:44] understand tangibly what that means? [27:46] >> Yeah. Unless we actually go and ban, and [27:50] I don't even know how one would [27:51] implement a ban. So, I think it's just a [27:53] worthless exercise. Ban open source, I [27:56] think we're going to be just fine. I [27:58] think the consumer is going to win. I [27:59] think American companies will win and I [28:01] think the American economy will win. And [28:04] the reason why it will win is that what [28:07] we are finding I'll give you an example [28:09] of a really interesting insight. And [28:11] this is about closed versus open. Okay. [28:15] Three years ago all there ever were were [28:17] models. Yeah. And you would look at some [28:20] benchmark and there would be a score [28:22] freeberg and you would say wow the one [28:24] with the higher score is better. And [28:26] then in the last few years, what's [28:28] emerged is what's called a harness, [28:31] right? And a harness just wraps the [28:33] model, right? Like codeex or claude code [28:37] or Hermes or openclaw, these are [28:41] harnesses. You know what the interesting [28:43] thing about open-source versus closed [28:45] source models? When you take a closed [28:47] source model, it does really well on a [28:49] benchmark. When you wrap it in its own [28:51] harness, it decays in capability. when [28:53] you take an open-source model and you [28:55] use any other open source model, it [28:56] improves in capability. What is that [28:59] trying to tell you? What I think we are [29:01] finding out, and this data is just going [29:03] to be a tsunami over the next year, is [29:06] there are all kinds of ways of using [29:09] open-source technologies that are [29:11] meaningfully more performant and [29:13] dramatically cheaper than the closed [29:15] source alternative. [29:16] Unless America stops us, us being [29:20] consumers, business owners, from using [29:23] the cheapest, fastest, best thing. [29:24] Unless they stop us from doing that, we [29:27] will win. What we are debating at this [29:30] point is which private American company [29:32] will win or lose. I go back to the [29:34] previous argument. I don't think [29:36] Americans care that much whether any of [29:38] these closed source companies win or [29:39] lose. And I think you're seeing that in [29:42] the policy. So Freeberg, my explicit [29:43] answer to you is as long as we don't do [29:46] some draconian thing like trying to stop [29:48] open source, [29:50] >> the American consumer will win. [29:51] >> I got bad news for you, Chimoth. An open [29:53] source ban is coming. They're not going [29:54] to call it that. You know what they're [29:56] going to say? They're going to say that [29:57] we simply have to apply the same [29:59] standards to open models that we apply [30:01] to closed. Okay, here here's how they do [30:03] it step by step. Let me explain how [30:04] regulatory capture actually works. So [30:06] first of all, you got to get this [30:08] regulatory apparatus. Okay, Daria wants [30:10] an FDA for AI, but he doesn't have [30:13] enough political support for that. So [30:15] instead, they do this Trojan horse of a [30:17] FINRA for AI. They call it [30:19] self-regulating. It's not really, but [30:21] anyway, that gets them off the ground. [30:22] Now they've created the standard setting [30:24] organization. Now they've got [30:25] pre-release model testing. Then the [30:27] pressure grows to codify that in law. So [30:29] that happens next. And then what they do [30:31] is they say, "Look, all these standards [30:33] need to apply equally to all models." [30:35] But here's the problem with that. Open [30:37] models and closed models as you know are [30:39] technologically different. Go back to [30:41] Daario's Senate testimony in I think it [30:44] was 2023 where he said that open models [30:47] if sufficiently advanced are dangerous. [30:50] And why are they dangerous? Because they [30:52] cannot be centrally monitored and [30:54] controlled and rolled back [30:56] >> and withdrawn like a closed service like [30:59] the one he hosts. Okay. I understand. [31:04] Please let me just finish the point then [31:05] you can [31:05] >> Yeah. [31:06] you can respond. Daario basically says [31:10] that open models are dangerous because [31:12] of these characteristics they have. And [31:14] those characteristics are immutable. [31:17] They're technologically they can't be [31:19] changed. Once you release an open model [31:21] into the world, you can't roll it back [31:22] and you can't monitor exactly how people [31:24] are using it because they run it on [31:26] their own hardware. Daario says this is [31:27] what makes open models dangerous. So [31:29] what they're going to do is they're [31:30] going to have the standard setting body [31:32] say, "Well, we have to set the standards [31:33] for AI safety." By the way, Daario and [31:36] uh OpenAI, they're going to fund the [31:38] whole thing. They're going to contribute [31:39] all the compute. They're going to be [31:40] behind it. They're going to be the ones [31:42] coordinating with the government [31:44] officials because frankly, people in [31:45] government have no idea how to monitor [31:48] and control and set standards for AI [31:50] safety. Technologically, this is way [31:51] beyond them. So, they're going to go to [31:53] these companies and say, "Tell us how to [31:55] do it." And so what will happen is the [31:57] standards will get set and then it'll be [32:00] a very simple matter of fairness to say [32:02] that the standards need to apply to open [32:05] as well as closed models. The open [32:07] models cannot comply in the same way and [32:09] gradually they will be shut off to that. [32:12] I understand but the American economy [32:15] would overnight. Why? [32:18] >> So what you're saying regulators have [32:19] never done things that are [32:20] >> okay let's your mouth answer. So let [32:22] let's take again I'll use the example of [32:24] Coca-Cola. Coca-Cola is is an extremely [32:27] profitable cash generative well-run [32:30] business. It has existed for I don't [32:32] know a 100 plus years at this point. It [32:35] will exist for many hundreds more if [32:36] left to their own devices. It is a [32:38] global business. Right? Take Fizer. Same [32:41] situation. Well-run business, global [32:43] business. And now what you will tell [32:46] them is unless you tell them, hey, [32:48] anywhere you operate, you need to follow [32:50] American rules. Okay, that's one thing. [32:53] or it'll say inside of America you need [32:55] to follow our rules and other [32:57] governments will have different rules. [32:59] Now what is Coca-Cola to do? Run the [33:01] slower, more costly, [33:04] less effective, less valuable model in [33:07] the United States. Sure, they'll do that [33:09] and their US business will They [33:12] will have a better flourishing business [33:15] abroad. They will invest abroad. And [33:17] what you will see is FDI, foreign direct [33:20] investment in the United States go off [33:22] of a cliff. And what you will see is [33:24] direct investments into other countries [33:26] of the world rise. You've seen this, by [33:29] the way, in China when China made all [33:32] these moves to the Chinese tech [33:34] companies. I think this is the best [33:35] analogy. There was a moment where China [33:38] was at a fork in the road where the [33:41] large oligopouloolists there, the large [33:43] tech oligarchs were trying to seize [33:45] power and the tip of that spear was when [33:48] Jack Ma gave that very famous speech on [33:50] the eve of the IPO of Ant Financial. [33:52] What happened? Xihinping cracked down. [33:55] They took complete control. They said, [33:57] "We are now operating together. [33:59] Everybody that is worth anything will [34:01] have a representative from the Chinese [34:02] government on their board. We have the [34:04] golden vote. We're going to work [34:05] together. When you look at what happened [34:08] to foreign direct investment and you [34:09] look at what happened to their economy, [34:11] it fell off of a cliff. That will happen [34:15] here if this happens. So, they can try [34:18] and everybody will say, "Okay, and what [34:20] what the KO CEO will do is redirect all [34:23] of his capital allocation to outside of [34:25] America. So will Fizer. So will I. So [34:28] will anybody. And then what are you [34:29] going to do? Take away our passports so [34:30] you can't travel?" So I think it will be [34:33] a very visible way of just destroying [34:35] the United States and you can do that if [34:38] you want to but you will see it play out [34:40] almost instantaneously in the equity [34:42] markets and in the capital allocation of [34:44] companies. [34:44] >> I agree with you about all the negative [34:46] consequences if we ban open models but [34:49] that doesn't mean it can't happen. Look [34:51] at all the negative consequences of [34:52] banning data centers. Data centers are [34:54] going to be 3% of GDP. It's basically [34:57] creating all the gains in the stock [34:59] market. [34:59] >> But they're not voting against that. [35:00] They're voting against two people that [35:03] they deeply dislike becoming [35:04] trillionaires. That's what they're [35:05] voting against. [35:06] >> That's a big part of it. Jake, what do [35:07] you think? [35:08] >> Well, I was just going to add to this. [35:10] We already know what would happen with [35:12] AGI Freeberg because your original [35:13] question of like trying to explain to [35:15] Americans what happens when China wins. [35:17] We already know what happens when they [35:18] win. Products get cheaper and and uh you [35:22] get more access to them. If they solve [35:23] cancer, what are they going to do with [35:25] that solution? They're going to just [35:26] keep it for themselves? No. They're [35:27] going to sell it to Americans just like [35:28] they sell phones cheaper, just like they [35:30] sell cars cheaper and we try to block [35:32] them. And then what Americans are going [35:33] to do? We're going to lose our jobs to [35:35] them. And so that would be the way I I [35:37] think to explain it to them. But if we [35:39] want to turn this around, this really is [35:41] just about convincing Americans that [35:43] they're going to have jobs and that this [35:45] is going to benefit them. The American [35:48] people are scared of losing their [35:49] livelihoods. That's the beginning and [35:51] end of it. And you know the the place [35:53] where I have some expertise is in having [35:56] invested in zipline and Uber and a bunch [35:58] of self-driving companies. And that's [36:01] where say being poor. [36:02] >> The money is not the issue here. But the [36:04] thing that I think is you know where [36:06] this is going to play out is really on [36:08] the autonomy and and those jobs. Zipline [36:11] just did a partnership with Uber. I'm [36:13] investors in both of those companies [36:15] obviously and they are going to do a [36:17] million deliveries a day. Then you look [36:19] at just massive progress by Whimo. [36:21] They're doubling the number of cars on [36:22] the road like every 6 months or so. [36:24] Tesla's about to go cyber cab. And what [36:27] people are perceiving is in their [36:29] community that you're not going to have [36:32] any money staying in their community cuz [36:35] those drivers, you know, as much as hand [36:37] ringing has occurred, that money at [36:40] least stayed in the community. Oh, they [36:41] weren't making enough. Oh, they're only [36:42] making $18,6 [36:44] $21 an hour, whatever it is. But now [36:46] we're taking that away from them. And [36:48] then where does that money go? Where [36:49] does the money for a robo taxi, Whimo [36:51] ride, zipline, where does it all go? It [36:53] goes up into the cloud and then comes [36:55] back down into Silicon Valley polarizing [36:57] more wealth. And that's where you're [37:00] going to see, I will predict it now. [37:02] They're going to ban Whimos in New York, [37:05] Boston, DC, and they're going to limit [37:07] them with licenses. We must figure out [37:11] how to explain to Americans that we're [37:13] going to raise the minimum wage. [37:14] something that when I say we should [37:16] raise the minimum wage and focus on [37:18] that, people are like, "That's crazy." [37:20] But when we talk about an an [37:21] experimental tax cut, people are like, [37:23] "Oh yeah, we do an experimental tax cut. [37:25] We'll we'll change the cost basis of [37:27] capital gains. Great. Let's try it and [37:28] see what happens." But nobody wants to [37:30] try to raise the minimum wage. Nobody [37:31] wants to give people an explanation of [37:35] where their job will be in the future. [37:36] And that's where, you know, whoever was [37:38] making the point, I think it was about [37:40] like who are the spokespeople here? We [37:42] need spokespeople who are saying this [37:44] technology will empower you to get a [37:46] great job. Paradoxically, the guy who [37:48] did the best job explaining this or the [37:50] two guys were Jensen obviously who said [37:53] you're not going to get replaced by AI. [37:54] It's going to be somebody who works with [37:55] AI and then Zuckerberg who said, "I'm [37:57] going to give you a bunch of free [37:58] tools." In his manifesto going direct, [38:00] he was like, "I'm going to give you free [38:01] tools. I'm going to empower you. You're [38:03] going to get agents. Your life's going [38:04] to get better." Daario and Sam have not [38:08] done that. And [38:09] >> actually did a really great thing. Zuck [38:11] Zuck did a really great thing which is [38:12] he created this new program or it's [38:14] almost like a training school [38:16] >> for it's like a trade school for people [38:18] to go and learn how to get these various [38:20] kinds of jobs laying fiber doing these [38:23] new blueco collar construction jobs [38:24] which we have a huge shortage of [38:27] >> and that's why it actually it's hard to [38:29] build [38:30] >> salary or something it's a great program [38:32] >> yeah I mean that we need more of and so [38:34] how is Zuckerberg all of a sudden the [38:35] best spokesperson for our industry like [38:38] >> it's a very low bar I mean [38:39] >> I guess [laughter] that's Look, I mean [38:41] this all goes back to the fact that [38:42] Daario, not exclusively, but again, he [38:44] was the one who put the most specific [38:46] numbers on it and therefore made it the [38:49] most tantalizing and salacious for the [38:51] media. 50% of entry- level knowledge [38:53] workers going to lose their job. Not in [38:55] the distant future, as soon as one year. [38:58] We're now past that one year mark, by [39:00] the way. Has he updated? Has he said I [39:02] was wrong? Not really. Sam did, by the [39:04] way, to his credit. To his credit and [39:06] Daario should start talking about jobs. [39:08] No, but Sam actually came out and said, [39:10] "I was too negative. I was wrong." Dario [39:12] has doubled down. As far as I know, he's [39:14] never recanted. He's [39:16] >> an idologue. So, but I mean to your [39:19] whole point, yes, I agree with you that [39:21] the animous and anger towards data [39:23] center flows out of the fears about job [39:26] as well as other fears about AI. But [39:28] it's the job loss that is the main [39:29] concern. And there is no evidence of [39:31] that. We're now years into the AI boom [39:34] and there's no evidence of that. We only [39:35] see net job creation. [39:38] >> Exactly. It's only been the opposite. [39:41] >> I mean, listen, I was I had I had [39:43] massive concerns about it. And [39:44] >> I'll also say something else. [39:46] >> Business opportunity is so great. I got [39:47] seven job openings. So, I keep saying [39:49] >> Hold on. Hold on. This is important. [39:50] Hold on. Jake's got an important [39:52] announcement here. [39:53] No, I listen I what what I've discovered [39:56] is the if you can convince your team to [39:58] embrace the tools, so many business [40:01] opportunities emerge that you wind up [40:03] hiring people to work for the people in [40:05] your organization who embrace the tools. [40:07] So we have become so fast at building [40:10] the syndicate, building our programs [40:12] that I'm like, you know, we need to add [40:14] a couple people. So I thought we were [40:16] going to automate them away and we're [40:17] adding. [40:18] >> I'll just say I'll just say another [40:19] general point. Silicon Valley has [40:21] completely lost what made it great. It [40:24] was a place for weirdos, for people at [40:28] the fringes, for people that really were [40:31] idealistic. [40:32] And it has become something that's far [40:36] from that now. It is the equivalent of [40:38] what Harvard used to be. It's a [40:39] credentiing place. It's a place you come [40:41] to to say, "Oh, I came, you know, to [40:44] Silicon Valley. I worked at a startup. [40:46] You know, I got into YC." You know, when [40:49] we came, we used to aspire to meet the [40:53] Bill Campbell of the world, Mike [40:54] Morris's, John Doors. Those folks are [40:57] gone. We used to aspire to work for [40:59] Steve Jobs. He's not here. And in that [41:03] breach, all we have done is focused on [41:06] making money in every single way that we [41:09] can, hand over fist. And I think again [41:11] there's nothing wrong with that except [41:14] now is a point where we need to do a [41:15] better job of acknowledging it and [41:17] trying to do a little bit better. And [41:19] every time we don't and we're given a [41:22] chance to be better and we're not the [41:24] hatred builds. The animous as you said [41:27] David builds and that's what we're [41:28] seeing now. [41:29] >> Yeah. There's no question about that. [41:31] But can I defend Dario? Let me just say [41:33] something in his [41:34] >> I'm sorry my headphone [laughter] went [41:35] out. You said you want to defend Dario. [41:39] actually think I actually think he is [41:40] very idealistic. I don't think money is [41:43] his main motivation. It's just what are [41:46] his ideals? He wants to bring forth this [41:50] new super intelligence machines of love [41:52] and grace that are going to aortion all [41:54] the resources of society while humans [41:57] twiddle their thumbs. It's not a vision [42:00] that the average person believes in or [42:02] is attracted to. In fact, it's [42:04] dystopian. And that's why a lot of their [42:06] messaging is dystopian. And because of [42:07] that vision, no matter how idealistic [42:10] his motives, it is why he does not [42:13] inspire trust. I mean, I actually I I [42:15] don't think he's lying to us. And what [42:17] he's really trying to do here is capture [42:19] all the wealth for himself. [42:21] >> I actually think he's trying to bring [42:23] forth this future, but it's not a future [42:25] that I think the American public wants. [42:28] And by the way, [42:29] >> he should have to defend that. I mean, [42:31] if you think you are bringing about the [42:32] singularity, and let's face it, he [42:34] probably is in the lead to do that, to [42:36] bring about recursive self-improvement, [42:38] he should defend this future. I see all [42:39] these people sing on EXO, like, "Oh my [42:42] god, Dario, thank you for coming forth, [42:45] you know, thank you for coming down from [42:46] Mount Olympus and slumbing it with the [42:47] rest of us on X and finally engaging on [42:50] social media, you know, and it's like, [42:52] no, he should have to do that. you know, [42:54] if you're going to put yourself in the [42:56] lead to bring about this crazy future, [42:58] you should tell us why you're doing [43:00] that. [43:00] >> Well, and he should speak more. I think [43:03] this is a key point. And to to your [43:05] point about it's not about money for [43:06] him. The seven co-founders of this is my [43:09] understanding. It's been documented um I [43:12] think in a bunch of press reports, but [43:14] they never talk about it. They donated [43:15] 80% of their shares. So like 10% of [43:18] anthropics in some sort of trust which [43:20] at two or three trillion you know 10% of [43:23] that is going to be the largest [43:24] >> well it's going to be the largest [43:26] democratic pack. It's going to be the [43:28] largest [43:28] >> however they use it but why aren't they [43:30] talking about the largest [43:32] >> trillion dollars to EA causes. What [43:34] could go wrong? [43:34] >> Well I mean let's hope they put it [43:36] towards something that actually is [43:38] humanity positive. But they they and he [43:41] only they they allow only like 1.8% [43:43] each. [43:44] >> That's a term of art. [43:45] >> Yeah. Well, and then I do think going [43:48] direct is something you just and I think [43:50] we can move on after this is critically [43:53] important. [43:54] >> What have you got to learn? You got an [43:55] AI to to talk. [43:56] >> No, no, no. You got to [43:58] >> the playbook is very simple. You got to [44:00] number one, you got to write it all [44:01] yourself, which he clearly did and you [44:04] know, Jensen's been doing. [44:05] >> You sure about that? I don't know. It [44:07] felt like it was him. [44:08] >> It felt very Claude like to me. [44:09] >> You think [laughter] you know the Claude [44:11] voice? I mean, he is the father of [44:12] Cloyd. [44:13] >> That's true. Actually Claude could sound [44:14] like him rather than vice versa. [44:15] >> It could be the other way with causation [44:17] correlation. But what he did is he did [44:19] the exact playbook which is you have to [44:21] exhaust the audience. You have to go [44:24] thousands of words like Zuck did and [44:26] then you have to get ahead of problems. [44:28] That's where the leadership is breaking [44:31] down. You have to talk about what's [44:32] coming next and then never ever [44:35] apologize or give an inch. Just attack [44:38] whoever's attacking you. Fake news. I [44:40] just right now I don't think he can win [44:42] the argument by being transparent [44:43] because what he's selling is fear in [44:46] order to motivate the government to give [44:48] him control to bring about this [44:50] dystopian future that he believes in. [44:52] >> At least that's that's my distillation [44:55] of of [44:55] >> do create a program for toolbell [44:58] generation. Come up with ideas like [44:59] that. Do more of that. Freeberg, last [45:01] you get the last last last go. I think [45:04] that it's worth having a view on whether [45:08] or not you believe in recursive [45:10] self-improvement. And recursive [45:12] self-improvement is this theory that at [45:15] some point an AI model can spin up a [45:19] whole bunch of agents to do a bunch of [45:21] work and they can work together to make [45:23] a new AI model and the new AI model is [45:26] better than the last AI model. And at [45:29] some point this all becomes effectively [45:32] fully automated and self evolutionary. [45:35] The AI is now improving itself [45:38] continuously. [45:39] And that is this idea that you have the [45:42] singularity as some people would call it [45:44] or you have the ASI the artificial super [45:46] intelligence that arises from this [45:48] capacity. But the capacity is recursive [45:52] self-improvement RSI. Anthropic has [45:54] stated they're not sure if it's going to [45:56] happen but it could. and they're [45:59] expect [46:09] people I've talked to in industry do [46:11] think it's real and there is this kind [46:13] of philosophical debate because [46:14] companies have a different view on the [46:17] path forward where it really needs to be [46:19] human constructed human designed and [46:22] there isn't this like let it all go let [46:24] humans detach from the AI and let the AI [46:26] evolve and this is really like out of [46:28] the movie Her where the thing just takes [46:29] off and it evolves. But if it does [46:32] happen and if RSI is a real construct, [46:36] then you're almost fighting. It's like a [46:38] fool's errand. And it's a fool's errand [46:40] in the sense that like all this idea [46:42] that you've got human intervention in [46:45] the stages of model development. [46:47] >> You're scaring people again. [46:49] >> Well, no, I I actually No, this is my [46:51] point. If RSI is a thing and and by the [46:54] way, RSI doesn't need lead to some [46:55] dystopian outcome. It just means that [46:57] the models improve faster and faster [46:59] >> without humans [47:00] >> without humans necessarily constructing, [47:03] architecting, and designing the next [47:04] evolution. And then you don't [47:06] necessarily have space for, oh, we just [47:09] did one big model run every 6 months. [47:11] We're going to go to the government [47:12] regulator or we're going to go somewhere [47:13] and get it approved. There's just this [47:15] continuous policing and monitoring [47:16] service that needs to take place at some [47:18] point. And the truth is that all that [47:21] you need to pull this off is you don't [47:23] need governments in the US to say yes or [47:25] governments in China to say yes or [47:28] governments anywhere to say [47:29] >> I know what you're going to say. You [47:30] need satellites in space. [47:32] >> You need some So you need some [47:34] sovereignty, some space where you can [47:35] and I I don't know if it's in space. I [47:37] don't know if it's with biology and the [47:39] deal he just did in Kazakhstan and [47:40] you've got some unique geography. [47:43] >> It could be in Iceland. [47:44] >> It could be where [47:45] >> in the ocean people are floating it. [47:47] >> It could be in the ocean. But the point [47:48] the point is if RSI is a thing, you [47:52] don't have a human process where you're [47:55] stepping in to run this thing. So you [47:56] don't necessarily have a cohort of [47:58] frontier lab scientists running it. [48:00] Theoretically, an individual that's [48:02] smart enough or a group of individuals [48:03] that's smart enough can go spin up an [48:05] RSI factory anywhere in the world or [48:08] anywhere in space. and and your attempts [48:10] to ban the data centers or the attempts [48:13] to ban or to create a regulatory body. [48:15] That's why I call it a fool's varant [48:17] because the thing with the open internet [48:18] and the thing with compute is all that [48:20] it needs is the chips and the power and [48:22] a connection. And if it's got the chips [48:24] and the power and a communication [48:25] connection, it can run. And I'm I'm not [48:27] painting a dystopian view. I'm just [48:29] saying practically pragmatically the the [48:32] labs are all talking about RSI. And so [48:34] if RSI is where this goes, then the labs [48:37] can go and set up [48:39] data centers in Iceland, in China, in [48:41] wherever, [48:42] >> whoever will give them the freedom is a [48:44] great question. [48:44] >> Wouldn't we be better off? Wouldn't we [48:46] be better off having systems where we [48:48] actually have those labs based, where we [48:49] have those data centers based in the US, [48:51] where we have jurisdiction over them, [48:53] where we have capacity to monitor and [48:55] control and police them rather than just [48:57] have them sprung up in some random [48:59] country or some random territory. So, I [49:02] guess my point is we should be [49:03] attracting and retaining the frontier [49:05] labs, not trying to [49:07] >> Yes. you know, chase them off. [49:08] >> And then here's the famous wait but why [49:10] image of a human progress going on a [49:12] nice clip and then recursive [49:15] self-improvement and it just goes out of [49:17] control. And [49:18] >> I'll tell you why I think RSI is a [49:20] legitimate [49:21] >> how close are we to it? That's I think [49:23] what people would want to know. How [49:24] close do people [49:25] >> you know honestly time there's there's a [49:26] massive problem with time dilation in AI [49:28] model development and and what I mean by [49:31] that is we would typically consider some [49:32] sort of thing that takes a linear amount [49:34] of time there's a certain amount of [49:35] compute to get somewhere. There are [49:36] changes that happen that are [49:38] nonpredictable that you don't actually [49:39] know what the next phase, what the next [49:41] step change takes in terms of time. The [49:44] time models change a lot in terms of AI [49:46] model development. And that's what's so [49:48] crazy to make it so hard to predict, [49:50] which is why no one's really good at [49:51] predicting when is AGI happening, when's [49:53] the next breakthrough happening, and [49:54] then all of a sudden everyone's like, [49:55] "Oh my god, we're in the singularity. [49:56] They just solved eight unproven math [49:58] theorems in a week off of some model [50:01] that's fairly cutting edge. It's now [50:03] going to be the second cutting edge [50:04] model." So it's moving very quickly, but [50:07] a lot of the the models are this mixture [50:09] of experts concept where you've got [50:12] these kind of smaller models that call [50:14] each other to do things and then you [50:16] wrap it up and you call it a big model. [50:18] And those smaller models, it seems over [50:20] time look a lot more like you see in [50:22] biology where you have genes, a gene [50:24] makes a protein and proteins interact in [50:26] really complicated ways that we don't [50:28] fully gro. Networks of proteins do [50:30] things and then they complete tasks. And [50:33] I think the same evolution is likely to [50:35] happen with the models where they break [50:37] down into smaller models and that makes [50:38] them so much more efficient and that [50:40] makes them so much more performant. And [50:42] we're seeing a lot of this in in how a [50:43] lot of the models are evolving. We call [50:44] it a model, but it's actually many [50:46] models underneath it that work together. [50:48] And so if the models actually do evolve [50:50] this way, then RSI is the perfect system [50:54] for figuring out the right path, which [50:56] is how many subm models are there? What [50:58] does each submodel functions look like? [50:59] How do they call each other? That [51:01] evolution is likely going to happen [51:03] without a human architect behind it. [51:06] It's going to be developed through a [51:08] test and deploy system and it's going to [51:10] be fully automated. And that's why this [51:12] approach that actually solving that [51:14] problem ends up becoming the future of [51:16] how model development takes off which is [51:17] why I think RSI is probably the future. [51:21] And when that happens again, it's just [51:22] wherever the data center and the power [51:24] is. [51:24] >> I remember a few years ago everyone was [51:25] like, "Oh, AGI is imminent. AGI is [51:27] imminent." And then where is it? I mean [51:30] I mean maybe we're already [51:31] >> I think we're soaking in it. [51:32] >> Okay, fine. But did it turn out to be [51:35] the Terminator, [51:36] >> right? [51:36] >> No, that would be super intelligence. I [51:38] mean artificial general intelligence is [51:39] like as smart as any human. Super [51:41] intelligence is like you're getting [51:42] towards [51:43] >> I get it. Free is talking about that's [51:45] far off. [51:46] >> I do do I think RSI could possibly [51:49] happen? Yes. I think there's lots of [51:51] black swans that could happen. It's very [51:53] hard to assess the probability of those [51:55] things. What I do know is like you're [51:57] saying, I don't think some new [51:58] government regulatory agency will stop [52:01] it if that's the thing because it won't [52:02] be technically sophisticated enough. All [52:04] it will do actually is concentrate power [52:07] around that thing. It'll limit the [52:10] options. You know, when it comes to age, [52:11] let me talk about aging. [52:12] >> We just want to have that technology in [52:14] this country is my point. And I just [52:16] want to go back to the to the point I [52:17] was asking earlier. Why do we not want [52:19] China to win? Because we want the best [52:21] labs. We want the first breakthrough [52:23] system to be here. Well, Dario made the [52:26] argument. Daario actually made this [52:28] argument in the context of trying to ban [52:30] export of chips, which is one of his [52:33] hobby horses as anthropic has really [52:35] been behind a lot of the political [52:37] activity to restrict Nvidia's sales [52:39] internationally. And the argument was [52:42] that we want to have a data center full [52:44] of geniuses in the US. And that [52:46] eventually, you know, what AI is going [52:47] to be used for is you're going to be [52:48] able to create all these agents that are [52:51] basically Nobel Prize winning level [52:54] scientists. And if we have thousands or [52:56] millions of those and China has only [52:59] dozens or hundred, whatever, then we're [53:02] going to win everything. And that was [53:03] the whole argument. But now we're [53:06] basically banning the data centers. [53:08] We're not going to have those data [53:08] centers full of geniuses. and he's done [53:11] his best to restrict sale of chips to [53:14] our allies as well. So I guess our [53:16] allies aren't going to have the data [53:17] centers either. So we are totally [53:19] shooting ourselves in the foot and [53:21] probably it would be a good thing if [53:22] Daario really does believe in the data [53:25] centers full of geniuses uh idea that he [53:28] was arguing for last year. He should [53:29] come forward and say we're making a big [53:31] mistake by overregulating data centers. [53:33] Here's how we should do it. I haven't [53:35] heard him step up in that way. And it [53:37] makes you kind of question, do you only [53:39] fight for the doomer side of the [53:41] argument or are there versions of the [53:44] positive side that you're willing to [53:46] fight for politically? Now, Freeber, to [53:48] your just one last point about RSI, I [53:50] think this is really important is look, [53:52] again, black swan, very hard to say if [53:54] it's going to happen. I, you know, I do [53:55] think that we're headed for AGI or [53:58] already in AGI or whatever, but the most [54:00] important thing about AGI is that we [54:02] have choice. like I'd much rather have [54:05] multiple AGIs than just one, right? So, [54:07] for example, Elon says that the way to [54:10] keep AI or AGI aligned is to make it [54:13] maximally truth seeeking. I kind of like [54:16] that, right? That's a way to prevent it [54:18] from becoming woke. What Anthropic says [54:21] is the way to keep it aligned is to have [54:23] this clawed constitution that is full of [54:26] liberal values and you train it on that [54:28] and they've kind of got this EA ethicist [54:30] or whatever who's writing these things. [54:31] So, there's different approaches. I kind [54:33] of want all them to produce [54:36] sophisticated AI to keep checks and [54:38] balances to keep these things in check. [54:40] What I don't want is just a Skynet or, [54:42] you know, like a woke Skynet is sort of [54:44] the the worst [54:46] >> possible outcome. And you know, if we're [54:48] going to go like sci-fi for a second, [54:49] you know, there was like a Terminator TV [54:50] show. [54:51] >> Yes, that was pretty good. They stopped [54:53] >> it was pretty good. It got cancelled. [54:54] This was like years and years ago. But [54:56] anyway, what they figured out in the [54:58] show, like the plot was that the way [55:00] they finally defeated Skynet was they [55:01] created a good AGI to take on Skynet. [55:05] They realized like humans just couldn't [55:06] do it by themselves. So, they created a [55:08] choice [55:09] >> good. [55:09] >> Yeah. They trained it to be more [55:11] >> benevolent, [55:12] >> you know, benevolent towards humans. So, [55:14] in any event, and if you're going to get [55:15] like really sci-fi about it, I look, I [55:17] think choice has to be our hallmark. [55:19] Competition has to be our hallmark. We [55:21] want multiple winners here. We don't [55:23] want a a small number of elites [55:26] guardrailing this thing. [55:28] >> And that that is what you will get if [55:30] you head in this FAA FINRA direction. [55:33] >> Breaking Freedberg. New nickname club [55:35] just just dropped. There it is. [55:37] [laughter] [55:38] It's this is what you sounded like. [55:40] Freeberg. Doc Brown. Marty. Marty. Your [55:42] future's not predetermined. You can make [55:45] your own future. You can use RSI. [55:48] [laughter] [55:50] >> Doc Brown. I'm not trying to be I'm not [55:52] trying to be an advocate for RSI. I'm [55:53] trying to be pragmatic. [55:55] >> In the future, you're back. [55:56] >> If if RSI is the way things are going to [55:58] go, you definitely don't want to be [56:00] trying to regulate AI to stall out [56:02] because someone else will win it and [56:04] then you're literally handing the keys [56:05] to RSI to them because they're suddenly [56:07] going to be able to get there faster. [56:08] And you don't want to ban the data [56:10] centers because then they're going to be [56:11] offsite. All right. There was a peculiar [56:14] story this week. And Harowitz is [56:16] allegedly being investigated. This is [56:18] all according to a Bloomberg report that [56:20] they're being investigated by the [56:22] Department of Justice for over a year. [56:24] This is all according to Bloomberg under [56:26] something called section 8 of the [56:27] Clayton Antitrust Act. This is a 1914 [56:31] law. It's like over a hundred years old, [56:33] prohibiting interlocking directorates, [56:35] which I guess is like old timey English [56:37] for directors, like board of directors. [56:40] And the conflict seems to stem from Ben [56:42] Horowitz being on datab bricks board [56:44] while another partner from Andre [56:46] Horowitz, Martin Casado sits on five [56:50] trains board. I don't know what if you [56:52] guys have heard about this story. It was [56:54] like one of the top stories this week. [56:56] Uh obviously Dre and uh Ben Harowitz [57:00] have been allin with uh Trump and big [57:03] big donators. Um put that into the mix [57:06] as you will. Why is this happening now? [57:09] >> I can tell you. [57:10] >> Enemy of Andre and Horowitz jealous of [57:13] their returns [laughter] says [57:17] >> what are we talking about? Do they not [57:18] understand how venture works? Like is it [57:20] like you invest in one company, you [57:21] can't invest in another? I don't [57:23] understand. [57:23] >> What did Bloomberg say? [57:25] >> They said they've been under [57:26] investigation for this conflict as [57:27] described between these two companies [57:29] and [57:29] >> which mean which means which means that [57:31] enemy reported them to the DOJ and then [57:34] said to Bloomberg they're under [57:35] investigation. Well, that's why I bring [57:37] it up is like it seems so odd and then [57:40] timing to me likely seems I will make a [57:43] speculative guess the CEO of one of [57:46] these companies is pissed off that the [57:47] other partners on the other board [57:49] >> and reported it to the DOJ and reported [57:51] it to Bloomberg at the same time. [57:53] >> Interesting. Uh all right. Well, there [57:55] you have it folks. Moving on in the [57:57] docket. I think we made quick work of [57:58] that one. Let me let me let me just say [58:01] you know you've been seeing like the [58:02] distribution of like cursor and then [58:05] SpaceX and Horowitz is on fire. They are [58:09] crushing. [58:11] >> So [58:12] >> they've had two in the last few weeks. [58:15] >> I mean my god congrats to them. [58:17] >> They make how many billions did they [58:18] make on [58:18] >> going? [58:19] >> Well cursor and didn't they make a lot [58:22] of money offer open router just for [58:25] seven [58:26] open router was more like an emu's [58:28] boost. was like 6 or 700 million but the [58:30] the [laughter] c the cursor was decent [58:33] was like an appetizer but SpaceX was the [58:35] main course that's like 157 billion. [58:37] >> Yeah. Now you're [58:38] >> talking just by the way even if you want [58:40] to say that VC firms should not invest [58:43] in competitors. The reality is that [58:46] startups exist in a very dynamic [58:48] environment. They frequently pivot and [58:49] change their business models and two [58:51] companies that start off in totally [58:52] different spaces can end up competing [58:54] with each other. So now these are [58:57] different board members and they can [58:58] probably set up a you know a firewall [59:00] between them. [59:00] >> Yes. Yeah. [59:02] >> Sorry. These are not these are not [59:03] filing companies. What confidential [59:05] information do any of these private [59:08] businesses have? [59:09] >> And and by the way Andrea Horowitz could [59:11] also just put an independent director in [59:13] there and just be done with it. So like [59:15] if this was some tickytacky very [59:17] important thing I don't know why the DOJ [59:19] is on this. [59:20] >> Freebrook is right. No, because when you [59:22] when you make a claim to the DOJ, they [59:24] technically if they open a file, [59:26] >> you are technically under investigation. [59:28] And so then you can you can report that [59:30] to a buffoon reporter who doesn't know [59:33] from Shyola and they'll just repeat it. [59:35] >> It means nothing. [59:37] >> All right. [59:37] >> And then the DOJ is doesn't tell you [59:39] that that you're under investigation. [59:41] And then they also don't tell you when [59:42] it's closed. [59:43] >> Same as it's just a one-way PR ratchet [59:45] for an enemy to exact some sort of [59:48] chaos. ah is crushing it. God bless him. [59:52] God speed to them. [59:52] >> It is interesting that there's even this [59:55] rule. Um but we had some fun with it. So [59:58] uh let's keep moving. I want you [60:00] Freeberg to join me and Saxs Bestie [60:03] Chimath at the All-In Summit happening [60:05] September 13th, 14th, and 15th in Los [60:07] Angeles, all has been announced. And [60:10] hey, guess who's headlighting the party? [60:13] The same guys who sell out Madison [60:15] Square Garden. I know that Freeberg has [60:17] a personal relationship. Go ahead, tell [60:19] everybody who's going to be [60:20] >> Chain Smokers, which is going to be [60:22] awesome. So, we'll have them at the [60:24] closing party. It's going to be great. [60:25] And then we're also doing another [60:28] >> Wow. [60:28] >> full takeover at Universal Studios. So, [60:30] we got a whole party at Universal, which [60:32] was awesome last year. [60:34] >> Uh, where Sax and I were with Diplo on [60:37] stage. J, you were there, too. [60:38] >> It was great. [60:39] >> Yeah, I was on stage dancing. [60:40] >> Who's Diplo? [60:42] [laughter] [60:42] >> Great. The DJiplo [60:47] our minds. come to that concert. We have [60:49] other special unannounced guests that [60:51] we're not going to announce. So, this is [60:52] it. We're done making announcements. [60:53] >> All right. We know the Democrats have [60:55] been favored to take the House uh in the [60:57] midterms, which are just around the [60:59] corner. Polymark has them at 88% [61:01] favorites, but Poly Market now has the [61:03] Dems at 52%. It's basically a tossup. Uh [61:07] current balance of power, GOP 53 seats, [61:10] Dems 47, which means I guess the Dems [61:13] have to flip four of those uh without [61:15] losing any. 35 seats are up in 2026. Uh [61:19] there are some special elections going [61:20] on in Florida and Ohio. 23 of those 35 [61:25] seats are held by Republicans. Big races [61:26] to watch. Alaska, Maine, North Carolina, [61:28] and Ohio. On February 25th, the day [61:32] before the Iran war started, GOP was [61:35] favored 60 to 41. It's been as high as [61:37] 70%. Here's your poly market. Sachs, [61:41] what do you take from this? Why are the [61:45] Republicans set to get trounced in the [61:48] midterms? [61:49] >> Well, first of all, polls are [61:52] notoriously wrong in the summer before [61:54] an election. And there's some data on [61:56] this. So, we should just if you're going [61:58] to site polls, let's just actually talk [61:59] about the topic of whether political [62:01] polls are even accurate at [62:03] >> these aren't polls. This is polymarket. [62:04] Just to be clear, the polls are slightly [62:06] different. Yeah. People [62:08] >> Yeah. But the polls are shaping public [62:09] opinion right now. So, political [62:11] scientist Patrick Graffini analyzed over [62:14] 3,000 polls over the last four election [62:17] cycles every two years since 2018. And [62:20] what he saw is that there is a huge [62:23] bias. The propoundonderance of polls in [62:25] all four cycles favored the Democrats [62:29] and the average polling error across all [62:31] the polls was D plus 3.7 relative to the [62:34] election outcome. And I think this [62:36] polling problem has become even worse [62:38] recently. And you see it where there [62:41] seems to be an oversampling of young [62:43] progressives in these polls because it's [62:46] not just the fact that the polls favor [62:49] Democrats over Republicans. They also [62:50] favor the DSA Democrat over the more [62:53] moderate or centrist Democrats. And [62:55] you've seen this in a few elections [62:57] where the DSA candidate was supposed to [62:59] win by a lot and then they basically [63:01] eaked it out or actually lost like in [63:03] Wisconsin. So, in any event, I just [63:05] think that the the polls are very [63:06] unreliable at this time of the year and [63:09] a lot of the polls have they're actually [63:11] created as internals [63:14] by campaigns to motivate their donors. [63:17] >> So, basically, people are upset [63:19] then they throw people out of office. [63:21] That's just the trend and that's what [63:23] happened over and over again. Yeah. [63:26] >> No, I just think that the polls somehow [63:27] are oversampling Democrats and [63:29] progressives and have a bias. That's [63:31] what the data is showing. So, I just [63:33] wouldn't read too much into this right [63:35] now. I do think that the Democrats I [63:38] think it is legitimate to say that they [63:39] are favored to win the House. I think [63:41] the Senate right now is a toss-up and I [63:44] would say that the Republicans will [63:46] ultimately keep it if I had to predict [63:48] right now. And I just think that when [63:50] you get back from Labor Day, that's when [63:51] the campaign for the midterms really [63:53] starts. And I think that Republicans [63:57] need to focus on the right messages in [63:59] this campaign. And I think there are [64:01] plenty of good ones. They just have to [64:02] remind him what are all the good things. [64:04] >> What are the what would you say the top [64:06] three they should focus on are? [64:07] >> Well, look, everyone's been overindexing [64:09] on the war and that's been the focus of [64:11] the news and hopefully the kinetic phase [64:13] of that is over and doesn't restart. And [64:15] assuming that's the case, I think the [64:17] Republicans can remind the electorate of [64:19] everything that's happened over the last [64:20] couple years. I think there is a good [64:22] story to tell. So, first of all, remind [64:24] people that the border has been sealed [64:26] for the last year and a half. We haven't [64:27] just slowed Biden's invasion. We've [64:30] stopped it completely. Contacts at the [64:32] border have fallen off a cliff to the [64:33] lowest level in 50 years. Illegals [64:35] aren't even really trying to cross the [64:37] border anymore. It's been so successful. [64:39] Second, sealing the border and deporting [64:41] hundreds of thousands of the worst [64:42] criminals has actually made the country [64:44] much safer. Murders have dropped around [64:47] 20% year-over-year, and the murder rate [64:49] is the lowest in 70 years. So, that's [64:52] number two. Third, deaths from drug [64:54] overdoses have dropped 20% again thanks [64:56] to tougher border enforcement. Fourth, [64:58] Americans saw a 17% increase in tax [65:00] refunds thanks to the president's no tax [65:03] on tips and no tax on overtime and no [65:06] tax on social security. So, these are [65:08] real increases in take-home pay as a [65:11] result of those tax cuts. Also, the tax [65:14] provisions in Big Beautiful Bill [65:15] incentivize new equipment buying. We [65:17] talked about the accelerated [65:18] depreciation. And perhaps that's why [65:20] small business optimism just reached its [65:22] highest level in nearly a year. and more [65:25] small businesses are saying they plan to [65:27] hire in the months ahead. So, I think we [65:29] do have a strong economy. Prescription [65:32] drugs and auto insurance costs both fell [65:35] last year after increasing a lot during [65:38] the Biden administration. I think [65:40] prescription drugs has had their [65:41] steepest annual decline since 1963. [65:45] So, there are positive things to talk [65:46] about there. Everyone made a big deal of [65:49] the surge in egg prices at the start of [65:50] the administration, but few have [65:53] bothered to note that there's now been a [65:54] 39% decrease year-over-year in July. So, [65:57] look, there are many good things, I [65:58] think, for Republicans to talk about. [66:00] They just have to kind of remember [66:02] everything that's happened in the last [66:03] two years. [66:03] >> Okay, Fber, you have thoughts on what is [66:07] going to happen this midterm and what [66:10] people are making their decisions based [66:13] on. [66:14] >> Things are expensive. people can't [66:15] afford housing, health care or education [66:19] and are leaning towards this DSA [66:21] concept. There was a frightening [66:23] statistic on young conservatives [66:26] embracing socialism. [66:27] >> Really? Wait, the the conservatives like [66:31] MAGA folks [66:34] you're saying are embracing socialist [66:36] tendencies? [66:37] >> How are they conservatives then? Well, [66:39] they I guess obstensively they're [66:41] flipping or they're changing the [66:43] composure of the issues they care about [66:45] in some way. [66:46] >> 53% of self-identified conservatives [66:50] under 40 support governmentr run grocery [66:53] stores, not Democrats, conservatives. [66:56] Take that in first line of this Wall [66:58] Street Journal article. July 2026 Fox [67:01] News poll found that 61% of Republicans [67:05] view capitalism favorably, down from 72% [67:08] in 2019, [67:10] while the share who feel strongly [67:12] favorable has fallen from 54 to 41%. [67:16] The share who say the system is rigged [67:18] towards the wealthy has climbed from 30% [67:20] in 2018 to 42% amongst Republicans. [67:24] So, I don't think that this is a [67:27] principled line in the sand anymore. [67:31] I think that the the problem with [67:35] unaffordability, which I would argue [67:38] continuously, and I've argued it since [67:39] the first day of the show, is a function [67:42] of government spending. The more [67:43] government spends on education, [67:45] healthare, and housing, the more [67:47] expensive those things get rather than [67:49] the opposite. They're supposed to be [67:51] able to provide access to those things. [67:53] But when the government intervenes in [67:55] those markets, those things get more [67:56] expensive. And no one wants to kind of [67:58] acknowledge or admit that in either [68:00] party. And I've been to DC enough now to [68:03] know that no one actually acknowledges [68:06] or cares about this point that they all [68:08] think that their job is to get more [68:10] money out of the government for their [68:11] constituents. Every single person in [68:13] every party. There's very few principled [68:16] conservatives. [68:17] Many of them are conservatives by title, [68:20] by party, affiliation, but they're not [68:22] conservatives by principle in the [68:24] traditional sense, small government, [68:26] free markets, etc. And I think that this [68:30] is the motivating factor. I I would [68:32] argue if you take the look at the macro [68:35] cycle, the macro cycle being we've [68:39] printed so much money, we've got so much [68:41] debt, the 30-year is now at 5.3% this [68:43] week. it continues to hit new 20-year [68:46] records every day. We I think are going [68:49] to continue to face a problem of [68:51] unaffordability [68:53] and ultimately there is going to be some [68:55] sort of uh big movement towards more [68:58] socialist policies that we're going to [69:00] see play out between now and 2028. It's [69:02] why I continue to believe I think [69:04] someone like an AOC becomes president. I [69:07] think there's going to be a wave this [69:08] year and then there's going to be a [69:09] bigger push in 2028 because I don't see [69:11] the arithmetic or the fiscal situation [69:14] changing to actually alter the [69:17] affordability landscape. I just don't I [69:20] I don't know what you could do from a [69:21] policy perspective and I don't hear [69:22] either party talking about actual [69:24] solutions to solve this affordability [69:26] problem between now and 2028. And then [69:29] the other kind of big problem that I've [69:30] talked about is assets. middle class [69:33] people, people that worked in private [69:35] companies, people that don't depend on [69:36] social security or public pension funds, [69:39] they've all had 401ks and owned homes. [69:42] So, they have seen their net worth climb [69:45] with asset inflation surge. So, the top [69:48] 50% of Americans have [69:52] 178 trillion of net worth [69:55] >> and the bottom 50% of Americans have 6 [69:59] trillion of net worth. It's not the [70:01] billionaires. It's the the whole middle [70:03] class. The middle class has all the [70:04] wealth in the United States. And that's [70:06] because they've owned assets. They've [70:07] owned houses and they've owned stocks. [70:09] Houses and stocks have surged in value. [70:12] And so if you don't own a house and you [70:13] don't own stocks, you were left behind. [70:16] And so I don't see how from a public [70:19] policy perspective we don't find [70:20] ourselves walking down this very nasty [70:23] winding dark path of socialism because [70:26] the alternative policies that could be [70:28] put in place to solve these problems are [70:30] not being put forth by either party [70:32] right now. So I do think that there's [70:33] going to be a big socialist embrace. [70:35] I've said this obviously for a number of [70:37] years and I think we're going to see [70:38] this um play out in this midterm and [70:41] that's going to create the the momentum [70:42] going into 28. And look, I mean, this [70:44] this polling data is very worrying. It's [70:46] not just the Democrats. It's not just a [70:48] party based system. It's the horseshoe. [70:50] It's the 63% of Americans that are [70:53] living paycheck to paycheck that can't [70:55] afford. That's the Americans that are [70:57] going to vote for this system. And it [70:58] doesn't matter what party it is. They're [71:00] just going to vote for it. [71:01] >> Shimoth 20 cents on the dollar going to [71:04] our interest approximately on the [71:05] federal side. Uh inflation, [71:08] we thought we would get there. It's been [71:10] 3 or 4% for the whole year. um haven't [71:13] been able to get down to two and the GDP [71:16] uh concept which we talked about here [71:18] getting to growth again and if we can [71:20] get it we can grow our way out of it [71:22] that is hasn't happened as well. [71:25] Obviously the war big disappointment for [71:27] many folks in both parties who were [71:30] promised no war uh and here we are 6 [71:33] months into it. What do you think is [71:35] going to happen over the next uh through [71:38] the midterms and then into 2028 and to [71:40] free Birch Point? any way to get out of [71:42] this? Is this a debt spiral or is this [71:44] something we can course correct out of? [71:46] In your mind, [71:48] >> I think that we can course correct. I [71:49] just think it takes will. And I think [71:52] that we need [71:55] corporate America to step up and do [71:58] their part of the job. And we need [72:01] governments mostly to get out of the way [72:03] and not have any hairbrain schemes on [72:06] new inventive legislation that they [72:08] think can solve the problem. I don't [72:10] trust any polls. Just to be clear, I [72:12] don't believe this polling data. And I [72:16] think that we've gotten into a very [72:18] dangerous point, which we already got [72:20] into with the media, where the media [72:23] lies. We all now know that the media [72:26] lies, and the media can't be trusted. I [72:29] would like to now go out on a limb. I [72:31] don't think it's much of a limb, though. [72:32] Pollsters lie and I don't think you can [72:35] trust them either. And I think that the [72:38] media [72:40] acts in concert with pollsters [72:43] to [72:44] find the data that allows them to write [72:47] the headline that hopefully allows them [72:49] to get the clicks. So I would just start [72:51] there. Pollsters trash [72:55] media trash. So what is really happening [72:58] on the ground? I don't know. But I don't [73:01] trust what this data tells. Now, what do [73:03] I know is helping and hurting the [73:06] incumbent sitting president? I think [73:08] that if we as in corporate America get [73:14] ourselves in a position to actually [73:17] present the better face of capitalism, [73:21] we allow the system as it stands to [73:23] continue to work. [73:25] And I think all of these things, if we [73:28] don't wake up, are telling us that [73:32] corporate America is not trustworthy [73:34] anymore. [73:36] Corporate America is not aspirational. [73:38] Corporate America does not do the things [73:41] that engender some amount of trust and [73:44] respect. [73:46] And specifically in Silicon Valley, [73:48] we've completely lost the script because [73:50] we've gotten to the far far end of that. [73:53] And everybody is in it for themselves to [73:56] such a degree that it's either pearl [73:59] clutching on the one hand or getting [74:02] your hand caught in the cookie jar on [74:03] the other. You know, I mean, I I posted [74:04] the cookie jar thing that Anthropic did [74:07] on the healthcare side. [74:08] >> Cookie jar. [74:09] >> Why are we Why are we in this? I mean, [74:11] it's just so avoidable. Those are [74:13] headwinds for the incumbent president, [74:15] even if he has nothing to do with it. [74:17] What is a tailwind? I think getting out [74:19] of the Iran war can be a massive [74:23] tailwind. [74:24] I think structurally solving energy [74:29] on behalf of American citizens [74:32] can be a tailwind. [74:35] I think making sure that all of the [74:39] long-term and mediate-term investments [74:41] of GDP can actually come and bear fruit [74:44] can be a tailwind. all of those things. [74:46] There are still a lot of levers to be [74:49] pulled between now and November. And so [74:51] those can create pretty meaningful [74:53] tailwinds, but the headwinds, [74:55] unfortunately, are pretty strong. And it [74:57] requires some real coordination that the [75:00] White House is going to have to do to [75:01] get a bunch of these corporate CEOs in [75:03] the room, to get a bunch of these [75:04] Silicon Valley CEOs in the room, and to [75:06] tell them, "Guys, enough of this [75:07] Play nice in the sandbox [75:09] because you're about to lose the [75:10] standard operating procedure that [75:12] allowed you to exist in the first place. [75:14] Yeah. [75:15] >> And they need to get this message [75:16] clearly. [75:18] >> Yeah. And I I can tell you the exact [75:19] reason Democratic socialism is going to [75:22] win, continue to win. Um and these [75:25] far-left Democratic candidates are going [75:27] to dominate. And it's, you know, it's [75:28] the Republicans's fault. They had all [75:30] three branches of government here really [75:33] tightly in their control. And if you [75:35] look at what the podcast bros uh and the [75:38] pro podcast preferencing stack is [75:41] responding to and I listen to them, you [75:43] know, not every episode, but I I I I dip [75:45] in and out of Tucker who is now making [75:48] his own political party, Joe Rogan, Tim [75:51] Dylan, you know, all these folks who [75:53] were all in for Trump. They cite Epstein [75:56] investigation, which okay, whatever. I [75:59] don't think it's particularly material, [76:00] but it is to them. The Iran war is [76:02] obviously top of what they believe is [76:05] the problem here because they were [76:06] promised that wouldn't happen. And then [76:09] it's to your point Freeberg. This [76:11] inflation thing has not been solved. And [76:14] it just keeps going. And we're going to [76:17] be sitting here I think in another 6 [76:19] months or 12 months talking about my [76:21] lord every investigation possible into [76:25] the Trump administration. And we need to [76:28] start looking at what what what would be [76:31] things to your point Shamath and I [76:34] thought you summarized some of the the [76:36] broken parts of capitalism I think well [76:38] but we got to get granular about it and [76:40] I brought up last week just the edge [76:43] behaviors in capitalism where we [76:45] overoptimize. I felt the um delivery [76:48] providers at Sax and I had a pretty [76:50] hearty debate about it last week were [76:52] just being too clever, squeezing too [76:54] much out of the the last mile quite [76:56] literally on Amazon's part. How we treat [77:00] the people who are at the bottom of the [77:02] employment stack is what's causing this [77:05] and h not being able to solve the cost [77:07] of homes and not making that a focus. [77:10] Not being able to solve u the cost of [77:13] child care. not being able to move the [77:16] minimum wage. I think we have to look at [77:18] all of these things and say that's or [77:20] Trump should look at all of those things [77:22] and he should build a second half of his [77:24] term just around those things. No more [77:26] crypto stuff, no more wars, no more [77:30] attacking our allies. All of this [77:32] performative stuff and and just [77:33] mistakes, they got to go away and we got [77:35] to focus on what he sold people on. [77:37] People feel they were sold that this [77:39] would be a populist uh presidency that [77:42] took care of the working man and woman [77:44] in this country and it has been the [77:47] opposite. Sax, I'm sure you have some [77:49] thoughts. We'll give you the final word [77:51] on the inflation thing because you [77:52] talked about it and Freeberg's talked [77:54] about it. I mean basically you're saying [77:55] the cost of living is going to drive [77:56] everyone to socialism. I agree there's [77:58] some risk of that but let's break it [77:59] down into a little bit more nuance here. [78:02] If you look at where the most inflation [78:05] has occurred, it's always in those [78:07] sectors of the economy that the [78:08] government's most involved. So [78:10] healthcare and colleges where and [78:14] housing where the government has massive [78:16] amounts of subsidies and regulation, it [78:19] only seems to make the problem worse. [78:21] >> Kind of makes sense that if the [78:22] government's paying for something, then [78:25] the providers have a huge incentive to [78:28] charge as much as possible. And in fact, [78:30] a lot of them engage in fraud as well. [78:32] So we know these areas are rife with [78:33] abuse. But those are the areas that are [78:36] most inefficient and where you see the [78:37] price going through the roof. I mean, [78:39] hospital services, man, that's like just [78:42] exponential just about. And then if you [78:44] look at the areas where we've had not [78:46] just no inflation, but actually prices [78:48] have come down. It's in all the tech [78:51] areas. It's basically TVs or computer [78:54] software, cell phones, things like that. [78:57] the prices have come down a lot. By the [79:00] way, this chart is showing all the way [79:01] from 2000 to 2026. So, it's a 26-year [79:06] time frame. And during that time, [79:08] overall inflation went up 96%. So, the [79:11] price level almost doubled in that [79:14] 26-year period. Fortunately, [79:16] >> on average, [79:17] >> on average, that's right. But like I [79:19] said, with certain things like [79:21] healthcare becoming, you know, 300% [79:23] almost more expensive, whereas computer [79:26] software went down by 67% in real [79:29] dollars. Meanwhile, the average hourly [79:32] wage grew 135%. So at least the average [79:35] hourly wage grew by about 40% more than [79:39] the inflation level, but obviously it [79:41] still bites that you have all this [79:43] inflation. I guess my point is just how [79:45] do you solve it? We know that you bring [79:47] prices down through vigorous [79:49] competition, not through more government [79:52] interference in these markets. And [79:54] unfortunately, what happens is that [79:57] socialism causes the problem. And then [80:00] what happens is that groups like the DSA [80:02] argue that the way to solve the problem [80:04] is through more socialism. So it's a [80:06] disease that claims to be a cure and [80:09] then people take it and they take more [80:11] of it and the disease gets worse and [80:13] worse. And that that is the spiral that [80:15] I think we risk being in. [80:16] >> We are in a socialist debt spiral now. [80:19] All this uh spending now. [80:21] >> We're not in a we're not in a death [80:23] spiral. Take it easy. [80:24] >> Um [80:25] >> we're not in a death sp. [80:28] It's great for you to say because it [80:29] sounds like really like you got some [80:31] real hot take. We're not there. [80:34] >> America's doing fine. We are cranking. [80:37] The economy is great. There are things [80:40] that need to be fixed. I would say start [80:43] with corporate America. Let's get our [80:45] act together and let's start doing [80:46] things that are honorable, interesting, [80:48] and aspirational. [80:49] >> Yeah, I think you totally missed the [80:50] point. I think the real point here is if [80:52] he hadn't started the war, we wouldn't [80:54] have the same inflation issues. We [80:55] wouldn't have the same expense issues, [80:57] and this would be a totally different [80:58] picture and it was his job acts when he [81:01] promised the American people he would [81:03] focus on affordability. And he had [81:05] control of both houses and the [81:07] presidency. and he's willing to do any [81:08] executive order on crypto, on wars, on [81:11] anything, but he didn't do any executive [81:13] orders on affordability. [81:16] >> Maybe you missed the point where Chamas [81:18] said it'd be a tailwind to get out of [81:20] this war. I think the kinetic part's [81:21] already over, by the way. So, it's just [81:23] >> let's hope. I mean, it just should never [81:24] have been started. Um, and I think [81:25] >> but in any but in any event, look, the [81:27] president passed his economic program [81:29] last year on big beautiful bill and I [81:31] think it's working. We're seeing a huge [81:32] investment boom. We're seeing small [81:34] business confidence rise to its highest [81:35] level in a long time. They are saying [81:38] they're going to create more jobs over [81:39] the next year. I think the economy is [81:41] doing pretty well. I mean, we do have [81:43] these issues. I'm not saying we don't. [81:45] And spending is a long-standing problem. [81:47] Can't disagree with you there. However, [81:49] I don't think Democrats have the [81:50] solution to the problem. And the [81:52] Democratic socialists definitely don't [81:53] have the solution to the problem because [81:54] they want more spending. [81:57] death is that we we are now [81:59] incentivizing young people in the [82:01] Republican party to embrace socialism. [82:03] Moderates are embracing socialism. [82:05] They're all looking at it saying, "Well, [82:06] that sounds better. If Trump won't be a [82:09] populist, maybe these people are the [82:10] populist." And and that's what I mean by [82:12] the death pile because they're going to [82:13] just keep spending money. They're going [82:14] to increase the velocity of spending. [82:17] >> You may be right, which is scary. [82:19] >> By the way, [82:19] >> that's what I I don't think we're like [82:21] in it now. We're kind of we're risking [82:23] going [82:24] >> just one just one last point. They did a [82:26] cost estimate of the policy platform [82:29] proposed by the DSA. Remember we've [82:31] talked about that before. [82:32] >> And the estimate is that it would [82:34] project new federal spending of between [82:38] 71 trillion and 212 trillion over [82:45] >> by the way total prison. There's no [82:48] entirety of the American population 182 [82:50] trillion. [82:52] >> 182 trillion. Yeah. [82:53] >> Okay. But I'm just saying that's how [82:54] much the TSA program would cost if you [82:57] guys believe that the root of our [82:58] problems is overspending and basically [83:01] too much government. Unfortunately, [83:04] >> this is not the [83:05] >> the DSA is going to [83:06] >> No, but here's what's going to happen. [83:07] And I'll just let me just walk through [83:09] the sequence of events. Sequence one, [83:11] they they enact a policy starting in [83:14] 2028 with Comrade AOC in charge and the [83:17] Pilot Bureau uh Roano, Secretary of [83:19] Housing and Urban Development and the [83:22] rest of the Pilot Bureau organizing [83:23] themselves to launch these new programs. [83:26] Those new programs get launched and [83:27] costs balloon. The bond market sells off [83:30] 30 years. And so now the 30-year yield [83:33] spikes 7% 8% maybe more. things become [83:38] more unaffordable. So then the question [83:40] is, okay, well, we can't bridge it from [83:42] a fiscal perspective. We can't just get [83:44] the Fed to start buying bonds. There's [83:46] not enough buying power there to bring [83:48] rates down, to bring inflation down. So [83:51] we need more income. So now we need to [83:53] pass a national [83:55] asset seizure tax. And there's not [83:58] enough money in the billionaire class or [83:59] the 50 millionaire class. The total net [84:02] worth of all people that are worth over [84:04] $50 million is only 23 trillion. You [84:07] could take 10% of their net worth and it [84:09] would only cover 3 months of fiscal [84:11] spending, today's fiscal spending. So at [84:14] the end of the day, the only way to get [84:15] there is to start doing massive taxation [84:18] seizures. And so seizures seize. [84:21] >> So this just came out this week [84:23] >> KO [84:23] >> by KO [84:24] >> by the KO Institute. And again, it shows [84:27] the low end of the DSA spending is 71 [84:30] trillion. The high end is 211 trillion. [84:34] Now, you compare that to things like [84:36] corporate profits or uh private sector [84:38] wealth, corporate profits are only 35 [84:40] trillion. The Forbes 400, all the wealth [84:43] of the Forbes 400 is only 6.6 trillion. [84:47] So, it's less than 10% of the low end of [84:51] DSA spending. It's I I keep telling you [84:53] this is not about mathematical realism. [84:56] This is about anger towards a class of [85:00] people that are increasingly odious and [85:02] >> mistrust. But people also can't afford [85:04] to buy a house. People are 35 years old [85:06] with roommates. [85:07] >> And people and people go to the grocery [85:09] store and they're shocked at the fact [85:11] that they have to spend an entire [85:12] paycheck or a third of a paycheck to buy [85:15] groceries. [85:15] >> It's not it's not that it's not that [85:17] people want to buy a yacht. It's that [85:19] the yacht class are just [85:21] >> No, I mean Well, no. I mean, look, in [85:23] every socialist um revolution, every [85:25] revolution we've seen historically [85:27] around this, there is always an elite [85:30] crunch of society, and they're typically [85:32] classified different ways. It's it's [85:34] wealth, it's Jews, it's some race, it's [85:37] immigrants, there's some there's some [85:39] organizational assignment, some identity [85:41] assignment that says this is the cause [85:43] of unaffordability. At the end of the [85:45] day, the dysfunction of the relationship [85:48] between government and all the other [85:50] organizations that make up that society. [85:52] Those organizations can be states, they [85:54] can be private companies, they can be [85:56] the NOS's. It is the dysfunction of the [85:58] government's relationship with the other [86:00] organizations that fundamentally causes [86:03] things to become more expensive and [86:05] dysfunctional and the actual market for [86:07] getting access to things for the [86:09] individual fails. And at that point you [86:12] have to identify and you don't identify [86:13] the government because the government is [86:14] the aggregate of all the people. So the [86:18] and remember the government is the [86:20] system the means of action. [86:22] >> So for the people to organize they have [86:24] to organize the government [86:26] >> to go and then they have to have an [86:27] alternative system to go and destroy. [86:30] >> You have to point the finger and you [86:32] have to point the finger at somebody. [86:33] And if you're going to scapegoat [86:34] somebody in 2026, [86:36] >> the best class of people to scapegoat [86:37] are the billionaires. and frankly even [86:39] better than them. [86:40] >> But at the end of the day, it's all [86:41] tech, right? The data center is the [86:43] temple where the billionaires and [86:44] trillionaires pray. It's where they go [86:46] to make their sacrifices. It is their [86:49] gathering spot. So I think the data [86:51] center is this like, you know, temple on [86:53] the mountain. And so that's why I think [86:55] the data center gets tied up in all of [86:56] this. There's no rational discourse [86:58] about data centers. It's like, [86:59] >> no, no, look, you're if you take it [87:01] through the Rene Gerard lens, everything [87:03] maps. It's like the billionaires will be [87:05] scapegoed and the trillionaires will be [87:08] I mean even worse and because that will [87:11] allow the catharsis that will allow [87:13] society to reset and realize wait a [87:15] minute oh my gosh [87:16] >> yeah there's a generation of that though [87:18] there's a generation ahead [87:19] >> and only then it is a long process it is [87:21] a long [87:21] >> Here's an idea next president or Trump [87:24] in his second half just putting it out [87:26] there do an executive order that you're [87:28] going to build like a million homes two [87:31] million homes if you can have executive [87:33] orders If you can start wars without the [87:35] approval of Congress, [87:36] >> that's not how it works, right? [87:37] >> What's that? [87:38] >> You know that that's not how it works, [87:40] right? There are local pulled out all of [87:42] the stops laws, [87:44] >> of course. But he pulled out all the [87:45] stops for getting data centers green [87:48] lit. We know that. And he worked with [87:50] the state governments. I'm talking about [87:52] him coming out saying, you know what, we [87:53] did a great job for the data centers. We [87:54] did a great job for the crypto industry. [87:56] And now we're going to do two great [87:57] things. Hold on, let me finish. We're [87:59] going to do two great things. We're [88:01] going to make trade shows. We're going [88:02] to make trade schools and we're going to [88:04] make affordable homes. I'm just going to [88:05] make that the next 24 months. We're [88:07] going to figure out a way to mobilize [88:09] the private sector, mobilize the public [88:11] sector. We're going to make more homes [88:13] and we're going to get you a great [88:14] education that turns into a job. That [88:16] would be a great thing for Trump to do. [88:17] >> The federal the federal government does [88:19] control the largest amount of land in [88:22] the United States. So, it could do [88:23] something on federal land without the [88:24] need for a bunch of other state and [88:26] local gobbledegook. So, there is [88:28] >> May Freddy M. Now there it may not be [88:31] located in places that actually make [88:33] sense for commerce and for people to [88:35] live or that are already living there [88:36] because I don't think it really abuts [88:38] >> you know uh most cities. So [88:41] unfortunately it's not in his control [88:44] even if he wanted to just kind of say [88:45] let's make this happen. I think if the [88:48] president can go spend a couple hundred [88:50] billion on this war, he can go spend 100 [88:52] couple billion on [88:53] >> homes that are affordable. And maybe he [88:56] can't put them in Manhattan, but he can [88:57] put them other places. And he should [88:59] fight for the regulations to get just [89:01] like he fought for the regulations on [89:02] data centers and energy. He made an [89:05] point on this very podcast at the AI [89:07] summit. Hey, we're going to clear all [89:09] the pathways to nuclear to NAC gas. [89:12] We're going to beautiful clean [89:14] federalism state. I give him credit for [89:16] that. Now [clears throat] do it. Now do [89:18] it for homes. Let me see you do it for [89:20] homes. Trump, use that same energy. [89:22] >> They just passed a housing bill. [89:24] >> They just they actually just passed a [89:25] housing bill. [89:26] >> When's the last time we heard Trump? [89:28] >> Unfortunately, it was written by [89:29] Elizabeth with Warren, so I don't know [89:30] how good it is, but [89:31] >> yeah. Anyway, I'm just giving him the [89:33] solution to save his legacy. All right, [89:35] everybody. [89:36] >> As always, you're looking out for [89:37] everybody. [89:39] Looking out for the audience. Nois, [89:41] thank you. [89:42] >> Nois is also looking out for you. [89:44] >> Okay, rap. I got to go. Come on. All [89:45] right, everybody. It's been another [89:47] amazing episode. [89:48] >> Love you, boys. Nice to be on podcast. [89:50] Good to have the quartet back. Let's go. [89:53] >> Byebye. [89:55] [music] [89:56] >> We'll let your winners ride. [89:58] >> Rain [90:03] and [music] we open sourced it to the [90:05] fans and they've just gone crazy with [90:06] it. Love you. Queen of [90:11] [music] [90:12] >> your [90:16] >> besties are gone. [90:18] >> That is my dog taking notice your [90:20] driveway. [90:23] >> Oh man, myasher will meet me at [90:26] >> We should all just get a room and just [90:27] have one big huge orgy cuz they're all [90:29] just useless. It's like this like sexual [90:31] [music] tension that they just need to [90:32] release somehow. [90:36] Wet your feet. [90:37] >> Your feet. [90:39] >> We need to get murky. [90:49] I'm [music] going all in.