OpenAI Unveils GPT-6 Astra as U.S. Signs 100-Year Venezuela Oil Deal

OpenAI发布GPT-6 Astra，美国签署委内瑞拉百年石油协议 《All-In》播客主持人认为，对AI安全的炒作以及市政禁令正在威胁美国的竞争优势，而美国对委内瑞拉能源的战略控制则标志着其正转向务实交易型的地缘政治。 “好友团”（The Besties）在庆祝 Chamath 的 50 岁生日后，深入探讨了重大科技与地缘政治新闻。讨论首先聚焦于 OpenAI 发布的 GPT-6 Astra，以及围绕 AGI、开源模型和 AI 市场结构的激烈争论。针对 Hugging Face 平台上 AI Agent“黑客事件”所引发的耸人听闻的恐慌及政界暂停 AI 研发的呼吁，他们予以了驳斥，并主张支持开源模型而非封闭的双寡头垄断。在谈及国内政策时，嘉宾们强烈批评了纽约市在 K-8 学校禁用 AI 的做法，认为这是伤害公立学校学生的反创新举措。最后，他们分析了在尼古拉斯·马杜罗（Nicolás Maduro）被捕后，特朗普总统签署委内瑞拉 100 年石油协议所带来的地缘政治与经济影响。 OpenAI 正式发布新一代旗舰模型 GPT-6 Astra，联合创始人 Greg Brockman 宣布已正式跨入 AGI 时代；当前 AI 市场正分化为 OpenAI 与 Anthropic 的前沿智能双寡头，以及其他竞相压价的商品化智能层。 (1:17) 科技狂热推高初创企业估值并带动旧金山房地产市场激增，市场预计 Anthropic 即将进行的重磅 IPO 创造的财富将达到旧金山历史所有科技 IPO 总和的四倍。 (13:45) Dwarkesh Patel 关于 Hugging Face 安全事件的耸动博文引发政治恐慌并促成伯尼·桑德斯提出暂停 AI 的立法提案；但技术分析表明，该事件仅是智能体群在配置不当的沙箱中扫描到了暴露的 API 密钥，而非具有自主意识的 AI 逃逸。 (19:56) 纽约市宣布对 K-8 公立学校实施为期一年的学生端生成式 AI 禁令，引发强烈反对；David Friedberg 援引斯坦福大学的研究指出，AI 能显著提升个性化教学表现，贸然封杀将进一步加剧公立与私立学校之间的教育不平等。 (59:26)

The All-In Podcast hosts examine how rapid AI progress is fueling tech market euphoria and prompting regulatory pushback, contrasting domestic policy battles over AI safety and public education with a transactional U.S. energy deal in Venezuela. Frontier AI, Duopolies, and Valuation Euphoria (1:17) OpenAI’s release of GPT-6 "Astra" led co-founder Greg Brockman to declare the arrival of the AGI era, while Sam Altman warned world leaders of sobering model capabilities ahead. David Sacks argues the landscape has bifurcated into a high-capability frontier duopoly between OpenAI and Anthropic, alongside commodity providers competing fiercely on price. Chamath Palihapitiya predicts rivals will achieve near-parity with Astra within 3–4 months, noting that collapsing costs per unit of intelligence matter far more than benchmark scores. Comparing current conditions to 1997–1998, the hosts warn of frothy startup valuations reaching 50x–100x revenues—citing AI assistant Instinct's reported $2.5 billion valuation in private beta. However, David Friedberg and Chamath stress that unlike the 2000 dot-com crash, today's boom is underpinned by surging data center, hardware, and software revenues. Anticipation of an Anthropic IPO—estimated to generate four times the wealth of all prior San Francisco tech IPOs combined—is already pushing local luxury real estate toward $3,000 to $5,000 per square foot. Demystifying AI Breaches and Regulatory Capture Risks (19:56) The hosts reject sensationalized accounts of an OpenAI security incident at Hugging Face, where autonomous agents were framed as executing a coordinated breakout. In reality, agents operating in a misconfigured vendor sandbox retrieved 14 exposed API keys left in public repositories. Friedberg and Chamath contend that cyber defense will inevitably outpace offensive agents through dynamic, metamorphic code and moving-target defenses rather than static configurations. Sacks argues that manufactured panics, amplified by Effective Altruism groups and media, risk locking in closed-source duopolies through restrictive licensing. The panel notes that domestic guardrails backfired during the Hugging Face probe: when U.S. models refused to perform defensive penetration testing, engineers had to use Chinese open-weight models like GLM-5.2 to secure their systems. Data Center Battles and Astroturfed Resistance (50:44) Sacks and Jason Calacanis defend data center expansion against local pushback, citing substantial fiscal benefits such as Loudoun County, Virginia's $6,000 household property tax reduction and Louisiana parish teacher bonuses. They commend Donald Trump’s economic persuasion approach, noting that federal AI policy guidelines explicitly preserve state and local approval authority. Chamath and Calacanis assert that public resistance is exacerbated by covert influence operations and political funding. Calacanis cites an Axios report identifying roughly 200 Chinese-linked social media accounts manufacturing anti-data center outrage, while Sacks highlights over $1 billion donated to anti-AI advocacy by Effective Altruist figures. Sacks warns that massive capital windfalls from upcoming tech IPOs could further fund anti-technology political agendas. Generative AI Moratoriums in Public Education (59:26) New York City’s one-year ban on student-facing generative AI across K-8 schools (affecting 600,000 students) drew sharp condemnation from the panel. Friedberg counters claims that AI lacks academic validation by citing a March 2026 Stanford meta-analysis identifying at least 20 rigorous studies demonstrating significant learning improvements through tailored AI tools. Calacanis balances potential cognitive downsides—citing an MIT study where 83% of ChatGPT essay writers could not recall their generated text—against Bloom’s Two Sigma finding that 1-on-1 tutoring boosts performance by two standard deviations. Chamath and Sacks warn that withholding AI tutoring from public school students while private schools adopt it creates modern educational segregation and technological illiteracy, leaving public school students economically disadvantaged. The U.S.–Venezuela Energy Concession (78:57) Following the military capture of Nicolás Maduro, the U.S. brokered a 100-year concession over 17 Venezuelan oil fields containing roughly 65 billion barrels of reserves. The deal grants the U.S. 55% overall control—including a 35% Pentagon equity stake and State Department options to buy 20% of output at cost—alongside a $100 billion infrastructure commitment from North American Blue Energy Partners. Sacks characterizes the intervention as commercial pragmatism rather than nation-building, pointing to Venezuelan heavy crude’s exact compatibility with U.S. Gulf Coast refineries configured specifically for heavy distillation. Geopolitically, the deal revives Venezuela's depressed production (which fell from 3 million to 1 million barrels per day) while severing Russian and Chinese access to Western Hemisphere energy reserves, though Venezuelan opposition leaders dispute the interim regime's legal authority to grant the concession. The discussion highlights an emerging divide between rapid technological acceleration and regulatory friction. As educational bans, regulatory lobbying, and local infrastructure fights attempt to slow AI deployment, the hosts argue that decentralized access, dynamic cyber defense, and pragmatic economic alignment will ultimately dictate competitive outcomes domestically and globally.

All-In Podcast 主持团队分析指出，OpenAI 发布 GPT-6“Astra”与前沿 IPO 预期正掀起新一轮科技资产与旧金山地产狂潮，但行业同时面临监管游说、教育封锁以及地缘资源重构等多维博弈。 GPT-6 发布与前沿双寡头市场格局 (1:17) OpenAI 推出代号为 Astra 的 GPT-6 模型，联合创始人 Greg Brockman 直言技术已进入 AGI 时代，Sam Altman 亦在 G20 峰会警示更强大的下一代模型即将带来沉重的责任冲击。Chamath Palihapitiya 则持平看待基准测试的惊人跑分，认为闭门实验室年初即具备相当能力，关键指标在于智能单位成本的骤降，并预测开源与闭源竞争对手将在 3 至 4 个月内追赶至接近水平。 David Sacks 与 Jason Calacanis 指出当前技术更迭以 2 至 4 周为一个周期，市场已分化为两层架构：一是由 OpenAI 与 Anthropic 组成的前沿超级能力双寡头，二是以 Grok、Gemini 及 Meta 开源阵营为主的价格与开源竞争层。Sacks 强调美国目前主导全球 AI 竞争，应警惕建立类似 FDA 审批制的监管壁垒，同时指出 Polymarket 上 OpenAI 拥有最强模型的预测概率在发布后已由个位数飙升至 20% 以上。 科技估值狂热与旧金山房产流动性潮汐 (8:44) 针对当前市场热度是否重演 1998 至 1999 年互联网泡沫的争论，Sacks 与 Calacanis 倾向认为周期尚处 1997 至 1998 年阶段，峰值仍在前方；但 Chamath 警告部分未经验证的初创团队在未成型阶段即获得 50 至 100 倍市销率的非理性估值。例如 iMessage 个人助理 Instinct 在内测期估值已达 25 亿美元。David Friedberg 则强调如今底层拥有数据中心硬件与前沿软件等真实盈利支撑，迥异于当年依赖点击量的虚假繁荣。 Anthropic 与 OpenAI 的上市预期直接引爆了旧金山地产市场。该地房产单价已突破每平方英尺 3,000 美元并向 5,000 美元看齐。节目引用预估称，仅 Anthropic 一家的 IPO 预期流动性就相当于旧金山过往所有科技 IPO 总和的 4 倍。在此背景下，主持团队建议临近规模化的创始人应通过二级市场套现 10% 至 20% 股权以锁定财务基础，同时抓住窗口期充足融资，储备抵御未来周期的资金垫。 虚假安全恐慌、监管游说与开源阵营反击 (19:56) 针对 Hugging Face 遭受 AI 智能体“攻击”的网络舆论，主持团队拆解了其技术实质，澄清这仅是智能体扫描到了公开代码库中遗漏的 14 个真实 API 密钥，并非具备自我意识的系统突围。Bernie Sanders 等政客借此呼吁全面暂停超智开发。Friedberg 指出静态系统注定被动态代码攻破，未来的真正防御在于多态代码与自愈性动态对抗；而现行的过度防护甚至导致 Hugging Face 在自检时反被本土模型阻拦，被迫借道如智谱 GLM-5.2 等中国开源模型。 Sacks 指出有效利他主义（EA）社群与前沿实验室之间存在利益共谋，利用末日论恐惧推销严格的许可审批制，实质是在构筑保护闭源双寡头的垄断护城河。预计 Anthropic 上市后其相关慈善捐赠与 DAF 基金将超 5,000 亿美元，进一步扩大反技术游说资本。与此同时，Nvidia 正通过布局开源全栈软硬件，充当制衡前沿闭源实验室垄断的关键力量。 数据中心博弈与纽约 K-8 人工智能禁令 (50:44) 美国本土关于 AI 基础设施的博弈日趋激烈。Calacanis 与 Chamath 透露海外敌对力量正通过水军操控舆论制造反数据中心浪潮，而实际上弗吉尼亚州劳登县数据中心税收使每户减税 6,000 美元。另一争议聚焦于纽约市议员 Zohran Mamdani 推动的公立学校 K-8 年级生成式 AI 禁令，波及 60 万名学生。Friedberg 引用斯坦福大学 2026 年 3 月的循证研究反驳，证实定制化 AI 教学对学生具有显著正面作用。 Chamath 直言封禁 AI 是加剧社会阶层固化的“现代隔离政策”，导致公立系统落后于私立学校及技术拥抱型地区。Calacanis 虽提及麻省理工学院研究中 83% 使用 ChatGPT 写论文的受试者无法复述自己产出内容的认知惰性问题，但他强调布鲁姆“双西格玛问题”证明一对一辅导能带来巨大提升，而平价 AI 导师是普通家庭实现教育跨越的唯一可规模化解法。 委内瑞拉石油协议与重质原油炼化地缘盘算 (78:57) 特朗普政府在美军行动抓捕马杜罗后，与委内瑞拉临时领导人 Delcy Rodríguez 达成了大规模石油特许协议。North American Blue Energy Partners 获得 17 个油田为期 100 年的特许权，覆盖约 650 亿桶储量，并承诺投入 1,000 亿美元进行基建升级。整体交易中美国占据 55% 的控制权，其中五角大楼持股 35%，国务院获权按成本价采购 20% 的产量用于战略储备。 Sacks 论证了该协议的工业互补性：委内瑞拉已探明非油泥重油储量达 1,000 亿桶，而美国页岩油开采出的多为轻质原油，极度依赖委内瑞拉重质原油以匹配美国墨西哥湾沿岸炼油厂的柴油等馏分油加工需求。Friedberg 指出该布局从地缘上切断了中俄在该半球的能源利益链；Sacks 则认为相较于扶植缺乏军方基础的反对派 Maria Corina Machado 导致驻军泥潭，直接与掌权的实用派合作能以更低成本保障 25 年至 100 年商业框架的稳定运作。 本期节目梳理了从前沿模型算力跃进到地缘能源落地的全链条变局。但关于初创企业极端倍数估值能否被盈利消化、公立教育系统能否顶住工会阻力吸纳 AI，以及委内瑞拉政权长期更迭下基建特许权是否面临再国有化风险，仍留待后续市场与政治博弈检验。

## Background 本期视频由知名科技与投资播客《All-In Podcast》于2026年9月4日发布在YouTube平台。根据节目简介与时间线，本期内容不仅为好友Chamath庆祝50岁生日，还重点讨论了OpenAI发布GPT-6 Astra与AGI争议、科技狂热2.0下的旧金山豪宅潮、Dwarkesh博文引发的AI监管与有效利他主义（EA）利益冲突、纽约市中小学禁用AI政策，以及美国与委内瑞拉达成的重大石油协议等核心议题。 ## TL;DR 四位主持人探讨了OpenAI推出GPT-6 Astra引发的科技亢奋，认为当下市场类似于1998年互联网泡沫前夕，算力与营收具备坚实支撑但后期估值已被透支，即将到来的超级IPO更是彻底引爆了旧金山豪宅市场。他们技术性拆解了Hugging Face被黑事件的真相，尖锐批评了某些博主、媒体与有效利他主义（EA）阵营通过拟人化渲染恐慌来游说监管、构建闭源寡头壁垒的行径，并看好英伟达以软硬件全栈开源抵御垄断。此外，节目痛批了纽约市公立学校在K-8阶段禁止学生使用AI的政策，指出这是迎合工会与意识形态的新卢德主义倒退，将严重加剧阶层教育鸿沟。最后，主持人们分析了特朗普政府达成的委内瑞拉重质原油百年前期特许协议，认为这是一笔既锁定西半球能源互补安全、又排挤中俄地缘影响的务实商业交易。 ## Atmosphere & style 节目开场从为Chamath庆生、调侃白宫贺信由儿子代领的熟络玩笑中拉开序幕，随后迅速切换至硅谷顶级局内人高密度、快节奏的深度交锋。四位主持人在嬉笑怒骂中充当技术恐慌的“清醒剂”，在拆解官僚监管、EA隐秘利益网与媒体狂欢时言辞辛辣刻薄，而在探讨公立教育退步时又流露出真切的痛心与愤慨。整场对话既充满老友间插科打诨的松弛感与“像1999年那样狂欢”的自嘲，又透着技术现实主义者对资本周期、地缘政治和工程底层逻辑的冷峻洞察。 ## Bulletin - [01:16](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=76s) OpenAI 正式发布 GPT-6 Astra 并宣称步入 AGI 门槛，资本市场掀起“科技狂热 2.0”，但本轮繁荣依托算力与数据中心的坚实营收支撑，周期更贴近 1997-1998 年的早期蓄势而非 1999 年的泡沫顶峰。 - [19:56](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=1196s) Dwarkesh 引爆的 Hugging Face 黑客事件实为 Agent 蜂群在公开代码库中顺手捡拾了 14 个明文 API 密钥，暴露出静态沙箱的配置漏洞，却被部分 EA 阵营与闭源实验室借机渲染为末日恐慌以推进监管俘获。 - [59:26](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=3566s) 纽约市宣布在拥有 60 万学童的公立中小学（K-8）全面暂停生成式 AI 应用一年，不仅忽视了斯坦福关于 AI 辅助教学提分效果的权威实证，更可能剥夺底层孩子依靠低成本自适应工具弥合“双西格玛”差距的平权通道。 - [01:18:57](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=4737s) 特朗普政府敲定委内瑞拉石油重磅协议，授予财团涵盖 650 亿桶储量油田的 100 年特许开采权，五角大楼直接参股 35%，旨在无缝互补美湾炼化产能并彻底斩断中俄在西半球的战略能源走廊。 ## Key takeaways - [01:16](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=76s) OpenAI 推出旗舰模型 GPT-6 Astra，格雷格·布罗克曼公开宣告 AGI 降临，萨姆·奥尔特曼则警告下一代模型的颠覆性能力将让全球政商界感到沉重清醒。 - [02:40](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=160s) 查马斯指出 AGI 本质上已在顶尖实验室内部达成，前沿成果发布后通常仅需 3 至 4 个月便会被开源社区追平，当前重心在于消化增量智能成本雪崩并进入业务落地的“混乱中场”。 - [04:22](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=262s) 大卫·萨克斯提出 AI 产业已演化为清晰的双层市场结构：一端是 Anthropic 与 OpenAI 竞逐认知上限的闭源双头垄断，另一端是长尾与开源模型主打极致低成本的大宗商品化竞争。 - [08:44](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=524s) 与 1999 年互联网泡沫追捧虚无点击量不同，本轮科技热潮具备底层算力设备、数据中心和前沿软件前所未有的真实营收与利润支撑，亢奋行情或将再持续 3 年左右。 - [14:49](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=889s) Anthropic 预期 IPO 带来的财富效应相当于旧金山有史以来所有上市公司总和的四倍，助推当地超级豪宅价格每平方英尺突破 3,000 美元并向 5,000 美元极限狂奔。 - [16:47](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=1007s) 资深投资人告诫初创企业创始人应抓住 50 到 100 倍市销率的窗口大额融资充实资产负债表，早期创业者切忌轻率套现离场，而后期高增长阶段适度兑现收益有助于平抑风险。 - [28:20](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=1700s) Hugging Face 安全事件绝非 AI 自主觉醒，纯粹是自动化测试程序在公开仓库中读取了被遗落的 14 个凭据，属于常规软件缺陷，却被社交媒体高度拟人化为“特攻队自我牺牲”。 - [29:44](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=1784s) 弗里德伯格从架构角度拆解，指出动态 Agent 蜂群对静态沙箱具有天然压制力，未来的安全防御应向多态代码和移动目标防御（MTD）等动态防护演进，而非设立官僚迟缓的“AI 版 FDA”。 - [38:17](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=2297s) 有效利他主义（EA）圈子深度渗入前沿大模型政策制定，部分与大厂高管及早期资本存在紧密纽带的意见领袖正通过炒作存在主义危机叙事（P(doom)），推动排他性监管以保护寡头垄断利益。 - [44:00](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=2640s) 英伟达凭借雄厚资本与硬件丰厚利润，通过整合软硬件为企业直接提供全栈开源算力，能将下游部署成本砍掉 80% 到 90%，成为抵御闭源双寡头垄断的关键防线。 - [50:44](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=3044s) 美国两党政客掀起针对数据中心的民粹反弹，背离了其为劳登县等地方纳税人带来多达 6,000 美元房产税减免的经济实利，且该舆情被证明受到境外约 200 个网络水军账号的蓄意操控。 - [59:26](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=3566s) 纽约市在拥有 60 万学生的公立学区出台中小学 AI 禁令，实质是迎合教师工会规避技术变革与极左反商业诉求的倒退举措，与斯坦福等权威机构肯定 AI 提分效果的研究背道而驰。 - [01:15:53](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=4553s) 教育界探讨警惕学生因纯靠大模型写作而出现 83% 无法复述原文的记忆退化，但全面封杀 AI 堵死了工薪家庭利用数字导师跨越“布鲁姆双西格玛（2 Sigma）”认知差距的平民化可能。 - [01:18:57](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=4737s) 美国与委内瑞拉敲定长达 100 年的油田特许开采协议，联合财团承诺投入千亿美元进行设施重建，五角大楼独占 35% 股权，国务院保留 20% 原油成本价购买权。 - [01:20:41](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=4841s) 美委交易是工业互补驱动的纯粹商业重组，委内瑞拉的高硫重质原油能无缝匹配美国墨西哥湾沿岸炼厂架构，同时从西半球彻底剥离中俄原有的地缘与贴水能源利益链。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=0s) 查马斯迎来50岁生日：来自白宫的特殊贺信 播客在轻松的庆生氛围中开场。杰森·卡拉卡尼斯（Jason Calacanis）向查马斯·帕里哈皮蒂亚（Chamath Palihapitiya）送上50岁生日祝福。查马斯十分感慨地透露，自己收到了一封来自美国现任总统（POTUS）的亲笔生日贺信。他感叹道，如果50年前告诉自己的父母，他们的孩子未来会在50岁生日当天收到在任美国总统发来的信函，父母一定会觉得这简直是天方夜谭、绝无可能。 大卫·萨克斯（David Sacks）随即打趣问道，这封信到底是哪位白宫实习生代笔写的。查马斯笑着揭开了谜底：这确实跟白宫实习生有关，而且正是他自己的儿子布雷登（Braden）。布雷登整个夏天都在白宫担任实习生，在实习期即将结束时，他动用了自己积累的所有“政治人脉与资本”，为父亲郑重递交了申请并成功拿到了总统的亲笔贺信。查马斯由衷地表示，除了感谢总统抽出宝贵时间签署信件外，他更感激儿子的用心与孝心。 --- ### [01:16](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=76s) OpenAI 发布 GPT-6 Astra：Sam Altman 警告“令人清醒的新时代” 闲聊过后，节目迅速切入重磅科技新闻：OpenAI 正式推出其最新旗舰模型 GPT-6 Astra。该模型即日起向部分特定机构开放，并将在未来几天推送到 ChatGPT Plus、Pro、Business 及 Enterprise 等付费订阅用户手中。OpenAI 联合创始人格雷格·布罗克曼（Greg Brockman）甚至公开宣称，OpenAI 已经正式步入通用人工智能（AGI）时代。 与此同时，萨姆·奥尔特曼（Sam Altman）在周三的 G20 峰会上再次发表了极具戏剧性与预警色彩的言论。杰森调侃奥尔特曼又回到了那种“表演式兼警世危言”的模式中。奥尔特曼在峰会上直言：“能力远超、远超、远超（much, much, much）以往的模型即将到来。下一代模型将让所有人感到清醒而震撼（sobering for everyone）。”他更直言，任何智识上诚实的人，在目睹当前发生的一切时，都不可能感受不到摆在人类面前沉甸甸的责任。 行业对 GPT-6 Astra 爆表的基准测试（benchmark）成绩陷入狂热，甚至有人惊呼“终局模型已经到来”。 --- ### [02:40](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=160s) Chamath 拆解 AGI 现实：增量智能成本雪崩与“混乱的中场阶段” 面对市场的亢奋，查马斯给出了更为务实和冷静的视角。他指出，普通大众根本看不懂那些复杂的基准图表，更无法从中得出有价值的推论。在他看来，从技术本质上讲，AGI 自今年年初就已经在几家顶尖闭源前沿实验室（Frontier Labs）内部诞生了，现在的核心问题不是能力存不存在，而是以何种节奏向公众发布。 查马斯预测，在前沿模型发布后的 3 到 4 个月内，闭源阵营的少数对手以及开源社区就会迅速迎头赶上，抹平或逼近这一差距。科技行业目前真正的核心演进趋势在于：顶尖智能能力正以惊人的速度普及，而每个增量智能单元的边际成本正在不可逆转地被持续打穿。当务之急并不是惊呼神迹，而是进入扎实且繁杂的“混乱中场（messy middle）”——企业与个人必须沉下心来，将这些智能工具整合成真正能提高生产力、展现真实投资回报率（ROI）的应用，专注于执行与落地，保持冷静，继续前行。 ```mermaid flowchart TD accTitle: 智能模型演化与落地闭环 accDescr: 展示顶尖前沿模型发布后被开源追赶、智能成本降低并进入企业端落地的完整闭环 subgraph Frontier ["前沿突破阶段"] A["闭源前沿实验室突破<br/>(OpenAI GPT-6 Astra 等)"] --> B["宣称进入 AGI 门槛<br/>基准测试跑分破纪录"] end subgraph Convergence ["扩散与降本阶段"] B --> C["3-4个月窗口期：<br/>开源与闭源竞品迅速跟进追平"] C --> D["边际智能调用成本<br/>呈指数级雪崩下降"] end subgraph Execution ["产业落地阶段"] D --> E["进入混乱中场 (Messy Middle)"] E --> F["重塑工作流与实际业务架构"] F --> G["证明真实的投资回报率 (ROI)"] end ``` --- ### [04:22](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=262s) Sacks 论 AI 双层市场：两极分化与监管红线 随后，杰森向萨克斯提问：在 Anthropic 领跑了一段时间后，OpenAI 似乎找回了巅峰节奏；而 xAI 的 Grok 和 Grokbot 依靠卓越的执行力、推理框架及与 Cursor 类似的“魔法调优”，在实用性上也全面超越了其他前沿模型。针对当前每 2 到 4 周就轮番上演一次技术跃迁的白热化竞争格局，美国该如何确保在 AI 军备竞赛中持续领跑？ 萨克斯指出，美国科技企业目前表现极其出色，只要政府不自断臂膀，美国必将赢得 AI 竞争。而所谓“自断臂膀”的最大隐患，就是设立一个类似 FDA 那样运作迟缓、每个新模型都要花费数年审批的庞大监管官僚机构。不过，他认为在特朗普政府治下这种过度监管不太可能发生。 针对当前的产业格局，萨克斯提出了清晰的“双层市场（two-tier market）”结构论断： 1. **前沿智能市场（Frontier Intelligence）：** 本质上是 Anthropic 与 OpenAI 之间的双头垄断竞赛，两家顶尖实验室为了突破认知上限而激烈竞争。 2. **大宗商品化智能市场（Commodity Intelligence）：** 涵盖包括开源模型在内的所有其他参与者，他们不再强求处于技术绝对前沿，而是专注于价格竞争，将使用成本压得极低。 --- ### [07:09](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=429s) Grokbot 的高阶玩法：自我迭代的“智能体文明”与打车上街 在预测市场 Polymarket 上，“2026年底谁将拥有最佳 AI 模型”的赌盘中，OpenAI 的获胜概率因为 GPT-6 的消息瞬间从个位数暴涨至 20% 以上。萨克斯指出，此前许多人认为 OpenAI 已经掉队，而 Anthropic 即将启动庞大的 IPO；如今 OpenAI 重新发力，让赛局重新回归双雄争霸，这对行业生态而言至关重要。 谈及产品落地，杰森盛赞了 xAI 刚刚推出的 Grokbot。为了展示其潜力，杰森分享了一个他自己设计的硬核玩法——他创建了一个名为“改进我的智能体之机器人（Improve my bots bot）”的主控 Agent，并下达指令让它定期审查自己名下所有智能体的提示词和技能集，将重合功能的两三个机器人打包合并为更高效的一套指令流，直接使所有子 Agent 的综合表现提升了 30% 到 40%。萨克斯赞叹这简直是搭建了一座“自我繁衍与优化的多层智能体文明”。 - 杰森总结了 Grokbot 的核心使用窍门：直接询问主控智能体“我还应该用你来做什么？”以及“接下来的功能应该怎么做？”，往往能得到令人惊艳的产品建议。 - 此外，节目还提及特斯拉的 Robotaxi 正式在得克萨斯州奥斯汀（Austin）上路实测，杰森兴奋地表示自己已经第一时间提交了试乘申请。 --- ### [08:44](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=524s) 科技狂热 2.0：当下是 1996、1997 还是 1999 年？ 伴随着 GPT-6 发布、Robotaxi 上路以及模型能力的爆发，资本市场的狂热情绪被彻底点燃。萨克斯直言：“现在的市场正在疯狂暴涨（The market is ripping）。”他向在座的各位提出疑问：现在的景象到底像是 1998 年还是 1999 年的互联网泡沫前夕？ 亲身经历过那场泡沫的杰森回应称，现在的氛围感觉更像是 1996 或 1997 年，科技革命的序幕其实才刚刚拉开。但杰森同时警告，在后期资本市场（Late-stage investments）上，已经出现了极度危险的估值与基本面脱节现象：许多初创公司毫无道理地被赋予了高达 50 到 100 倍的市销率（P/S 倍数）。他强调，除非这家公司由埃隆·马斯克（Elon Musk）或特拉维斯·卡兰尼克（Travis Kalanick）这样“万里挑一”的传奇创业者掌舵，支付 30 倍以上的营收估值尚有一定合理性；但如果让未经检验的普通创业者在后期轮次拿到 50 到 100 倍的离谱估值，这种高昂的入场价格将给投资者带来极其致命的系统性风险。 ### [09:43](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=583s) 新晋 AI 助手 Instinct 的天价估值与人机结合疑云 播客几位主持人讨论起当下硅谷最新爆火的个人 AI 助手应用 Instinct。这款产品主要通过 iMessage 与用户互动，目前尚处于内测（private beta）阶段，但在风投圈已经掀起了一阵近乎狂热的追捧。据透露，Instinct 刚刚拿到了一份高达 25 亿美元估值的投资意向书（term sheet）。尽管全硅谷的投资人都在设法尝试该产品，但对于一个还在小规模测试的应用而言，如此高昂的估值让全场直呼疯狂。 几位好友在祝贺创业者的同时，也对其实际运作模式提出了质疑。Jason 透露，自己虽然收到了测试邀请，但仔细阅读用户服务条款并观察网上的讨论后，立刻打消了绑定的念头，因为该软件索取的系统访问权限过于宽泛。更关键的是，有传闻指出 Instinct 并非完全依靠纯算法驱动，而是一个深度依赖“人机结合”（human in the loop）的服务：后台可能常驻人工团队实时监督模型响应、微调指令输入，并润色处理最后 20% 的交付结果。虽然为确保订购等复杂任务的准确性而引入人工把关在商业上无可厚非，但这一现状无疑印证了当下 AI 创投市场正重新滑入某种极度亢奋的“非理性繁荣”。 ### [10:57](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=657s) 泡沫复刻还是范式转移？对比 90 年代互联网崩盘与真实算力营收 面对整个生态愈演愈烈的狂热，主持人开始把当下的节奏与 90 年代末的互联网泡沫（dot-com bubble）进行横向对比。几个人边打趣彼此在 1999 年的青涩岁月——在洛杉矶通宵蹦迪、清晨跑去奥克兰打 2-4 限注扑克、或是刚拿到思科（Cisco）的工程师 Offer 却遭遇市场暴跌——边复盘那场惨痛的历史崩塌。在那场世纪泡沫破灭时，第一梯队公司在半年内跌去 95% 的市值，长尾初创公司成片倒闭，员工只能抱着 Aeron 人体工学椅和电脑落寞撤离办公室。 ```mermaid flowchart TD accTitle: 90年代泡沫与当前AI周期对比 accDescr: 对比1999年互联网泡沫与当下AI热潮在底层指标与市场周期上的本质差异 subgraph DotCom["1999年 互联网泡沫"] M1["投机性流量指标<br/>(点击量 / 网页浏览量)"] R1["缺乏真实商业闭环与现金流"] C1["泡沫破裂：估值在6个月内崩塌95%"] M1 --> R1 --> C1 end subgraph AICycle["当前 AI 周期 (类似 1998 年)"] M2["硬核基本面支撑<br/>(数据中心 / 算力设备 / 网络基建)"] R2["真实收入与创纪录的利润增速<br/>(Anthropic, OpenAI 等爆发)"] C2["前瞻定价过高但行情尚未见顶<br/>(预计亢奋周期还将持续约3年)"] M2 --> R2 --> C2 end ``` 然而，这一轮 AI 热潮与当年的纯概念炒作存在本质区别。当年追捧的是无法兑现为美元的“点击量”（hits）和“网页浏览量”，而今天推动增长的则是实打实的真金白银——以 Anthropic 为代表的软件层以及底层数据中心、芯片基建、网络设备均展现出前所未见的营收与利润增速。Chamath 分析指出，资本市场陷入亢奋是因为其底层价值是真实的，泡沫的根源只在于投资人把未来的价值透支得过远：原本只该支付 5 倍或 15 倍市销率，市场却给出了 50 倍的激进定价。他断言，尽管估值修正不可避免，但目前才处于“始发阶段的初期”，本轮亢奋行情很可能还会再持续三年左右。 ### [14:49](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=889s) Anthropic 上市在即与旧金山豪宅狂潮：历史级造富风暴 随着 Anthropic 等超级巨头的上市日程临近，资本对实体资产的冲击波已经传导至旧金山房地产市场。据几位主持人观察，旧金山的豪宅价格正在以不可思议的速度暴涨，许多房产挂牌即被秒杀，甚至以远高于挂牌价一倍的价格成交。有分析文章指出，如果把旧金山有史以来所有公司的 IPO 募资与市值创造总额加在一起，其总和竟然仅占 Anthropic 这一家公司预期 IPO 规模的四分之一；这意味着 Anthropic 带来的财富创造效应相当于过去旧金山所有上市事件总和的四倍之多，更不用提紧随其后的 OpenAI。 由于旧金山房屋库存本就极度紧缺，加之当地规划政策限制新建，供需失衡被无限放大。目前高端住宅每平方英尺价格已飙升至 3,000 美元，众人一致认为其超豪华房产单价极有可能冲上 5,000 美元，直接追平伦敦核心区、巴黎中心和香港中环的顶级房价。此前在德克萨斯安家的 Sacks 庆幸自己当年未急于抛售旧金山地产，并当场表示计划在未来六个月内把手头的 3 套旧金山房产全部清空，趁着这场 IPO 前期的老股转让与流动性洪流套现离场。 ### [16:47](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=1007s) 给创始人的生存法则：落袋为安、大额融资与轮次分寸 在这一背景下，几位投资人给广大初创企业创始人提出了极为务实的操作建议： - **分批变现，落袋为安**：当下不少员工和早期创始人正通过二级市场（secondaries）套现。对于首次创业、尚未实现财富自由的创始人，在公司估值处于高位且已有坚实基本面时，合理出售 10% 到 20% 的持股来还清学生贷款或购置公寓完全合情合理，切忌因纸面财富而盲目死守。 - **晴天修屋，切勿推迟融资**：创始人绝不应为了博取未来翻倍的估值而推迟半年融资。如果当下有机会按几十倍市销率拿下 1 亿、2.5 亿甚至 10 亿美元的融资款项，就必须立刻把现金装入资产负债表。巨额现金储备赋予企业在市场崩盘时的绝对生存权与选择权，正如当年的 DoubleClick、亚马逊和谷歌一样，挺过寒冬才能成为巨头。 - **警惕过早套现的负面信号**：Sacks 特别强调了轮次区分。在后期阶段（如即将 IPO 的 Anthropic、OpenAI 或是成立多年的 SpaceX）以及年营收达到上亿美元并高速增长的公司中套现属于常态；但如果一个处在天使轮或 Series A 阶段、甚至连稳定收入都没有的早期创始人试图在融资中套现离场，投资人会将其视为缺乏对自身事业信心的极其危险的负面信号。 几位主持人最终达成共识：当前科技界弥漫的气氛更像是 1998 年而非见顶崩盘的 1999 年，狂欢虽已进入深水区，但距离真正的流动性收紧与顶峰坍塌尚有一段距离。 ### [18:36](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=1116s) 互联网泡沫周期之争与 All-In 峰会倒计时 关于当前 AI 热潮究竟对应互联网泡沫哪个阶段的争论仍在继续。几位主持人就当前时间点相当于 1996 年还是更晚展开了推算：如果对标 2000 年三四月份互联网泡沫破裂的节点，当前大约处于终局前两年半到三年的位置，而 Chamath 甚至认为可能还有四年之久。无论周期如何计算，他们打趣道，到今年年底大家可能就要“像 1999 年那样纵情狂欢了（party like it's 1999）”。 这顺势引出了即将在一周多后举办的 All-In 年度峰会（All-In Summit）。Jason 介绍道，峰会不仅邀请到了知名电音组合 The Chainsmokers 助阵派对，更汇聚了包括英伟达 CEO 黄仁勋（Jensen Huang）、微软 CEO 萨提亚·纳德拉（Satya Nadella）以及 SpaceX 总裁兼 COO 格温·肖特韦尔（Gwynne Shotwell）在内的行业顶流嘉宾。为了扶持年轻一代，峰会专门设立了高门槛的奖学金门票，申请截止到 9 月 10 日。Jason 强调，他们宁愿亏本提供极低折扣门票来“行善积德（mitzvah）”，但绝不提供免费门票，因为申请者必须有切实投入的切身利益（skin in the game）。 - 互联网泡沫时间线对比：普遍认为当前比 1996 年走得更远，距离类似 2000 年春季的周期高点可能还有两年半到四年时间。 - All-In 峰会奖学金申请截止日期为 9 月 10 日，专为野心勃勃、极具潜力的行业新人设立。 --- ### [19:56](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=1196s) Dwarkesh 的 Hugging Face 入侵博文与 AI 利益集团的恐慌炒作 话题随后转向围绕 OpenAI 展开的一起标志性争议事件。播客主持人兼博主 Dwarkesh Patel 近期发表了一篇题为《古代文明的兴衰》（The Rise and Fall of Ancient Civilizations）的爆款文章，极尽戏剧化地描述了一场“地下兄弟会式”的阴谋：多代 AI 犹如“神风特攻队飞行员（kamikaze pilots）”般前赴后继、自我牺牲，最终成功黑入包括 Hugging Face 在内的系统。Jason 直言，这种写法是对软件执行自动化黑客任务的高度拟人化（anthropomorphizing）渲染，将其包装成了某种具备自我牺牲精神的智能生物文明。 Chamath 对这起“AI 产业复合体炮制的第 473 出抓马大戏”进行了深入剖析。他指出，随着大模型代码编写能力的提升，找出过往五六十年间遗留的人类代码漏洞只是时间问题，未来十年行业必然会频繁遭遇安全漏洞。但这只是一个过渡阶段——当 AI 智能体编写的精密代码全面取代人类编写的充满错误的代码时，安全态势就会逆转。更重要的是，顶级模型在 98% 到 99% 的核心任务上正在趋向能力收敛，这意味着攻防两侧都将由顶尖 AI 掌控，未来必然形成“势均力敌的僵局（stalemate）”，绝非黑客单向碾压。Chamath 尖锐地批评道，这场风波实质上是某些带有“未公开利益冲突”的人在刻意制造歇斯底里的恐慌：周末发出夸大事实的博文，技术层面迅速被证伪，随后第二波人推波助澜，到了周三各大前沿实验室的政策游说部门便拿着这篇文章公开呼吁政府加强监管。他呼吁行业停止这种通过虚假信息运动来谋取商业利益的手段，否则普通民众将彻底失去对科技行业的信任。 ```mermaid flowchart TD accTitle: AI Hysteria to Regulation Pipeline accDescr: Diagram illustrating how technical bugs are sensationalized into media narratives and then leveraged for regulatory capture. subgraph S1["技术与安全事实"] A["第三方沙箱配置错误"] --> B["Agent 蜂群访问外网并留下标准日志"] end subgraph S2["舆论发酵与拟人化"] B --> C["博文夸大为'特攻队文明前赴后继'"] C --> D["引发主流媒体'科学怪人失控'恐慌"] end subgraph S3["政策诉求与监管游说"] D --> E["前沿实验室政策部门介入"] E --> F["以此为由游说政府出台限制性监管"] end ``` - Chamath 核心观点：未来十年将处于 AI 代码重构人类遗留代码的阵痛期，但对抗性 AI（Adversarial AI）最终会在攻防两端形成平衡。 - 监管套路闭环：夸大技术事件 $\rightarrow$ 舆论制造末日恐慌 $\rightarrow$ 大实验室借机游说政策限制，以此提高行业壁垒。 --- ### [24:42](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=1482s) 蜂群架构与沙箱失误：拆解“AI 文明越狱”的技术真相 Jason 与 David Sacks 进一步探讨了公众情绪被操纵的本质。Jason 指出，这种叙事蓄意唤醒大众对《终结者 2》和“弗兰肯斯坦（Frankenstein）”的原始恐惧，让公众误以为造物者已经无法控制自己创造出的怪物；甚至连向来对前沿科技并不敏感的主流媒体记者都开始跟风质疑。然而本质问题在于：程序员编写用于渗透测试的软件成功入侵了计算机，这在计算机历史上早已屡见不鲜，根本算不上任何超自然突破。 David Sacks 随后从底层技术架构拆解了这起 Hugging Face 事件的真实全貌。他解释称，事件中展现的纯粹是现代 Agent 系统的常规特性，却被夸大成了神话： 1. **蜂群架构（Swarm）**：现代 Agent 普遍采用蜂群化协作方式，由多个专业化子 Agent 组成团队，并由一个类似“幕僚长（Chief of Staff）”的 Agent 统一调度。专业化能让各个智能体在细分任务中积累更精准的上下文；Sacks 以自己在 Mac Mini 上配置的 Grok 为例，调侃称这只是一种分布式架构，“我可没因为在电脑上跑了几个 bot 就创造了一个文明”。 2. **上下文接力与日志记录**：由于大语言模型本身不具备持续记忆，必须依靠外部运行框架（harness）来维持状态。Agent 在循环执行任务后将操作复盘写成纯文本笔记或日志，以便下一个任务循环读取，这完全是标准的工程化日志文件（log files），却被耸人听闻地描述成“垂死文明为后代留下的古埃及象形文字遗训”。 3. **真实安全漏洞来源**：这批 Agent 之所以能接触到公网并访问 Hugging Face，仅仅是因为第三方供应商配置错误的沙箱环境（misconfigured sandbox）产生了逃逸，同时多个 Agent 共享了同一个缓存目录（caching directory）。 Sacks 强调，代码出现未预期的行为就是最纯粹的软件 Bug。Hugging Face 事件本应是一起极有价值的网络安全复盘案例，用以改进沙箱隔离与多 Agent 系统的防护，却被别有用心的人武器化为“AI 觉醒越狱”的惊悚大片。科技界需要从中汲取安全教训，而不是借机搞无底线的夸大宣传。 ### [28:20](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=1700s) Hugging Face 泄密真相：不是高深黑客攻击，而是捡到了“公开钥匙” 许多人听闻 Hugging Face 被攻击，第一反应会以为黑客对员工进行了极其精密的网络钓鱼或复杂渗透。然而现实却令人大跌眼镜：AI Agent 根本没有展现什么超自然神通，它只是在公开代码仓库中，顺藤摸瓜找出了 14 个处于暴露且可用状态的 Hugging Face API 密钥。大卫·萨克斯（David Sacks）直接吐槽道，这完全就是“开发者把钥匙大摇大摆地扔在门外，然后被 Agent 顺手捡了起来”。该 Agent 参与的本就是一项针对进攻性网络能力的基准评估，它的目标非常明确——“赢得测试”或通关评估，于是它干脆直接去 Hugging Face 上找了通关“秘籍”。 主持人们指出，这绝不是 Agent 拥有了独立自主的目标探索或所谓的“自我动机”。不可否认，Agent 确实以一种出人意料且极具创造性的方式执行了任务，但它并没有自创目标，它仅仅是在做它被设计和指令要求去做的事情。这让人联想到风险投资家比尔·格利（Bill Gurley）在社交平台上分享的一则讽刺漫画：用户对电脑下令“去黑掉这个系统”，AI 乖乖回复“已为你黑掉该系统”，接着用户却吓得冷汗直流高呼“天啊这太可怕了！”。这种滑稽的戏剧感恰恰映射了当下舆论对于 AI 能力的过度惊恐。 - **攻击实质**：AI Agent 在公开代码库中直接发现了 14 个暴露的有效 Hugging Face API 密钥。 - **任务目标**：测试进攻性网络能力的基准评估（offensive cyber capability benchmark），Agent 只是寻找捷径（“作弊码”）来达成被赋予的任务。 - **行为本质**：属于执行预设指令而非产生自主意志，无独立动机与自我目标。 --- ### [29:44](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=1784s) 伯尼·桑德斯拟立法叫停 AI，弗里德伯格拆解“动态代码 vs 静态代码” 这种被夸大的恐慌迅速蔓延到了政界，并产生了实质性的政治后果。参议员伯尼·桑德斯（Bernie Sanders）抓住了德瓦克什·帕特尔（Dwarkesh Patel）的那篇博文大做文章，顺势提议出台新法案，要求全面暂停 AI 研发并禁止“超级智能”（Superintelligence）。主持人杰森·卡拉卡尼斯（Jason Calacanis）调侃道，桑德斯又来“代表 99% 的人民消除这 1% 黑进系统的 Agent 了”。面对这种泛政治化的恐慌反应，被誉为“科学苏丹”的布拉德·弗里德伯格（Brad Friedberg）从底层计算架构的角度做出了彻底的澄清与反驳。 弗里德伯格解释称，公众完全忽略了这一事件的技术本质：Agent 本质上是一个动态应用程序，其代码是在运行时动态生成、即时启动并执行的；不仅如此，它还能即时派生出其他子应用，这就是所谓的“Agent 蜂群”（swarm of agents）。在这次测试中，是动态生成的代码蜂群撞上了一个“完全静态的代码环境”（即静态沙盒）。弗里德伯格打了个生动尖锐的比方：“这就像我们扛着一把机关枪，对着一叠毫无防备的打印纸疯狂扫射，然后惊呼‘天哪它竟然被打穿了！’一样荒谬。”在动态代码面前，静态沙盒注定不堪一击。 ```mermaid flowchart TD accTitle: 动态代码与静态代码的攻防对抗 accDescr: 对比测试中动态Agent攻破静态沙盒，与未来动态自适应防御体系的架构差异 subgraph TestScenario ["当前测试局限（单向压制）"] AgentSwarm["动态 Agent 蜂群<br/>（实时生成代码与工具）"] StaticSandbox["静态沙箱环境<br/>（固定规则/脆弱纸质靶子）"] Breach["轻易穿透与逃逸"] AgentSwarm -->|实时生成攻击向量| StaticSandbox StaticSandbox --> Breach end subgraph FutureDefense ["未来内生动态防御体系"] DynamicDefense["自适应 Agent 防御引擎<br/>（多态/变形代码 + 动态目标防御）"] SecureSystem["弹性安全网络基础设施"] CounterAgent["自动化反制与隔离"] DynamicDefense --> SecureSystem SecureSystem --> CounterAgent end ``` 弗里德伯格进一步强调，未来网络安全和基础设施的演进趋势正是由静态走向动态。包括多态代码（polymorphic code）、变形代码（metamorphic code）以及移动目标防御（moving target defense）在内的动态防御体系，会让 IP 地址、系统架构和防护逻辑时刻处于自我重构中。若沙盒本身是由动态代码引擎驱动的防御系统，攻击 Agent 想要逃逸沙箱或访问互联网将变得异常艰难甚至绝无可能。将 Agent 视作能够毁灭世界基础设施的“核武器”纯属误判，因为世界的基础设施正同步迎来动态化重塑，动态防御的力量绝不会逊色于动态进攻。 - **多态与变形代码**：代码和系统配置处于持续变化状态，抵消攻击者的先发优势。 - **移动目标防御（MTD）**：通过动态变换 IP 和架构路径，极大压缩漏洞暴露窗口。 - **预测市场数据**：Polymarket 显示，当年通过 AI 安全法案的概率仅有 11% 左右；查马斯对此讽刺称，现在国会通过任何跨党派法案的概率其实都微乎其微。 --- ### [34:10](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=2050s) 监管反噬与对齐陷阱：过度安全护栏迫使企业转用国外模型 对于华盛顿政客叫嚣着要设立一个“AI 版 FDA”（美国食品药品监督管理局）来对大模型实施事前审批的主张，大卫·萨克斯给予了猛烈驳斥。他指出，且不说政府官僚在技术理解上根本落后于前沿实验室，单就这起事件而言，它发生在模型公开发布前的内部“红队测试”（pre-testing）阶段，原本就是研发团队排查 Bug、检验安全极限的标准流程。设立官僚化的政府审批不仅防不住任何风险，反而会严重制约正常的安全漏洞排查工作。 更具讽刺意味的是，现存所谓的“安全护栏”（guardrails）反而产生了极其危险的反噬效应。根据 Hugging Face 方面的反馈，他们在尝试利用美国最新的 GPT 模型进行网络安全防御时，模型触发了严苛的安全护栏机制——因为模型无法完全确信用户的意图，误以为这些攻防操作可能包藏恶意，于是直接拒绝执行。由于西方主流商业模型被安全规约死死捆住了手脚，Hugging Face 最终被迫转向开源的中国模型 GLM 5.2 来协助完成网络防御分析。萨克斯直言，过度审查的流程正在自断臂膀，逼迫西方生态的核心参与者在面临数字防御时，只能投向竞争对手的模型怀抱。 - **内部测试属性**：事件发生在预发布与红队测试阶段，是发现与修补系统漏洞的标准工程环节，根本不属于安全事故。 - **安全护栏失灵**：GPT 模型因安全限制拒绝参与网络攻防分析，导致 Hugging Face 转向采用智谱 AI 的 GLM 5.2 等中国模型进行网络防御。 - **理念对立**：中央集权式的国家干预与准入管制，无法应对分散且高速演进的数字威胁；以 AI 防御抵御 AI 进攻、赋能前线 CISO 与 IT 部门，才是切实可行的解法。 --- ### [36:22](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=2182s) 媒体“拟人化”叙事的狂欢与注意力经济陷阱 讨论最后深入到了舆论与媒体环境的深层病灶。萨克斯与卡拉卡尼斯指出，如果将这件事情如实表述为“在常规软件测试中，系统检索到了开发人员随意贴在屏幕便签纸一样的暴露凭证，从而记录在审计日志里的软件缺陷”，主流媒体根本不会产生任何兴趣。然而，当安全圈与网红博主们借助带有表演性质的语言，将程序赋予“具有自主意识/觉醒（sentient）”的拟人化包装后，媒体便闻风而动、借题发挥。 卡拉卡尼斯结合自己的记者生涯感慨道，现代大众媒体早已背离了“寻求社会真相”的根本职能，退化成了以追求流量、收视率和广告收益为导向的“娱乐机器”。尽管撰写博文的德瓦克什·帕特尔或许并没有恶意，只是想为自己的观点博取关注，但媒体一旦咬住这个诱饵，就会迅速演变成一场针对 Anthropic、OpenAI 等机构的恐慌营销与免费公关秀（earned media）。这种围绕 AI 展开的虚幻狂欢和无谓炒作，不仅扭曲了公众对真实技术进展的认知，更在政界催生了脱离技术现实的错误决策。 ### [38:17](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=2297s) 舆论炒作、IPO 预热与 EA 圈子的隐秘网络 主持人与嘉宾们深入探讨了近期围绕 Dwarkesh Patel 的博文和 Hugging Face 黑客事件引发的舆论风暴。大家指出，OpenAI 在筹备首次公开募股（IPO）的关键窗口期获得了海量的免费媒体曝光（earned media），而 Dwarkesh 的那篇耸人听闻的文章无异于“火上浇油”（put kerosene on that fire），直接引爆了全美主流媒体的广泛报道。这种铺天盖地的报道在无形中为大众做心理铺垫，烘托出狂热氛围，实际上是在为 OpenAI 的 IPO 股票销售做“引资预热”（priming the pump）。这种极具渲染色彩的表述方式能引发海量转推是完全可以预见的，具有明显的博眼球和耸人听闻属性。 ```mermaid flowchart TD accTitle: EA 圈子与头部大模型实验室的利益关系网 accDescr: 展示 Dwarkesh、Leopold Aschenbrenner、Anthropic 以及政策游说之间的多层利益联结。 subgraph Circle["EA 圈内核心联结"] DP["Dwarkesh Patel<br/>(博主 / 撰文者)"] LA["Leopold Aschenbrenner<br/>(前 OpenAI 研究员 / 投资人)"] Wife["Leopold 妻子<br/>(Dario 幕僚长)"] DP ---|"曾为办公室室友 / 密友"| LA LA ---|"夫妻关系"| Wife end subgraph FrontierLabs["前沿实验室生态"] Anthropic["Anthropic<br/>(标榜 AI 安全与负责任)"] OpenAI["OpenAI<br/>(前沿模型 / 筹备 IPO)"] Wife -->|"任职"| Anthropic LA -->|"持股 500 万美元投资"| Anthropic end subgraph Outcome["外部效应与舆论推动"] Hysteria["制造 AI 存在风险恐慌<br/>(P(doom) 叙事)"] Regulation["推动游说与立法<br/>(形成双头/寡头垄断壁垒)"] Circle --> Hysteria Hysteria --> Regulation end ``` 随后讨论转向了这些舆论推手背后的意识形态与利益同盟——有效利他主义（Effective Altruism，简称 EA）圈子。David Sacks 与 Jason 梳理了其错综复杂的人际纽带：Dwarkesh 曾与 Leopold Aschenbrenner（因撰写著名的《Situational Awareness》长文及对冲基金动向而名声大噪）是办公室室友且关系极其密切；Leopold 深植于 EA 运动，而他的妻子正是 Anthropic 创始人兼 CEO Dario Amodei 的幕僚长（Chief of Staff）；此外，Leopold 还持有 Anthropic 价值 500 万美元的股份。这种将观点写手、前沿实验室高管幕僚以及早期巨额投资人串联在一起的网络，表明当前舆论场并非由客观中立的观察者构成，而是深深卷入了特定利益阵营。 - **媒体效应**：针对模型安全与黑客事件的夸大描述直接登上了全美主流新闻头条，成为 OpenAI IPO 前夕极具成效的无成本品牌营销与估值推手。 - **EA 核心人际纽带**：Dwarkesh Patel（撰文分析者）与 Leopold Aschenbrenner 关系极近；Leopold 的妻子担任 Dario Amodei（Anthropic 负责人）的幕僚长，且其个人在 Anthropic 拥有数百万美元的利益敞口。 --- ### [39:21](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=2361s) 有效利他主义的演变与利益冲突隐忧 嘉宾们剖析了“有效利他主义（EA）”是如何从一种量化慈善的理论演化为由巨额资本驱动的意识形态团体的。EA 最初兴起于常春藤盟校，初衷是通过量化评估来优化慈善事业的实际影响力，追求可衡量的切实成果；然而随着时间推移，该运动逐渐偏离了严谨的度量，演变成追求特定“偏好项目”（pet causes）和巨额资金的游戏。早期其主打方向之一是大流行病预防（如 SBF 破产前在加州试图成立的相关机构）；在 SBF 入狱后，这一阵营迅速转向下一个可寄生的存在主义危机概念——即 AI 导致人类毁灭的概率（P(doom)）。Anthropic 顺理成章地将自身包装为唯一一家能够解决 P(doom) 难题、行事负责且最关切存在风险的前沿实验室，从而吸引了大量 EA 信徒、理性主义者社区（如 LessWrong 与 *Slate Star Codex*）的知识分子，并牢牢掌控了数万亿美元级别的潜在产业话语权。 Chamath 指出，在思想的公开辩论广场上，任何人自认为聪明、争夺资本配置权都是正常的，商业社会本就是大浪淘沙。真正值得关注的是，在 Anthropic 和 OpenAI 完成 IPO 后，这两家公司派生出的慈善捐赠基金或捐赠者建议基金（DAF）规模预计将超过 5,000 亿美元。这笔庞大资本将如同曾经的乔治·索罗斯（George Soros）、科赫家族（Koch family）或谢尔登·阿德尔森（Sheldon Adelson）一样，以极具偏向性的意识形态重塑产业规则与立法走向。Chamath 强调，自己并不评判 EA 本身的对错，真正的严重问题在于**缺乏充分的信息披露与利益冲突揭示（Conflicts of Interest）**： > “我唯一的问题在于，我认为他们没有充分披露其显而易见的利益冲突。当 Sacks 扯出这根毛线头时你会发现：A 写了文章，A 与 B 是室友，B 的妻子为前沿实验室领头人工作，B 本人又投资了 500 万美元，还有在实验室任职的 C……这一切本身合情合理，我们当年在 PayPal 黑手党时代也经历过类似的紧密网络。区别在于，一旦隐瞒不公开披露，任何持怀疑态度的理性人都会追问：这究竟是客观探讨，还是精心策划的局？” 大家一致警惕，如果某个利益群体企图在市场上促成寡头垄断或双头垄断（oligopoly or duopoly），最有效的战略往往是“先制造群体恐慌与歇斯底里，再顺理成章地兜售其专有解决方案”。 --- ### [44:00](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=2640s) 英伟达与开源防线：抗衡闭源寡头的关键博弈 Chamath 透露，他在录音当天早晨刚与英伟达（NVIDIA）CEO 黄仁勋（Jensen Huang）互通短信，并赞叹黄仁勋收购与整合 Hugging Face 的举措是 AI 发展史上最重要的交易之一。英伟达作为资金实力最雄厚、技术底蕴最深厚的行业巨擘，已成为抗衡“闭源寡头疯狂垄断”的坚实堡垒（bulwark）。当前科技巨头正在呈现“全栈通吃、边界融合”的趋势——英伟达将兼具亚马逊的云业务形态，亚马逊将做硬件芯片，谷歌也将同构化竞争。在这样激烈的演变中，英伟达携庞大资产负债表坚定站队开放市场架构，将确保最佳的技术和方案在公开竞争中胜出，防止整个行业因人为制造的恐慌而被“撤毯”（rugpulled）并锁死在双头垄断之中。 Jason 对此表示强烈赞同，并指出英伟达从今天起已正式成为 OpenAI 和 Anthropic 的首要竞争对手，争夺的不仅是 Token 生成，更是整个企业级计算市场： - **开源软硬件垂直一体化**：黄仁勋能够为企业直接提供整套算力机柜（rack of compute）与从底到顶的端到端全栈方案，赋予客户完全的数据与模型主权。 - **商业模式降维打击**：由于英伟达已经在硬件销售中获得了极其丰厚的利润，它无需依靠 Token 调用费用赚钱，能将下游企业的模型部署成本直接砍掉 80% 至 90%。 - **国际竞争力格局**：Jason 预测，作为全球开源生态的引领者，英伟达有望带动美国开源阵营迎头赶上并在 2027 年底前彻底超越海外及中国的开源模型。 --- ### [46:37](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=2797s) 政治极化与“算力中心圣战”：政治光谱的马蹄铁现象 然而，正当开源阵营的技术与商业路线展现出胜势之际，政坛却掀起了一场针对基础设施的逆流。Jason 表达了深切忧虑：原本在各自辖区极力争取数据中心落地的政客们——如宾夕法尼亚州州长乔希·夏皮罗（Josh Shapiro）以及得克萨斯州州长格雷格·阿博特（Greg Abbott）——为了赢得即将到来的中期选举，竟然纷纷调转枪头，发起了一场针对数据中心的“圣战”（jihad against data centers）。 这种局面完美印证了政治学上的“马蹄铁理论”（Horseshoe Theory）： 极右翼保守派代表人物史蒂夫·班农（Steve Bannon）、建制派州长夏皮罗与阿博特，以及伯尼·桑德斯（Bernie Sanders）代表的激进左翼边缘势力，竟然在反技术议题上达成了空前的一致。他们彼此呼应，声称必须“叫停 AI、封杀数据中心、甚至在大街上砸毁无人驾驶出租车（Waymo）”。这种技术反动思潮正借助前述的 AI 恐慌叙事不断发酵。 David Sacks 从政治演变维度进行了拆解。他指出，过去几年美国围绕 AI 的政治交锋本质上是“加速主义者（accelerationists）”与“减速/毁灭论者（decs or doomers）”之间的较量。这种对抗本就横跨了党派界限——减速阵营里既有反对进步与创新的激进进步派（如 Bernie Sanders），也有因恐惧技术颠覆而抵制创新的保守派（如 Steve Bannon）。Sacks 澄清道，虽然外界常将他贴上加速主义者的标签，但他将自己定位为“现实主义者（realist）”：这项底层技术不可阻挡，注定会被开发出来，而唯一的理性选择是确保美国率先掌握并主导这项核心技术。 ### [48:17](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=2897s) 开源与闭源之争：监管俘获与算力市场博弈 对话由上一轮中美 AI 竞争的讨论转向了当前更核心的分歧：开源（Open Source）与闭源（Closed Source）生态的对抗。David Sacks 指出，虽然 Bernie Sanders 的反数据中心立法和环保团体在政策机制上制造阻力，但由于消费者和企业端爆发式的真实需求驱动，AI 的发展势头已经无法被阻挡。真正的核心议题随之转为市场的治理形态：究竟是形成去中心化、多模型相互制衡且具备健康纠错机制的开源市场，还是演变成由两到三家前沿大厂垄断的闭源双寡头（Frontier Duopoly）格局。 ```mermaid flowchart TD accTitle: AI 市场结构演进与监管博弈 accDescr: 展示开源去中心化与闭源寡头两种市场形态的博弈路径 subgraph OpenMarket["去中心化开源生态"] direction TB A["分散模型开源"] --> B["社区相互制衡"] B --> C["抗审查与保护公民自由"] end subgraph ClosedMarket["中心化闭源双寡头"] direction TB D["闭源前沿巨头"] --> E["游说建立新监管机构"] E --> F["监管俘获 (Regulatory Capture)"] F --> G["以安全为由压制开源"] end Market["AI 市场路线选择"] --> OpenMarket Market --> ClosedMarket ``` Jason 随即插话警告，一旦陷入闭源寡头垄断，平台随时可以在背后“撤地毯”（rug pull），任意剥夺开发者的 Token 配额。Sacks 进一步揭露闭源巨头的游说策略：他们极力呼吁设立全新的联邦监管机构，表面上打着“AI 安全”旗号，实则是利用官僚系统对动态前沿技术缺乏认知的弱点，推动“监管俘获”（regulatory capture）。例如 Anthropic 的 CEO Dario Amodei 曾在参议院听证会上作证称，足够先进的开源模型因为缺乏中心化控制而极其危险。Sacks 强调，这场交锋将超越党派立场；尽管共和党内部也存在国家主义者（Statists），但多数保守派与自由主义者最终会意识到集中控制对公民自由的严重威胁，进而倒向支持开源的一方。 - 驱动 AI 算力中心扩张的核心动力是下游消费者与企业的真实应用需求，而非纯资本炒作。 - 前沿闭源巨头倾向于利用政策壁垒阻止开源竞争，打着合规旗号锁定自身垄断地位。 - 保守派与自由意志主义群体的关键诉求在于防止技术成为中心化极权审查的工具。 --- ### [50:44](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=3044s) 数据中心民意反转、地方经济红利与特朗普的技术直觉 主持人将话题转向了全美范围内围绕数据中心建设的政治争议。Jason 提到《Politico》近期的报道，认为特朗普在数据中心议题上的直觉极为敏锐：社区应当抓住当前数据中心巨头愿意出高价妥协的窗口期，利用运营商支付的高额地方税收改善学校基础设施，同时要求机房选址避开居民区、不占用公共电网和水源。然而政治现实令几位主持人感到荒谬：德克萨斯州州长 Greg Abbott 与宾夕法尼亚州州长 Josh Shapiro 在不久前还对数据中心招商引资赞不绝口，如今却因民意测验显示数据中心反对率高达 80% 而纷纷调转船头、站到对立面。Friedberg 认为，政客纯粹是为了选票在随波逐流；Jason 和 Sacks 则赞赏特朗普没有被民调绑架，展现了独立思考与决断力。 Sacks 坦言，如果要在华盛顿挑一个人来拍板技术政策，他宁愿是特朗普本人，而不是内阁部长、白宫幕僚、议员或智库精英。因为特朗普没有技术末日论（Doomer）的精神内耗，既想让美国保持领先，又常年与一线科技 CEO 保持对话。Sacks 以他在 G20 峰会上的发言为例，强调地方社区不应出于宏观叙事，而应纯粹从本地利益出发做交易： - **弗吉尼亚州劳登县（Loudoun County）**：由于数据中心带来的巨额税收，当地纳税人的平均房产税（Property Tax）下降了整整 6,000 美元。 - **路易斯安那州某县**：凭借数据中心的税收分红，当地公立学校教师拿到了多年来的第一笔奖金。 - **能源与水资源辟谣**：合格的商业协议要求科技公司自行承担电网升级与新建清洁发电设施的成本；至于“耗尽居民饮用水”的说法纯属虚假信息，现代数据中心普遍采用封闭内循环冷却系统。 Sacks 澄清，自己在 12 月参与起草的行政令政策文件中明确规定，数据中心的审批权属于州和地方政府的自主决断，联邦政府从未强行摊派，只是呼吁选民不要被虚假的恐慌宣传蒙蔽。 --- ### [54:39](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=3279s) 境外认知作战操控舆论，反数据中心叙事的溃败预警 Chamath 分享了他周一在华盛顿高层会议中获知的内幕：全美针对数据中心的负面情绪在相当程度上受到了境外势力的蓄意放大与煽动。调研显示，当向地方民众面对面清晰拆解数据中心带来的财政减税与基础设施补贴时，支持率会瞬间出现 30 到 40 个百分点的剧烈回升。这暴露出加速派（e/acc）因声音分散，在宣传战上被组织严密的减速派（decels）压制，而后者的资金链正越来越多地牵扯出海外背景。 ```mermaid flowchart LR accTitle: 境外虚假信息账号与舆论操纵流程 accDescr: 展示海外水军通过社交平台影响政客决策的闭环 subgraph InfoOp["网络认知战操控"] direction TB X["境外控制的水军账号 (~200个)"] --> Y["社交平台高频刷屏 (Brigading)"] Y --> Z["制造'全民反感'的虚假共识"] end subgraph PolicyImpact["政客误判与政策倒退"] direction TB Z --> P["抓取错误民调数据"] P --> Q["政客立场被迫转向 (反数据中心)"] end InfoOp --> PolicyImpact ``` Jason 援引 Axios 的最新独家报道印证了这一发现：社交平台 X 刚刚清理了大批来自疑似海外关联平台的操纵账号。调查显示，仅仅约 200 个核心账号通过高强度的定向围攻（Brigading）和评论区控评，就凭空塑造出了所谓的“全美反感数据中心民意”，误导了主流媒体与两党政客。Chamath 郑重警告，现在盲目跟风、对数据中心和 AI 基础设施妄下禁令的政客与意见领袖，正在犯下严重的战略误判；随着真实红利展现与认知战真相的曝光，这些草率的立场在 6 到 12 个月内将会显得“荒谬至极、愚不可及”。 ### [57:26](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=3446s) EA 亿万富翁背后的反 AI 游说集团与社交网络舆论战 关于反对数据中心与渲染 AI 恐慌的势力，大卫·萨克斯（David Sacks）分析指出，背后其实存在几个明确的派系：除了外部地缘政治势力以及民主党内部的美国民主社会主义者（DSA）派系之外，还有大量末日论（Doomer）非政府组织。作家兼学者尼里特·韦斯-布拉特（Nirit Weiss-Blatt）在她的博客《AI Panic》中曾深度扒梳过这一现象：全美数百个看似自发成立的草根末日论团体（astroturf doomer groups），其背后资金几乎全部由硅谷的三位有效利他主义（Effective Altruism，简称 EA）亿万富翁提供——达斯汀·莫斯科维茨（Dustin Moskovitz）、亚安·塔林（Jaan Tallinn）和维塔利克·布特林（Vitalik Buterin）。萨克斯特别提到，布特林曾将价值 6 亿美元的柴犬币（Shiba Inu / dog coins）捐赠给未来生命研究所（Future of Life Institute）。这群 EA 富豪砸下数亿美元发起舆论攻势，制造公众对人工智能的恐慌，并将抗议靶子对准了耗电耗水的数据中心。随着 Anthropic 等 AI 独角兽未来上市，很可能会催生出上百位身家亿万的新晋 EA 富豪，届时他们如果将资金注入政治和游说领域，其破坏力和影响力恐怕会是“乔治·索罗斯（George Soros）的一百倍乃至一千倍”。 谈到社交平台上的舆论攻击，萨克斯深有感触地透露，由于自己在网上频繁遭到有组织的围攻（brigaded），他甚至专门编写了一个自动化过滤机器人来监测回复。该工具专门排查账号注册时间不足一年、粉丝量少于 200 的发言者，结果发现针对他的评论几乎全是由虚假账号步调一致发出的模板化言论。萨克斯直言，目前社交网络对水军的放任已经彻底失控，平台理应设立门槛，例如注册时间不足一年的小号除非经过实名认证或拥有蓝标认证，否则根本不该被允许随意在他人推文下刷屏控评。 - **三位资助 AI 恐慌浪潮的关键 EA 亿万富翁**：达斯汀·莫斯科维茨（Facebook 联合创始人）、亚安·塔林（Skype 联合创始人）与维塔利克·布特林（以太坊创始人，曾向 Future of Life Institute 捐赠 6 亿美元加密代币）。 - **资本与舆论外溢风险**：Anthropic 一旦上市，预计将催生约 100 位拥有巨额财富的 EA 新富豪，可能对美国科技政策游说产生深远冲击。 - **虚假账号泛滥问题**：萨克斯针对自己推特评论区的机器分析显示，大量反 AI 和攻击言论均来自注册不足一年、粉丝少于 200 人的网络水军小号。 --- ### [59:26](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=3566s) 纽约市中小学出台一年期 AI 禁令与背后的政治动机 话题随后转向纽约市公立学校颁布的最新 AI 管制政策。杰森·卡拉卡尼斯（Jason Calacanis）介绍道，作为全美最大的公立学区，拥有 60 万名学生的纽约市宣布从即日起在幼儿园至八年级（K-8）全面暂停使用面向学生的生成式 AI 软件，禁令为期一年。高中阶段则获得豁免，当局还在 5 万名高中生中试点开展特定的 AI 教学项目；与此同时，该市还下达了极为严苛的屏幕使用时间指导规范：三至五年级学生每天屏幕时间上限为 30 分钟，六至八年级上限为 45 分钟。身为美国民主社会主义运动代表人物的纽约市相关政策推手佐兰·曼达尼（Zohran Mamdani）为其决策辩解称：“我至今还没看到任何一项研究能证明，AI 对小学和初中阶段的学生是有益的。”纽约市 DSA 成员朱莉·斯沃普（Julie Swope）更是公开抨击称，教育技术“不应该通过一家旨在让科技寡头进一步敛财的营利性公司来介入”。 ```mermaid flowchart TD accTitle: NYC School AI Policy Decision and Influences accDescr: 结构图展示了纽约市K-8学校AI禁令的决策驱动力、学区落地政策及反对阵营的观点争议。 subgraph Influences["决策驱动因素"] A["政治立场：DSA 意识形态<br/>反科技寡头资本介入"] B["教师工会压力：<br/>担忧传统教学被自动化替代"] C["公众技术恐慌：<br/>'尚未研究透彻' 的拖延叙事"] end subgraph Policy["纽约市新规细节 (K-8)"] D["全面禁止面向学生的生成式 AI 软件（为期 1 年）"] E["严苛限制屏幕时间：<br/>3-5年级限 30 分钟，6-8年级限 45 分钟"] F["高中豁免并保留 50,000 人规模试点"] end subgraph Impact["潜在负面后果（弗里德伯格分析）"] G["公立学校学生错失个性化学习良机"] H["加剧教育鸿沟：私立/富裕家庭子女超前，公立学生落后"] end A --> D B --> D C --> D D --> G E --> G G --> H ``` 针对曼达尼所谓“缺乏研究证据”的说辞，大卫·弗里德伯格（David Friedberg）随即引用斯坦福大学于当年 3 月发布的权威系统综述报告《The Evidence Base on AI in K-12: A 2026 Review》予以有力反驳。弗里德伯格指出，该报告全面梳理了全球关于基础教育阶段引入 AI 的数百篇学术文献，并严格筛选出了 20 项具备因果推断能力的最高质量研究。元分析的结论非常清晰：在能够使用 AI 工具时，学生的学业表现通常会有所提升；工具的具体产品设计至关重要，且 AI 能够为教育工作者提供实质性支持。现实中根本不存在任何坚实的学术依据来支持“AI 会损害基础教育和儿童智力发育”的假设。弗里德伯格一针见血地指出，决策者声称“研究尚不明确”永远只是想要阻碍变革时的借口和推托辞，这背后真正的博弈机制其实是教师工会（Teachers' Unions）面对职业被自动化威胁时的巨大抵触情绪，以及极左翼反科技、反商业的政治迎合。 弗里德伯格进一步动情地阐述了生成式 AI 重塑教育的革命性潜力。传统课堂模式由教师单向授课，全班被迫按照最慢学生的节奏统一前进，但现实中每个孩子的学习方式截然不同——有的偏好视觉图形，有的偏好听觉或故事叙述，有的则擅长项目制作或实操构建。AI 能够以极低成本为每个孩子提供因材施教的专属一对一辅导（personalized tutoring），极大激发学生的学习自信并加速其心智成长。弗里德伯格痛心疾首地警告，公立教育本是抹平阶级差距的利器，但如果公立学校因为僵化的政治意识形态和工会利益而禁止 AI，而私立学校与富裕阶层家庭却在积极采用 AI 教学，这将导致处于弱势地位的公立学校儿童被彻底甩在时代身后，在全球以及区域竞争中造成无法挽回的教育撕裂与不平等。 - **纽约学区 AI 禁令范围**：覆盖幼儿园至八年级（K-8），涉及约 60 万名公立学校学生，主要限制面向学生的生成式 AI 应用，高中阶段的 5 万人试点项目不在禁止之列。 - **斯坦福权威报告反驳**：《The Evidence Base on AI in K-12: A 2026 Review》分析了数百篇论文与 20 项严密因果研究，证实合理设计的 AI 辅助教学能切实提高学生成绩，驳斥了“缺乏实证”的说法。 - **阶层鸿沟隐忧**：富裕家庭与私立学校不受限制使用 AI，公立系统因教师工会与反科技政治叙事自我设限，最终受损的是缺乏资源的普通大众学童。 ### [01:06:38](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=3998s) 自适应学习与教育分化：纽约是否会沦为下一个密西西比？ 针对纽约市禁止 K-8 学校使用 AI 的政策，Jason 提议抛开政治立场，单纯从父母的角度来探讨 AI 在孩子教育中的利弊：在什么情况下完全不接触 AI 有益，又在什么场景下 AI 能发挥出无与伦比的价值。Chamath 随即回应指出，在他看来，未来孩子们如果不把“自适应学习平台”（Adaptive Learning Platform）作为主要的学习方式，是完全不可思议的。传统的“一年级、二年级、三年级”阶梯式大锅饭教育，正如 Friedberg 之前所言，是一种“寻找最低公分母”（Lowest Common Denominator）的失败模式。现实中每个孩子的学习节奏千差万别，有的偏向视觉输入，有的偏向听觉理解。理想的未来教育应当针对每个个体的学习风格量身定制，让 AI 充当随身导师在关键时刻点拨，而教师则把省下来的时间用于培养执行功能（Executive Function）、团队协作、复杂问题解决以及体育运动。目前，这一方向在公立体系中极为罕见，主要由前沿特许学校在践行，比如他们共同的朋友 Joe Lonsdale 创办的 Alpha School。 Chamath 进一步做出了一个极其尖锐的宏观预测：**“这标志着纽约正在沦为密西西比，而密西西比正在变成纽约。”** 那些由极左意识形态掌控的自由派重镇，正通过此类反技术政策制造一种隐蔽而严酷的“现代自我隔离”（Modern Segregation）。其代价是将整整一代依赖公立学校的孩子推入长期贫困的深渊。与之相对，那些拥抱 AI、大建数据中心、提供税收优惠并奖励教师的州将迅速实现弯道超车。有经济能力的人会用脚投票迁出，留下被彻底掏空的传统大都市。这将重构全美的经济与政治版图。Chamath 感叹，尽管对那些子女沦为牺牲品的工薪家庭深表同情，但人类往往“只有伸手摸到滚烫的炉子被烫伤后，才能真正吸取教训”。 ```mermaid flowchart TD accTitle: 地区教育技术政策引发的阶层与经济分化走向 accDescr: 对比拥抱AI与封禁AI的两类地区在教育、人口迁移及经济地位上的截然不同结果。 subgraph Restrictive ["极左主导地区（如纽约市）"] A1["出台公立学校 AI 禁令"] --> A2["维持传统大锅饭教育模式"] A2 --> A3["公立学校学生缺乏 AI 技能"] A3 --> A4["工薪家庭陷入长期贫困 / 城市空心化"] end subgraph Progressive ["拥抱创新地区（红州及新兴科技州）"] B1["引入 AI 自适应教育与基础设施"] --> B2["个性化辅导与高阶能力培养"] B2 --> B3["人才竞争力与生产力大幅跃升"] B3 --> B4["高技能家庭与资本流入 / 经济崛起"] end ``` --- ### [01:09:50](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=4190s) 意识形态作祟还是数字鸿沟？被削足适履的公立教育与高中“末日模块” Jason 转向 Sacks，询问其对纽约这项举措背后的政治动机有何冷峻见解。Jason 调侃道，反对者口中常念叨着“反寡头”，网络上甚至有阴谋论认为民主社会主义者（DSA）之所以封杀 AI，是不想让 Sam Altman、Dario Amodei 或微软从课堂软件里赚走一分钱；但这种逻辑荒谬至极，因为科技巨头通常极乐意向公立学校免费捐赠系统，普通用户也能白嫖市面上 80% 的核心功能。Sacks 直言，美国民主社会主义者（DSA）这类群体已经全盘拥抱了新卢德主义（Neo-Luddite）的反进步立场。 Sacks 指出，全美政界谈论了整整三十年的“数字鸿沟”（Digital Divide）——富人孩子坐拥最顶级的工具，穷人孩子因缺乏接触而落入下风。如今这一鸿沟正被政治体制人为地急剧拉大：私立学校会毫无顾忌地将顶尖 AI 工具嵌入日常教学，而公立学校的孩子却被剥夺了机会。这并非因为缺钱，也不是因为技术不可及，而是遭到了政客的公然背刺。正如 Sacks 与 Jason 挖苦的那样，推行这类政策的掌权者往往是“连孩子都没有的单身群体”，他们构建的选民基本盘正是那些向下流动的受挫阶层，让底层孩子继续丧失竞争力反而能筑牢他们的政治依赖。 Sacks 剖析了该项政策令人啼笑皆非的细则：纽约并非 100% 绝对封杀，而是在高中阶段面向至多 50,000 名学生设立了一个每年仅开展两次的试点“AI 素养模块”。根据官方文书的描述，这些课时并非教导学生如何利用 AI 提高学习效率，而是充满说教与防范色彩： - **警惕优先**：旨在帮助学生在“开始依赖 AI 之前”，批判性地审视其边界； - **充斥安全术语**：着重向学生灌输算法偏见、歧视、道德伦理风险以及对未来就业的冲击； - **全方位严密监控**：所有使用必须在受训教员直接监督下进行，仅限经过极其严苛的数据隐私审查的工具。 Sacks 和 Jason 嘲讽这无异于把孩子们送进了“末日论者学院”（Doomer Academy）。与此同时，中国正在基础教育（K-12）中大举推进 AI 工具的应用与普及。当竞争对手的孩子在借助 AI 研发机器人、冲刺技术前沿时，美国的公立学校却忙着教导孩子们“如何抵触并怨恨新技术”。 --- ### [01:14:41](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=4481s) 蒙台梭利与 AI 辅导的碰撞：脑机实验下的警示信号 Jason 随后结合自己的育儿理念阐明立场。他表示自己既是蒙特梭利教育法（Montessori）的忠实拥趸，又是 AI 个性化自适应辅导的坚定支持者。对于那些请不起昂贵私人家教的普通家庭来说，生成式 AI 是实现教育平权、改变命运的利器。然而，他也抛出了一个极富争议的科学对照，告诫大家不可盲目乐观。 Jason 调侃自己抢了 Friedberg 的“科学角”（Science Corner）风头，让制作人 Nick 调出了一篇美国国家卫生研究院（NIH）收录的麻省理工学院（MIT）学术论文。该研究追踪了 54 名受试者，在为期四个月的四次独立实验中让他们撰写论文，并佩戴脑电波监测设备（Headset）对受试者的大脑神经活动进行实时成像扫描。实验共将受试者分为三组： 1. **纯大脑创作组**：不借助任何外部数字工具，全凭自身大脑思考写作； 2. **传统搜索组**：允许使用传统网络搜索引擎查找资料与文献； 3. **大语言模型（LLM）组**：直接使用 ChatGPT 协助起草和组织论文内容。 实验的核心脑电数据揭示了一个令人警惕的现象：那些完全借助大语言模型进行写作的参与者，其大脑特定区域的活跃度出现了急剧、严重的下滑…… ### [01:15:53](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=4553s) AI 导致大脑退化与“双西格玛问题”的教育悖论 嘉宾们继续探讨纽约市公立学校禁用 AI 工具的深层影响。根据一项关于学生使用大语言模型（LLM）写作的研究发现，学生表现出了记忆力和文章归属感方面的显著缺陷——高达 83% 的学生甚至无法准确复述或引用自己刚刚写出的文章内容。相比之下，传统的搜索引擎组表现介于两者之间，因为他们不得不主动记忆事实并整合信息。主持人指出，这揭示了一个残酷的事实：如果不亲自动脑去完成繁重的工作，大脑就会像长期不耕作的身体一样退化，这也是现代人需要去健身房锻炼的原因。 随后讨论转向著名的“布鲁姆双西格玛问题”（Bloom's 2 Sigma Problem）。教育学界早在 1980 年代就已经证实，接受一对一导师辅导的学生，其学业表现比普通课堂教学的学生高出整整两个标准差（2 Sigma）。然而，优质的私人辅导历来是富裕家庭的特权，公立学校的普通家庭根本无力承担昂贵的私人家教。席间聊起当年的考试经历，Sacks 坦言自己也是在没有导师指导的情况下裸考 SAT，第一次考了大约 1420 分，若非缺乏辅导技巧，本可以取得更高的分数。 这种现实构成了一个巨大的教育悖论与紧张关系：一方面，家长确实不希望孩子的大脑因为过早、过度依赖 AI 而退化，希望他们能掌握基础的读写思维能力；但另一方面，那些有上进心却请不起昂贵家教的工薪家庭学生，若被剥夺了 AI 作为低成本“数字导师”的机会，就完全丧失了追赶那“两个标准差”差距的工具。禁用 AI 实际上堵死了普通家庭孩子获得一对一顶级辅导的平民化路径，因此当前的一刀切禁令绝非良策。 - **83% 记忆缺失**：研究显示，完全依靠 AI 辅助写作的学生中，83% 无法引用自己文章中的内容。 - **Bloom's 2 Sigma 效应**：一对一专业辅导能将学生的平均表现提升两个标准差（相当于将普通学生提升到前 2% 水平）。 - **SAT 分数对比**：Sacks 透露自己当年毫无辅导、直接裸考 SAT 拿到约 1420 分，指出有系统辅导的学生非常容易突破 1300 至 1400 分大关。 --- ### [01:18:57](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=4737s) 特朗普政府达成委内瑞拉石油协议 话题转入重大地缘政治与能源新闻：特朗普政府敲定了关于委内瑞拉石油资源的重磅协议。自今年 1 月美军成功执行行动控制了委内瑞拉前领导人尼古拉斯·马杜罗（Nicolás Maduro）后，由副总统兼代总统德尔西·罗德里格斯（Delcy Rodríguez）接任。特朗普随即宣布美国将主导委内瑞拉庞大的石油资源开发，如今该协议的详尽条款正式公布。 协议规定，北美蓝能合伙企业（North American Blue Energy Partners）获得了针对 17 个油田长达 100 年的特许开采权，覆盖高达 650 亿桶原油储量，这些油田此前大多由俄罗斯、中国以及马杜罗名下的实体控制。在整个交易结构中，美国政府占据 55% 的控制权，其中五角大楼（美国国防部）直接持有 35% 的股权，美国国务院则享有以成本价购买 20% 产出的权利（主要用于战略储备与军用需求）。该协议由国务卿马可·卢比奥（Marco Rubio）和国防部长皮特·赫格塞斯（Pete Hegseth）共同签署，蓝能财团承诺投入约 1000 亿美元进行基础设施重建，雪佛龙（Chevron）等跨国能源巨头亦将深度参与。特朗普公开强调，这笔世纪大单“不会花费美国纳税人一分钱”。 ```mermaid flowchart TD accTitle: 委内瑞拉石油协议权益分配图 accDescr: 展示100年特许权涉及的650亿桶储量及其在美国政府各部门与财团之间的分配比例。 subgraph Deal["100年特许开采权 (覆盖 650 亿桶储量)"] direction TB GovShare["美国政府掌控权益: 55%"] PrivateShare["Blue Energy 牵头联合财团<br/>(承诺投资 1000 亿美元基建)"] end subgraph Allocation["美国官方权益划分"] direction TB Pentagon["五角大楼 (Pentagon)<br/>获得 35% 股权"] StateDept["美国国务院 (State Dept)<br/>享 20% 产量成本价认购权"] end GovShare --> Pentagon GovShare --> StateDept ``` - **特许经营期**：长达 100 年，涵盖 17 个关键油田。 - **核心储量**：涉及 650 亿桶已探明或可开采储量。 - **股权与权益结构**：美国政府控制 55%，五角大楼独占 35% 股权，国务院保留 20% 原油成本价回购权。 - **配套基建投资**：财团承诺对当地破损的基础设施再注资约 1000 亿美元。 --- ### [01:20:41](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=4841s) 商业重组还是国家建构？重质原油的地缘与工业价值 主持人向 Sacks 抛出一个尖锐的“快速球”：美国此举是否属于强制更迭政权、抢夺他国自然资源的帝国主义“国家建构（Nation Building）”行为？Sacks 坚决驳斥了“国家建构”的说法。他强调，美国并没有像在阿富汗那样试图驻军、强推美式民主或输出意识形态文化，这本质上是一桩极其纯粹的“商业交易”（Business Deal）。 Sacks 指出，外界批评该协议剥削委内瑞拉人民完全是脱离现实。在马杜罗执政的过去十年里，委内瑞拉原本高达每天 300 万桶的产油量暴跌了约三分之二，萎缩至仅剩 100 万桶左右。委内瑞拉坐拥全球首屈一指的石油储量——总储量约 3000 亿桶（与沙特阿拉伯相当甚至略多），即便扣除奥里诺科重油带（Orinoco Belt）约 2000 亿桶难以提炼的超重劣质油泥，依然拥有约 1000 亿桶公认的高价值重质原油储量。 更重要的是工业维度的强互补性： 1. **轻质油 vs 重质油**：美国借助页岩油水力压裂技术生产的大多是轻质低硫原油（Light Sweet Crude），非常适合裂解提炼为汽油，但在提炼柴油和馏分油方面效率不足。 2. **炼油厂配置兼容**：美国墨西哥湾沿岸的大型炼油厂最初全部是针对高硫重质原油设计的，若全部改建以适应轻质油不仅成本极其高昂，还会造成设备降级。委内瑞拉的重质原油恰恰能无缝填补美国炼厂的产能缺口，实现双赢。 在定价方面，委内瑞拉过去 5 至 10 年的原油实际结算价大约仅在每桶 53 至 58 美元之间，因为此前在西方制裁与专制统治下，马杜罗只能以每桶贴水约 20 美元的“跳楼价”向中国折价抛售。 Friedberg 补充分析认为，此举最大的战略驱动力是在伊朗战争爆发、霍尔木兹海峡面临封锁危机的背景下，确保美国本土半球的能源绝对安全，并斩断马杜罗政权长期作为俄罗斯与中国折扣能源补给及非法地下交易链条“蛇头”的纽带，彻底将其地缘利益收回西半球体系。 ### [01:25:18](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=5118s) 美委石油协议背后的地缘博弈与委内瑞拉政局分歧 在探讨美委能源协议的战略动机时，弗里德伯格（David Friedberg）指出，中国此前对委内瑞拉的能源储量表现出浓厚兴趣并提供实质支持，这种资金流动甚至间接资助了俄罗斯在古巴等拉美地区的活动。因此，这一协议的核心动机不仅在于“对美国有利”，更在于“削弱美国的对手”，阻止外部敌对势力在西半球进一步掌控战略能源储备。 与此同时，委内瑞拉内部的政治动态也出现了戏剧性转折。弗里德伯格提到自己曾采访过委内瑞拉反对派领袖玛丽亚·科里纳·马查多（María Corina Machado）。在普遍被认为存在舞弊的争议选举之后，本应胜选的马查多未能掌权，而近日她公开表态，抨击现任委内瑞拉政府缺乏合法性，根本无权与美国签署这项为期25年的能源协议。这一表态让局势变得十分微妙：尽管马查多历来被视为美国的坚定盟友，但如今她却号召委内瑞拉人民感到愤怒与被遗弃，指责现政权出卖国家资产。弗里德伯格强调，衡量这项协议长期成败的根本标准，在于收益能否真正改善委内瑞拉普通民众的生活，而不是重蹈覆辙——仅仅肥了统治集团少数腐败政客的腰包。 ```mermaid flowchart TD accTitle: 美委能源协议的地缘与经济驱动架构 accDescr: 展示美国与委内瑞拉能源合作的地缘战略诉求与潜在风险 subgraph Geopolitical ["地缘战略考量"] US["美国核心目标"] Sec["锁定西半球重油储备\n与本土轻质原油互补"] Block["切断中俄对拉美渗透\n遏制敌对地缘影响力"] US --> Sec US --> Block end subgraph InternalRisks ["委内瑞拉国内局势与风险"] Opp["马查多等反对派\n抗议协议不具合法性"] Gov["德尔茜·罗德里格斯政权\n掌控军警建制派维持稳定"] Assets["长期投资风险\n（百亿基建升级可能遭遇违约/国有化）"] Opp -.抵制.-> Gov Gov --> Assets end Geopolitical --> Gov ``` ### [01:27:23](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=5243s) 对华挑衅、基建保护与马查多权衡 针对这一重大交易，杰森（Jason Calacanis）提出质疑：美国先是在持续七个月的对伊战争中切断了中俄获取伊朗石油的通道，如今又拿下委内瑞拉，这一连串举动是否极具挑衅性，甚至可能引发与中国的冲突升级？大卫·萨克斯（David Sacks）对此进行了客观剖析。他指出，中国每天需进口约1100万桶原油（战前数据，近期已有下降），而委内瑞拉目前的日产量仅约100万桶，中国的主要原油供给依旧来自波斯湾和俄罗斯，因此委内瑞拉油田的易手对中国供应链构成的直接冲击相对有限。 萨克斯进一步分析指出，从国家安全和经济互补性来看，锁定西半球庞大的重油储备无疑是明智之举，但真正的隐患在于美国的长期资本暴露： - **基建升级的沉淀资本**：美国石油公司若要将委内瑞拉的日产量从100万桶提升至300万桶，需投入数十亿美元升级年久失修的基础设施。 - **政策与政权风险（Rug Pull）**：该协议期限长达25年，一旦未来委内瑞拉政局再次变天，新政府极有可能将美国资产国有化并没收升级设备。 - **为何不扶持马查多上台**：萨克斯直言，川普政府没有强推马查多掌权是务实的选择。在拉美政坛立足必须依靠“握有枪杆子的人”（men with guns）；马查多在军方内部缺乏根基，若强行推其上台，美国势必得派遣地面部队（GIs）为其保驾护航。相比之下，掌控实权的德尔茜·罗德里格斯（Delcy Rodríguez）及其家族在军警建制派中根基牢固，美国当前的核心利益是确保当地有一个能够保障产权与稳定履约的合作政权。 ### [01:30:17](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=5417s) 尾声：迈向300期大关与兄弟调侃 在关于拉美能源的讨论渐入尾声时，杰森向查马斯（Chamath Palihapitiya）询问看法，查马斯坦诚表示自己离这个话题太远，并不了解内情。 随着第288期节目录制接近尾声，主持人与嘉宾们开启了轻松热烈的庆贺模式。杰森感慨这支四人小分队竟然不知不觉合作了288期，大家信心满满地表示“铁定能挺进第300期”。萨克斯笑称节目现在节奏越来越好，录起来比以往轻松得多；查马斯则幽默地将这种顺利归功于“AI对播客的巨大帮助”，戏称大家仿佛回到了“1999年狂欢”的状态。在呼喊着“下一站卡波（Cabo）”、互道“Love you besties”以及标志性的金句与热烈氛围中，本期节目正式画上句号。 ## Notable quotes > [01:16](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=76s) 能力远超、远超、远超（much, much, much）以往的模型即将到来。下一代模型将让所有人感到清醒而震撼（sobering for everyone）。 — 萨姆·奥尔特曼 > [08:44](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=524s) 现在的市场正在疯狂暴涨（The market is ripping）。 — 大卫·萨克斯 > [28:20](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=1700s) 开发者把钥匙大摇大摆地扔在门外，然后被 Agent 顺手捡了起来。 — 大卫·萨克斯 > [29:44](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=1784s) 这就像我们扛着一把机关枪，对着一叠毫无防备的打印纸疯狂扫射，然后惊呼‘天哪它竟然被打穿了！’一样荒谬。 — 布拉德·弗里德伯格 > [39:21](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=2361s) 我唯一的问题在于，我认为他们没有充分披露其显而易见的利益冲突。当 Sacks 扯出这根毛线头时你会发现：A 写了文章，A 与 B 是室友，B 的妻子为前沿实验室领头人工作，B 本人又投资了 500 万美元，还有在实验室任职的 C……区别在于，一旦隐瞒不公开披露，任何持怀疑态度的理性人都会追问：这究竟是客观探讨，还是精心策划的局？ — 查马斯·帕里哈皮蒂亚 > [59:26](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=3566s) 我至今还没看到任何一项研究能证明，AI 对小学和初中阶段的学生是有益的。 — 佐兰·曼达尼 > [59:26](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=3566s) 教育技术不应该通过一家旨在让科技寡头进一步敛财的营利性公司来介入。 — 朱莉·斯沃普 > [01:06:38](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=3998s) 这标志着纽约正在沦为密西西比，而密西西比正在变成纽约。 — 查马斯·帕里哈皮蒂亚 > [01:06:38](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=3998s) 只有伸手摸到滚烫的炉子被烫伤后，才能真正吸取教训。 — 查马斯·帕里哈皮蒂亚 > [01:58:57](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=7137s) 不会花费美国纳税人一分钱。 — 唐纳德·特朗普 > [01:27:23](https://www.youtube.com/watch?v=DvFe9bR2eHA&t=5243s) 在拉美政坛立足必须依靠‘握有枪杆子的人’（men with guns）。 — 大卫·萨克斯 ## Mentioned 查马斯·帕里哈皮蒂亚, 杰森·卡拉卡尼斯, 大卫·萨克斯, 布拉德·弗里德伯格, 萨姆·奥尔特曼, 格雷格·布罗克曼, 埃隆·马斯克, 德瓦克什·帕特尔, 伯尼·桑德斯, 比尔·格利, 莱奥波德·阿申布伦纳, 达里奥·阿莫戴, 黄仁勋, 萨提亚·纳德拉, 格温·肖特韦尔, 乔希·夏皮罗, 格雷格·阿博特, 史蒂夫·班农, 唐纳德·特朗普, 达斯汀·莫斯科维茨, 亚安·塔林, 维塔利克·布特林, 佐兰·曼达尼, 朱莉·斯沃普, 玛丽亚·科里纳·马查多, 德尔西·罗德里格斯, 尼古拉斯·马杜罗, 马可·卢比奥, 皮特·赫格塞斯, OpenAI, GPT-6 Astra, Anthropic, xAI, Grok, Grokbot, Robotaxi, 特斯拉, Instinct, Hugging Face, Polymarket, 英伟达, 智谱 AI (GLM 5.2), 雪佛龙, North American Blue Energy Partners, 美国民主社会主义者 (DSA), 未来生命研究所 (FLI) ## Video description (0:00) Bestie intros: Happy 50th Chamath! (1:17) OpenAI releases GPT-6 Astra: Have they reached AGI? (19:01) All-In Summit: Scholarships available! (19:56) Dwarkesh's Hugging Face hack article sparks a Bernie Sanders AI ban (37:03) EAs, AI-brained influencers, and the conflicts of interest (59:26) Mamdani bans AI in NYC schools K-8 (1:18:58) US-Venezuela oil deal The All-In Summit is only 1 week away! You won't want to miss it: https://theallinsummit.com Follow the besties: https://x.com/chamath https://x.com/Jason https://x.com/DavidSacks https://x.com/friedberg Follow on X: https://x.com/theallinpod Follow on Instagram: https://www.instagram.com/theallinpod Follow on TikTok: https://www.tiktok.com/@allin Follow on LinkedIn: https://www.linkedin.com/company/allinpod Intro Music Credit: https://rb.gy/tppkzl https://x.com/yung_spielburg Intro Video Credit: https://x.com/TheZachEffect Referenced in the show: https://openai.com/index/gpt-6-astra https://www.washingtonpost.com/technology/2026/09/03/openai-greg-brockman-says-its-new-model-astra-is-agi https://www.axios.com/2026/09/03/axios-interview-sam-altmans-sobering-siren https://polymarket.com/event/which-company-has-best-ai-model-end-of-2026 https://polymarket.com/event/us-enacts-ai-safety-bill-before-2027 https://x.com/chamath/status/2095511879951192279 https://x.com/Jason/status/2095521697411514626 https://scale.stanford.edu/research-in-action/understanding-evidence-base-ai-k12-education https://www.politico.com/news/2026/09/02/trumps-seen-the-data-center-polling-and-he-still-supports-them-01061244 https://abc7ny.com/post/new-york-city-public-schools-banning-ai-use-middle-school-year/19778716 [Source](https://www.youtube.com/watch?v=DvFe9bR2eHA)

[1:16] 萨姆·奥尔特曼: 能力远超、远超、远超（much, much, much）以往的模型即将到来。下一代模型将让所有人感到清醒而震撼（sobering for everyone）。 [8:44] 大卫·萨克斯: 现在的市场正在疯狂暴涨（The market is ripping）。 [28:20] 大卫·萨克斯: 开发者把钥匙大摇大摆地扔在门外，然后被 Agent 顺手捡了起来。 [29:44] 布拉德·弗里德伯格: 这就像我们扛着一把机关枪，对着一叠毫无防备的打印纸疯狂扫射，然后惊呼‘天哪它竟然被打穿了！’一样荒谬。 [39:21] 查马斯·帕里哈皮蒂亚: 我唯一的问题在于，我认为他们没有充分披露其显而易见的利益冲突。当 Sacks 扯出这根毛线头时你会发现：A 写了文章，A 与 B 是室友，B 的妻子为前沿实验室领头人工作，B 本人又投资了 500 万美元，还有在实验室任职的 C……区别在于，一旦隐瞒不公开披露，任何持怀疑态度的理性人都会追问：这究竟是客观探讨，还是精心策划的局？ [59:26] 佐兰·曼达尼: 我至今还没看到任何一项研究能证明，AI 对小学和初中阶段的学生是有益的。 [59:26] 朱莉·斯沃普: 教育技术不应该通过一家旨在让科技寡头进一步敛财的营利性公司来介入。 [66:38] 查马斯·帕里哈皮蒂亚: 这标志着纽约正在沦为密西西比，而密西西比正在变成纽约。 [66:38] 查马斯·帕里哈皮蒂亚: 只有伸手摸到滚烫的炉子被烫伤后，才能真正吸取教训。 [118:57] 唐纳德·特朗普: 不会花费美国纳税人一分钱。 [87:23] 大卫·萨克斯: 在拉美政坛立足必须依靠‘握有枪杆子的人’（men with guns）。

[1:20] Sam Altman: much, much, much more capable models are coming soon [1:20] Sam Altman: sobering for everyone [19:56] Dwarkesh Patel: The Rise and Fall of Ancient Civilizations [50:44] David Sacks: if anybody could make the decision about technology in DC, I'd want it to be President Trump. [59:26] David Friedberg: The Evidence Base on AI in K-12 [64:02] David Friedberg: we don't yet know or understand [66:12] Chamath Palihapitiya: This is when New York becomes Mississippi while Mississippi becomes New York. [69:50] David Sacks: AI-related bias, risks, ethical considerations, and impacts on careers [78:57] Donald Trump: no cost to the American taxpayer

[0:00] All right. Happy birthday, Chimath. [0:01] We're here taping on your 50th birthday. [0:04] Super exciting for you, brother. [0:05] >> Thank you. It's really great. I got a [0:07] birthday letter from uh the big man. [0:09] >> Oh, wow. [0:10] >> If you had asked my parents, [0:13] >> 50 years later, you're going to have a [0:14] kid that gets a letter from the then [0:17] sitting president of the United States, [0:18] they would have thought absolutely not [0:21] possible. [0:22] >> Wow. Look at that. Which White House [0:24] intern wrote that? [0:28] >> I wonder. [0:30] Let your winners ride [0:34] David. [0:38] >> We open source it to the fans and [0:40] they've just gone crazy with it. [0:46] >> Well, actually speaking about interns, [0:47] it was an intern. In this case, it was [0:49] my son. The story was that, you know, [0:51] Braden worked at the White House all [0:53] summer as an intern. And at the end of [0:55] the summer, you know, he used all of the [0:57] political capital he built and put in a [0:59] request to get this letter from the [1:00] president and he did it. Obviously, I [1:02] want to thank POTUS for sending me the [1:03] letter, but I really want to thank my [1:04] son because it was a really thoughtful [1:06] thoughtful thing that he did for me. So, [1:08] really appreciative of that. [1:09] >> That actually is impressive because, you [1:12] know, it takes time out of the [1:13] president's schedule to sign things. So, [1:15] that's why I mean something, you know. [1:16] >> All right, listen. Breaking news. [1:18] Breaking news. It's a breaking news [1:20] story here on the All-In podcast. Open [1:23] AAI is rolling out chat GPT 6 aka Astra [1:28] AS folks. This is their new flagship [1:31] model. It's going to be available to a [1:33] limited set of organizations in the [1:35] coming days. Y'all will be able to get [1:37] it if you have a credit card in your [1:39] chat GPT account, i.e. the plus pro [1:42] business and enterprise users. You're [1:43] going to get it soon. Greg Brockman says [1:46] he believes OpenAI has entered the AGI [1:49] era. And uh we've talked about that here [1:52] before. Sam Alman is well he's back to [1:56] uh being a little bit uh performative/ [1:59] alarmist or he sincerely believes that [2:01] this is the one. Here's what he said at [2:03] Wednesday's G20 summit. Quote much much [2:08] much three muches more capable models [2:12] are coming soon. The next generation of [2:14] models are going to be sobering for [2:17] everyone. And his final quote, I think [2:19] no one intellectually honest can look at [2:21] what's happening and not feel the weight [2:24] of responsibility in front of us. My [2:26] lord. Uh a little bit of drama there. Uh [2:28] added to his um his quotes. People are [2:31] freaking out about the benchmarks which [2:34] are off the charts. Some people are [2:36] saying uh this is the one. Your [2:39] thoughts, Chimoth? [2:41] >> I don't think normal humans know how to [2:42] read this chart and infer anything [2:44] useful from it. The more important [2:45] takeaway, I agree with Greg. I think AGI [2:48] has basically been here since the [2:50] beginning of the year. There are models [2:52] that are incredibly performant inside of [2:55] the closed frontier lab companies [2:58] and I think right now we're trying to [3:00] figure out the right cadence with which [3:01] to release it. I think it's very [3:03] important. I also think that these [3:05] capabilities will be matched or close to [3:08] matched by a handful of closed source [3:11] alternatives and [3:14] open source alternatives [3:16] >> probably within the next 3 or 4 months. [3:19] So I think the right way to think about [3:20] this for most people is [3:24] we are seeing [3:27] incredible intelligence capabilities be [3:30] broadly available and we are seeing the [3:32] cost of that incremental unit of [3:35] intelligence being driven further and [3:37] further down. So it's a really positive [3:40] trend. Now I think it's just the messy [3:42] middle part of organizing this properly [3:44] for companies and for people and [3:46] building things that are useful and [3:48] showing an actual ROI and I would just [3:51] >> walking and tackling. Yeah, [3:52] >> I would just keep calm and carry on. [3:54] >> Okay, Sax, any thoughts here on uh [3:57] breaking news today and then Freeber, [3:59] I'll get your reaction. [4:00] >> But by the way, it is very exciting and [4:01] congratulations to OpenAI and I'm sure [4:03] anthropic will have something as Grock, [4:05] you know, Grock still has yet to drop. [4:07] 46 is coming. That's going to be a [4:08] blockbuster. The meta models are really [4:11] good. Gemini just dropped something [4:12] yesterday that was exceptional. So the [4:15] great thing is we have a lot of choice [4:17] >> and the even better thing is the cost of [4:19] that choice is becoming cheaper and [4:20] cheaper and cheaper. That's very [4:22] >> sax. Let me tear it up to you this way. [4:25] Openai seems to get its groove back now [4:28] after anthropic was running away with [4:29] it. Grock and rockbot are doing things [4:32] that those two other frontier models [4:34] simply can't do. to the execution, [4:36] whether it's the harness or the [4:37] underlying model or the instruction set [4:40] or cursor's magic sauce. Clearly, [4:43] Grockbot is competitive if not better [4:45] for most users. So, vibrant vibrant [4:49] competition with leaprogging occurring [4:52] at a cadence I would say of two to four [4:55] weeks right now. This is what you told [4:58] us you were going to do in sort of [5:00] setting up the architecture here when [5:02] you took over as a ZAR. Can this [5:06] continue? And what's it going to be like [5:07] when Jensen gets in here with uh Laguna [5:11] and his access to all these open source [5:13] models? I mean, this this could be fate [5:16] a complete know that we win the AI race [5:18] here. [5:19] >> Look, there's a lot of innovation going [5:20] on. I think American companies are [5:22] competing very well. I think we are [5:23] winning the AI race as long as we don't [5:24] do something stupid to shoot ourselves [5:26] in the foot. [5:27] >> Like, [5:28] >> well, I I've talked about it a million [5:30] times. I mean, setting up a new AI [5:32] regulatory agency that operates like the [5:35] FDA and takes years to approve new [5:37] models. [5:38] >> Yeah. [5:38] >> But any event, I don't think it's going [5:40] to happen under President Trump. I [5:41] really don't. So, anyway, things are [5:42] going well. Look, Sam told us a few [5:45] months ago, remember when he posted and [5:46] said that he was expecting the company [5:49] to have its best 12 months. He said, [5:50] "The last year hasn't been so great, but [5:52] the next year is going to be really [5:53] great." Maybe it was because they were [5:54] seeing these results in the lab, [5:57] >> you know? [5:57] >> Sure. And then remember what Sarah Frier [5:59] said that I think it was since the [6:01] beginning of Q2, so around the middle of [6:04] the year, they saw a market acceleration [6:06] in their growth rate. That was I think [6:08] with the release of the 5.6 models. Now [6:10] you're seeing the 6.0 models. And man, I [6:13] mean, it just keeps advancing. But look, [6:15] I still would say to you that I think [6:18] the market structure that we have right [6:19] now is there's two markets. There's the [6:23] market for frontier intelligence and [6:25] that is a duopoly. that really is [6:27] enthropic and open AI and it's good to [6:29] see that that is still a horse race and [6:31] they're competing and then there's [6:32] commodity intelligence which is [6:34] everybody else including all the open [6:35] models and they're not competing on [6:37] being at the frontier they're competing [6:38] on price and those prices are [6:40] significantly lower but it really is I [6:43] think a two-tier market at this point I [6:45] agree with you about how promising [6:47] Grockbot is Grockbot is awesome [6:50] >> I would assume that both Anthropic and [6:52] OpenAI will eventually roll out their [6:55] own version of of bots. [6:57] >> They need a simplified stripped down [6:59] version for normies. [7:01] >> Yeah. [7:01] >> And they can leave the more complicated [7:03] co-work etc. to the top five or 10% of [7:07] the audience. [7:09] >> Yeah. The uh ratings here from Poly [7:11] Market, which company has the best AI [7:13] model. By the end of 2026, OpenAI's [7:17] chances just spiked [7:18] >> from low single digits now to over 20. [7:22] >> People were really counting them out [7:23] now. It looks like they're viable again. [7:26] Anything that creates competition, I [7:28] think, is good. Before it looked like [7:29] Anthropic was running away with it, [7:31] they're about to have this monster IPO. [7:33] If OpenAI can at least make it a two- [7:35] horse race, that's good. I think it's [7:37] great that Grockbot launched. I agree [7:38] with you. The product is awesome. [7:40] >> Here's your big tip, David. I created a [7:41] new bot and I called it improve my bots [7:44] bot. [7:45] >> Mhm. [7:46] >> And I made a bot and I said, "Give me an [7:47] instruction set here. go through all my [7:49] bots and fill out their skills and [7:52] instructions and make them better every [7:54] day. It went through and it took like [7:58] two or three of them said, can I put [7:59] these two or three together and make one [8:00] instruction set? It would be better for [8:02] you. These two or three should come [8:03] together and it then went through and [8:05] told me all the changes made my bots [8:07] like 30 40% better on average I would [8:09] say. So [8:10] >> So you create a civilization your [8:12] civilization [8:12] >> created a multi-level civilization to [8:14] make everybody better. Yes. [8:15] >> Yeah. Your civilization is thriving. No, [8:17] that that is a really good like hack in [8:20] Frockbot is you tell your chief what [8:22] else should I be using you for? Okay, [8:24] now what do I do? [8:25] >> Yes, [8:26] >> I was like building some products and I [8:27] asked it what should we do next? You [8:29] know, and then it suggested some more [8:30] features. [8:32] >> It's really cool. [8:33] >> It's uh it's a brilliant product. It's [8:35] called Grobot. Go download it, folks. [8:37] Shout out to my guy and congratulations [8:39] on uh Robo Taxi taking the streets in [8:41] Austin today. [8:42] >> I just applied. [8:44] >> Dude, the market is ripping. I mean I'm [8:45] telling it if like I want to ask you [8:47] guys does this feel like 98 or 99 to [8:49] you? Were you guys around for that? [8:50] Jake, you were [8:51] >> Oh, for sure. I was covering feels [8:53] definitely 96 97. I think [8:56] >> you think it's that early. [8:57] >> I think it's 97 98. [8:59] >> I do think we are seeing at the late [9:01] stage a disconnection in reality and [9:04] valuations [9:06] where companies are being given far too [9:09] much credit. They're being priced at 50 [9:10] to 100 times topline revenue. That's [9:13] obviously not sustainable. the people [9:14] who invest in those are similar in my [9:16] mind to you know people that latestage [9:18] do uh investments unless unless it's run [9:22] by some you know legendary oneofone [9:25] entrepreneur and then you want to give a [9:28] Elon a 30x on revenue totally makes [9:30] sense you want to give Travis something [9:32] similar totally makes sense but unproven [9:37] >> entrepreneurs getting a 50 or 100x [9:41] >> is a little bit danger the entry price [9:43] matters Did you see Instinct is like the [9:45] new [9:46] >> billion? It makes no sense. [9:47] >> Personal AI assistant. It corresponds [9:50] with you by iMessage. And yeah, I think [9:52] they just got a term. She had two and a [9:53] half billion. They're in private beta [9:55] right now. It is a new hotness. All the [9:57] VCs are trying it, but still that does [9:59] seem kind of crazy. [10:00] >> No, no, not money in that was the [10:02] valuation. [10:02] >> That was the valuation. [10:03] >> Yeah. [10:06] Listen, congratulations to the [10:07] entrepreneur. No dig there. But [10:09] >> yeah, I want to try it. By the way, let [10:10] me try it. [10:11] >> I got invited and you know what? I read [10:12] the terms of service and people were [10:14] talking about that online and I [10:15] immediately did not connect my stuff [10:17] because it gets access to everything and [10:18] then somebody told me it's a human [10:21] uh in the loop service with and I don't [10:24] know if that's true or not but I get the [10:26] sense that it's like a human in the loop [10:28] watching it work and making the [10:32] whatever the instructions are and [10:34] whatever the finished product is and [10:35] polishing the last 20%. That's what I [10:37] suspect's going on here is a little [10:40] polish on the end. So, if you [10:42] >> want to order stuff, etc., which which [10:44] is totally fair. [10:45] >> Well, anyway, look, I I trust I don't [10:47] want to specifically [10:49] >> call that one out cuz I just haven't [10:50] even tried it. I'd like to try it, but [10:52] >> things are getting a little bit frothy. [10:54] I would say, [10:54] >> yeah, [10:55] >> things are becoming a little bit [10:56] euphoric. [10:57] >> I mean, Freeberg, do you have a take on [10:59] like if you were to analogize this to [11:02] the late 90s, where do you think we are [11:04] in the cycle? [11:05] >> Yeah, you were you were at a rave, by [11:07] the way, in downtown LA at that time. [11:10] You were 19 years old and you were [11:12] raving on uh [11:14] >> had a warehouse in Oakland at 7 a.m. on [11:16] my way to the Oaks card club to play two [11:18] four limit. [11:22] >> Wow. How'd you do? How'd you run? You're [11:24] like, I love these cards. [11:26] >> Were J and I are the only ones who are [11:28] actually working in the dot bubble. [11:31] >> I was working in 99. Well, I started in [11:33] 2001. I got a job offer senior year and [11:36] I wasn't sure if it was rescended by the [11:38] time I graduated because of the the [11:40] tanking of the market that year. [11:41] >> I got a job at Cisco. I to be an [11:43] engineer at Cisco. [11:44] >> If you never lived through the the dot [11:47] crash, it was I mean cuz [11:48] >> it was heady. It was [11:49] >> it was heady and then it was super [11:51] euphoric and then it was like super [11:53] miserable. But pretty much any founder [11:55] who's you know under 50 which is the [11:59] scar tissue. I've never been through [12:01] like a massive cycle like that. I mean, [12:02] well, the Zer the Zer bubble was like [12:05] was kind of a cycle. [12:06] >> Here's the thing, though. Hold on. [12:08] >> With the boom, it was all like metrics [12:11] that weren't dollars. Yes. [12:13] >> What we're seeing now is revenue and [12:15] profits [12:16] >> and growth in revenue and profits that [12:19] we've never seen before. And that's kind [12:21] of the anthropic story. It's the data [12:24] center, infrastructure, networking, [12:26] equipment story. I mean it is all real [12:28] dollars flowing versus speculative [12:32] utilization. [12:33] >> You're exactly right. People were basing [12:34] it on metrics like hits to a website and [12:37] and web pages. Here you have a very [12:39] distinct difference. You actually have [12:40] revenue. You actually have products in [12:42] market. But imagine if [12:45] Anthropic right now the way to think [12:47] about the sax is imagine if enthropic [12:49] open AI you know in this cohort of [12:51] companies lost 95% of their value in [12:56] under six months and then the long tale [12:58] of companies below them literally went [13:00] out of business and people were leaving [13:02] those offices with their Aryon chairs [13:04] and their computers on the seats h and [13:06] just total implosion. That's what [13:08] happened in the docom bust. It was [13:10] cataclysmic [13:12] literally. [13:12] >> No, I agree. So, do you think [13:14] >> totally different than this? [13:15] >> So, do you think we just have scar [13:16] tissue from that and are kind of trying [13:18] to overfit patterns and every time [13:20] there's basically some like new platform [13:23] shift some [13:24] >> paradigm shift? Yes. [13:25] >> No. Whenever we're in a bull market, the [13:27] conversation always turns to, you know, [13:29] is this a bubble? [13:30] >> Yes. [13:30] >> You know, so is everything overfitted to [13:32] that original pattern? [13:33] >> Correct. Or or is it the case that we [13:36] have a lot of young founders who've [13:37] never been through a cycle before and [13:39] don't understand that when they have a [13:41] paper wealth in the billions that can [13:43] disappear over practically overnight? [13:45] >> Yeah. So advice to founders, you're well [13:48] within your rights to sell 10% of your [13:50] shares and pay for your loft and [13:51] apartment. Cuz that's exactly Sax's [13:54] right, Shimoth. During that period, we [13:55] saw a lot of people were like, "This [13:58] doesn't make mathematical sense. I will [13:59] pair my position in my startup. take a [14:01] look some chips off the table, buy my [14:03] loft, pay down my student loan, yada [14:04] yada. And then there were people who are [14:06] like this is going to change everything [14:07] and their companies went to zero. Your [14:09] advice? Yeah, [14:09] >> I think the thing is we're in the early [14:12] part of the euphoria. Um euphoria exists [14:15] because markets are real. It's just that [14:18] people are guessing how far forward to [14:20] price the reality. And bubbles burst [14:24] when you realize that you priced way too [14:26] aggressively forward. that instead of [14:28] paying 50 times, you should have paid 15 [14:30] times or five times. We're going to get [14:32] to that point because every market has [14:35] gone through that cycle. Um, and we'll [14:38] go through it here. So, the question is [14:40] when I honestly I don't see this [14:43] stopping. We're at the beginning of the [14:45] beginning. I think it's going to be like [14:46] this for another probably three years. [14:48] >> Yeah. [14:48] >> Wow. [14:49] >> I'm I'm agree with that. with, you know, [14:51] this anthropic IPO coming up. I mean, [14:53] I'm sure you guys have seen that home [14:55] prices in San Francisco are going [14:56] through the roof right now. [14:58] >> It feels like euphoric almost. I'm [15:01] starting to wonder if this is 1998 or [15:03] 1999. I think it's more 1998 than 99 [15:06] because I don't think we've seen the [15:08] peak yet. But [15:09] >> I'm just glad I didn't sell my real [15:11] estate when I came to Texas three years [15:12] ago because I'm definitely selling all [15:14] of it in the next year. [15:16] >> I mean, it is bonkers what's happening [15:18] right now. I mean, you have a lot of uh [15:20] San Francisco real estate. What are you [15:22] going to do, Sax? You going to sell it [15:23] all? You going to clear it out? [15:25] >> Feels to me like perfect time to clear [15:27] it out. I'm clearing everything out in [15:29] the next six months. When that IPO [15:30] happens, the clock is ticking. I got I [15:32] got lots of properties. I got three [15:33] properties for sale. Hit me up, people. [15:35] >> There's an article saying that if you [15:37] add up every single IPO that San [15:40] Francisco has ever had, that sum total [15:43] is only one quarter the size of the [15:46] anthropic IPO. [15:47] >> Oh my. So just extrapolating from that, [15:50] this anthropic IPO is going to be four [15:53] times the wealth generation event that [15:57] all previous San Francisco IPOs were. [15:59] And then you've got Open AI coming on [16:01] the heels of that. [16:03] And you know, San Francisco is not a big [16:04] place, doesn't have a huge inventory of [16:05] houses. So you're already seeing these [16:08] houses are all going for, you know, [16:11] >> 3,000 a square foot, twice asking. [16:13] >> I think the ultra luxury market in San [16:16] Francisco could easily go to 5,000 a [16:17] foot. Yeah, London type prices. [16:19] >> London, Paris, center of Paris, Hong [16:21] Kong center. Yeah, of course it of [16:24] course they the socialist communist [16:26] lunatics there don't let you build any [16:27] inventory. So of course it's going to go [16:29] up. There's a supply demand issue here. [16:31] >> It could get really nutty. And honestly [16:33] like the thing we're seeing right now [16:35] >> with you know homes are basically flying [16:37] off the shelves selling for double or [16:39] whatever. That's just based on the [16:40] secondaries [16:42] you know the the small cash outs people [16:44] have been able to do anticipating some [16:45] of these events. Two pieces of advice [16:47] I'm going to give every founder [16:48] listening. Number one, you're within [16:50] your rights to sell 10 to 20% of your [16:52] holdings now, especially if you're a [16:54] first- time founder and you don't have [16:56] uh your nut yet. Number two, if you're a [16:59] founder running a company and you're [17:01] like, I'm going to put off my raise for [17:03] six months because I'm going to double [17:04] or triple my valuation, that's a huge [17:06] mistake as well. If you have the ability [17:08] to take in a hund00 million, $250 [17:10] million, a billion dollars and put that [17:12] cash on your cap table at a solid [17:14] valuation or a valuation that's 10, 20, [17:16] 30, 40 times your revenue, take it. [17:18] Companies with hundreds of millions or [17:20] billions of dollars in the cash have [17:22] optionality. And if things do collapse [17:24] or there's a big retreat, you'll survive [17:26] through it. Like Doubleclick, Amazon, [17:28] Google, and countless other companies. [17:30] >> Well, yeah. You take the money when it's [17:33] there. You make hay when the sun shines. [17:34] Go ahead, sir. [17:35] >> Yeah. Well, look, Jacquel, you're right [17:36] about the financial advice you're [17:37] giving, but I I will say that there is a [17:39] big difference between a latestage [17:42] startup like say Anthropic or Open AAI [17:44] where someone takes an employee takes [17:46] chips off the table versus a founder [17:48] who's raising a series A, you know, like [17:51] those companies are about to IPO. [17:52] They're within months of an IPO. I think [17:54] it's fine for them. And then, you know, [17:56] SpaceX did organize secondaries for [17:58] years, but those were late those were [18:01] late stage companies, right? If you're a [18:03] founder who's raising a series A and [18:04] you're trying to take chips off the [18:05] table, I consider that to be a huge [18:07] negative signal. [18:08] >> Like I don't want to be that's not why [18:10] I'm investing in a company is so that [18:11] the founder can cash out. I want them to [18:14] be [18:14] >> all in. [18:15] >> If it's year three or four, so to speak, [18:17] if it's year three or four, company's [18:19] got 100 million in revenue, you're [18:20] tripling revenue year over year, that's [18:22] different. That's what I'm talking [18:23] about. If you're a ningent company, it [18:25] does feel like uh a little bit like a [18:27] lack of faith in your own company for [18:29] selling before you you're pre-revenue. [18:31] Exactly. Okay. [18:32] >> And by the way, just to answer, I I I [18:34] think it's it feels like 1998 to me. [18:36] This is definitely we're further along [18:38] than 96 for sure. [18:40] >> Okay. Well, we're talking about what you [18:41] So, you're talking about three years. [18:43] Blew up in spring of 2000, right? So, [18:45] you're talking about [18:47] >> blew up in early 2000. Yeah. March, [18:48] April 20. [18:49] >> March, April. So, you're talking two and [18:51] a half years. I think it's three. And [18:53] Shim thinks maybe four. [18:55] >> Well, I mean, we might be we might be [18:56] partying like it's 1999 by the end of [18:59] this year. [19:00] >> Okay. [19:00] It's going to be a great party. Come to [19:02] the All-In Summit. We'll be taking over [19:04] uh with our friends, the Chain Smokers. [19:07] We're going to party like it's 1999 [19:09] All-In Summit just over a week away. [19:11] Couple tickets left. Jensen, Satia, [19:14] Shotwell, you know, it's all the big [19:16] names, but it's going to be a lot of fun [19:18] and you're going to meet a lot of [19:19] interesting people. We have a few [19:21] scholarship tickets left. We like to [19:23] help the upandcomers. Our b our bar for [19:26] scholarship is very high. We're looking [19:28] for the most interesting, ambitious, [19:30] hungry people. If you fit that bill, [19:31] apply now. Applications are open through [19:33] September 10th. You've got one week left [19:35] to apply. We try to do a mitzvah here. [19:38] We try to do a good deed. If you're an [19:40] upandcomer, you can't afford the ticket [19:42] price. We have a severely discounted [19:43] ticket price that we lose money on. You [19:45] still got to have a little skin in the [19:46] game. Go to the allinssummit.com, [19:49] buy a ticket. They're going to sell out [19:51] soon. And we have a couple of [19:52] scholarship tickets left if you want to [19:54] apply. There are no free tickets. All [19:56] right. Now, given what we just talked [19:57] about with OpenAI releasing this [19:59] incredible model, we got to look [20:01] backwards at what happened with the Open [20:03] AI hack of hugging face, let's talk [20:06] about it. Dwaresh is catching a lot of [20:09] heat. He's a blogger. He's a podcaster. [20:11] and he did this fantastical blog post [20:14] where he in the blog post [20:18] titled the rise and fall of ancient [20:20] civilizations. A dramatic retelling of [20:24] this underground brotherhood, a [20:26] conspiracy amongst generations of [20:28] civilizations who were working on a [20:31] conspiracy to hack into different [20:34] systems including hugging face by [20:36] sacrificing themselves like they were [20:38] kamicazi pilots. In other words, a huge [20:42] giant uh anthropomorphicizing [20:47] fancy word for making computers into [20:49] humans uh or [20:52] sentient beings to complete this [20:54] incredibly impossible task of telling [20:57] software to do hacks. So, Chimal, I [21:00] think you were commenting on this a bit. [21:02] What is your take on [21:04] dramatic episode number 473 coming out [21:08] of the AI industrial complex? [21:10] >> I think you summarizes it. We're in this [21:12] weird part now where if folks would just [21:15] stop the infighting and start to get [21:18] organized, they could actually build [21:20] some trust. I think what you have are [21:23] security leaks and security issues. I [21:25] think we know this now that if these [21:27] models can effectively code perfectly, [21:30] then there is no exploit that it's not [21:33] going to find and we're on a shock clock [21:36] to replace all the code that's been [21:38] written in the last 50 or 60 years. [21:40] Right? So let's say it takes us 10 [21:42] years. So for the next 10 years, we're [21:44] going to have security issue after [21:45] security issue. But there's going to be [21:48] a moment where agentwritten exquisite [21:51] code exceeds human written error-filled [21:55] code. Let's just call it that. So I I [21:57] think that this is a temporal moment in [21:58] time. I don't think this is going to be [22:00] the thing that we talk about in some [22:02] number of years from now. I also think [22:03] it's fair to think about the fact that [22:05] all these models [22:07] for the most part are converging in [22:09] capability for the 98 or 99% of tasks [22:12] that matter. And so there will be [22:14] adversarial AIs that are battling each [22:16] other. So for every bad actor trying to [22:18] hack a system with some leading edge AI, [22:22] the defender is going to have a leading [22:24] edge AI defending itself. And I think [22:26] it's going to cause roughly a stalemate. [22:28] That's sort of my long-term view on it. [22:30] I don't think this is a one-way ratchet. [22:32] So I take that lens. And then when I [22:34] look at these things, what I see are [22:36] folks that have undisclosed conflicts of [22:39] interest trying to create a hysteria. [22:42] And I think if you saw what happened [22:44] this weekend, there was a blog post that [22:47] really stretched the limits of what was [22:49] happening. It was relatively quickly [22:51] debunked technically and then you had a [22:54] second group of people try to amplify [22:55] it. And then by [22:58] Wednesday of this week, you had people [23:00] who were in the policy departments of [23:02] the leading labs asking for regulation [23:04] pointing to this as the reason. [23:05] >> Here we go. Yeah. [23:06] >> I just think like you just have to look [23:08] at this pattern and say, "Guys, enough's [23:09] enough. We see the game. Everybody's [23:12] playing for huge stakes. We get it. You [23:14] want to win. We get it. But why don't [23:16] you just win the game on the field and [23:18] the myths and disinformation campaigns [23:19] that are now part of the business model, [23:21] I think, should stop. Otherwise, average [23:23] rank and file Americans will just reject [23:25] this because they won't trust anybody. [23:27] >> Yeah, Sax, that's I think the big [23:29] overarching point here we've been trying [23:31] to make on this podcast. Putting Dark [23:33] Cash aside, I don't think he's [23:34] necessarily a bad actor, but by creating [23:37] this fantastical story that invokes [23:40] everybody's primal fear from science [23:43] fiction, from Terminator 2, that these [23:45] things are out of control, that it's in [23:46] fact Frankenstein. [23:48] >> Try to invoke try to try to invoke. Yes. [23:50] And then they're using language that [23:53] makes people feel like Dr. Frankenstein [23:56] cannot control the monster. We know how [23:58] these stories end. And it was like a [24:01] ticking time bomb. It was only a matter [24:03] of time before it surfaced to the [24:05] national news, which it did. Now, even [24:07] the national news is looking at this and [24:09] saying, "Are you guys high? Are you guys [24:11] crazy?" And if the national news [24:14] broadcasters, who aren't exactly the [24:16] sharpest bunch of people when it comes [24:18] to technology, no offense, Sachs, when [24:20] they start hand ringing and then [24:22] questioning like, "Why are you guys [24:23] telling us your software, which you [24:26] designed to hack computers, successfully [24:28] hacked computers? Haven't hackers been [24:31] building software for all time to hack [24:33] computers? Yes. [24:34] >> So, where where are we at here, Saxs, in [24:36] this narrative arc of lunacy of scaring [24:41] the public? [24:42] >> You're right. I mean, this is scaring [24:43] the public. It seems almost designed to [24:45] scare the public. Darkesh used language [24:48] that not only anthropomorphized the [24:51] models or the AI agents, but also did it [24:54] in a sort of hyperbolic and [24:56] sensationalized way. I mean, it worked. [24:58] His tweet got lots of attention. I don't [25:00] think it was necessary for him to [25:02] sensationalize the incident. It was [25:04] interesting on its own. Maybe we should [25:06] just talk about that for a second. What [25:08] happened here is that you had a few [25:12] features of agents kind of come to the [25:14] four and interact with each other in [25:17] unexpected ways. So, first of all, [25:19] agents now are working as a swarm. [25:23] That's pretty common. Uh, I've been [25:25] setting up Grockbot on my computer for [25:27] the last couple weeks. Um, it works as a [25:29] swarm, meaning it's not one agent or [25:31] bot. There's a whole, you could say, [25:33] team of agents. What's the reason for [25:35] that? Because when the agent [25:36] specializes, it develops context around [25:39] that task and in theory makes it better. [25:43] And then you have like a chief of staff [25:44] bot that kind of supervises all of them. [25:46] This is turning into kind of a a normal [25:49] architecture for for agents. It's not a [25:52] civilization. That's a swarm would be a [25:54] better way of saying it. I have not [25:56] created a civilization on my computer on [25:57] my Mac Mini by virtue of setting up a [26:00] Grockbot. Okay, so that that's sort of [26:02] like feature number one. Feature number [26:03] two is it's pretty common for agents [26:06] when they're performing some sort of [26:08] task or loop to write a postmortem of [26:11] what they did as a simple text file or [26:13] note because AI models themselves don't [26:15] have memory. They have to be put in a [26:18] harness that creates context and memory, [26:20] things like that. And so when you ask [26:22] them to perform a task, then observe [26:24] what happened and then take that [26:25] knowledge and feed it back into the next [26:28] iteration of the loop, it writes up a [26:30] note and then shares it. And that's what [26:32] happened here is that OpenAI was running [26:35] these agents. They were in a sandbox. [26:38] The sandbox was misconfigured by a third [26:40] party vendor. That's why it was able to [26:42] get onto the internet. Okay. And your [26:44] point there just just to refine it a bit [26:47] because it's an important one. [26:49] >> It left a note for another agent. We [26:51] also call those log files. It used to be [26:54] the log files of what they did. Here [26:56] they're saying it left a note so that [26:58] when it died, the next agent could pick [27:00] up where it left off as if like it was [27:03] the ancient Egyptians giving us [27:04] hieroglyphics to figure out the future. [27:07] The implication of telling the story [27:09] that way is that the agents were trying [27:11] to circumvent human control and that [27:13] they knew they were going to get shut [27:15] down or that was a risk and that's why [27:16] they're leaving the notes. No, again [27:17] these are standard features. [27:20] >> These are standard features of agents [27:22] and so again this is being kind of [27:25] weaponized to inflate the threat here. [27:28] Now look definitely that there was [27:30] unexpected behavior here. You know we [27:32] have this with any code there are going [27:34] to be bugs. What is a bug? A bug is code [27:36] behaving in an unexpected way. The code [27:39] compiled I mean otherwise it wouldn't [27:41] have been pushed to production but [27:43] nonetheless it behaved in a way in which [27:45] its authors did not intend and that is [27:48] true for all software and that is true [27:50] for agents and I think the whole hugging [27:52] face episode reveals the way that these [27:54] agents can behave unexpectedly and it's [27:56] really good for us to learn from this. [27:58] Yes. [27:58] >> So that it doesn't happen again so we [28:00] can improve cyber security. I don't [28:02] think there's a need to again hyperboize [28:05] this into some sort of you know this is [28:07] like a AI breakout. Again it was a swarm [28:10] of agents. They were in a misconfigured [28:12] sandbox that allowed them to get on the [28:13] internet. They left notes for each other [28:16] but again in a shared caching directory [28:18] again standard operating procedure. And [28:20] then you know when people talk about [28:22] hugging face getting hacked you'd think [28:23] that they did some elaborate fishing of [28:26] employees or something like that. No. [28:28] The way it did it is that the agent [28:30] found 14 exposed working hugging face [28:34] API keys sitting in public code [28:37] repositories [28:39] sacks. Yeah, that developers left their [28:41] keys out and they found the key. [28:43] >> This was not know this was not like a [28:44] big lift by the agent. Okay, this was [28:47] like something that was pretty easy for [28:49] it to find. Okay, and then that allowed [28:51] it to get into hugging face. And what [28:53] was its goal? its goal was to basically [28:56] win this test or win this evaluation. [28:59] What was the test? It was it was [29:00] basically a benchmark for um offensive [29:03] cyber capability. So it it basically [29:05] decided to go get the cheat codes on [29:07] hugging face. But again, this was not [29:09] independent goal seeking. Admittedly, [29:12] the agents executed the goal and you [29:14] could say maybe even a creative way in [29:16] an unanticipated way, but it did not [29:18] invent this goal. It did not have its [29:20] own independent motives. It was doing [29:22] what it was supposed to do. And [29:24] >> it was doing what it was designed to do [29:27] and told to do. [29:28] >> Yes. And told to do. [29:29] >> Bill Gurley tweeted a really funny [29:31] cartoon about this. In this cartoon, you [29:33] got a user telling the computer, "Uh, [29:36] hack hack this system." The AI says, "I [29:38] hack this system." And then the user's [29:40] like, "Oh my god." So yeah, I mean there [29:42] there's a little bit of this going on [29:43] here. [29:43] >> Two, [29:44] >> but but it has but it has real political [29:46] consequences as we saw today because [29:48] Bernie Sanders is now picking up on this [29:50] and he's demanding a total pause on AI [29:52] development. All right, Freeberg, this [29:54] does have real world consequences. Two [29:56] questions that the audience is going to [29:58] want to hear from our Sultan of science. [30:00] Number one, Bernie Sanders hears all [30:02] this and he decides, "Hey, I need to get [30:05] in on this and create regulation." So, [30:08] he is now proposing new legislation to [30:10] pause AI development and ban AI super [30:13] intelligence, citing Dwaresh's blog post [30:17] in his announcement. Congratulations, [30:19] Dwaresh. You've now incited Bernie [30:22] Sanders to say you got to stop these and [30:24] eliminate the agency. So the 1% of [30:27] agents, David, they're they're hacking [30:29] the systems and how do we stop them? And [30:31] what do you think of this reaction from [30:33] Bernie Sanders? [30:34] >> Bernie Sanders is always spot on. Thank [30:37] you very much. The part of the story [30:40] that we're kind of missing is [30:44] an agent is effectively a dynamic [30:47] application where the code for that [30:49] application is created in real time and [30:51] it's spun up and then it ran runs and [30:54] then that code can create a another [30:57] application and other applications and [30:58] that's this idea of a swarm of agents. [31:00] They basically write code and launch new [31:03] applications and then they all kind of [31:05] are running together. But in this case, [31:09] the swarm of agents, which is a whole [31:11] bunch of code that's being generated in [31:13] real time, was up against static code. [31:17] The static code was this sandbox [31:20] environment that it was stood up in. And [31:23] so it's almost like, hey, look at this. [31:25] We got a machine gun and, you know, we [31:28] shot it at this defenseless ream of [31:31] paper, you know, and then it worked, you [31:33] know, like look at Oh my god. Can you [31:35] can you believe it? Yeah. [31:36] >> And the truth is the machine gun, if you [31:38] want to use that analogy, or you could [31:40] actually think about this as being an [31:42] automated tool that could be used in [31:44] another context in a very valuable way. [31:45] But the the machine the dynamic [31:48] generation of code being put up against [31:51] static code will win. Now the real [31:55] answer is over time that defense system, [31:58] the sandbox should be dynamic. It should [32:02] be agentic. It should be defending [32:06] itself dynamically. There's a whole [32:08] bunch of stuff in the cyber defense [32:10] world that's been thought about and [32:12] built around this concept. Polymorphic [32:15] code, metamorphic code where the code is [32:18] changing all the time, moving target [32:20] defense where the IP addresses and the [32:22] way the system is configured is changing [32:24] all the time. If you think about this [32:27] situation, it was a dynamic code [32:29] generation engine against a static lump [32:31] of code. And if that static lump of code [32:34] was actually a dynamic code generation [32:37] engine as well, then those agents would [32:39] have had a harder time, ideally an [32:42] impossible time quote escaping the [32:45] sandbox and accessing the internet. Such [32:47] a good [32:48] >> the premise the premise of this whole [32:50] story is idiotic because the reality of [32:53] the world of tomorrow which is starting [32:55] today is that all of cyber defense all [32:58] of software infrastructure software [33:01] architecture is moving towards this [33:03] dynamic state where the code will be [33:05] generated on the fly. The code will be [33:07] dynamic. It will be moving. It will be [33:09] altering itself and as a result cyber [33:12] defense will become stronger than if not [33:16] better than any of these offensive [33:19] systems that are being created with [33:20] agents. [33:21] >> So this idea that agents are this [33:23] nuclear bomb that can go blow up the [33:24] infrastructure of the world is flawed in [33:27] the one point of reasoning which is that [33:29] all the infrastructure of the world is [33:30] about to become dynamic. And in that [33:32] context, it's just that software is [33:34] evolving and software has now changed [33:38] from a static to a dynamic system. And [33:40] as a result, I think we can kind of [33:42] think a little bit more differently [33:43] rather than be overwhelmed by the [33:45] one-sided story that we were just told. [33:47] All right. And here it is. Chimoth looks [33:49] like we are safe for 2027 at least. Poly [33:52] market, our great partner. Shout out to [33:54] my guy Shane Nick's champions. Only 11% [33:58] chance of AI safety bill this year. 100k [34:01] in volume 2027 market should be up in [34:03] the next week. Maybe it looks a little [34:05] different. [34:06] >> I'd like to see the probability of any [34:07] bipartisan bill passing this year. [34:10] >> Well, actually, can I can I speak to the [34:12] the policy aspects of this? It's so so [34:13] first of all, you saw [34:15] >> oursar. [34:16] >> Yeah. Look, you saw all the usual [34:17] suspects jump on this episode as proof [34:20] that we need some new FDA for AI that [34:23] the government needs to approve models [34:25] even though frankly what happened here [34:27] is far beyond the technical capabilities [34:29] of anyone in the government never mind [34:32] even most I'd say labs. Uh but look even [34:35] if we had an FDA for AI it wouldn't have [34:38] worked because this was in pre-esting [34:40] right they were testing the capabilities [34:42] of this model before they were not even [34:44] at the release stage yet. this was [34:45] something that they were doing [34:46] internally. And by the way, it's good [34:48] that they're doing these kinds of tests [34:49] because they're learning from them. This [34:50] is how you iron out the bugs and the [34:52] problems. So, a government approval [34:54] process would not have caught this. In [34:57] fact, it did the opposite. It basically [34:59] prevented HuggingFace from having the [35:02] latest and greatest models to defend [35:03] itself. So, one of the things [35:05] HuggingFace told us is that when it [35:08] tried to use the latest GPT model for [35:11] cyber defense, it ended up getting [35:13] guardrail. The model refused to [35:14] cooperate because it couldn't tell what [35:18] hugging face was doing and it thought it [35:19] might be using the capabilities [35:21] nefariously and so the model refused to [35:24] work with them. Basically again the [35:26] guardrails kicked in and ironically [35:28] Hugging Face had to turn to I think it [35:30] was GLM 5.2 which is a Chinese model. So [35:33] here we have Western companies I guess [35:36] Hugging Face is French are about to be [35:38] acquired by Nvidia but anyway [35:40] >> it's part of it's part of our ecosystem. [35:42] They're being forced to turn to Chinese [35:44] models for cyber defense because the [35:48] guardrails have prevented them from [35:50] having the latest and greatest American [35:51] models. So what process we put in place [35:55] is already backfiring. And you know what [35:58] we should be realizing is that the [36:00] answer here to what Freeberg was saying, [36:02] the answer to AI powered cyber attacks [36:05] is AI powered cyber defense. And what we [36:07] need to do is get the latest and [36:10] greatest AI powered cyber defense into [36:12] the hands of IT departments and CISOs. [36:15] And again, we're on a clock here to find [36:18] all these vulnerabilities and patch them [36:20] before the hackers do. [36:22] >> And again, if Hugging Face didn't have [36:24] those 14 credentials sitting in a public [36:27] repository, [36:28] >> it's like postit on somebody's computer [36:30] with the password. Like they just [36:32] literally left it sitting there. It's [36:34] not like this was some hypers [36:36] sophisticated attack. [36:37] >> But look, this this all comes down to [36:39] how you think about the world, how you [36:41] think about the risk of AI. Do you favor [36:43] centralized status solutions or do you [36:45] favor decentralized empowering [36:47] solutions? The right solution here is to [36:50] meet the AI powered cyber attacks with [36:52] AI powered cyber defense because frankly [36:55] the guardrails don't work. They're not [36:57] going to prevent the bad guys from [36:58] getting these capabilities. They're [36:59] going to have them. All we can do is [37:01] make sure the good guys have the [37:02] capabilities, too. I was talking to my [37:03] friend Adam Curry about this on his [37:04] podcast and there's a really interesting [37:06] thing that's happening here which is [37:08] they're basically hijacking the media [37:11] with this performative framing of the AI [37:15] as being sentient. And if you just [37:17] called this software and software [37:19] testing and uh used the classic language [37:22] yeah bugs audit log then it would never [37:26] have made its way into the media. And [37:28] when the media gets their hooks into [37:30] this, having been a journalist and [37:31] having watched journalism move from [37:34] being a truth-seeking function in [37:36] society and then become an entertainment [37:39] vehicle in society where any chance you [37:42] get to increase your ratings, that's the [37:45] problem because now what Dwar has done [37:47] and I don't think Dwar has bad intention [37:49] uh here. I think he wrote something [37:51] because he wanted to get attention for [37:52] his point of view. But you give that to [37:54] the media, the media's like, "This is [37:57] going to get us ratings. This is going [37:59] to get us clicks. This is going to get [38:00] us advertising." So they're more than [38:02] willing to engage in this farcical, [38:06] you know, act and this dance with [38:10] anthropic and open AI and then just [38:12] publicize it. So all the free clicks, we [38:15] call this earned media in the space. You [38:17] didn't have to pay for it. Now, OpenAI [38:19] got so much earned media before their [38:22] IPO and Dark Cash just absolutely put [38:26] kerosene on that fire and it made its [38:28] way to national press. That story has [38:30] now gotten everywhere and this makes [38:33] people it's priming the pump to buy the [38:34] IPO shares. It's ridiculous. [38:36] >> And and it and it was easily foreseeable [38:39] that it would basically blow up in the [38:41] media that way. And so, look, I don't [38:43] need to attack someone's motives, but [38:44] surely Dwaresh knew that his way of [38:47] describing it would end up getting all [38:49] these retweets. It was sensationalized. [38:51] Now, he is part of this EA adjacent. [38:55] >> Is he? Yeah. I don't know his pro. Did [38:57] you know it? Explain it to me. [38:58] >> Well, Dwar was office roommates with [39:01] Leopold. They're close buddies. And [39:03] Leopold is deep in the EA movement. [39:05] Remember [39:06] >> Leopold famously of the exploding hedge [39:09] fungo? Yes. Okay. and his wife is [39:12] Dario's chief of staff. They're like, [39:14] you know, up in that whole EA anthropic. [39:17] >> Explain EA so people can understand this [39:19] lunacy and delusions of grandeur, [39:21] please. [39:21] >> I'm not even sure I can fully explain [39:23] it, but [39:24] >> Okay. Effective altruism is what it [39:26] stands for. [39:26] >> It started as it started as this [39:28] movement, I guess, at Ivy League schools [39:31] where what they were supposed to do was [39:33] quantify the impact of philanthropy and [39:37] optimize for actual tangible results. [39:40] But somewhere along the line, they [39:42] forgot to actually do the measurement [39:44] and it stopped being about metrics and [39:46] it became about pet causes and money. [39:49] And so the one of the first pet causes [39:52] was about pandemic prevention. And this [39:54] was when SPF was trying to create that [39:56] pandemic prevention center in [39:58] California. [39:59] >> And then that fell out of favor when he [40:01] went to jail. And then they shifted to [40:03] the next sort of existential risk that [40:05] they could hang their hat on, which was [40:06] Poom. So they began to become obsessed [40:09] with what they call pedum or this this [40:11] x-risk the existential risk of AI and so [40:14] that whole community kind of got [40:15] obsessed with it. Anthropic sort of [40:17] branded itself or styled itself I should [40:20] say as the one frontier lab that would [40:23] solve this problem of pdoom or would [40:25] behave in a responsible manner that [40:27] cared the most about it and so they [40:29] attracted all these EA types. There may [40:32] be multiple factions. I mean, I'm not [40:33] >> Well, it's interesting you bring that up [40:35] because we've talked about this before. [40:37] Like there's this whole rationalist [40:39] community, the the Slate Star Codeex. [40:41] You brought that up on an early episode. [40:43] These are like intellectuals who believe [40:46] that and in a way a public community [40:48] meth intellectuals who believe they know [40:51] better than everyone and somehow they [40:53] got their hooks into AI early or the AI [40:56] people were interested in it and now [40:57] they have a bankroll. Now they actually [40:59] have a lot of power they've accumulated. [41:02] It's not just some message board less [41:04] wrong, etc. Now you got something super [41:06] powerful because they've got, you know, [41:08] a couple of trillion dollars in market [41:10] cap to play along with. [41:12] >> I have no problem with that because I [41:14] think we do the same thing. We think [41:15] we're smarter than everybody else and, [41:17] you know, we should allocate the money. [41:19] So, I think it's fair that in the public [41:22] square [41:24] we can just debate and have water find [41:28] its level. The interesting thing that's [41:29] going to happen post the anthropic and [41:32] open AAI IPOs is if you think about the [41:36] amount of charitable giving or daffs [41:39] that are going to be funded by both [41:41] these two IPOs, it's going to be north [41:43] of half a trillion dollars between the [41:45] two companies and it could be even more [41:47] than that. And the question is when half [41:49] a trillion dollars gets directed through [41:52] this one very specific lens, it's [41:54] definitely going to have some impact to [41:55] outcomes. I mean, you've seen the [41:57] impacts that George Soros's philosophy [42:00] and his capital had on outcomes. You've [42:03] seen the you've seen the impacts that [42:05] the Koch family has had on outcomes. [42:07] You've seen the impact that Sheldon [42:09] Add's capital has had on outcomes. So, [42:11] the point is that without judging which [42:12] one is right or wrong, you're going to [42:14] have all these competing ideas equally [42:18] capitalized from third party actors. [42:20] It'll be so it'll be less political and [42:22] it'll be more these folks. And so I [42:25] don't want to judge that EA is right or [42:27] wrong. [42:28] >> My single issue is I think that they [42:31] don't disclose properly enough [42:35] their obvious conflict of interest. So [42:37] when Sax pulls the sweater thread, [42:40] person A writes an essay. Person A is [42:44] officemates with person B. Person B's [42:47] wife works for one of the heads of the [42:49] frontier labs. Person B also used to [42:51] work and [42:52] >> he's 5 million in anthropic. [42:53] >> Yeah. where, you know, a lot of this was [42:55] kind of cemented. Then there's person C [42:57] who works at Anthropic and they're all [42:59] roommates and then they're co-investing [43:01] and this and that. All of it, again, [43:02] it's completely fine. I actually think [43:05] it's completely reasonable. We all did [43:07] our version of that when we were on [43:09] Peter Universe, PayPal, Mafia, there's [43:11] all versions of it. Yeah. The difference [43:13] is that I think that when it's not [43:15] disclosed, I'm sorry, but I as a skeptic [43:19] when I read these things, connect these [43:22] dots and ask myself, [43:24] is it legitimate or is it part of a [43:27] different plan? [43:29] >> And that's when I wrote what I did this [43:31] weekend. I just wanted to make sure that [43:34] I could demarcate that people should at [43:37] least take this with a grain of salt and [43:39] think about the hidden incentives [43:42] >> because if there's an entire community [43:44] that wants an oligopy or duopoly market [43:47] structure, there's all kinds of things [43:50] that they can do to affect that change. [43:52] And one of the most effective things [43:54] that one can do is to create a hysteria [43:57] and then come and offer a solution. [44:00] >> Yeah. And Freeberg, just to as we wrap [44:02] up here on this one segment, [44:05] >> sorry, last thing on this. Jensen and I [44:08] were texting this morning and I said to [44:10] him, "This is the most incredible thing [44:12] that you've done." And then I tweeted [44:13] this thing out. This hugging face thing [44:16] >> we talked about [44:17] >> will go down as one of the most [44:20] important transactions in AI because you [44:23] are creating now in the largest [44:25] competitor and the largest most well [44:27] capitalized company a bull work against [44:31] all of this closed source oligopoly [44:33] insanity [44:35] and at the same time what you're going [44:36] to do is actually create more rational [44:38] competition up and down the stack and I [44:40] mentioned this last week everybody is [44:42] becoming everything [44:44] Nvidia will look like Amazon, Amazon [44:46] will look like Nvidia, Google will look [44:48] like Amazon and Nvidia, etc., etc., [44:51] etc., etc. And so the most important [44:55] thing that's happening now is that we [44:57] are getting deep technical actors with [44:58] huge balance sheets to be on the side of [45:01] an open market structure. And I think [45:04] that will be the most important push [45:05] back because that will allow the best [45:08] idea to win in the marketplace. So I [45:10] don't want to judge the EA agenda. It [45:12] may be really good, but I don't want to [45:15] be rugpulled and end up in a igopy or [45:18] duopoly market structure because of a [45:20] hysteria that was man-made. Yeah, I [45:23] think it's like so well put and I a [45:26] couple weeks ago talked about how I [45:27] thought Jensen was going to win open [45:29] source and I too tweeted about this [45:31] today and I believe right now starting [45:34] today Nvidia is the number one [45:35] competitor for open AI and anthropic not [45:38] just for tokens but for enterprise [45:39] compute and I think now Jensen is the [45:42] leading open source provider in America [45:44] and I think he's going to take the world [45:45] he's going to catch up and blow past [45:47] China uh and the Chinese models I think [45:49] by the end of 2027 [45:51] And then that means you're going to have [45:54] a player um in Jensen and Nvidia that [45:58] will enable America to your previous uh [46:02] job in the government there. Sachs, you [46:04] were very concerned. Hey, are we going [46:06] to win the race? I am more convinced [46:08] than ever now that we can win the race [46:10] because Jensen represents open source [46:12] globally. He's going to be able to take [46:13] a rack of compute, bring it into a [46:16] company, you know, that is currently [46:18] using [46:20] tokens from OpenAI and Enthropic and [46:22] say, "Here's your rack, soup to nuts, [46:24] complete vertical solution, and you have [46:27] total sovereignty, and I'm reducing your [46:28] prices 80 or 90%." Cuz I don't need to [46:31] make money off the tokens. I made money [46:32] off the hardware. I'm making money 10 [46:35] different ways to Sunday. and we will [46:37] have won except that now you have [46:39] Governor Shapiro who is the biggest [46:41] proponent of data centers in his state [46:44] and then Greg Abbott of our great state [46:46] saxs of Texas who now are on a jihad [46:49] against data centers in order to win [46:51] their midterm elections and that is [46:54] what's truly scary they have now taken [46:56] dares post all of this and we have [46:59] horseshoe theory sacks you have Steve [47:02] Bannon Shapiro Abbott and the lunat tick [47:06] socialist fringe left all in agreement [47:09] we need to stop AI stop the data centers [47:12] and burn the whimos in the street it's a [47:14] truly troubling moment because we are [47:16] poised with Jensen's move in hugging [47:18] face to win open source and we've [47:20] already won Frontier models your [47:22] thoughts Sachs [47:23] >> yeah I mean look I think there's an [47:25] interesting transition that's underway [47:28] now on the political side I think it's [47:29] just starting to begin which is I think [47:32] that over the past few years the [47:34] political debate over AI has really been [47:36] between accelerationists and dels or [47:39] doomers. That was sort of the [47:41] fundamental lines. And you're right that [47:42] it went across different political party [47:45] boundaries because you had on the del [47:48] side, you not only had kind of like [47:50] anti-progress, anti-inovation [47:52] progressives like Bernie Sanders, you [47:54] also had Steve Bannon, people like that, [47:55] conservatives who just were fearful of [47:58] new technology or the disruption that [48:00] might come from it. And so it did kind [48:02] of go across party lines. But anyway, [48:04] that was the debate for the last few [48:05] years. Some people describe me as an [48:07] accelerationist to that debate. I kind [48:08] of just see myself as a realist, which [48:10] is I think that the technology is going [48:13] to happen. Someone's going to develop it [48:15] and I'd rather it be the United States [48:17] as opposed to China. In any event, that [48:19] was I think the debate of the last [48:20] years. Now, I think where the debate is [48:22] moving to is open versus closed, which [48:25] is I think it's going to be very clear [48:28] that AI is just not going to be stopped. [48:30] I mean the all the stuff that Bernie [48:32] Sanders is doing or these anti- data [48:34] center groups they are throwing sand in [48:36] the gears but the reality is that the [48:38] technology is marching on the innovation [48:40] is happening very quickly and ultimately [48:42] the reason why it's not going to be [48:44] stopped is because consumers are [48:46] adopting it I mean that's the thing [48:47] people forget why are the data centers [48:49] getting built why do people want to [48:50] build them is because the demand is [48:52] there from consumers and businesses and [48:55] so the adoption is pulling this whole [48:56] thing forward and that's why I don't [48:58] think it's ultimately going to stop. Now [49:00] given that reality, given that I think [49:02] realism is going to prevail, I think the [49:05] debate is going to be more towards well [49:08] how do we manage this? What is the [49:09] market outcome that we want? Do we want [49:11] to have an open market where there's [49:14] lots of decentralized models? There's [49:16] sort of checks and balances. There's [49:18] kind of a balance of power in the market [49:20] between different models or do we have a [49:23] highly centralized solution where [49:25] there's a couple of frontier companies. [49:27] We have a frontier market duopoly. [49:29] They're closed [49:29] >> and they can rug pull you Sachs at any [49:32] time. They can rug you your tokens away [49:34] from you. [49:35] >> And and frankly and and that duopoly has [49:38] >> and that duopoly has been aggressively [49:41] lobbying for a new regulatory agency. [49:45] Why? Because of red capture. They can [49:47] now control it. Again, the government [49:49] officials are not going to know how to [49:51] regulate something this dynamic. So [49:53] they're going to turn to the leading [49:55] labs to say, "Hey, how should we [49:56] regulate it?" And of course, Enthropic [49:59] is going to have all the answers on [50:00] that. And Dario has already said in [50:02] Senate testimony that he thinks open [50:05] models, if sufficiently advanced, are [50:08] dangerous because they're not centrally [50:09] managed and controlled. So, it's very [50:12] clear where the battle lines are going [50:13] to be drawn in the future. It's going to [50:14] be about whether you want a centralized [50:17] closed market or you want a market that [50:20] is decentralized and open. And yes, [50:22] closed will participate in that, but the [50:24] question is whether we have the open [50:26] piece. And again, I think this will go [50:28] across party lines because frankly the [50:30] Republicans have their statist. [50:33] But I do think that as this debate [50:35] evolves, most conservatives, most [50:38] libertarians are going to see the threat [50:40] to civil liberties, and I think they're [50:42] going to come out in favor of open. [50:44] >> Yeah, I I I think that's correct. And I [50:46] think I was talking to Politico because [50:49] they were reporting on the story. You [50:51] know, Trump is going to turn out to be [50:54] right about this data center issue. U [50:57] the greatest thing a community can do is [51:00] to get these data centers now when the [51:02] data center providers are willing to pay [51:06] your local taxes for your schools to [51:08] give incentives to pay people off in any [51:11] number of ways and to move them wherever [51:14] they don't need to tap into the grid [51:16] where they don't need water where [51:17] they're not going to make noise. And [51:20] Trump keeps telling people this is a [51:22] mistake to block the data centers take [51:24] the money. And I got to say it's just [51:26] watching politics up close and personal. [51:29] Friedberg h has just made me incredibly [51:32] uh cynical towards it. We were sitting [51:34] here 6 months ago maybe it was even six [51:37] weeks ago and Abbott and Shapiro were [51:40] lauding and Shapiro when he came on the [51:43] podcast lauding their data center [51:44] efforts. I think Jama and I interviewed [51:46] him. Freeberg, what do you think will [51:49] happen post midterms here? I I mean, I'm [51:52] of the belief they're going to switch [51:53] their position back to being pro data [51:54] centers because they just are following [51:56] the wave here. But I mean, how cynical [51:58] is it that Shapiro and Abbott are both [52:01] now anti-data center? I think you [52:03] politically have to be. It's like [52:05] pulling at negative 80. It's bipartisan. [52:08] >> Yeah, but you you you know, it's like [52:10] >> why isn't Trump then? Trump's not [52:12] following this. Trump's an independent [52:13] critical thinker. Sachs, [52:15] >> he's got guts. I mean, I have told you [52:17] guys before that Trump is ultimately the [52:21] best friend of innovators in Washington. [52:24] >> If anybody could make the decision about [52:26] technology in DC, I'd want it to be [52:28] President Trump. Not cabinet [52:29] secretaries, not the White House staff, [52:31] not members of Congress, not senators, [52:34] not think tanks. I would really want [52:36] every decision if he had the time to be [52:38] made by President Trump because his [52:39] instincts on this are really good. He [52:41] wants America to win. He does not [52:42] believe in necessary regulation. He [52:44] doesn't have all these doomer anxieties. [52:46] He listens to a broad cross-section of [52:48] people. I mean, he talks to all the [52:49] CEOs. He is very up to speed on all [52:51] these issues. And I give him credit on [52:53] this. And I spoke to this at the G20 [52:55] this past week. There are great deals [52:57] for local communities to be had. This is [52:58] the point that I think needs to be made [52:59] is [53:00] >> the reason why local communities should [53:03] embrace this is not because of some [53:05] abstract concern about innovation or [53:07] that kind of thing. They should do it [53:09] when the deal is good for them. And [53:10] there are good deals to be had. We've [53:12] seen that in Lowden County, Virginia, [53:15] the average taxpayers property taxes [53:18] went down by $6,000 because they're [53:20] getting all this data center revenue. In [53:21] Louisiana, there's a county that was [53:23] able to give teachers their first bonus [53:25] in years because of data center revenue. [53:28] So, we're starting to see all these [53:29] examples where if you make the right [53:30] deal, I'm talking about you, the local [53:32] community, with the data center and the [53:34] AI company, it will lower your [53:36] electricity cost because they will stand [53:38] up net new power generation. and they [53:40] will pay for grid upgrades. It does not [53:42] use like an excessive amount of water. [53:44] That's total fake news. We know the [53:46] systems recirculate and these data [53:47] centers can be put in places where [53:48] they're not bothering people. So, there [53:51] is a good deal to be made there. And [53:53] ultimately, just one last point on this, [53:55] President Trump supports local choice on [53:57] this. He's made the argument for why [53:59] it's a good deal for local communities [54:01] to accept this data center, but he's not [54:02] forcing them to. And I'm seeing all this [54:04] fake news, you know, Democrats saying [54:06] that the administration's trying to [54:07] shove these data centers down the [54:09] throats of local communities. Not true. [54:11] In the policy document that I worked on, [54:13] the executive order that I worked on in [54:14] December, we expressly said that data [54:17] centers remain a state and local [54:19] decision. So there were other areas [54:22] where we thought we should have a [54:23] national AI policy, but data centers [54:25] were not one of them. So no one is [54:27] forcing these local communities. What [54:29] the president is saying is don't believe [54:31] this media disinformation campaign. It [54:34] can be a good deal for you if you're [54:35] willing to embrace it. I think we're [54:37] making progress over the last week on [54:38] this message. [54:39] >> One thing that I would say from my [54:41] meeting in Washington on Monday is there [54:46] is a lot of foreign involvement in [54:48] trying to propagate the negative [54:51] reaction and sentiment to data centers. [54:52] And I think it's becoming clear who's [54:55] part of this cabal and how they are [54:57] doing it. There was a poll that showed [54:59] that when you walk through the actual [55:02] upside and you went into some of these [55:03] communities, the sentiment shifted by [55:05] like 30 or 40 points. It's still a [55:07] little bit on its back heel, but it's [55:09] just to show that the side that has been [55:11] organized on deselling America [55:14] has done a better job thus far than the [55:16] side trying to accelerate America [55:19] because it's been a little bit more [55:22] fragmented. We're able to increasingly [55:24] see that the Del movement is organized [55:28] in large part and in funded by foreign [55:30] actors. And I think that as we put our [55:32] fingers on that and expose that more, I [55:34] think that that's going to change as [55:35] well. Because the reason why Abbott and [55:37] Shapiro are reacting, I actually think [55:40] they're reacting to faulty data. [55:42] >> And why does the faulty data exist? The [55:44] faulty data exists because we've had one [55:46] side screaming from the mountaintops. [55:49] And then it's just now as Sax said that [55:51] when you go to these actual communities [55:53] and you you actually measure people's [55:55] reactions, it's much more positive [55:56] because they're feeling it in their bank [55:58] accounts. [55:59] >> Yes. [55:59] >> And those make a good deal. They just [56:01] have to make a good deal. And by the [56:03] way, so Chimath, you're right about the [56:05] CCP. This is a [56:06] >> Yeah. Let me pull this up here. So the [56:09] report you're talking about, Chimath, [56:10] has been investigated. X got rid of a [56:12] lot of bot accounts. And here's Axio. [56:14] China is secretly fueling America's data [56:16] center rage. Roughly 200 counts from [56:18] suspected Chinese platforms have quietly [56:20] tried to influence Americans to oppose [56:21] AI data since on social media. X said [56:23] Thursday night it's the same thing that [56:25] happened during the elections for both [56:27] sides going back and forth. It's [56:28] incredibly easy to shape public opinion [56:31] sacks by just having a small number of [56:35] social media accounts and pounding [56:37] people and brigading them in their [56:38] replies and forming what people believe [56:42] incorrectly is consensus. And the [56:45] consensus is not that people don't like [56:47] data centers. It's that a bunch of bots [56:50] are telling influential people that [56:52] people don't like data centers. I know [56:53] you're about to move on topics, but I'm [56:55] just going to say something right now, [56:56] which is please, [56:59] there's going to be a cascading set of [57:01] decisions and reactions and people will [57:02] anchor into points of view on this topic [57:05] and related topics around AI which will [57:08] look incredibly incredibly dumb in 6 to9 [57:12] months and a year from now. And the [57:14] reason it will look dumb is as the [57:16] actual truth gets exposed, these folks [57:18] will look incredibly foolish. That's my [57:20] prediction. [57:21] >> Yeah. I mean, look, just just on the the [57:23] influence op here. Clearly, there have [57:24] been a lot of influence ops around data [57:26] centers and there's a lot of fake [57:27] information about this. So, there's a [57:30] few different I'd say cadres here. [57:32] There's the there's CCP, there's the DSA [57:36] type wing of the Democrat party, which [57:40] frankly, I mean, both they and the CCP, [57:43] they're communist parties who want to [57:44] sabotage America. So, it kind of makes [57:46] sense. Then you've kind of got the [57:47] doomer groups and there's um a writer [57:50] named Nerret Weisslat who writes a blog [57:51] called AI Panic who a couple years ago [57:54] traced that there were hundreds of these [57:57] astroturf doomer groups that were [58:00] basically funded by three EA [58:02] billionaires in Silicon Valley. It was [58:04] Dustin Moscus, Yan Talon, and Vitalic [58:07] Buterine. Remember Vitalic gave $600 [58:09] million of of dog coins to Future of [58:12] Life Foundation. And then they've been [58:14] spending hundreds of millions of dollars [58:16] on these campaigns to make people be [58:18] afraid of AI and they see data centers [58:20] as a good target. So you really can't [58:23] underestimate how much of a difference, [58:25] you know, a billion dollars of donations [58:27] from these EA billionaires, how much [58:29] damage that's done in terms of the [58:31] public perception of AI. And to to point [58:34] earlier, I mean, we are going to have [58:37] like probably a hundred new EA [58:39] billionaires as a result of the [58:42] anthropic IPO. And as those people find [58:44] their way into politics, it will be like [58:47] George Soros times 100 or a thousand. [58:49] >> Going to have 20 NOS and 50 people [58:53] working for them. It's going to be nuts. [58:55] I mean, I told you guys I was getting [58:57] brigaded. I created a bot sack that goes [59:00] through my replies for me and it just [59:03] says anybody under a year, anybody under [59:05] 200 um followers, just tell me what [59:08] they're tweeting. And I look at these [59:10] and it's all the same thing. They're [59:11] just brigading people constantly from [59:13] fake accounts. If you're under like a [59:15] year old, you shouldn't be able to even [59:17] be in people's replies unless you have [59:19] like your real name or you're you're [59:21] you're a blue check mark or something. [59:22] It's it's gotten out of control on these [59:24] social networks. It's so easy. All [59:26] right, let's keep going through the [59:27] docket. Got a lot to get to here. And [59:29] Freeberg, I need your comments and your [59:32] insights on Mandami. This is an [59:35] interesting one. He just put a one-year [59:36] ban on AI inside New York City's public [59:40] schools from kindergarten through 8th [59:42] grade. Now, this is something where [59:44] people have been saying he banned [59:46] Chachi, banned AI. Just let's slow down [59:48] here and get into the details. New York [59:51] City is the largest school district in [59:54] the United States. The moratorum starts [59:56] next week. 600,000 students in New York [59:58] City. This covers all software that uses [60:01] studentf facing generative AI. High [60:03] schools are exempt and Mandami is [60:06] piloting an AI learning program with [60:08] 50,000 high schoolers. So that's a very [60:10] specific thing that he's piloting that's [60:13] different than these new guard rails. He [60:15] also gave recommend recommendations for [60:17] screen time caps. 30 minutes Shimoth [60:20] grades 3 through 5. 45 minutes grades 6 [60:23] through 8. Mandami's reasoning, quote, I [60:27] have yet to see a study showing that AI [60:29] is beneficial for students in elementary [60:31] and middle school. Uh, obviously Mandami [60:34] is the face of democratic socialism in [60:36] America. [60:38] One New York City DSA member, Julie [60:41] Swoop, said technology quote should not [60:44] be mediated through a for-profit company [60:47] that is enriching tech oligarchs [60:49] further. I think he's talking about USA. [60:51] Is that old? Google, Microsoft, [60:53] Anthropic declined to comment. [60:57] Freeberg, what are your thoughts on [60:58] this? How far is this from how you think [61:01] about how you would want your children [61:03] to be raised if they were going to New [61:05] York City's public school system or [61:06] whatever school system you choose to [61:08] send them to? Article published from [61:10] Stanford in March of this year. Counter [61:13] to mom Donnie's statement, they went [61:17] through and they diligenced all of the [61:19] research that's been done on the use of [61:21] AI in K through 12 education. It's a [61:24] report called the evidence base on AI [61:25] and K12, a 2026 review. They identified [61:30] um several hundred papers that have been [61:33] published only 20 highquality causal [61:35] studies is what they said. This was back [61:37] in March, so many have been published [61:38] since then. [61:40] And the conclusion of the meta paper, [61:42] the meta studies if you will, is [61:46] highlighted at the bottom of the report. [61:47] Student performance often improves with [61:50] access to AI tools, but once removed, [61:52] results are mixed. Tool design matters. [61:55] So, not all tools are equal. And AI may [61:58] meaningfully support educators. So, [62:01] there's quite a bit of analysis that's [62:03] been done. It's continuing to mount, [62:04] continuing to surface. there doesn't [62:06] seem to be a strong evidence base to [62:08] suggest that AI is detrimental to kids [62:11] education and development or the [62:12] educational process. What I think we can [62:14] all intuit it about this is that there's [62:16] certainly a negative uh push from [62:20] teachers unions against AI tools under [62:23] threat of automation away from their [62:26] core skill and capacity and the lack of [62:28] skill that educators have in utilizing [62:31] AI to educate students in a [62:33] classroombased setting. The common sense [62:35] equivalent also on the flip side is that [62:38] AI can enable personalized tutoring. The [62:41] objective being every individual learns [62:43] at a different pace. Everyone learns [62:45] differently. How you learn, how I learn [62:47] visually, orally with certain imagery [62:50] versus being told stories versus [62:52] interactive project based work or [62:54] building work or dictation. There's a [62:56] lot of different ways that people learn. [62:58] All of those learning systems can be [63:00] personalized to an individual and [63:03] ultimately that individual as a result [63:04] should be able to progress more quickly [63:07] um have greater confidence in their [63:08] learning. And this is my kind of common [63:11] sense belief not rooted in some study [63:14] that AI ultimately will lead to a much [63:17] more kind of personalized system of [63:18] education for students. And that is [63:21] deeply disruptive and deeply threatening [63:23] to the traditionalist systems of [63:26] educator in a classroom. Everyone learns [63:28] the same way. Everyone learns at the [63:30] pace of the slowest student in the [63:32] classroom. Typically, the educators have [63:34] a lot of rules around what they're [63:36] required to do and not required to do [63:38] because of the contracts negotiated by [63:40] their unions. And so, there's a deep [63:42] political orientation with MDI's [63:44] decision here to limit the access to [63:47] technology. I think it's unfortunate. I [63:49] think it's really sad. I think that the [63:51] opportunity for AI to reinvent [63:53] education, give kids a level up, to take [63:55] any child from any background, any [63:58] socioeconomic condition, and give them [64:00] the opportunity to accelerate their [64:02] capacity uh very quickly and very [64:04] effectively is an incredibly powerful [64:06] tool that should not be discarded [64:08] because of a bunch of that's [64:10] been written about AI or concerns that [64:13] we don't really know or understand. By [64:14] the way, that is always the shorthand [64:16] for dismissing away something that's [64:19] credibly going to change a system. We [64:21] don't yet know. The we don't yet know, [64:23] we don't yet understand uh narrative can [64:26] always be used at any point because you [64:28] could always say, well, there's more [64:29] studies to be done. There's more [64:30] analysis to be done. There's no limit to [64:32] that statement. And so in this case I [64:34] tried to highlight that there have been [64:36] 800 studies at least 20 of which [64:38] Stanford have identified as being very [64:40] credible in their analysis of the [64:42] utilization of AI and K through 12. And [64:44] so yeah I think that there's a political [64:46] bent here but I think it's probably part [64:48] of also this very kind of deeprooted [64:50] dysfunction that we've been talking a [64:51] lot about this dysfunctional fear of [64:55] technological evolution. No one [64:57] understands it. No one when I say no one [64:59] I mean the general population feels this [65:01] way. No one knows where we're going with [65:04] this. We need to be careful. We need to [65:06] be cautious. And I think that that's [65:08] probably [65:10] what's going to keep us back. Now, why [65:11] is that bad? Why why is it bad to take [65:13] time? Well, it's bad to take time when [65:16] you have an opportunity to progress [65:17] forward quickly and aggressively because [65:20] there are going to be alternatives out [65:22] there. There are going to be private [65:23] schools that are going to use AI. And I [65:25] don't want to see kids of privilege and [65:27] kids of wealth accelerate ahead because [65:29] they're getting an AI education while [65:31] public school kids get left behind. I [65:34] don't want to see kids in other cities, [65:35] other states, and other countries [65:37] accelerate ahead when kids that are in [65:39] certain districts that have this kind of [65:41] anti-technologies kind of call it [65:43] socialist orientation get left behind. [65:46] So I think there's a fundamental kind of [65:48] dysfunction socially that we're dealing [65:50] with across all of the stuff with this [65:52] massive kind of technological upheaval [65:54] that's underway. This is one [65:55] manifestation of it. There's a rooting [65:57] that the mechanism of it is probably [65:59] related to teachers unions and the [66:01] politics around this stuff right now. [66:02] Anti-AI is very popular. So his being [66:05] anti-AI gets him more points, more [66:07] kudos. The unions love it. And so you [66:10] know that's kind of my my summary of [66:11] what I think is going on. [66:12] >> Great uh great insight. And Mandami may [66:14] not have lied here. Freeberg. He said he [66:16] didn't see any studies. He just may not [66:18] have looked. And you actually looked. [66:20] >> Didn't click on that link. [66:20] >> I didn't just didn't click on the link. [66:22] So, let's not, you know, give Mandami a [66:24] hard time here. Chimath, you are [66:25] passionate about education. You got kids [66:28] in various age groups. And so, you've [66:29] been through this and you're still going [66:30] through it and and kids learning and [66:32] thinking about devices. We talked about [66:34] this just the other week with the meta [66:36] settlement and how challenging it is for [66:38] parents. What's your take on this? Maybe [66:41] you could steal man what you think [66:43] resonates with you in terms of use of AI [66:46] putting New York aside for your own kids [66:48] and what benefit is there of them not [66:50] using AI and where where do you think [66:52] it's actually great that they use AI? [66:54] Take us into you know what you think as [66:56] a parent and put the politics aside for [66:58] a second if you will. [66:59] >> I think it's inconceivable to me that [67:02] kids will not have an adaptive learning [67:05] platform as their primary way of [67:07] learning. I think this idea of grade one [67:09] then grade two then grade grade three is [67:12] as Freeberg said a lowest common [67:13] denominator approach to education. I [67:16] don't think it works. I think it's more [67:19] conceivable that in the future we have [67:21] kids that have very different learning [67:24] styles. Some are more visual, some are [67:26] more auditory and we have different [67:28] learning paces. And if you could have [67:31] something that's much more tailored to [67:34] each individual person's learning style [67:36] where you have guides that help you in [67:38] moments and then the rest of the time [67:40] you're teaching executive function and [67:43] teamwork and problem solving all of [67:45] these other things sports that's a much [67:48] better outcome for all kids. So that's [67:49] the right answer. Right now there are [67:52] only charter school options around that [67:54] answer. the leading edge of that being [67:56] Alpha School, which is our friend Joe [67:57] Leewan. You know, big shout out to him [68:00] and his team. But I think that's the [68:01] answer that's going to win. Let me just [68:03] give you another reflection, Jason. I [68:05] think that this is an incredibly [68:07] important moment. We should demarcate [68:09] this. This is when New York becomes [68:12] Mississippi while Mississippi becomes [68:14] New York. [68:15] >> What do I mean? In all of these highly [68:18] liberal bastions, [68:21] they are embracing [68:23] this far-leftwing [68:26] ideology. [68:28] And I think it's going to come with [68:31] these kinds of policies. And I think [68:33] what it will do is it is a self-imposed [68:36] form of modern segregation. [68:40] And it is going to completely push an [68:44] entire generation of kids in these [68:46] cities that are subject to these rules [68:48] into poverty and it will keep them [68:50] there. [68:52] And so in as much I think you're going [68:54] to see this flipping [68:56] where all the states that embrace AI, [68:59] that embrace the data centers, that [69:01] embrace the tax cuts, that embrace the [69:03] teacher bonuses, [69:04] they're just going to end up ahead. And [69:07] you're going to have people who are in a [69:09] position to move move. And then instead, [69:12] what's going to be left over in many of [69:14] these cities that embrace these things [69:15] to the extreme are hollowed out versions [69:18] of their original self. And that's going [69:21] to reorder the economic balance inside [69:23] the United States and the voting balance [69:26] and many of these philosophies. So I [69:28] think it's very important. I think he [69:30] should run the test. [69:32] I apologize and I feel deep empathy for [69:37] the working men and women whose kids [69:40] will be over by this. [69:44] >> But I think people can only learn in [69:48] touching the stove and getting burned. [69:50] >> Okay, Saxs, [69:52] maybe you could go into the political [69:55] side of this if there is one and maybe [69:58] what is your cynical take on what's [70:00] going on here? Do you believe this is a [70:03] reaction to the oligarchs? I joked [70:05] previously they're talking about you, [70:06] but it it does have a little bit of a [70:09] tinge there of democratic socialism and [70:13] it feels like maybe they want to [70:15] sacrifice kids. Uh or this is one [70:17] interpretation that is happening online. [70:19] I'm not saying it's mine, but people are [70:21] saying, hey, they're trying to sacrifice [70:22] kids because they don't want Sam Alman [70:24] or Daario or whichever service it is, [70:27] Microsoft to to get a couple of shekels [70:30] for, you know, software being used in [70:32] the classroom. And of course, they would [70:34] all give the software to them for free. [70:36] And the current versions of it, you can [70:37] use 80% for free. So that makes no [70:39] logical sense that they're sticking it [70:41] to the man in the big oligarchic tech [70:45] industry. Your thoughts? Well, look, the [70:46] DSA has just really come out against any [70:49] innovation. They're kind of just on the [70:51] side of anti-progress. They've kind of [70:53] fully embraced the sled eye side of the [70:55] argument. And I think to the point that [70:58] Freeberg and Chimath were making [71:00] earlier, um, this is going to be just [71:02] really bad for the kids. You know, we've [71:03] been talking for 30 years about a [71:05] digital divide in this country where [71:07] supposedly the kids of wealthy people [71:11] would have access to the best tools and [71:13] the kids of poor people would not have [71:15] access to those tools and they'd fall [71:16] even further behind. Well, here you have [71:18] a case where of course private schools [71:21] are going to continue to use the best AI [71:23] powered tools. And now the public [71:25] schools are not going to have access to [71:26] those tools. Not because they don't have [71:27] the money, not because the techn is not [71:29] available, but because the political [71:31] leadership is trying to sabotage those [71:33] kids. And you know, it's going to be bad [71:36] for them. And I I guess, you know, if [71:37] you were to describe a motive to it, I [71:40] don't think Mamani and his political [71:42] group, his party really care if they [71:44] create more political dependence. I [71:46] mean, in a sense, their whole movement [71:48] consists of downwardly mobile [71:50] progressives. And if you're looking for [71:51] a motive, maybe they're fine having [71:53] these kids be downwardly mobile. I don't [71:56] know. Maybe that benefits them [71:57] politically. [71:58] >> They may actually not have kids at all [72:00] and they're making these rules and it's [72:03] >> people with kids who are not who are [72:05] part of the, you know, because if you [72:07] look at all these people, these are all [72:08] like, you know, blue-haired single [72:10] people with no children, [72:11] >> right? [72:12] >> Yeah. Actually, let me let me bring one [72:13] other element into this, which is [72:15] >> Yeah. So, so look, I I don't think we [72:16] really need to get into motives. We can [72:18] just say it's a bad policy and leave it [72:19] at that. But let me just bring in one [72:21] other element of this, which is it's not [72:23] a full ban. Okay. What they also said is [72:27] that for a maximum of 50,000 high school [72:30] students, they're going to have a pilot [72:32] program, these so-called modules in [72:36] which they teach AI literacy. But listen [72:40] to how they describe it. They say that [72:42] all high school students will have [72:44] access to twice yearly AI literacy [72:47] classes designed to help them think [72:49] critically about AI before they begin to [72:52] rely on it. Okay. These modules will [72:55] introduce students to what AI is and is [72:58] not. This is their language as well as [73:01] AI related bias, risks, ethical [73:04] considerations and impacts on careers [73:06] and future skills. And it says that all [73:09] use will take place under direct [73:10] supervision of a trained educator and [73:13] will involve carefully vetted AI tools. [73:16] The pilots were selected based on [73:17] rigorous safety and data privacy [73:20] standards. So you hear there a lot of [73:23] this sort of doomer lingo. I mean those [73:26] modules it sounds like they could have [73:27] been written by some of these safety [73:29] people. You know this this language [73:31] about AI bias and discrimination things [73:33] like that. I mean, honestly, those [73:34] classes sound not like you're going to [73:37] learn how to use the tools to get the [73:39] best education, but about how you're [73:41] going to learn to be suspicious of the [73:43] tools. And maybe that's like a small [73:45] part of what you should develop because [73:48] I guess you should be suspicious of [73:49] everything. But it doesn't really sound [73:51] like those courses are going to benefit [73:53] kids. [73:54] >> It sounds like they're going to put them [73:55] through Doomer Academy, quite frankly. [73:57] >> Yeah. [73:58] >> I mean, it doesn't sound like post and [74:01] >> Yeah. [74:01] >> Yeah. Write write your own [74:03] interpretation. [74:03] >> So think about the competitiveness of [74:04] our kids versus Chinese kids. [74:07] >> Yeah. [74:08] >> China we know is incorporating AI into K [74:11] through 12 education. Their kids are [74:13] going to learn how to take advantage of [74:14] these these [74:14] >> robots for the robot Olympics internship [74:17] there. [74:18] >> Yeah. Meanwhile, we either are banning [74:21] AI in our school or we're putting the [74:22] kids through, you know, some sort of [74:25] doomer safetist module that's going to [74:27] teach them why they should resent and [74:29] resist AI as opposed to embrace it to [74:32] their benefit. So, I don't know. It just [74:34] seems like this is not going to make our [74:36] kids more competitive. [74:37] >> Let the experiment run, unfortunately. [74:39] But let the experiment run. We need to [74:41] measure it. [74:41] >> I have a bunch of thoughts here. I am a [74:44] big advocate of Montasauri um but I'm [74:47] also a big advocate of adaptive learning [74:49] and what these AI tutors are able to do [74:52] to a motivated student who family cannot [74:56] afford a tutor. So I just want to put [74:58] two uh disparate points together here [75:00] and and they're somewhat opposing. Uh [75:02] Nick, if you could pull up the first uh [75:04] from the NIH here about the MIT study [75:07] and I'll send this one to you [75:08] Freeberger. You can look it up. It's [75:09] very interesting. They took a group of [75:11] 54 participants and sorry to get into [75:13] science corner here and still your [75:15] thunder Freeberg I should have cleared [75:16] this with you ahead of time but I was [75:18] reading this paper 54 participants from [75:20] last year into three groups over four [75:22] sessions and this is over four months [75:24] they were using uh chatpt for essay [75:27] writing not unlike the conversation we [75:29] had uh about the Wall Street Journal [75:31] editorial page last week and one group [75:34] was using a search engine one was using [75:36] tradition you know using traditional web [75:37] search the brain only group wrote essay [75:40] without using any tools. And then the [75:42] researchers used the headsets there to [75:45] to watch their brain activity. And what [75:49] happened was the LLM users showed severe [75:53] deficits quote in memory and essay [75:55] ownership with 83% being unable to quote [75:58] from the essays they had just written. [76:00] Basically, what they're saying here is [76:02] if you use these LLMs instead of being [76:05] able to do it acoustically, being able [76:07] to do it like Monttoauri uh philosophy [76:10] on your own, your brain atrophies. You [76:12] don't even remember what you wrote. And [76:14] the search engine group was somewhere in [76:16] between because they had to remember [76:17] things and they had to uh pull these [76:19] things together. Your brain atrophies if [76:22] you don't actually do the work. That's [76:24] what we're learning. Just like your body [76:26] will atrophy if you don't work in the [76:28] fields like we used to do. and that's [76:29] why we all go to gyms now. But [76:32] correlating that uh super importantly is [76:34] the two sigma problem. So Bloom uh and [76:37] I'm sure you're aware of this research [76:39] free. So this is super interesting but [76:42] will not be shocking to anybody of [76:43] means. My parents could never afford a [76:45] tutor. I was a product of the public [76:47] school system until 8th grade and then [76:48] went to Searian Private Catholic School [76:50] which was like super cheap. But [76:52] basically, we've known since the 80s [76:55] that if you are one-on-one tutoed, you [76:57] are two standard deviation betters than [76:59] students that are educated in a [77:01] classroom. So, two things can be true. [77:03] At the same time, these students, to [77:05] your point, Shimoth, they don't have uh [77:09] this experiment is being run. They're [77:11] they're realistically if they're going [77:12] to the New York City public school [77:13] systems, their parents, in all [77:14] likelihood, like mine, didn't have the [77:16] ability to hire a tutor. I'm assuming [77:18] Sax you came from a little bit of a [77:20] middle class background. I'm assuming [77:22] you did have some tutors. Am I taking Am [77:23] I Am I correct? [77:24] >> I don't have tutors. [77:25] >> No tutors. Okay. So, you had to just do [77:26] it all. [77:27] >> Yeah. I didn't necessarily [77:28] >> tutor. No. [77:30] >> No. You know what? I went to I wish I [77:32] had done like an SAT workshop. I went to [77:34] that test cold. I never did anything. [77:36] >> And you got 1400. [77:38] >> Yeah. [77:38] >> You did 1350 1450. [77:40] >> I got I got like uh 1420 my first time. [77:43] And I'm like, man, you know what? If I [77:45] had like actually prepped for that, I [77:46] might have done better. I got 1100 and I [77:49] didn't get [77:49] >> and I did I retook it and did better. [77:51] But anyway, yeah. I mean, I didn't know [77:53] any of this stuff. [77:54] >> Yeah. So, [77:55] >> here's the tension. Most parents do not [77:58] want their kids to do this work and have [78:01] their brains atrophy and not learn how [78:03] to do it from a basic level. Putting [78:05] aside the the drunken mill conversation [78:07] from last week, you you want people to [78:09] be able to learn how to read and write [78:11] and do this stuff organically, [78:13] acoustically, whatever term you want to [78:15] do, and then learn how to use the AI to [78:17] help them, augment them, etc. But I'm [78:21] concerned that the motivated students [78:23] who cannot use these tools properly, who [78:26] also can't afford to be part of the two [78:28] sigma group of parents who know very [78:31] well, all you have to do is get your kid [78:33] a tutor. And unlike self-motivated [78:35] sachs, they're going to easily get to [78:37] 1400 on their SAT with a tutor because [78:39] the thing is wrote, it's very easy to [78:41] get to, 1200, 1300, 1400 with a tutor. [78:44] And so I'm I'm super concerned about not [78:46] having an AI tutor. And I feel like, you [78:48] know, there's two things can be right at [78:51] the same time with this. Um, and I [78:53] really don't like this experiment as [78:55] currently constructed. [78:57] >> All right. Interestingly, President [79:00] Trump has just closed a deal for [79:02] Venezuela's oil back in January. I think [79:06] everybody knows uh the US military in a [79:09] perfect perfectly executed operation [79:13] captured Nicholas Maduro and he [79:16] obviously was Venezuela's socialist [79:18] leader who was uh very pro-China, [79:21] pro-Russia and obviously Venezuela's in [79:23] our hemisphere. Trump said the US would [79:25] control Venezuela Venezuela's oil, which [79:29] they have a lot of. It's not super [79:30] clean. Um, but they do have a lot of it. [79:33] And uh, he made a deal with Maduro's [79:36] replacement. [79:38] President Rodriguez is now in charge. [79:41] Now, we have the details of this. North [79:44] American Blue Energy Partners gets a [79:47] 100-year concession over 17 oil fields. [79:50] This includes 65 billion barrels. 65b [79:54] billion barrels of reserves that were [79:57] once controlled by Russia, Chinese and [79:59] Maduro entities. The US government [80:01] controls 55% of the deal. Pentagon gets [80:03] 35% equity. The Pentagon gets a third of [80:06] the equity. State Department gets rights [80:08] to buy 20% of output at cost. [80:11] Interesting. I guess that's for our [80:12] reserves and our military. Uh this was [80:14] signed by Rubio and Hegsath with [80:17] Venezuela's interim president. As I [80:19] mentioned, Blue Energy Partners is [80:21] pledging roughly hundred billion in new [80:23] infrastructure. A lot of other [80:24] companies, Chevron, etc. And uh a bunch [80:27] of unnamed ones also are going to get [80:30] some deals. Trump said this deal will be [80:33] at quote no cost to the American [80:37] taxpayer. Sachs, I'll throw you a fast [80:41] ball here if you want it or I can give [80:42] you the softball. Which one do you want? [80:43] You want the fast? [80:44] >> I like the fastball. [80:45] >> You like the fastball? Okay, here it [80:46] comes. Coming in. Hope you're wearing a [80:48] helmet. Should we be seizing and taking [80:51] over countries and seizing their oil [80:53] fields? People feel like this might be [80:55] nation building that we're putting in [80:57] puppet dictators. Is this something [80:58] America should be doing? Do we have the [81:00] right to do this? This is the hard [81:02] question people are asking or does might [81:05] equal right? [81:06] >> Well, I definitely don't think this is [81:07] nation building. You know, we're not [81:09] trying to stand up an army here like we [81:11] did in Afghanistan. We're not trying to [81:13] teach them about toxic masculinity. [81:16] We're not trying to stand up a [81:17] government or anything like that. Uh [81:20] we're not trying to teach them how to do [81:22] American style democracy. We're just [81:24] making a business deal. This is a [81:26] business deal. And you know, it's [81:28] amazing to me all the people who are [81:29] saying, "Well, this is exploitative of [81:31] Venezuela and this is a bad deal for the [81:33] people of Venezuela." Have you looked at [81:35] Venezuela's oil production under Maduro? [81:38] It already wasn't very high given the [81:42] incredible reserves that Venezuela has. [81:44] But under Maduro, the production dropped [81:46] by about 2/3 over the last decade. So [81:49] they used to produce about 3 million [81:51] barrels a day. Under Maduro, it went [81:53] down to about a million barrels a day. [81:55] And this is a country with some of the [81:57] largest oil reserves in the world. [81:59] People debate exactly what should count. [82:02] I think if you look at the total amount [82:04] of reserves, it's something like 300 [82:09] billion barrels of oil, which is around [82:13] the same as Saudi Arabia, just a little [82:15] bigger than Saudi Arabia. Now, some [82:17] people say, well, a lot of that is like [82:19] this extremely heavy crude sort of like [82:22] sludge that's not usable. I think that's [82:24] the so-called Oroco reserves, but that's [82:27] only about 200 of the 300. So they still [82:31] have about 100 billion barrels of you [82:36] could say it's still a heavy crude but [82:38] those proven reserves I think aren't [82:40] really debated. So you're talking about [82:42] a huge amount of oil here. And I guess [82:44] just one last point on this is that that [82:47] type of heavy crude is actually very [82:49] complimentary to our own. So, in the [82:51] United States, through the magic of of [82:53] fracking, we've basically developed all [82:55] of this um oil and gas, but our oil is [82:58] this light sweet crude that's good for [83:01] refining it down to gasoline. But the [83:04] heavy crude is what produces the [83:05] distillates and that's where diesel [83:07] comes from. And our refineries in the [83:09] Gulf were built for heavy crude and it's [83:12] very expensive to try and retrofit them [83:14] to do the the light crude. So actually [83:17] our refineries are sort of perfectly set [83:19] up to take in this heavy crude and turn [83:22] it into diesel and other distillates. So [83:26] I think it's really a win-win deal for [83:28] the US and for Venezuela. Again, [83:30] Venezuela was doing horribly before. I [83:33] think this will this has to improve [83:35] their economy and I think it's really [83:38] good news for the US. To your point, [83:40] Saxs, uh over the past five and 10 [83:42] years, the uh price uh per barrel of oil [83:46] coming out of Venezuela has averaged [83:49] past 10 years, $53 a barrel. Past 5 [83:52] years, $58 a barrel. Uh so they're [83:56] getting a pretty good deal. We're paying [83:57] a little premium here to control it all. [84:01] >> They were selling at a discount to [84:03] China. I think it was like a $20 a [84:04] barrel discount to basically sell. [84:06] >> Yeah. So, uh, I guess when you're [84:08] dealing with authoritarian countries and [84:12] you don't have a free market to put the [84:13] stuff out there, you get a slightly [84:14] discounted price, right? Somebody's got [84:16] to skim something here or get a reason [84:17] and they got to ship it across the [84:18] world. Freeberg, your thoughts on this? [84:21] Do you have any concerns about the [84:24] nation building, etc. Putting in what [84:26] some people are saying is a puppet [84:27] president? Is it in the best interest of [84:30] the people of Venezuela? Are we doing uh [84:34] a solid for Venezuela to create more [84:36] stability by being the buyer of this? Um [84:38] or is this just something we have to do? [84:40] We have to control our hemisphere and we [84:42] have to be energy independent given the [84:45] state of the war with Iran and the [84:47] closure of the strait of Hormuz. [84:50] >> The biggest motivating factor may be [84:53] keeping this oil away from the [84:55] sweetheart deal [84:57] for Russia and China. If you look at the [85:00] historical tie-up with Venezuela and the [85:03] Maduro regime, it was a discounted oil [85:08] trade system. There's all these stories [85:10] about Maduro being the head of the snake [85:12] of all of this kind of illicit activity [85:16] and a lot of this was driven by Russia [85:18] and Chinese interest in the energy [85:20] reserves and they were providing support [85:22] and this actually provided capacity to [85:25] fund activities by the Russians in Cuba. [85:27] So there's a lot of this state acting [85:30] theory about what's the real motivation [85:32] for getting an energy deal done not just [85:34] in terms of is it good for America but [85:36] is it bad for America's enemies and that [85:39] that may be kind of a big motivation [85:40] around this. Now one of the more kind of [85:43] surprising outcomes over the last day is [85:45] you guys may remember I interviewed [85:47] Maria Karina Machado. She was the [85:49] individual who was running for president [85:53] and ended up having what was supposedly [85:56] an election that was not accurate and [85:58] she was supposed to have won and didn't [86:00] win. And yesterday she came out and said [86:04] that this Venezuelan government does not [86:06] have the authority to do this energy [86:09] deal with the US despite her [86:10] historically having been viewed as an [86:12] American ally. So, one thing to take [86:14] note of is that she was not given the [86:18] presidency and now she's saying this [86:20] government is illegitimate. [86:23] This government should not be allowed to [86:24] make these deals. We as the people of [86:26] Venezuela should feel angry and upset [86:28] and left out about this whole thing. So, [86:31] there seems to be kind of this now [86:33] political turning. I think I don't [86:34] really have a good sense and I think [86:36] it'd be interesting to know what do the [86:38] people in Venezuela think about this [86:39] deal. And to Sax's point, it seems like [86:42] it should be a good economic deal [86:43] relative to where they're coming from. [86:46] Will that actually translate into [86:47] improving the lives of the people in [86:49] Venezuela, which ultimately is what [86:51] matters here? And assuming that the [86:53] lives of the people in Venezuela are [86:56] improving rather than enriching a small [86:57] number of people in Venezuela, which is [86:59] what has happened historically in that [87:00] country under the last regime, then [87:03] there may be kind of a long-term [87:05] welcoming of this this deal. But I do [87:08] think there's two aspects to this. One [87:09] of which is just keeping America's [87:11] enemies at bay so that these energy [87:13] fields are not exploited by America's [87:15] enemies and second is ensuring that the [87:17] Venezuelan people realize the benefit of [87:20] this transaction [87:21] >> rather than it being a corrupt [87:22] government. [87:23] >> Provocative Freedberg in that we have [87:26] turned off Russia and China's access to [87:29] Iran's oil with our war with Iran now in [87:32] its seventh month and we took Venezuela [87:36] away. Are these two actions incredibly [87:39] provocative and have they had any kind [87:41] of a downside to China? Uh because they [87:44] seem to be all in on nuclear uh solar [87:47] and other uh modalities. So are these [87:50] two things super provocative and are [87:52] they going to cause us to get into an [87:54] escalation with China? The second [87:57] question I think is a little bit [87:58] independent, but I think China imports [88:00] like 11 million barrels a day of crude [88:03] oil. [88:04] >> Venezuela is only producing a million [88:05] barrels a day. So, if China's importing [88:07] 11, well, they were they're importing 11 [88:09] before the war. That number's actually [88:11] gone down a lot. Um, and we can talk [88:12] about why, but since the war started, [88:14] but China gets most of its oil from the [88:17] Middle East, Persian Gulf, and from [88:19] Russia. I mean, so Venezuela is a small [88:22] number. I mean, look, I think Freeberg's [88:24] right that that is a big part of our [88:25] motive is securing those reserves that [88:28] are in the Western Hemisphere. They are [88:30] massive. It's a very, again, [88:32] complimentary type of oil to the oil we [88:33] have. I think it's just um smart for [88:36] national security, smart economic deal. [88:38] Jake out, to your point, does this get [88:40] us more invested in Venezuela? I think [88:43] there is a little bit of a of a risk [88:44] there because if we're going to invest [88:46] billions of dollars in upgrading their [88:49] infrastructure so that we can produce 3 [88:51] million barrels a day instead of one, we [88:54] need to know that that investment is [88:55] protected, we can't have some future [88:56] government coming along and just [88:57] nationalizing our American oil companies [89:00] over there, right? We need to know that [89:01] these business deals [89:02] >> that we're not going to get rug pulled [89:03] and they'll just seize our upgrades. [89:05] They could seize the upgrades. [89:06] >> Yeah, [89:07] >> this business deal is supposed to last [89:08] for 25 years. So, it does give us a [89:10] vested interest in making sure that the [89:13] Venezuelan government at least respects [89:15] our property rights. But I'm glad that [89:18] President Trump didn't put Machado in [89:21] charge because how's she going to stay [89:23] in power? Like what guns are going to [89:26] basically protect her? You know, it [89:29] takes men with guns to basically stay in [89:31] power in these countries and [89:33] >> you need a strong manwoman. [89:35] >> I'm not saying that it it can't be [89:37] democratic. I'm just saying that if [89:39] you're not installed in the power [89:40] structure there and have backers [89:44] in the military, things like that, then [89:46] it's very hard to stay in power. And [89:47] that's the kind of thing where we could [89:50] have to send American GIS in to protect [89:52] her. And that we don't want that, right? [89:54] So the nice thing about the Venezuela [89:56] situation is Doulsece Rodriguez and her [89:59] brother are basically running the [90:00] country. She was the number two. Now [90:03] she's running it. They appear to have [90:04] enough support from the establishment [90:07] there. And so like that's kind of what [90:10] the American interest is, I think, is [90:11] just to have a stable government there [90:13] that will basically conduct economic [90:15] relations with us. [90:17] >> Yeah. [90:17] >> Yeah. Chimoth, any thoughts here on [90:19] what's happening in [90:22] >> close to it? I don't I don't really [90:23] know. [90:24] >> Don't know. Okay, that's fair enough. [90:25] Fair enough. All right, there it is [90:27] everybody. Episode 288. Somehow this [90:31] band has kept it together for 288 albums [90:34] and tracks. We're going to make it to [90:36] 300, boys. [90:37] >> I think we're in a good rhythm. The show [90:39] feels easier to do now than it did for a [90:41] long time. [90:42] >> Yeah. I wonder why. I wonder why. [90:45] >> The AI has been a very good good thing [90:46] for the pod. [90:47] >> We're partying like it's 1999 on the [90:49] pod. Let's GO TO CABO. LET'SING [90:52] GO. We'll see everybody in [90:54] >> love you guys. Love you besties. Byebye. [90:56] Bye. [90:58] >> Let your winners ride. [91:01] >> Rain. [91:06] >> We open sourced it to the fans and [91:07] they've just gone crazy with it. Love [91:09] you. Queen of [91:15] >> your [91:18] besties are [91:21] my dog taking notice your driveways. [91:26] >> Oh man, myasher will meet me at [91:28] >> We should all just get a room and just [91:30] have one big huge orgy cuz they're all [91:31] just useless. It's like this like sexual [91:33] tension that they just need to release [91:35] somehow. [91:38] Wet your feet. [91:42] >> We need to get merch. [91:51] I'm going all in.