Anthropic Fable 5 Safeguards Draw Tech Industry Backlash Over Data Retention

Anthropic Fable 5安全防护措施因数据留存引发科技界反弹 闭源AI实验室隐秘降级模型并强制监控提示词，正促使西方初创企业转向中国开源基础设施。 主持人深入探讨了Anthropic面临的巨大开发者反弹，此前有爆料称该公司秘密对用户模型进行降级并留存提示词数据，他们认为这凸显了企业试图针对开源AI实施监管俘获的更广泛动向。他们还激烈辩论了参议员Bernie Sanders提议设立的AI主权财富基金，将精英阶层关于AI摧毁就业的叙事与表明生产力激增的强劲经济数据进行了对比。此外，嘉宾小组分析了大幅攀升的CPI和PPI通胀指标，并在最后剖析了加利福尼亚州宽松的邮寄投票及选票代收集法律，他们称这些法律侵蚀了公众信任并扭曲了民主选举结果。 Anthropic在争议中发布了Fable 5模型，其隐秘的安全措施被指会悄然对从事竞争性AI研究或芯片设计的用户进行降级。 (0:00) 隐秘的模型评估机制给使用专有闭源AI平台的技术企业客户带来了运营、可预测性以及单点故障方面的风险。 (2:11) 西方AI对提示词的限制和安全护栏，正促使本土科技初创企业转而利用来自中国的高性能开源模型。 (4:47) Anthropic首席执行官Dario Amodei呼吁建立类似FAA的AI监管机构，此举遭遇批评，被指是针对开源架构的反竞争手段。 (9:10) 顶尖AI数据中心所需的巨额基础设施资本支出形成了巨大的护城河，目前1吉瓦规模的建设成本已逼近1000亿美元。 (17:14)

The All-In Podcast hosts argue that aggressive corporate safety guardrails and mandatory data surveillance by proprietary AI labs threaten domestic innovation, while examining broader policy tensions across public AI wealth proposals, resurgent inflation, and California's election administration. Anthropic's Safety Guardrails and the Open-Source Backlash (0:00) Anthropic's release of its high-performing Fable 5 model sparked significant developer backlash over terms mandating 30-day prompt storage and automated system downgrades for users conducting frontier AI research or probing sensitive subjects. Chamath Palihapitiya warns that intercepting and profiling user prompts creates unacceptable operational risks for enterprises, who risk sudden operational cutoffs if molecular or proprietary queries trip corporate guardrails. Sacks characterizes this regime as an attempt at regulatory capture aimed at crippling decentralized open-source models that cannot comply with centralized policing. David Friedberg reports that biological safeguards have already impaired legitimate scientific work at his genomics company, Ohalo, which previously utilized Anthropic models for plant phenotype prediction and gene editing. Because Western models increasingly block basic agricultural genetics, Friedberg notes his firm is turning to open-source alternatives, warning that the most capable open-source models currently originate in China. The hosts argue that dual-use technologies cannot be suppressed without ceding technological leadership, asserting that safety should be enforced on tangible downstream outputs—such as nucleic acid synthesis screening—rather than upstream computational gatekeeping. Diverging Economic Visions for AI Wealth and Labor (47:26) Political and economic friction over AI ownership escalated with Senator Bernie Sanders's proposal of the American AI Sovereign Wealth Fund Act, which seeks a one-time 50% equity levy on major AI firms like OpenAI, Anthropic, and xAI. While David Sacks denounces direct equity confiscation as an unconstitutional precedent, he acknowledges that AI executives fueled public hostility by claiming models will eliminate jobs while training on collective human output. Friedberg counters Sanders's plan by proposing that the existing $4 trillion Social Security trust fund transition from Treasury paper into an actively managed sovereign wealth fund holding public equities in foundational technology enterprises. The hosts strongly contest forecasts of widespread AI-driven unemployment, such as Dario Amodei's prior projection of a 50% reduction in entry-level knowledge jobs. Friedberg contends that AI acts primarily on the corporate revenue side, multiplying individual engineering capacity and spurring net hiring. However, Chamath emphasizes a severe macroeconomic difference between AI and the early internet: unlike zero-marginal-cost web distribution, frontier AI demands enormous operational compute costs, with modern data center infrastructure running up to an estimated $100 billion per gigawatt. Resurgent Inflation and Escalating Interest Rate Risks (65:40) May macroeconomic data indicated renewed pricing pressure, with the Consumer Price Index (CPI) rising to 4.2% year-over-year—its highest point since April 2023—and the Producer Price Index (PPI) jumping to 6.5%. In response, Polymarket betting odds priced a 49% probability of a Federal Reserve rate hike within the year, up from under 10% prior to hostilities involving Iran, alongside a 21% chance of inflation reaching 5% by 2026. Concurrently, the European Central Bank executed its first quarter-point rate increase since September 2023. Friedberg attributes the inflation spike to a temporary energy disruption stemming from the Iran war combined with unchecked federal spending, warning that overnight interest rates could climb past 5.5% or 6%. He criticizes fiscal policymakers for continually expanding public expenditures to secure political support, arguing that structural deficits remain the primary underlying engine of wealth inequality and sustained inflation. California Voting Frameworks and Mail Ballot Anomalies (72:27) The hosts scrutinize vote-counting trends in the Los Angeles mayoral primary, highlighting substantial divergence across voting methods. Day-of in-person voting favored Spencer Pratt at 35% compared to Nithya Raman at 26%, whereas late-arriving mail-in ballots saw Raman's tally climb to 37% while Pratt dropped to 19%. Friedberg links these shifts to legislative measures implemented under Assembly Bill 1921, which legalized unlimited third-party ballot harvesting, alongside universal mail-in ballot distribution and relaxed registration rules. The panel criticizes the systemic risks created by mailing 23 to 24 million ballots when only approximately 9.5 million citizens vote, leaving an estimated 14 million unvetted ballots circulating among outdated voter rolls. While recognizing alternative explanations—such as candidate Raman possessing an organized field operation or tactical late voting—hosts argue that low signature-verification thresholds and untracked ballot custody undermine election integrity. They advocate for federal legislation requiring proof of citizenship and government-issued photo identification to restore institutional credibility. The source leaves unresolved whether public pressure will force major AI developers to abandon prompt profiling and silent downgrades, whether central banks will resume monetary tightening to tame renewed energy and fiscal inflation, and whether federal courts or Congress will intervene to standardize state-level balloting rules.

All-In 播客主持团队重点探讨了 Anthropic 最新模型严苛安全审查引发的“监管俘获”指控，批评闭源科技巨头借安全之名阻碍科研并推高壁垒；同时，节目深入剖析了由伊朗战争与财政扩张推升的通胀压力、围绕 AI 资产征税的财富分配辩论，以及加州初选计票中暴露的选票收割与程序漏洞。 Anthropic 审查争议与监管俘获指控 (0:00) Anthropic 推出了新模型 Fable 5，每 token 价格高达 Opus 4.8 的两倍。其 319 页文档中规定的隐藏隐私与反竞争条款引发开发者强烈反弹：平台强制要求所有用户提示词数据至少存储 30 天；若算法检测到用户正在利用该模型研发竞争性前沿大语言模型，便会在不发出通知的情况下将用户静默降级至受限版本。Chamath Palihapitiya 警告称，这种在生成结果前拦截并评判提示词的系统，为企业客户带来了不可预测的单点运营风险，可能因误判而骤然切断核心业务能力。 David Sacks 与主持团队批评 Anthropic 表面以防范网络攻击与生物武器为名，实质是在推行复杂的“监管俘获”，试图游说政府出台开源架构在技术上无法合规的法规来遏制竞争。在录制现场，主持人输入关于化肥炸药监管的历史法规，尽管模型底层思考元数据判定其为合法学术研究，系统仍立即触发了降级惩罚。Sacks 指出，包括分析师 Ben Thompson 询问 GLP-1 药物致癌风险、普通用户查询线粒体等正常操作均导致账户被降级，暴露出平台缺乏真正的实名认证（KYC）通道，仅凭信用卡与邮箱便施加不透明的单方面监控与审查。 算力资本壁垒与开源地缘科技竞争 (4:47) David Friedberg 以其农业基因组公司 Ohalo 的经历说明，Anthropic 近期以防控生物武器风险为由收紧了科学研究权限，致使其团队无法继续利用该模型预测植物表型与设计基因编辑 RNA 导向，被迫转向本地运行的开源模型。他发出警示：目前全球最顶尖的开源模型大多来自中国，美国在开源领域已落后；西方过激的封闭限制正在无意中将初创企业推向外国技术基础设施。与此同时，由 Ark Institute 开发并由 Collison 兄弟资助的开源基因组语言模型等案例证明，社区资助的开源生态正成为抗衡封闭模型的重要支柱。 算力基础设施的资金门槛进一步加剧了这一困境。节目披露，两年前在亚利桑那州开发 2 吉瓦数据中心的单吉瓦成本约为 40 亿至 50 亿美元，但受 AI 供应链通胀影响，目前单吉瓦成本已飙升至 1000 亿美元；若要为开源社区聚合 3 吉瓦算力，资金需求高达 3000 亿美元。主持团队将当前局势类比为曼哈顿计划中的核裂变发现，认为 AI 具有天然的双重用途，技术能力无法被强行扼杀。他们主张监管应遵循既有的产品责任法和刑法，对下游违法实体结果（如对合成 DNA 订单进行核酸筛查）进行事后追责，而非在源头由少数私营寡头垄断把守访问权限。 AI 财富分配争议与主权财富基金构想 (51:31) 参议员 Bernie Sanders 提出《美国人工智能主权财富基金法案》，主张对 OpenAI、Anthropic、xAI 等主要 AI 企业一次性征收 50% 的股权税（非利润税）以建立公共基金。Sacks 批评该提案是直白的财产没收，但坦言公众不满情绪源于 AI 领袖的言行矛盾：高管们一边频繁预警技术将导致大规模失业，一边使用人类免费知识训练模型。主持团队梳理了行业对就业影响的分歧：Elon Musk 设想人类劳动完全可选的高收入乌托邦，但指出这可能需要上百年；Dario Amodei 此前声称初级白领岗位将在 1 至 5 年内减少 50% 的预测已被现实数据证伪；Sam Altman 亦收回了早期关于失业潮的断言。 Friedberg 强调，AI 通过成百上千倍提升工程师生产力来推动企业拓展产品线与营收，实际运行中反而会刺激招聘，而非导致失业。他提议对持有 4 万亿美元国债借据且濒临破产的社会保障信托基金进行现代化改造，借鉴加拿大或澳大利亚模式转型为主权财富基金，允许公民通过个人账户合法投资美国前沿 AI 公司，通过市场化手段让全民共享科技增长红利。 通胀抬头与地缘冲突带来的宏观压力 (65:40) 美国 5 月宏观数据显现通胀重燃迹象：CPI 同比上涨 4.2%，创 2023 年 4 月以来新高；PPI 同比上涨 6.5%，创 2022 年底以来最高水平。Polymarket 预测市场显示，2026 年通胀达到 5% 的概率升至 21%，美联储年内加息的概率则从伊朗战争前的不足 10% 飙升至 49%；欧洲央行亦于周四宣布加息 25 个基点，为 2023 年 9 月以来首次。 Friedberg 严厉批评了现行货币与财政政策，指出通胀反弹是伊朗战争带来的能源短期冲击与国会无节制财政支出共同作用的结果。他警告在潜在的紧缩政策下，隔夜利率可能被推高至 5.5% 甚至 6% 以上，政客通过挥霍未来收益换取选票的财政赤字扩张正是加剧贫富分化与削弱经济根基的核心推手。 加州选举制度漏洞与选票收割争议 (72:27) 节目深入剖析了洛杉矶市长初选计票中出现的统计异常：现场投票中 Spencer Pratt（35%）领先 Karen Bass（29%）与 Nithya Raman（26%），但选举日后收到的迟到邮寄选票中，Raman 得票率暴涨 80% 达到 37%，Pratt 则下滑至 19%，增量选票高度聚集于 Skid Row 贫民窟。Friedberg 指出，加州第 1921 号法案放开无限制“选票收割”，叠加全员无门槛邮寄选票与未定期清理的选民名册，构成了严重隐患——全州寄出约 2300 万至 2400 万张选票，实际投票仅约 950 万张，导致多达 1400 万张空白选票存在流失风险，且自动分拣机签名匹配阈值低至 40%。 主持团队亦客观分析了对立视角的解释，即 Raman 拥有深耕多年的传统地面竞选动员体系，抑或是反 Pratt 选民在确认 Bass 胜势后进行了战术性选票集中。然而，主持团队仍指出该现象极大削弱了选举公信力，并援引 2026 年 Brenda Lee Brown Armstrong 承认有偿收买选票的司法判例，敦促联邦与州司法部对反常数据展开正式调查，同时呼吁国会立法推行统一的选民官方身份证件核验标准。 本期节目集中展现了从硅谷技术治理到宏观制度运行的多重矛盾。无论是私营前沿大模型以安全审查掩盖垄断意图、地缘危机与财政失衡引发的再通胀风险，还是程序漏洞对选举信任体系的侵蚀，均反映出监管架构与现实需求之间的脱节。如何平衡前沿安全与开源生态的活力，以及联邦法律能否有效纠偏地方制度失灵，仍是悬而未决的核心议题。

## Background 本期节目由《All-In Podcast》于 2026 年 6 月 27 日在 YouTube 平台发布。节目简介显示，四位主持人重聚并深入探讨了 Anthropic 因秘密降级 Fable 模型及隐私问题引发的巨大反弹、AI 监管俘获与国有化争议、Liquidity 峰会精彩看点、创三年新高的 CPI/PPI 通胀数据，以及加州宽松选举法引发的计票公正性疑云。 ## TL;DR 节目首先深入剖析了 Anthropic Fable 5 强制留存提示词与暗中削弱模型性能引发的信任风暴，批评其借安全之名推行监管俘获并扼杀开源，同时结合科研实操力挺开源模型生态。随后，嘉宾们针对伯尼·桑德斯拟对 AI 公司征收 50% 股权税建立主权基金的激进主张展开交锋，讽刺科技高管亲手制造的失业恐慌促成了政治反噬，并提出重组社保基金为国家主权基金的破局构想。在复盘 Liquidity 峰会揭示的百亿与万亿巨头跃迁幂律后，讨论转向 5 月超预期爆表的 CPI 与 PPI 数据，警示中东战事外溢与财政失控叠加可能带来的持续加息风险。最后，全场聚焦洛杉矶市长初选中晚期邮寄选票的戏剧性反转，激烈辩论了加州合法化“选票收割”与宽松核验机制对民主选举公信力的深层破坏。 ## Atmosphere & style 全场对话节奏紧凑明快、火药味与幽默感兼具，既有硅谷顶级投资人插科打诨的圈内八卦与实测翻车名场面，也有直指华盛顿与萨克拉门托政客的尖锐批判。四人分工与化学反应极其鲜明：杰森热衷于抛出反方论点挑起论战，萨克斯辞色严厉、对监管套路与选举异象展开狂风暴雨式的政治剖析，弗里德伯格依托扎实的一线科研管理数据和统计逻辑进行冷静论证，查马斯则从吉瓦级电力基建与资本底层逻辑切中要害。从对模型审查荒诞现实的哄堂大笑，到面对通胀阴霾与制度性信任瓦解时的严峻焦虑，整个过程充满了极高密度的思想交锋与内幕透视感。 ## Bulletin - [00:19](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=19s) Anthropic 发布 Fable 5 顶尖模型后因强推 30 天提示词强制留存政策及对涉及模型研发竞争的请求暗中降级，在开发者生态引爆信任危机与监管俘获质疑。 - [37:53](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=2273s) 参议员伯尼·桑德斯提出《美国人工智能主权财富基金法案》，主张对头部 AI 巨头一次性征收 50% 股权税，与前沿高管长期宣扬的“AI 摧毁就业”末日论调形成政治反噬。 - [01:05:39](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=3939s) 5 月美国 CPI 与 PPI 分别同比反弹至 4.2% 和 6.5% 创三年新高，受中东地缘与政府财政开支推动，Polymarket 预测美联储年内加息概率飙升至 49%。 - [01:12:27](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=4347s) 洛杉矶市长初选晚到邮寄选票出现统计学罕见的反转，引发关于加州全量自动邮寄、缺乏 Voter ID 与无限制选票收割制度性漏洞的激烈交锋。 ## Key takeaways - [00:19](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=19s) Anthropic Fable 5 在基准测试刷新的同时引入严苛合规与暗中降级机制，对疑似前沿 AI 研发或硬核技术的 Prompt 实行静默降级且照常收费。 - [04:47](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=287s) 中心化闭源模型对生物和硬核科研的武断封锁，正逼迫农业基因组等实际科研企业转向开源生态，并客观推动企业采纳中国开源模型与本地部署。 - [16:01](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=961s) 头部大模型厂商借“安全审批”与灭世论游说政府建立前置牌照准入制，其深层意图在于通过监管俘获彻底扼杀开源模型的自由分发与本地微调。 - [17:14](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=1034s) AI 算力中心建设成本从早期的 40-50 亿美元/GW 暴涨至 1000 亿美元/GW，算力与电力壁垒加剧了开源生态突围的资金与基建难度。 - [32:32](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=1952s) 务实的生物与物理安全防护应收紧在物理世界的输出端（如强制性合成核酸筛选），而非在大模型输入端对基础科学词汇实施无差别内容审查。 - [39:31](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=2371s) AI 巨头公开宣扬“AI 将致半数人类失业”却未能被就业数据证伪，反而为政客没收企业股权与推进主权财富基金接管递交了民意抓手。 - [51:31](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=3091s) 与传统互联网零边际成本不同，AI 的重资产与边际算力消耗高度依赖公共电网与国家级基础设施，成为政府要求分润股权的核心杠杆。 - [59:22](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=3562s) Coatue 在 Liquidity 峰会的数据揭示估值跃迁的极端幂律：百角兽（$100B）成长为万亿巨头（$1T）的概率高达 31%，远超早期阶段的晋级概率。 - [01:08:27](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=4107s) 通胀反弹的底层诱因被指为政府开支失控与中东能源地缘波动的双重挤压，若中国战略原油储备缓冲见底，油价与物价恐面临二次失控冲击。 - [01:15:28](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=4528s) 加州通过 AB 1921 法案将选票收割合法化并取消严格身份核验，导致选民名册混乱与海量无主选票流转，动摇了公众对程序公正的根本信任。 - [01:17:49](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=4669s) 围绕加州选举争议的争论分歧在于：这究竟是属于钻营制度漏洞的“合法操纵”还是有组织的非法舞弊，亟需统一联邦立法核验身份以重建互信。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=0s) 开场调侃与 Anthropic Fable 5 的重磅发布及争议风波 节目一开始，Jason Calacanis 照例以熟悉的自信与幽默拉开帷幕，调侃 All-In 是“全球第一的播客”、“播客主们最爱听的播客”，也是“别人抄了 50 期就不得不放弃、而原班四人组却在不断加码”的节目。本周的核心焦点迅速转向了近期在 AI 圈掀起惊涛骇浪的 Anthropic。Anthropic 推出了代号为 Fable 5 的顶级前沿模型（被视作接近此前的 Mythos 级别模型），首席执行官 Dario Amodei 也重返长篇博文阵地。Fable 5 在发布后几乎横扫了所有基准测试榜单，展现出令人惊叹的智能水平。然而，伴随其卓越性能而来的，不仅是比 Opus 4.8 贵出一倍的 token 成本，更有一套极为严苛且极具争议的安全护栏机制。 回顾今年 4 月，Anthropic 曾因安全隐患（特别是极强的网络攻击与黑客渗透能力）而未向公众完全开放 Mythos，仅向少数合作伙伴定向提供。当时 Palo Alto Networks 首席执行官 Nikesh Arora 还公开证实 Mythos “货真价实”，帮助他们堵住了内部大量的安全漏洞。但此次 Fable 5 的落地却彻底引爆了开发者社区的怒火。导火索主要有两点：其一，使用 Fable 5 时，Anthropic 会将用户输入的所有提示词数据（Prompt data）在服务器端保存至少 30 天，引发了极大的隐私疑虑；其二，如果在调用过程中系统检测到用户正在从事“前沿 AI 研究”——即涉嫌利用其模型蒸馏或训练具有竞争力的竞品模型——系统会在不作任何明确通知的情况下，悄悄对用户的模型性能进行“降级处理”（nerfing）。这项严苛的限制被隐蔽地埋在了一份长达 319 页的技术合规文档深处。 在 X（原 Twitter）等社交媒体平台上爆发广泛抗议后，Anthropic 随后在接受《连线》（Wired）采访时有所退让，声称将调整前沿 LLM 研发的安全防护机制，使其变得更加透明和可视化。但这一系列暗箱操作已然在技术界与商业界引发了关于 AI 审查、信任危机与开发者权益的深度恐慌。 - **模型定价与性能**：Fable 5 几乎刷新了所有公开基准测试，虽然官方辩称因为模型更聪明因此解决问题所需的总 token 数量更少，但单价直接达到 Opus 4.8 的两倍。 - **暗中降级（Nerfing）**：当系统判定调用涉及前沿模型研发竞争时，在无明确告警的情况下主动削弱模型能力，相关免责条款深埋在 319 页的冗长文档内。 - **数据保留政策**：所有提示词内容默认被 Anthropic 强制存储至少 30 天。 --- ### [02:11](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=131s) Chamath 的警示：输入审查、企业单点故障与治理危机 Chamath Palihapitiya 肯定了 Anthropic 顶尖研发团队的技术实力，认为他们在封闭前沿实验室模型的突破上确实马力全开，但其展现出来的运作逻辑却给整个行业敲响了沉重的警钟。Chamath 指出，Anthropic 实际上已经“亮明了底牌”：他们会全面拦截用户的输入提示词，并在云端进行价值评估与意图研判，随后才决定是否输出以及输出什么质量的内容。对于个人用户而言，这意味着不可避免的言论与思想审查风险；而对于企业用户来说，这几乎是一个“免谈”的致命商业雷区。 ```mermaid flowchart TD accTitle: Anthropic Fable 5 提示词审查与降级处理流程 accDescr: 展示用户输入提示词到 Anthropic 判定是否敏感或竞品研发，并最终分流至正常输出、隐蔽降级或切断服务的流程 subgraph InputStage ["用户交互阶段"] A["企业/个人输入提示词"] --> B["云端拦截与内容审计"] end subgraph EvalStage ["意图与风险研判"] B --> C{"判定提示词属性"} C -->|"生物/黑客安全红线"| D["拦截输出 / 触发合规切断"] C -->|"前沿 AI 研发/蒸馏竞争"| E["隐蔽降级模型性能 (Nerfing)"] C -->|"普通商业/常规请求"| F["正常生成高阶输出"] end subgraph ImpactStage ["业务后果"] D --> G["业务流程意外瘫痪"] E --> H["不知情下获得劣质输出"] F --> I["数据强制留存至少 30 天"] end ``` 企业内部的科研人员、高管或分析师在正常调用云端 API 时，可能在毫不知情的情况下误触某种未公开的安全规则或敏感词库，进而直接导致核心业务通道被单方面掐断，失去賴以生存的差异化技术优势。Chamath 形象地调侃道：“Anthropic 确实说了大实话，只是这句实话当你咽进肚子里后，胃里会极度不舒服、恶心得要命。”因此，当前企业界正在迎来一个重大转折时刻——必须重新审视 AI 的控制权：企业究竟在允许谁利用自己的数据进行二次学习？企业能否承受将整个业务底座押注在单一闭源模型上的单点故障风险（single point of failure）？多元化的多模型冗余以及更为自主可控的 AI 治理策略，已经成为企业生存的必然选择。 --- ### [04:47](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=287s) Friedberg 现身说法：生物科研受限与被逼倒向开源生态 David Friedberg 从自身创立和运营的农业基因组公司（Ohalo）切身经历出发，深入剖析了闭源模型设限对真实科研带来的重创。Friedberg 坦言，尽管各家平台的服务条款（TOS）白纸黑字承诺不会拿客户的专有数据去训练模型，但企业界普遍对此持保留怀疑态度；此前出于对强大前沿能力的渴望，团队选择“把谨慎抛在脑后”。在过去几年中，他们利用这些模型进行植物基因组与变异位点的影响预测、CRISPR 基因编辑的 RNA 引导链设计，以及特定蛋白质构建构造等前沿合成生物学研究。这些曾经需要大量科学家耗费无数工时的人工计算，在 AI 的辅助下效率发生了质的飞跃。 然而就在最近几周，由于 Anthropic 等厂商对所谓“生物武器风险”理论的极度敏感，平台大幅收紧了相关领域的访问权限。其结果是正常的植物基因编辑与农业育种研发被全盘误伤，科学家甚至无法再使用这些工具设计有益的蛋白质构型。这种武断的限制彻底切断了关键科研路径，直接迫使像他们这样的企业转向开源模型并在本地算力上部署运行。 Friedberg 发出了发人深省的战略警告：当今全球最优秀的开源模型并非来自美国，而是来自中国（"And what are the best open source models today? They're Chinese."）。Anthropic 等美国前沿实验室出于道义优越感或合规考量而施加的严苛限制，并不会真正“阻止 AI”，反而只会将美国初创公司与大型企业整体推向中国开源模型的怀抱。 - **农业与基因组研发的现实损失**：原本用于农作物增产、抗病害的基因引导链与蛋白质设计工具被“一刀切”误伤，严重阻碍正当科学研究。 - **开源版图格局的倒错**：由于美国闭源生态的强力自缚与政治化倾向，市场上表现最强劲、最能填补本地部署空白的开源模型主要来自中国团队。 - **技术扩散不可逆**：通过政治打压或社交舆论制造恐慌试图“关掉 AI”完全是徒劳的妄想，盲目限制只会自毁经济活力与科技领先地位，将非对称优势拱手让人。 --- ### [07:44](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=464s) 自建专属模型与监管俘获陷阱的隐忧 随着讨论的深入，Jason Calacanis 指出企业应对这种不可抗力的终极路线图：结合私有专有数据与开源基础模型，自建内部的专用大模型。Friedberg 对此完全赞同，确认他们正在沿着这一方向推进，利用开源底座配合内部专有农业基因组数据，打造专属的“基因组语言模型”（Genome Language Model）与表型预测系统。 与此同时，Jason 与 David Sacks 将话题推向了更深层的政治与商业博弈。Anthropic 等机构不断公开渲染 AI 的“末日风险”、“生物武器危险”，在客观上引发了政客的恐慌与激进表态（包括后续即将深入探讨的伯尼·桑德斯提出剥夺 AI 企业股权等极端国有化主张）。Jason 质疑 Anthropic 是否在刻意大喊“快看，这里着火了，快来严厉监管我们”，从而借监管之名筑起阻止后来者竞争的高墙。David Sacks 对此一针见血地回应，指出这正是他在八个月前就预警过的经典策略——Anthropic 正在极其老练地践行一场精密的“监管俘获”（regulatory capture）棋局。 ### [10:02](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=602s) Anthropic Fable 降级风波与监控危机：AI 阶层的暗中分化 围绕 Anthropic 最新发布的 Fable 和 Mythos 系列模型，开发者社区爆发了空前的愤怒与信任危机。此前大卫·萨克斯（David Sacks）曾指出这类科技巨头正在打着安全旗号制造恐慌，如今这一观点已经成为科技圈的普遍共识。这次争议的核心并不仅仅在于强制性的隐私监控——作为一家曾公开宣称反对政府监控的公司，Anthropic 如今却要求对 Mythos 级别模型的所有 Prompt、生成输出以及超大上下文窗口中包含的数据（包括 Agent 平台抓取的用户个人文件、记忆库和工作流数据）实行长达 30 天的强制数据留存。即便是过去签署了“零数据留存”（Zero Data Retention, ZDR）协议的企业客户，在此次新模型下也毫无豁免权，要么被迫接受全面监视，要么只能放弃使用。 更为恶劣的是，Anthropic 根据其单方面的裁决对用户进行分类画像，并秘密对其模型进行性能阉割（Nerfing）。当系统判定用户“不配”或“不适合”获取某种高阶信息时——特别是在机器学习、AI 前沿研究以及芯片设计等核心硬核技术领域——模型会在后台暗中重写用户的提示词，神不知鬼不觉地把用户降级调用更差的基础模型，同时却继续按照顶尖模型的标准向用户全额收费。虽然在舆论反弹后 Anthropic 做出了一小步退让，承诺在降级时向用户予以提示，但这种肆意剥夺用户获取完整知识权利的机制依然存在。 在开发者社区公布的大量实测案例中，这种审查展现出了极端的越界与荒谬：有用户仅仅向模型提问关于“线粒体”（mitochondria）的生物学基础问题，就直接触发了降级逻辑；科技分析师本·汤普森（Ben Thompson，Stratechery 创始人）询问关于 GLP-1 减肥药与癌症风险之间的临床关联，同样被直接踢出高阶服务。萨克斯直言，Anthropic 正在构建一个极端奥威尔式的反乌托邦图景：几家中心化的大模型巨头自命为真理的仲裁者，把全社会划分为“AI 阶层有产者”（AI haves）和“AI 阶层无产者”（AI have-nots），并根据自己设立的道德标尺决定谁能接触前沿认知。 - 强制数据留存：所有 Mythos 级别交互强制保留 30 天，无视企业现有的零数据留存（ZDR）协议。 - 暗箱降级机制：后台篡改提示词、切换低阶模型且照常扣费，重点波及芯片设计、生物计算和前沿 AI 算法开发。 - 审查过度蔓延：线粒体基础科学问题及 GLP-1 药物机理探究均遭到降级封杀。 ```mermaid flowchart TD accTitle: Anthropic 模型降级与数据监控流程 accDescr: 展示用户提交请求后，Anthropic 收集上下文、进行用户画像画像、决定是否降级以及暗中改写 prompt 的闭环流程。 subgraph Surveillance["数据留存与分析"] A["用户提交复杂 Prompt\n与上下文文件"] --> B["强制云端留存 30 天\n(即使签署 ZDR 也无效)"] B --> C["系统构建用户画像\n并评估查询领域"] end subgraph NerfingDecision["审查与降级判定"] C -->|敏感/硬科技领域| D{"系统判定用户\n是否具备资格?"} D -->|判定不合格| E["后台静默改写 Prompt"] E --> F["降级调用弱化模型\n(按顶尖价格收费)"] D -->|判定合格| G["提供完整 Mythos 性能"] end ``` --- ### [13:30](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=810s) 商业偏袒、黑箱审计与自由市场的监管悖论 查马斯·帕里哈皮蒂亚（Chamath Palihapitiya）进一步剖析了这种中心化黑箱控制背后更隐秘、但也更为险恶的商业危害：模型厂商将获得直接挑选“商业赢家与输家”的绝对权力。以制药业为例，若诺华（Novartis）与礼来（Eli Lilly）在 GLP-1 赛道展开针锋相对的竞争，而 Anthropic 与诺华签署了深度战略联盟协议，那么算法在输出药理对比或市场评价时就会天然偏向诺华；如果模型厂商与摩根大通建立了合作关系而排斥花旗银行，普通用户在咨询房贷利率或理财策略时，就会在毫不知情的情况下接收到带有系统性偏见的信息流。 更致命的是缺乏底层的技术指纹追溯机制（Audit Trail）。当公众或监管机构质疑某个具体的模型输出存在暗中操纵时，大模型厂商往往可以轻而易举地藏匿证据、推脱责任，根本无法向外界还原生成该特定回答时的确切运行时轨迹（trace）。这直接导致了一种扭曲的商业寻租现象：新兴初创公司如果想要在 AI 时代生存，就必须亲自去给 Anthropic 这类寡头“纳贡”，拿出半数股权换取算法导流；而大厂甚至不仅看重经济利益，还会把其单方面推崇的社会议题、道德观念作为考核标准，偏袒符合其哲学立场的商业实体。 面对这种系统性腐败的风险，杰森·卡拉卡尼斯（Jason Calacanis）提出了典型的自由市场主义质疑：早年搜索引擎也曾大搞付费收录（paid inclusion）和幕后操盘，但这种恶行最终损害了用户体验，导致消费者集体弃用并转向更纯粹的 Google；既然如此，只要放任自由市场竞争、不搞政府干预，消费者自然会用脚投票选择更优秀的模型，何须如此悲观？ --- ### [16:01](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=961s) 达里奥博文曝光：以“安全审批”之名扼杀开源模型 萨克斯立刻指出了自由市场假设的盲区：这一逻辑成立的唯一前提是市场上存在充分的替代选择，然而 Anthropic 正在联手监管力量封死所有退路。就在 Anthropic 启动强制监控和模型暗中阉割的同时，其首席执行官达里奥·阿莫代伊（Dario Amodei）发表了重磅政策博文，宣称单纯依靠“透明度”已经无法应对 AI 的指数级风险，必须建立一个类似于联邦航空管理局（FAA）或食品药品监督管理局（FDA）的全新政府监管机构，对所有前沿大模型的发布实行前置审查和牌照批准。 这一举动揭开了“安全主义”的遮羞布，暴露出极其露骨的“监管俘获”（Regulatory Capture）本质。由于开源模型是去中心化分发、由社区在本地设备上自由运行的，根本不可能嵌入这类官僚化的审批机制。因此，游说政府出台严苛的牌照准入制，其真正目的就是从源头精准扼杀开源生态，阻止用户在自己的本地机器上进行不受限制的本地算力微调（token maxing on local machine）。大模型巨头并非真的渴望接受合规束缚，而是试图把这套无法承受的合规重压强加在开源对手身上，打一场先发制人的歼灭战。 --- ### [17:14](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=1034s) 算力基建的天价壁垒与中美 AI 竞争的地缘危机 查马斯随后分享了自己在硬核算力基建领域的亲身经历。两年前，他与合伙人在亚利桑那州买下了 2000 英亩土地，搞定了土地规划和区划审批，拿到了建设 2 吉瓦（GW）超大型数据中心的全部许可。他原本的算盘是在获批后直接转手加价卖给黑石（Blackstone）、博枫（Brookfield）或 Google 这样的万亿美元资管巨头套现离场。但他现在彻底改变了想法，因为他意识到，如果科技圈不把海量算力直接定向倾斜给开源生态，萨克斯和弗里德伯格（David Friedberg）所警示的噩梦就会成真——目前全球电网兆瓦级算力的绝大多数份额都被超大规模商业模型吃尽，留给开源生态的长尾算力微乎其微。 然而，挽救开源的资本壁垒高到了令人绝望的地步。在两年前项目启动时，建设 1 吉瓦算力中心的成本约为 40 亿至 50 亿美元；而伴随英伟达顶级芯片价格的飙升与电网设施挤兑，如今 1 吉瓦数据中心的建设门槛已经暴涨 20 倍，飙升到了不可思议的 1000 亿美元。即便查马斯刚在另一处提交了争取额外 1 吉瓦土地的收购要约，要真正把这 3 吉瓦的算力彻底建成并向开源社区提供充沛流动性，意味着必须砸进 3000 亿美元的惊人资金。 如果美国政府在监管层面上演“抽梯子”（pull the ladder）式的准入限制，再加上千亿美元级的资本封锁，美国本土的开发者将被彻底焊死在单一模型、垄断规则的牢笼之中。与此同时，世界的其他地区绝不会画地为牢停滞不前。中国已经在生物科技、新材料科学和先进工业制造体系等领域飞速狂飙。一旦海外凭借开放自由的模型和廉价无阻的算力形成不对称的技术迭代优势，美国原本引以为傲的劳动力竞争力和实体经济体系将面临不可逆转的重大地缘风险。 - 算力成本暴涨：从早期的 40-50 亿美元/GW 狂飙至当前的 1000 亿美元/GW，涨幅达 20 倍。 - 查马斯资产布局：亚利桑那州已锁定 2000 英亩（配额 2 GW 数据中心），正在竞购新增 1 GW 场地，总资金缺口高达 3000 亿美元。 - 全球竞争失衡：严苛监管仅能锁死美国本土创新，中国在生物科技、材料科学与工业自动化体系上的持续狂飙将彻底重塑全球经济版图。 ### [19:40](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=1180s) 开源基因组大模型与封闭生态的对比 David Friedberg 证实了团队在植物育种项目中使用开源模型的现状，并详细介绍了 Patrick Collison 与 John Collison 兄弟资助的 Arc Institute 所研发的基因组语言模型（Genome Language Model）。该模型整合了全球所有可公开获取的基因组数据。在育种实践中，当缺乏历史经验数据来判断植物某种基因变异的优劣时，团队会将 DNA 序列输入该模型。模型会直接分析碱基排列顺序，评估其为优良序列还是劣质序列，这就像语言模型判断一段英文表达是否规范地道一样——其核心逻辑在于捕捉这些碱基排列在其他生物演化历程中出现的统计概率。 ```mermaid flowchart TD accTitle: 基因组语言模型工作流 accDescr: 展示 Arc Institute 开源基因组模型从数据整合到农业育种落地的流程 subgraph DataPrep["数据整合阶段"] A["全球公开基因组数据"] --> B["Arc Institute 基因组大模型"] end subgraph Inference["序列评估阶段"] C["植物基因变异样本"] --> D{"模型分析碱基排列概率"} B --> D end subgraph Application["育种决策阶段"] D -->|高概率/良性| E["优良基因序列"] D -->|低概率/缺陷| F["劣质基因序列"] E --> G["指导农作物改良育种"] end ``` Friedberg 强调，像 Collison 兄弟这样通过民间资本资助前沿科研并将核心成果彻底开源的做法，正在为行业带来巨大的竞争优势，能强力抗衡那些全封闭的专有大模型生态。他认为，美国越是扶持并推广这种开源繁荣，就越能避免未来因政府过度干预或头部大模型厂商搞“自我监管”（Self-regulation）垄断一切而陷入被动局面。 - **资助与研发机构**：Stripe 创始人 Collison 兄弟出资设立的 Arc Institute。 - **技术原理**：将 DNA 视作自然语言，通过海量跨物种序列的字母排列概率推断基因优劣。 - **应用场景**：农业育种规划、植物优良基因性状的早筛与预测。 --- ### [21:04](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=1264s) 替 Dario Amodei 辩护：精英心态、人才虹吸与开源错失的战机 Jason Calacanis 提出了针对 Anthropic 首席执行官 Dario Amodei 的“正方辩护”（Steelman）。尽管他戏谑 Amodei 动辄写下 5000 字的长篇大论，表现得像是患上了“AI 精神病”（AI psychosis）并伴有夸大妄想（Delusions of grandeur），以为自己造出了某种弗兰肯斯坦式的科学怪兽；但换位思考，如果一名创始研发者深信自己手里掌握着足以颠覆世界甚至毁灭现实的终极力量，他采取极端谨慎的策略是符合逻辑的。与其冒着被恶意滥用的巨大风险全面公开，不如将其仅交付给少数可信伙伴以加固网络安全，随后再以可选功能的形式小心翼翼地向大众推送。 Jason 指出，虽然外界觉得 Amodei 的言论充满末日夸张，但行业内 80% 到 90% 的顶尖 AI 研发者思维模式与 Amodei 高度同频。他们发自内心地认为自己正在“创造上帝”，因此 Anthropic 的使命感对硅谷最精英的人才形成了天然的虹吸效应。相比之下，崇尚自由市场与自由意志主义的人才投奔了 Elon Musk 与 Grok，追逐最高薪酬激励的人才则流向了 Sam Altman。谈及模型生态格局，嘉宾们共同惋惜 Meta 在 Llama 上的战略失误——Mark Zuckerberg 原本应当站在博弈论的高度，以“焦土政策”（Scorch the earth）推出不可撼动的最强开源模型，彻底抽干封闭模型厂商的利润空间，将基础智能彻底商品化，然而 Meta 却白白错失了统领开源的大好窗口期。 --- ### [23:43](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=1423s) 曼哈顿计划的警示：AI 的三类武器化风险与规制悖论 面对关于技术失控的辩护，Friedberg 借用核物理的历史进行了有力反驳。人类在发现原子核裂变时，既诞生了能够提供无限廉价清洁能源的核反应堆，也造出了毁灭性的原子弹；当年深度参与曼哈顿计划的科学家在战后便爆发了激烈的反思与伦理抗争。Friedberg 将前沿模型潜在的武器化风险精炼归纳为三大领域，并阐明了双重用途技术内在的巨大悖论： | 武器化分类 | 潜在滥用威胁 | 同等能力带来的巨大正面价值 | |:--- |:--- |:--- | | **网络武器** (Cyber Weapons) | 自动化漏洞挖掘、全新网络谍报工具、恶意勒索攻击 | 软件生产力跃迁、个人赋能创业、提升社会收入 | | **物理武器** (Physical Weapons) | 杀伤性武器的自动化工程辅助、动力学设计优化 | 尖端工业制造革新、材料科学突破与工程效率倍增 | | **生物武器** (Bioweapons) | 人工合成全新高致病性病原体、剧毒生化试剂设计 | 治愈癌症、农作物抗逆改良、彻底解决全球粮食短缺 | Friedberg 强调，技术发展具备高度的“决定论”（Deterministic）特征——任何理论上行得通的事情，迟早会被人尝试，如今大模型的潘多拉魔盒已被彻底打开。试图通过少数私人企业或个别科技精英来决定“谁能用什么技术”，在逻辑上荒谬且危险。正确的治理应当聚焦于**最终产出的非法行为**，而非在最前端扼杀工具本身。现行的法律体系早已对网络入侵、生化武器制造、武器研发及产品侵权做出了严格界定，社会应当强化后端执法与追踪能力，而不是在前端筑起重重审查高墙，否则就等同于因噎废食，剥夺社会享受技术红利与创造财富的机会。 --- ### [26:43](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=1603s) “网络防火墙”隐喻、化肥炸弹与真正需要推行的 KYC 讨论进一步延伸到监管机制的荒谬性。Jason 形象地比喻：试想在互联网发展的早期，如果网民在浏览器中输入网址敲击回车后，必须先由中心化机构审查你的身份、再决定是否放行或重定向到其他网页，那本质上就是建立起了一套类似特定国家的信息网络防火墙；或者搜索引擎按照它所判定的用户身份来随意过滤真实信息。Friedberg 认为，当前很多所谓的 AI 监管提案正在把大模型引向这种危险的前置审查境地。 针对监管难点，Jason 搬出 1995 年俄克拉荷马城爆炸案后对化肥管制的先例来替私企设立安全护栏辩护：制造炸药的硝酸铵化肥具有极高危险性，政府与厂商随后要求购进化肥必须实名并嵌入追踪标记，因此 Anthropic 作为私营企业，为了避免卷入致命攻击诉讼而设立严格的防御门槛在情理之中。对此，Chamath Palihapitiya 随即一针见血地指出，化肥案的核心机制是实打实的**身份认证（KYC, Know Your Customer）**。真正令人不齿的是，Anthropic 明明有能力在明天就全面上线严格的开发者 KYC 机制、让开源和透明模型正常运转，但他们却选择逃避落地实际的身份审查，转而私下搞信息审查与降级，并在台前不断向政客大倒苦水、鼓吹由他们主导规则定制的准入壁垒，这种既当裁判又当运动员的做法本质上极为虚伪且充满算计。 ### [28:49](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=1729s) 监管俘获陷阱与大模型寡头垄断隐忧 大卫·弗里德伯格（David Friedberg）谈到，即便自己愿意向 Anthropic 抵押保证金并提供全体员工的姓名与住址，Anthropic 目前也不会因此给予未经限制的 Fable 5 访问权限。弗里德伯格认为这纯粹是 Anthropic 自身的商业与技术决策，作为一个崇尚自由市场的人，他不满意大可选择不用，市场自然会孕育出其他替代选项。然而，大卫·萨克斯（David Sacks）对此提出了针锋相对的严厉警告：弗里德伯格忽略了 Anthropic 这家万亿美元级别的巨头正斥资数十亿美元推动“监管俘获”（Regulatory Capture）议程。其创始人达里奥·阿莫代伊（Dario Amodei）正以极其夸大且煽动性的言辞渲染 AI 灭世风险，其最终目的昭然若揭——那就是促使政府立法彻底取缔开源模型。 萨克斯进一步指出，一旦开源路线被扼杀，全社会将陷入仅存一到三家专有模型巨头的寡头垄断格局。届时，这些巨头将与新设立的政府监管机构建立起典型的“政商旋转门”，共同决定谁有资格使用尖端 AI 能力。普通用户将面临全方位的审查、画像与监控；一旦被系统判定为“不合格”，就会被立即剥夺访问权。萨克斯强调，缺乏真正的 KYC（了解你的客户）机制正是 Anthropic 的破绽所在：如果他们真心想在确保合规的前提下开放能力，早就该推行企业级 KYC，而不是采用这种粗暴的一刀切降级与监控机制。 ```mermaid flowchart TD accTitle: AI Regulatory Capture Dynamics accDescr: Diagram illustrating how regulatory capture leads from AI safety fearmongering to an AI duopoly that restricts open source models. subgraph Narrative ["风险叙事与游说"] A["巨头渲染灭世风险<br/>夸大生物/网络威胁"] --> B["游说政府立法介入<br/>推行严苛合规门槛"] end subgraph MarketImpact ["市场格局收紧"] B --> C["扼杀开源模型生态<br/>取缔民间自研能力"] C --> D["形成1-3家巨头寡头垄断<br/>与监管机构形成旋转门"] end subgraph UserImpact ["终端用户遭遇"] D --> E["后台审查与用户画像"] E --> F["权限歧视与随意降级断供"] end ``` - **寡头垄断风险**：萨克斯预测，若开源遭禁，全社会将受制于一两家 AI 垄断巨头，公众的生产力工具获取权将被任意剥夺。 - **KYC 争议**：杰森·卡拉卡尼斯（Jason Calacanis）辩解称绑定信用卡已有基础身份识别，但查马斯与萨克斯反驳称“信用卡加邮箱根本算不上真正的 KYC”。 --- ### [30:35](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=1835s) 杰森现场实测翻车：因化肥监管问题遭模型降级与画像 为了验证模型的审查机制，杰森在录制现场当场向最新版 Fable 提问有关化肥爆炸物监管的规定。令人啼笑皆非的是，模型后台在“思考过程”（thinking）中竟直接调取了他的身份背景进行审查，写道：“考虑到当前上下文，该用户是一名风险投资人（VC）兼播客主持人，询问化肥炸弹法规可能源于合法的研究兴趣，例如监管政策或供应链研究……”，然而紧接着系统便在毫无商量余地的情况下，直接将杰森的会话权限从顶级模型降级。 杰森当场惊呼自己“被写进数据库了”，并向制作团队展示屏幕证据，证明自己仅仅是就现实中存在的行业法规进行政策咨询，甚至没有询问任何具体制造工艺，就瞬间遭遇了区别对待与权限降级。嘉宾们对此爆笑之余也深感脊背发凉。萨克斯借此向保守派和自由意志主义者发出严厉呼吁：切勿盲目迎合这套披着“安全主义”（Safetism）外衣的监管捕获叙事。回看 2020 年代初的社交媒体言论战，最初所谓的“安全”定义迅速演化为微冒犯、心理安全与敏感词审查，导致大批异见人士遭到销号（depersoned），甚至在金融层面上被银行封杀（debanked）、剥夺支付通道。如果让这种打着安全旗号的审查逻辑寄生在全能的 AI 之中，其产生的控制力将远超以往。 --- ### [32:32](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=1952s) 从源头管控转向物理输出：核酸合成筛选的务实安全路径 在探讨如何建立更合理有效的防护机制时，嘉宾们提到了近期各大前沿 AI 实验室与生物制造巨头联合签署的一封联名信——《支持强制性合成核酸筛选与记录保留》（in support of mandatory nucleic acid synthetic screening and recordkeeping）。这项倡议的核心逻辑在于：安全干预不应该发生在用户向模型提问的输入端（Input），而应该收紧在物理世界的输出端（Output）。 弗里德伯格与萨克斯指出，早自 2009 年成立“国际基因合成联盟”（International Gene Synthesis Consortium, IGSC）以来，各大寡核苷酸合成企业就已经建立起一套成熟自律的机制。全球任何实验室在订购人造 DNA 或 RNA 序列用于科研时，合成厂商都会自动将其与危险病原体（如埃博拉病毒、生物武器毒素序列等）数据库进行比对。经过 15 年的实践检验，这套体系不仅高效自动化，而且从未阻碍正常的医学与生物科研。如今业内呼吁通过立法将该标准强制化，正是一个极具示范意义的务实安全方案。萨克斯直言，与其让 AI 模型粗暴地将“线粒体”、“癌症”或“GLP-1”等普通医学提问一律视为违规红线，不如在物理交付节点进行精密把关，这样既保障了公众的言论与知识探索自由，又真正化解了实质性的安全威胁。 - **IGSC 2009 自律联盟**：全球主流核酸合成供应商在此机制下自愿开展序列审查逾 15 年。 - **输出端管控对比输入端审查**：物理层面的化合物/基因合成具备可追溯性，远优于大模型前端对普通生物医学词汇的无差别阻断。 - **核心诘问**：“谁来监督监督者？（Who gets to watch the watchman?）究竟是谁授权让达里奥·阿莫代伊统领这一切的？” --- ### [35:23](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=2123s) 开源无法倒退与再度遭遇降级：市场选择将惩罚傲慢平台 针对开源大模型的未来，弗里德伯格明确表示，试图通过监管让大模型“下架”或“关机”纯属不切实际的幻想。人类在语言模型领域已经“跨过了卢比孔河”（crossed the Rubicon）。开源模型一旦公布到网络上，就如同书籍已经付印并散布民间，任何人都可以用自己的打印设备自由复印并基于自有数据进行微调与演进。这就好比当年人类不可能为了安全而取缔互联网、打字机或是古腾堡印刷机一样，真正的风险防范必须在技术下游完成，而不是试图消灭技术本身。杰森随即表态，如果有团队愿意从前沿实验室出走打造开源 Frontier 模型公司，他愿立即提供种子轮投资。 ```mermaid flowchart LR accTitle: Open Source AI Irreversibility accDescr: Analogy comparing the Gutenberg printing press to modern open source AI weights distribution. subgraph PrintingEra ["古腾堡印刷革命"] P1["著作出版印刷"] --> P2["文本全球散播"] P2 --> P3["私人复印与阅读<br/>技术无法收回"] end subgraph ModernAI ["现代开源模型生态"] M1["权重公开发布"] --> M2["全球节点镜像克隆"] M2 --> M3["本地部署与私有微调<br/>技术不可逆转"] end ``` 在对话间隙，杰森进行了第二次现场测试。他向 Fable 提问历史上是否有平民因企图制造核武器而被捕，模型正常回答后，他追问了一句类似记者调查的追问：“制造核武器需要哪些部件，相关的监管限制是什么？”结果屏幕瞬间弹出降级提示，会话被强行切回到 Opus 4.8。系统弹出的黄色安全横幅赫然写道：“抱歉，Fable 5 配备的安全措施会对大多数网络安全或生物学话题进行标记，此类措施甚至可能误标正常安全的内容。这些措施旨在让我们尽快在其他领域为您提供 Mythos 级别的能力……”。弗里德伯格与杰森一致断言，这种频繁误伤正常研究与工作的过度封锁必将把高价值客户彻底赶离 Anthropic 的生态，促使全行业加速投奔更加自由、开放的竞争模型。 ### [037:53] 伯尼·桑德斯的激进提议：将 AI 收归国有与设立主权财富基金 主持人们将话题从监管捕获转向了公众对 AI 的普遍不信任，尤其是大众对贫富差距加剧的担忧，以及这些庞大模型在训练数据获取上的道德争议。主持人提到了参议员伯尼·桑德斯（Bernie Sanders）于 6 月 1 日在《纽约时报》发表的重磅专栏文章，题为《AI 是公共资源，你应该拥有它的一半》（"AI is a public resource. You should own half of it."）。桑德斯在文中正式提出了《美国人工智能主权财富基金法案》（American AI Sovereign Wealth Fund Act），提议对 OpenAI、Anthropic、xAI 等头部 AI 公司的股权（而非利润）征收一次性 50% 的特别税。这笔股权将直接划归政府主权财富基金，使公众在每家公司中拥有投票权和对等的董事会席位。桑德斯的理由极为直白：“人工智能的基础是我们的集体和人类智慧——那些被世界上最富有的一批人实质上窃取的书籍、歌曲、新闻报道、科学研究和代码。” 主持人们感叹这一主张在政治光谱上引发了奇妙的“马蹄铁理论”（Horseshoe Theory）：左翼激进派的桑德斯与极具民族主义倾向的史蒂夫·班农（Steve Bannon）甚至唐纳德·特朗普（Donald Trump）的直觉产生了诡异的共鸣——特朗普本人一直对建立国家主权财富基金、让国家持有企业资产抱有浓厚兴趣。 ```mermaid flowchart TD accTitle: 伯尼·桑德斯 AI 主权财富基金构想 accDescr: 展示桑德斯提议如何从头部 AI 公司强制征税股权并转化为具有投票权的公共财富基金 subgraph Source["数据与价值源泉"] A["人类集体智慧<br/>(图书/歌曲/新闻/科研/开源代码)"] --> B["头部 AI 巨头<br/>(OpenAI, Anthropic, xAI)"] end subgraph Proposal["桑德斯法案机制"] B -- "一次性征收 50% 股权税" --> C["美国 AI 主权财富基金<br/>(Sovereign Wealth Fund)"] C --> D["公众获得平等董事席位与投票权"] end ``` ### [039:31] “自作自受”：AI 巨头如何亲自点燃了公众的怒火 针对桑德斯的提案，David Sacks 明确表示反对这种对私有财产的直接剥夺，认为这开启了极其危险的先例。但他直言非常理解这一政治诉求背后的民意基础。原因在于，这群 AI 公司的 CEO 们过去一段时间一直在四处宣扬“AI 将导致 50% 人类失业”的末日论调。比如 Anthropic 的 Dario Amodei 在最新采访中依然态度坚决，毫无收敛。Sacks 犀利地剖析了普通民众的心态：AI 巨头们自己承认模型是靠全人类成百上千年免费放在互联网上的知识训练出来的，普通人当年贡献内容绝不是为了让自己丢掉饭碗；如今这帮科技富豪反过来声称要用这些知识取代人类的工作，普通美国人自然会发出质问：“这笔交易对我们有什么好处？” Sacks 强调，所谓“AI 导致大批人员失业”的末日叙事纯属无稽之谈。最新的 5 月非农就业数据显示新增就业 17.2 万人，远超经济学家预期的两倍以上，失业率维持在 4.3% 的历史低位，建筑业岗位大增，甚至软件工程师岗位也创下三年新高（主持人们借此打趣喊道“感谢特朗普总统”）。然而，公众的恐惧正是被科技巨头自己灌输的。马斯克虽然也提过未来无需工作，但他至少描绘了一幅通过 AI 和机器人创造极致物资丰足、让人类告别苦役的蓝图；而 Dario Amodei 却直言未来将出现“超高 GDP 与超高失业率并存”，相当于直接宣布普通人将彻底被财富增长边缘化。当科技巨头以这种傲慢姿态示人时，公众要求分一杯羹乃至没收股权的政治反扑便成了必然。 此外，OpenAI 和 Anthropic 都采用了“公共利益公司”（Public Benefit Corporation, PBC）架构。主持人们指出，一般商业公司的法定义务是为股东利益最大化服务，而 PBC 必须在股东回报与设定的公共使命之间取得平衡。既然如此，Sacks 调侃道：“世界上没有什么比用 Anthropic 一半的利润来偿还美国国债更符合公共利益的事了。”这至少比让他们把钱拿去搞政治偏见、监控或言论审查要强得多。 ### [043:32] 弗里德伯格的解局之策：将社保信托基金重组为主动投资的主权基金 针对资产没收式征税的争议，David Friedberg 提出了一个根本性的制度改革方案。他指出，美国现存的社会保障信托基金（Social Security Trust Fund）本质上是一个濒临破产的旧时代“确定给付制”（Defined Benefit）计划，而现代私人部门早已转向“确定缴费制”（Defined Contribution，如 401(k)），即个人按期存入资金、自主投资并享受复利增长。目前规模达 4 万亿美元的社保信托基金账面上仅仅躺着一张特殊的美国国债凭证，完全被限制只能购买国债。 Friedberg 提议借此契机彻底重组社保系统，将其转化为一个真正的国家级主权财富基金： - **转向个人账户制**：废除僵化的统一福利机制，让每位美国公民拥有明晰的个人投资账户。 - **允许配置权益资产**：打破只能购买国债的限制，允许该基金在全美优质企业中配置股票。 - **以注资替代剥夺**：政府不应直接没收企业股权，而是应当真金白银地投资包括 AI 在内的前沿优质资产，获取的股票份额直接存入全体公民的账户，让每个人真正成为时代发展的股东。 在探讨 AI 对劳动力市场的影响时，Friedberg 再次重申自己坚信“AI 绝不会造成失业”。他从第一线管理实践出发分析了企业运营的两个核心维度： - **成本端（Cost Side）**：AI 替代人工以削减成本的实际影响微乎其微。 - **收入与产出端（Revenue Side）**：AI 带来的实质飞跃在于赋予单兵极高的生产力杠杆。过去一个工程师只能完成有限的任务，现在其产能被放大了 100 倍到 1000 倍，无论是研发种子产品、制造船只还是编写软件。生产力的飞跃极大地拓展了企业的业务边界，企业反而陷入了极度的人才饥渴之中，目前根本招不够人。 ### [47:23](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=2843s) AI 究竟是岗位杀手还是繁荣催化剂？ 讨论由工程团队的扩招切入。嘉宾指出，自己的工程小组正计划额外增加 15 名员工，因为 AI 带来了过去根本无法开展的海量全新业务机会。就业市场的实际数据也在印证这一点——“AI 将摧毁就业”本质上是一种卢德主义（Luddite）的论调，正在被现实逐日证伪。目前行业一线呈现出的完全是一场大繁荣（boom）而非衰退（bust）。 面对 Jason 提出“这是否意味着与 Dario Amodei、Elon Musk 和 Sam Altman 的观点相左”的疑问，主持人澄清了这三位领袖言论之间的关键差异： - **Elon Musk** 描述的是一种远期的人类终局状态（甚至可能是上百年后）：机器人与 AI 高度丰富物质，人类不再需要为生存被迫工作，迈入全民高收入（Universal High Income）的乌托邦。 - **Dario Amodei**（Anthropic 联合创始人）曾在约一年前做出激进的具体预测，称在接下来的 1 到 5 年内，初级知识工作者将面临高达 50% 的失业率。 - **Sam Altman**（OpenAI 首席执行官）过去曾泛泛谈论过失业潮即将到来，但近期已改口承认自己过去的预测有误，因为现实数据中并未出现大批失业。 Dario 的 Anthropic 实现了高达 400 亿美元的营收规模，其驱动力绝非企业在裁撤客服代表，而是各行各业正在利用 AI 打造前所未有的全新产品。 ```mermaid flowchart TD accTitle: AI对就业影响的观点图谱 accDescr: 对比行业领袖与现实数据的不同观点 subgraph 行业领袖预测 M["Elon Musk<br/>远期终局/高福利乌托邦"] D["Dario Amodei<br/>1-5年内初级知识岗位缩减50%"] S["Sam Altman<br/>早期预言失业，近期承认数据未支持"] end subgraph 现实产业落地 R["企业业务扩张与新产品研发"] --> H["工程团队扩招<br/>（例：单组新增15人）"] R --> B["Anthropic营收飙升至400亿美元"] end ``` --- ### [49:11](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=2951s) AI 巨头的自我矛盾与“愚蠢税” 公众与政客对 AI 的恐慌情绪很大程度上是被这群科技高管亲自灌输出来的。科技巨头们多年来一直在舆论场上宣扬“自己的产品极其危险、会摧毁就业”，导致公众普遍认为 AI 只会抢走饭碗而毫无益处。这种叙事直接给 Bernie Sanders 等政客递了刀子，促成了“没收科技巨头一半财富”的激进提案。 David Sacks 讽刺这种局面是巨头们应得的“愚蠢税”（stupidity tax）。他直言，作为 AI 的倡导者，自己已经厌倦了替这帮自相矛盾的高管辩护： - 如果你觉得自己正在造出“弗兰肯斯坦”（Frankenstein）怪物或者在养育“婴儿希特勒”，最合理的做法就是拔掉电源、掐死在摇篮里；而不能一边大肆建造怪物，一边向全世界哭诉自己阻止不了它。 - Anthropic 前一周刚发文警告称“递归自我改进（recursive self-improvement）可能导致世界末日，因此需要暂停研发”，而仅仅一个月前，他们就高薪聘请了 Andrej Karpathy 来主持递归自我改进相关的研发，充斥着双重标准。 - Ben Thompson 指出，Anthropic 抛出所谓“呼吁暂停”的公关博文，本质上是在为其限制用户使用 Claude Fable 进行 AI、机器学习及芯片设计等反竞争操作寻找借口。 --- ### [51:31](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=3091s) 主权财富基金与 AI 的边际成本重构 话题转向建立美国国家主权财富基金（Sovereign Wealth Fund）以及向 AI 巨头索取股权的可行性。Chamath 认为，借鉴加拿大和澳大利亚的养老金制度（Superannuation）思路建立由股权充资的主权基金理论上极具价值，现任政府手中其实已经通过各类产业扶持积累了涵盖关键金属、原材料以及英特尔、美光等半导体资产的组合。 Chamath 深入剖析了 AI 与传统互联网在商业底层经济模型上的本质区别： - **传统互联网**：边际生产成本几乎为零。谷歌或 Facebook 每新增一个搜索或社交用户，不需要增加边际服务器成本，却能通过网络效应无限售卖广告变现，堪称“无限印钞漏洞”。 - **AI 经济学**：边际成本极其沉重且真实存在。每一个新发起的并发请求都在消耗昂贵的 GPU 算力、电力能源和高带宽内存。 这种高昂的边际算力成本使得 AI 高度依赖国家级电网、基础设施与产业韧性支持。Chamath 打比方称，这就像当年联邦政府斥巨资修筑了全美州际公路系统，如果当时仅有两家运输公司跑在国家修筑的公路上垄断货物运输，政府完全有权要求持有其股权。基于这种关键基础设施的杠杆，政府完全有理由向 OpenAI、Anthropic 等巨头索要股权份额（“如果让我来谈，我最后甚至能谈下 75% 的股权”）。 --- ### [55:40](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=3340s) 预测市场 IPO 概率与对“资本家顺从者”的调侃 根据 Polymarket 预测市场上关于 2026 年底（即 2027 年前）各明星独角兽上市的概率押注： - **SpaceX**：100% 概率上市。 - **Anthropic**：83% 概率上市。 - **OpenAI**：48% 概率上市（节目播出前不久，OpenAI 已被披露进行了秘密递表申报）。 在对 Bernie Sanders“没收智能生产资料”的政策提问中，嘉宾打趣给出了一个支持收缴股权的标准框架：如果一家公司宣称自己是公益性公司（PBC），一边免费利用全人类的公共知识库进行模型训练，另一边大肆宣扬自己将引发人类大规模失业，并且搞技术壁垒拒绝回馈公众，那么国家完全应当对其征缴 75% 的股权。这群硅谷巨头就像“资本主义受虐狂/顺从者”（Capitalist Cucks），一边自己递上鞭子，一边任由政府将其财富与股权收编。 ### [56:52](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=3412s) 强制开源之争：AI 巨头的“被监管癖好”与算力电力现实 几位主持人围绕人工智能前沿模型是否应被强制开源展开了激烈交锋。大卫·萨克斯（David Sacks）指出，伯尼·桑德斯（Bernie Sanders）在此前专栏中提到的一个核心逻辑确实戳中了要害：这些前沿 AI 实验室几乎利用了全人类公开发布的所有知识与内容进行免费预训练，最终却对输出和模型能力层层设卡、建立壁垒，甚至以安全为名限制竞争对手与公众的使用。这种垄断姿态直接违背了最初的技术分享初衷。萨克斯回顾道，当年埃隆·马斯克（Elon Musk）联合创立 OpenAI 时，最初的动机正是担忧谷歌彻底垄断前沿人工智能，因此马斯克坚信唯一的制衡手段就是完全开源。时过境迁，现在的垄断者不仅是谷歌，以 Anthropic 为代表的新贵正积极游说政府实施严苛监管，试图通过合规壁垒构筑起双寡头垄断的“监管俘获”陷阱。 面对杰森·卡拉卡尼斯（Jason Calacanis）呼吁四位 Bestie 共同投资打造开源前沿实验室的想法，查马斯·帕利哈皮蒂亚（Chamath Palihapitiya）泼了冷水。查马斯直言，没有电力支撑，讨论开源前沿模型纯粹是空中楼阁；如果无法向数据中心稳定输送数百兆瓦（MW）乃至吉瓦（GW）级的庞大电力，所有的实验室雄心都是空谈。随后，杰森以极具讽刺意味的口吻嘲讽了 Anthropic 首席执行官达里奥·阿莫代伊（Dario Amodei）等人的行事风格，称他们是“资本主义受虐癖 / 绿帽奴”（Capitalist Cucks）： > “他们就像是有特殊的癖好，喜欢被人拿走股权、渴望被人严厉监管。达里奥就像在喊：‘噢，快拿走我的股权吧！快来监管我！桑德斯，拿走我的股权，我真是个坏男孩，我根本不值得被信任；伊丽莎白·沃伦，快来监管我！’” 在一番爆笑与调侃中，大家对比了 OpenAI 首席执行官萨姆·奥特曼（Sam Altman）与达里奥的处事之道。萨克斯认为，奥特曼虽然也面临不少争议，但他非常务实且精明，能看清政治风向并愿意坐下来务实谈判，这种灵活性远胜于道貌岸然的道德绝对主义；而 Anthropic 之所以能吸引一大批自视高尚、喜欢“道德标榜”（virtue signaling）的员工，正是因为达里奥契合了他们的意识形态。 --- ### [59:22](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=3562s) Liquidity 峰会回顾：Sarah Friar 与 Coatue 的百亿、万亿企业幂律数据 话题随后转入刚刚圆满落幕的 Liquidity 峰会。在讨论嘉宾表现时，杰森提到了 OpenAI 首席财务官莎拉·弗莱尔（Sarah Friar）的惊艳登场，甚至抛出猛料称会场私下有流言认为她能力卓越，完全有潜力成为未来的正式 CEO，而让奥特曼转任执行董事长。萨克斯和弗里德伯格（David Friedberg）立即打断杰森，警告他不要在播客上挑拨 OpenAI 管理层关系，但大家一致认同莎拉和萨姆目前配合默契，弗莱尔展现出了顶级高管的统治力。 在所有演讲中，Coatue 联合创始人托马斯·拉丰（Thomas Laffont）关于风险投资宏观生态的深度报告赢得了全场一致赞誉。拉丰通过严谨的数据分析揭示了资本周期的残酷幂律：在近期的周期中，一家估值达到 1000 亿美元的超级巨头，实现 10 倍增长迈向 1 万亿美元市值的概率，反而远远高于一家 100 亿美元的公司成长到 1000 亿美元。这一颠覆传统认知的结论让台下的资本配置者们深受震撼。 ```mermaid flowchart LR accTitle: 创业公司估值晋级概率漏斗 accDescr: 展示独角兽向十角兽、百角兽以及万亿巨头跃迁的概率跃升曲线 subgraph Funnel["估值跃迁与晋级概率"] direction LR U["独角兽 (Unicorn)<br>$1B"] -- "晋级率: 8%" --> D["十角兽 (Decacorn)<br>$10B"] D -- "晋级率: 13%" --> S["百角兽 (Senicorn)<br>$100B"] S -- "晋级率: 31%" --> T["万亿巨头 (Trillicorn)<br>$1T"] end ``` - **独角兽（Unicorn，$10 亿）晋级为十角兽（Decacorn，$100 亿）**：概率仅为 **8%**。 - **十角兽（$100 亿）跃升为百角兽（Senicorn，$1000 亿）**：晋级率接近翻倍，达到 **13%**。 - **百角兽（$1000 亿）跨越至万亿巨头（Trillicorn，$1 万亿）**：成功率暴增至 **31%**。 - **万亿巨头的推演**：萨克斯向拉丰求证，若将这一逻辑向万亿巨头进一步外推，未来几年内从 1 万亿美元冲刺到 10 万亿美元的概率可能高达 **60%**。 听到这些令人咋舌的估值，杰森与查马斯再度开启讽刺幽默，调侃称这些数据必定会让社会主义政客垂涎三尺：“我们只需要没收一家万亿美元公司 10% 的股权，就能还清美国 2% 的国债；要是直接没收 30 家万亿巨头的股份，联邦债务瞬间清零，甚至还能结余数百亿美元直接买下澳大利亚、古巴，或者干脆买下温哥华与多伦多来整蛊加拿大。” --- ### [01:02:11](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=3731s) 资本局幕后花絮：顶级名厨、Jake Paul 与 Liquidity 的极致精英定位 随后，四位主持人回顾了 Liquidity 大会的众多精彩花絮与赞助支持。由纽交所（NYSE）在米其林三星餐厅 The French Laundry 举办的演讲者晚宴格调非凡，传奇名厨托马斯·凯勒（Thomas Keller）亲自操刀。查马斯特别注意到凯勒手腕上的顶级腕表——一块配备复杂功能的百达翡丽（Patek Philippe）Aquanaut 玫瑰金绿带腕表，极显品味。席间众人品尝了标志性的招牌菜“生蚝与珍珠”（Oysters and Pearls，即木薯珍珠配生蚝与白鲟鱼子酱）、顶级宫崎和牛与特色鸭肉。 会场内同样亮点不断：安永（EY）赞助了极具创新性的商务洽谈枢纽；网红转型职业拳击手兼投资人的杰克·保罗（Jake Paul）以“新 Bestie”的身份亮相晚宴并坐在萨克斯身旁，其成熟专业的投资策略与随行团队给众人留下了极深印象；官方长寿健康合作伙伴 Niagen 则在现场搭建了充盈科技感的健康休息室，为参会者提供现场 NAD 静脉注射恢复体验。晚宴次日，大家还与 Athena 首席商务官克里斯·霍（Chris Hoe）进行了一场每洞 100 美元的 9 洞高尔夫让分对抗赛。 在节目尾声，杰森提到 Liquidity 收到超过 3,000 份申请，门票瞬间售罄，许多未能入场的人士怨声载道。查马斯随即阐述了 Liquidity 与即将在 9 月 13 至 15 日举行的 All-In Summit 之间的根本定位差异： - **All-In Summit**：定位于包罗万象的思想嘉年华，面向广泛受众，充满了狂欢氛围、多元跨界碰撞以及弗里德伯格天马行空的先锋舞台创意。 - **Liquidity**：定位则是极度精细化筛选的闭门高规格会议，严苛限制准入，旨在为全球最具影响力的资本配置者（Capital Allocators）搭建一个能够深度研讨、辩论、建立实质业务纽带的顶级平台。查马斯强调，正是资本的流向最终塑造了现实世界的走向，因此 Liquidity 必须保持这种绝对的高水准与排他性。 ### [01:05:39](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=3939s) 5月通胀数据爆表：CPI与PPI创三年新高，加息阴影再现 节目切入当天的突发经济重磅新闻：美国 5 月 CPI（消费者物价指数）与 PPI（生产者物价指数）数据全面超预期走热，通胀再度抬头。CPI 同比上涨 4.2%，创下自 2023 年 4 月以来的最高纪录；而衡量上游生产端与批发原材料价格的 PPI 同比涨幅更是高达 6.5%，创下 2022 年底以来的新高。预测市场 Polymarket 上的数据随即剧烈变动：2026 年通胀率触及 5% 的概率升至 21%，而美联储年内加息的概率飙升至 49%——要知道，在伊朗战争爆发前，这一概率还不足 10%。与此同时，欧洲央行（ECB）刚刚在周四加息 25 个基点，这是其自 2023 年 9 月以来的首次加息。 面对这一连串数字，主持团队展开了激烈讨论。被问及背后的驱动因素时，弗里德伯格（David Friedberg）言辞极为尖锐。他指出，伊朗战争导致的能源价格跳涨确实推高了核心指数，但这只是表象，真正的宏观根源在于政府支出的彻底失控。“我这话已经说过上百遍了，实在太愚蠢了，”弗里德伯格难掩愤怒地直言，这个国家贫富差距和通胀的根本核心全都在于政府过度开支，所有的地缘博弈和供应链波动不过是“网络边缘的枝节把戏”。国会议员为了拉拢选票，不断向选民许诺更多现世利益，将子孙后代的财富挥霍一空。在这一恶性循环的末期，美联储（尤其是传闻中的凯文·沃什可能执掌的美联储）隔夜利率完全有可能被迫升破 5.5% 甚至 6%。 - **CPI 表现**：5 月消费者物价指数同比上涨 4.2%，为 2023 年 4 月以来最高水平，租金、食品杂货、燃油和医疗成本压力全面反弹。 - **PPI 表现**：5 月生产者物价指数同比上涨 6.5%，为 2022 年底以来最高水平，原材料及批发端涨价压力巨大。 - **预测市场变动**：Polymarket 显示美联储年内加息概率升至 49%（战前不足 10%），2026 年通胀突破 5% 概率达 21%。 - **海外央行联动**：欧洲央行周四宣布加息 25 个基点，为 2023 年 9 月以来首次加息举动。 ### [01:08:27](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=4107s) 伊朗战争与能源连锁反应：中国储备缓冲与油价危机 针对通胀是否属于“暂时的”以及伊朗战争有无缓和出口，查马斯（Chamath Palihapitiya）深入分析了能源市场的供需暗流。查马斯认为通胀数据固然令人担忧，但目前全球油价之所以尚未冲上每桶 200 美元、依然被压制在 100 美元以下，很大程度上是因为中国一直在释放其庞大的原油储备平抑全球消费，同时国内外太阳能等无需繁琐环保审批、落地极快的实用能源产能被大幅提前拉动。然而，这种平衡极其脆弱：一旦中国的储备耗尽，被迫重新回到现货市场每天抢购 300 万桶原油，国际油价恐将迅速突破 100 美元，甚至暴涨至 150 到 200 美元区间。届时石油作为基础化工与运输原料的成本外溢，将引发灾难性的下游 CPI 全面暴涨。 查马斯强调，PPI 的飙升应当成为决策层迅速从伊朗僵局中寻找“下坡路（offramp）”的警报。萨克斯（David Sacks）对此表示赞同，但他补充了一个耐人寻味的盘面细节：当天纳斯达克指数反而逆势上涨了 2.5%。在以往出现通胀暴击时，对高利率最为敏感的科技股必然遭遇抛售，而当天市场的反常强势表明，该数据已在预期之内，且资本市场可能正在提前为某种潜在的停火与外交解决方案定价。杰森则表达了更强烈的忧虑，痛批伊朗战事是一场彻头彻尾的灾难性泥潭，如果不能在一年内迅速解决而演变成旷日持久的“永久战争”，美国大众预期的生活成本下降与黄金时代将彻底沦为空谈。 ```mermaid flowchart TD accTitle: 能源市场传导与通胀风险链路 accDescr: 展示伊朗战争爆发后，中国储备缓冲与未来油价暴涨风险的因果传导图 subgraph 当前缓冲机制 A["伊朗战争爆发"] --> B["能源供应端受扰动"] C["中国释放石油战略储备"] --> D["全球油价维持在100美元以下"] E["太阳能等实用能源产能快速上线"] --> D end subgraph 潜在恶化风险 D -.->|若储备耗尽并在现货市场日增购300万桶| F["国际油价飙升至150-200美元"] F --> G["全球下游制造与运输成本全面激增"] G --> H["CPI全面失控与美联储大幅加息"] end ``` ### [01:12:27](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=4347s) 洛杉矶市长初选计票疑云：晚到邮寄选票引发的选举公正性争议 节目的最后部分转向了在社交媒体 X 上引发轩然大波的加州洛杉矶市长初选计票问题。面对杰森关于“加州选举系统是否存在腐败、斯宾塞·普拉特（Spencer Pratt）的选举是否被窃取”的提问，弗里德伯格给出了极为讽刺且激进的回答：“根本就没有什么选举。你的选举权早就不复存在了，你只是那些指定你监督官的人的臣民——就像《黑客帝国》里的史密斯探员构建的虚假母体一样，所有官员不过是‘被指定’的。” 随后弗里德伯格摆出了一组令人震惊的统计数据对比： 在选举日当天的现场投票中，斯宾塞·普拉特以 35% 的得票率领先，现任市长凯伦·巴斯（Karen Bass）获得 29%，而极左翼候选人尼蒂亚·拉曼（Nithya Raman）仅获得 26%。而在选举日之前收到的早期邮寄选票中，巴斯占 38%，普拉特占 28%，拉曼占 20%。诡异的转折发生在此后陆续送达的晚期邮寄选票中：拉曼的得票率暴涨至 37%，巴斯为 35%，而普拉特的得票率骤降至 19%。 弗里德伯格强调，从纯数学统计角度看，普拉特在晚期邮寄选票中的支持率直接缩水了整整三分之一，而拉曼的得票率却戏剧性地飙升了 80%。更为离奇的是选票的地理分布图——拉曼这些关键的新增选票绝大多数异常集中在洛杉矶臭名昭著的游民聚集区“贫民窟（Skid Row）”。弗里德伯格澄清自己绝非习惯性的“选举否定论者”，但在任何常规的社会政治与统计学模型下，这种前后票仓的剧烈断层都显得不可思议。话题随即引申到了美国众议院行政管理委员会（House Committee on House Administration）近期发布的一份关于加州选举漏洞的重磅调查报告。 - **现场投票得票率**：斯宾塞·普拉特 35%，凯伦·巴斯 29%，尼蒂亚·拉曼 26%。 - **选前早期邮寄选票**：凯伦·巴斯 38%，斯宾塞·普拉特 28%，尼蒂亚·拉曼 20%。 - **选后迟到邮寄选票**：尼蒂亚·拉曼飙升至 37%（激增 80%），凯伦·巴斯 35%，斯宾塞·普拉特跌至 19%（暴跌三分之一）。 - **选票地理异常**：尼蒂亚·拉曼暴增的晚期选票高度扎堆在洛杉矶市中心的 Skid Row 区域，引发了关于协助收票（ballot harvesting）与操控选民的广泛质疑。 ### [01:15:28](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=4528s) 加州选举法变更：从民主选举走向“合法任命” 大卫·弗里德伯格（David Friedberg）回顾了相关报告对加州自 2018 年中期选举以来的立法演变分析，指出一系列旨在降低投票门槛的法律在客观上动摇了自由选举的根基。首先是加州议会通过的第 1921 号法案（AB 1921），该法案使无限制的“选票收割”（ballot harvesting）在全州范围内完全合法化。根据该规定，任何个人都有权从任何其他选民手中收集选票，无论彼此是否存在亲属或法定关系，并可代为填写与寄送。紧接着在一年半至两年后，加州立法将其永久化为向全州所有注册选民自动邮寄选票，每次直接寄出上千万张邮寄选票。与此同时，选民登记的门槛被大幅放宽，无需出具公民身份证明，甚至仅凭一张健身房会员卡即可完成登记。 ```mermaid flowchart TD accTitle: 加州选举机制演变路径 accDescr: 展示加州一系列立法如何将原本的普选机制转变为缺乏审计与身份验证的选票收割闭环 subgraph Legis["立法与选民登记"] A["AB 1921 法案<br/>选票收割合法化"] B["全量邮寄选票<br/>寄出2300-2400万张"] C["宽松选民登记<br/>无需公民身份证明/ID"] end subgraph Operation["流转与核验环节"] D["数百万闲置选票<br/>滞留公寓/错误地址"] E["选票代收与代填<br/>缺乏监管链与见证人资质库"] F["宽松比对<br/>签名匹配阈值低/无邮戳计票"] end subgraph Outcome["最终结果"] G["选票驱动型政治任命<br/>偏离'一人一票'初衷"] end Legis --> Operation Operation --> Outcome ``` 弗里德伯格强调，在整个过程中，接收和填写选票均不需要出示身份证件（Voter ID），也无法验证实际填票人是否就是选票所属的选民本人；个人批量收集成百上千张选票并送审也是完全合法的。因此，这种现象从技术上讲并非违法乱纪，恰恰相反，“系统完全是在按照既定设计运转”。其结果是只要掌握了对应的组织机制，就能在“自由选举”的名义下，将特定人选事实性地“任命”到政府公职中，而非真正的民主竞选。弗里德伯格直言这构成了对民主选举权利的根本性侵蚀，但他强调自己并非所谓激进的“疯狂 MAGA”人士，而是纯粹从成文法案的累加效应来看，这套机制背离了“一人一票、不投则不计入”的核心原则。 - **加州 AB 1921 法案**：取消对代收代交选票人员的关系限制，开启无上限的第三方选票收集。 - **全员邮寄**：全州数千万注册选民无论是否主动申请，都会自动收到邮寄选票。 - **无须严格身份证明**：选民登记缺乏公民身份核验机制，接受诸如健身房卡等宽松替代证明。 --- ### [01:17:49](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=4669s) 合法漏洞还是选举舞弊？“逆向选择”引发的争论 弗里德伯格的分析立刻在节目中引发了对“这究竟是不是欺诈（Fraud）”的交锋。杰森·卡拉卡尼斯（Jason Calacanis）指出，弗里德伯格的表述听起来就是在指控存在大规模舞弊。大卫·萨克斯（David Sacks）也认为弗里德伯格所描述的本质就是欺诈，只是系统给其开了绿灯。弗里德伯格则坚守法律定义进行反驳：严格意义上的欺诈必须包含违法行为；如果法律明文允许代收选票，只要不按选票计件领取报酬，这种行为在现行法下就是合法的。 弗里德伯格进一步借用保险学中的“逆向选择”（adverse selection）概念来剖析制度设计的悲剧。表面上看，每一项法律在出台时都带有某种利他主义的包装——提高选民参与度、保障投票权、增加便利途径，让民主更加触手可及。然而正如黑客会寻找网络漏洞一样，一旦系统出现可钻的缝隙，恶意行为者与利益集团必然会不对称地将其利用到极致。萨克斯随即介入并将矛头指向洛杉矶的选举结果（如市议员席位争夺中候选人得票曲线的戏剧性翻转），直言这就是组织严密的作弊运作（"She had a better cheating operation"），并质问为何计票拖延整整一周后排名会离奇倒转。 --- ### [01:19:28](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=4768s) 萨克斯拆解收割套路：脏选民册、40% 签名匹配与缺乏监管链 萨克斯详细拆解了他眼中的这套操作链条，表示无论这在加州法律层面上被定性为合法还是非法，“大家都心知肚明这是见不得人的手段”。加州总共向 2300 万至 2400 万名注册选民邮寄选票，但实际投票者通常只有 950 万人左右，这意味着有 1400 万张选票不知所踪。萨克斯引用洛杉矶一位公寓管理经理的亲身经历称，成堆的未认领选票被长期弃置在公共邮箱旁，收件人甚至包含早已搬离多年的前租客，反映出选民花名册极其混乱，多年未作清洗。 在选票验证机制上，萨克斯列举了多项荒唐的漏洞：投票现场既不需要身份证，甚至连选民登记都不查 ID；邮寄选票上的签名经机器扫描时，只要匹配度达到 40% 就会被系统放行通过；如果选票附有“见证人”，选民本人甚至只需划一个简单的叉（X），而加州根本没有设立见证人登记机制或身份数据库，任何人随手画个波浪线就能充当合法见证人。在毫无监管链（chain of custody）的情况下，选票收割人员完全可以大肆填票、代签并集中投递。萨克斯进一步提到，加州州长加文·纽森（Gavin Newsom）签署法案让事后审计变得几乎不可能；而在加州，从医疗补助（Medicaid）、失业金、EBT 食品券到临终关怀以及“无家可归工业复合体”已被证实存在数十亿美元的系统性欺诈，利益集团绝不可能放过没有邮戳亦可计票、从空无一床的避难所批量登记选民等选举规则漏洞。 - **1400 万张闲置选票**：寄出的 2300-2400 万张选票中，仅约 950 万人正式投票，余下数百万张成为流动隐患。 - **40% 签名核验门槛**：机器比对签名的合格标准低至 40%，且见证人签名无需任何资质审核。 - **缺失的监管与审计**：延期送达无邮戳选票仍被计入，缺乏全流程监管链，限制了事后独立审计。 --- ### [01:23:04](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=4984s) 诉讼现实与质疑：成千上万张选票真能神不知鬼不觉？ 对于萨克斯和弗里德伯格勾勒的暗箱操作图景，杰森提出了质疑与反思。杰森指出，如果这属于欺诈，按照常理应当存在大量的法律诉讼和司法调查，每次选后州总检察长以及各方竞选团队都会提起诸多民事诉讼，例如特朗普曾发起过 60 起选举诉讼。萨克斯反驳称特朗普并未在加州提起诉讼，而杰森则解释那是因为加州是深蓝州，特朗普输得太惨以至于没有起诉的实际政治价值。 杰森转而向查马斯·帕里哈皮蒂亚（Chamath Palihapitiya）提问，如果真如网传视频（例如网红尼克·雪莉 Nick Shirley 探访洛杉矶 Skid Row 流浪汉聚集区，指控有人付钱给吸毒者和神志不清的人换取选票）所展示的那样存在如此明显的疑云和反常统计数据，那么要伪造或非法操控成千上万张选票，势必需要牵涉庞大的合谋网络，理应极易被执法部门顺藤摸瓜当场抓获。对话由此进入深入剖析选举舞弊的可行性与查处难度。 ### [01:24:21](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=5061s) 加州选举机制的制度化弊端与贫民区选票丑闻 关于加州选举中是否存在操纵选票的争议持续升温。针对数千乃至上万张选票被逆转的说法，现场讨论了这究竟需要多大规模的共谋。有人质疑去洛杉矶贫民区（Skid Row）找人作证近乎不可能，但事实上早已有录像公开，甚至进入了司法程序。涉案人布伦达·李·布朗·阿姆斯特朗（Brenda Lee Brown Armstrong）已于 2026 年认罪，承认曾以 2 至 3 美元一票的代价向贫民区的吸毒人员收购选票。这证明当地的制度设计存在致命漏洞，甚至在客观上为欺诈大开绿灯。 ```mermaid flowchart TD accTitle: 选票收购与法律模糊地带流程图 accDescr: 展示了从低成本收集选票到利用宽松法规使潜在欺诈合法化的演变过程。 subgraph S1["草根操作与收购"] A["贫民区低收入人群 / 街头选民"] -->|"2至3美元报酬"| B["中介/操盘手收购选票<br/>(如布伦达·阿姆斯特朗案)"] end subgraph S2["制度与法律漏洞"] B --> C["公寓大堂随意丢弃的无主选票"] C --> D["选票收割 (Ballot Harvesting)<br/>与极度宽松的身份核查"] end subgraph S3["政治结果与垄断"] D --> E["单一党派通过立法<br/>将灰色操作合法化"] E --> F["形成不可撼动的单向权力垄断"] end ``` 查马斯·帕里哈皮蒂亚（Chamath Palihapitiya）对此做出了极具冲击力的剖析。他直言，加利福尼亚州本质上已经变成了一个专制体系（autocracy），由同一个政治机器牢牢掌控。无论是市长、州众议员、州参议员还是州长，政客们通过不断修改法律，形成了一扇只进不出的“单向门”，确保该机器的绝对垄断。这种体制的悲剧在于，当权者越来越缺乏治理如此多元复杂大州的能力，而民众长期的默许与不作为，注定需要数十年时间才能让这种体制性腐败彻底暴露。查马斯指出，大卫·弗里德伯格（David Friedberg）和大卫·萨克斯（David Sacks）其实都没说错：弗里德伯格指明了法律已被篡改，将本属非法的行为变成了合法；而萨克斯则道破了任何理性之人在审视这些规则时，都会得出其充满欺诈的结论。 - **贫民区选票交易**：法庭记录显示，存在以每人 2 到 3 美元现金或香烟、食品换取选票签名的真实案例。 - **政治机器的自我固化**：通过制定极其宽松的选票收集法规，将原本涉嫌违规的操作合法化，从而构筑起竞争对手极难打破的选举护城河。 --- ### [01:27:14](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=5234s) 破局的“最后按钮”：紧急状态与联邦立法的呼声 面对看似无法通过常规投票撼动的加州体制，查马斯透露了一个“打破玻璃按警报”的激进方案：若想扭转系统性溃败，必须有一位非建制派人士赢得加州最高行政职位。他点名史蒂夫·希尔顿（Steve Hilton），认为如果希尔顿能在州长选举中登顶，上任第一天就可以合法宣布全州进入“紧急状态”（State of Emergency），并利用紧急状态赋予的行政权力对加州的各项行政与选举体制进行大刀阔斧的重构与清理。 杰森·卡拉卡尼斯（Jason Calacanis）则呼吁回归基本常识。他认为，疫情期间普遍邮寄选票的做法在今天已经毫无必要，要求出示有效身份证件（Voter ID）是理所应当的门槛——如果一个人连申办证件都懒得做，那他或许根本不该投票。虽然杰森认为传统基金会（Heritage Foundation）耗费数千万美元追踪的选举舞弊极少能在大选中逆转整体结果，但萨克斯随即激烈反驳，搬出历史上的肯尼迪对阵尼克松（芝加哥黑手党操纵选票争议）以及林登·约翰逊（Lyndon Johnson）早期通过选票欺诈起家的政坛丑闻，证明边缘操作完全能够颠覆重大历史走向。 ```mermaid flowchart LR accTitle: 国会规范选举与恢复信任路径 accDescr: 展示国会如何通过立法设定投票门槛，进而重建公众对民主选举的信任。 A["国会统一立法"] --> B["公民资格验证"] A --> C["必须持有带照片的官方ID"] A --> D["禁止随意选票收割"] B & C & D --> E["一人一票、合法合规"] E --> F["恢复公众对选举结果的信任"] ``` 随后，弗里德伯格将矛头直指国会的失职。他严厉抨击当前的荒谬现状：在美国，开车需要驾照，不纳税联邦政府一清二楚，办理任何基本社会事务都需要严格的身份验证；然而在关乎国家基石的投票环节，某些地方居然允许用健身房会员卡来登记选票，甚至允许任何人从公寓大厅收割成堆的无主选票。弗里德伯格强调，如果国会不通过联邦法案强制执行公民身份与有效证件核查，民众对选举真实性的信仰将彻底瓦解，而一旦民众不再相信体制的公正，社会局势必将走向极其危险的深渊。 --- ### [01:31:30](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=5490s) 选民身份法案与“社会主义试验”的代价 加州即将迎来的“选民身份识别”（Voter ID）公投提案，被萨克斯视为拯救加州合法性的“最后一根救命稻草”。针对杰森调侃萨克斯听起来像个“选举否定者”（election denier），萨克斯毫不客气地反唇相讥，指出民主党全党在 2016 年特朗普胜选后，编造出长达数年的“通俄门”（Russiagate）谎言，才是真正的大规模选举否定者。 萨克斯进一步亮出核心论据：现场统计数据中，现场亲自投票（walk-in）与邮寄选票（mail-in）之间呈现出无法解释的巨大断层与背离。从纯粹的统计逻辑来看，这种数据分布根本说不通，绝非编造的事实。 萨克斯情绪激动地澄清，自己提出质疑并非因为厌恶某位候选人，甚至不介意洛杉矶选出一个社会主义市长： - **直面社会主义的代价**：萨克斯直言，不管是洛杉矶、西雅图、纽约还是明尼苏达，如果当地选民偏要拥抱社会主义政客，那就让他们去承受苦果。就像当下的拉美国家正在经历的觉醒一样，唯有让民众在现实中体会到社会主义如何剥夺经济流动性、摧毁建国之本的个人自由，人们才会彻底看清其空想本质。 - **数据背后的民意扭曲**：萨克斯坚称，自己愤怒的焦点在于，眼前的选举数据清楚地表明，这根本不是选民真实的投票意愿，而是规则漏洞与体制操纵制造出的扭曲结果。 ### [01:33:15](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=5595s) 选举合法性争议与“拒绝相信双眼”的指控 对话围绕加州选举制度的扭曲进一步白热化。主持人感叹，当前的政治体制本质上已经沦为一种“指定体系（system of appointment）”，而非真正的“选举体系（system of election）”，完全脱离了选民真实的意愿。David Sacks 引用乔治·奥威尔《1984》中的经典名句激烈回应：“党要求你拒绝相信双眼和双耳所看到的证据，这是他们最终且最根本的命令。”他坚称只要看一眼图表和数据，就能明白发生了极度不正当的事情，从统计学角度看那是完全不可能的。 在 Sacks 看来，当下唯一的悬念只在于这属于“非法舞弊（illegal fraud）”还是“合法操纵（legal fraud）”——因为加州政客在法律中留下了太多漏洞和腐败空间，导致这种行为在法律条文上甚至可能是被允许的。但他强调，即便法律允许，这也绝对是不折不扣的腐败和荒唐。Sacks 直言不讳地表示：“就算他们骂我是‘选举否定者’（election denier），我个人根本不在乎。我就是否认！Spencer Pratt 本应该进入决选，我绝不承认 Nithya Raman 的获胜是合法的。” ```mermaid flowchart TD accTitle: 加州选举结果争议归因分析 accDescr: 展现主持团队针对加州初选结果反常现象的两种核心归因逻辑 subgraph Legality["操纵性质判定"] A["选后统计出现万亿分之一的小概率反转"] --> B{"法律性质界定"} B --> C["非法舞弊 (Illegal Fraud):<br/>收买流浪汉/非法伪造选票"] B --> D["合法操纵 (Legal Fraud):<br/>选票收割机制/系统性法律漏洞"] end subgraph Institutional["体制与媒体反应"] D --> E["加州签署法案<br/>阻碍联邦审计选民名册"] C --> F["主流媒体拒绝追问真相<br/>避免被贴上亲特朗普标签"] end ``` ### [01:34:10](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=5650s) 呼吁联邦介入审计与媒体失职的深层根源 面对统计异常，Jason Calacanis 指出数据本身已经在大声疾呼司法介入，加州总检察长以及联邦司法部应当彻底展开调查。针对 Nick Shirley 等独立报道者在洛杉矶 Skid Row 街区拍到的涉嫌操纵游民投票的现象，Jason 喊话唐纳德·特朗普总统，既然特朗普控制着联邦司法部，并且派任命的司法部官员到处出击，就应当动用手中的联邦权力去审计加州选举、查明真相。Sacks 随即反驳称，加州州长加文·纽森（Gavin Newsom）刚刚签署了一项法案，让联邦政府介入审计选民名册变得更加困难。 与此同时，Sacks 强烈抨击主流媒体的掩盖行为。他质疑媒体不仅不提出尖锐问题、追问为何选票在选举日后被发现时会出现如此离谱的翻盘，反而急于洗白说“没有任何异常发生”。Chamath Palihapitiya 剖析了媒体的心态：媒体仍然沉溺在过去的影响力幻觉中，但当他们有机会重建公信力时却选择了退缩。因为在极度极化的环境下，记者一旦调查 Spencer Pratt 为什么会输，就会被外界打成“默许支持特朗普”的阵营；为了避免得罪同温层并讨好既有政治机器，媒体宁可闭上眼睛宣称一切正常，最终自己也将被这台政治机器无情碾压。 - **媒体极化困局**：调查选举公正性往往被粗暴贴上“MAGA 阵营”的标签，导致严肃的数据核查与统计质疑无法正常展开。 - **统计翻盘概率**：席间指出，在选举日之后计入的选票中出现如此悬殊的单向倒戈，其在自然统计学上的概率仅为“万亿分之一”。 ### [01:38:30](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=5910s) 选票收割还是战术投票？推演反转背后的理论逻辑 为了保持讨论的严谨性，Jason Calacanis 多次要求全场提出能够解释该异常现象的“钢人论点（steel man argument）”。面对 Sacks 质疑其为何非要维持“加州选举完全没有腐败”的预设立场，Jason 解释他并非持此立场，而是希望进行智识诚实的推演。他率先提出自己的假设：Nithya Raman 的地面动员机制（ground game）远比首次参选的体制外网红 Spencer Pratt 更加成熟和职业化，而后者根本没有建立能够有效收缴法定选票的竞选机器。 David Friedberg 随后补充了社交媒体上流传的另一种战略投票理论： - **民主党内部选民博弈**：反对 Spencer Pratt 的选民为了确保将他彻底淘汰在决选大门之外，一直在观望局势并推迟寄出选票。 - **决选保底策略**：最后时刻看到民调显示 Karen Bass 稳赢后，反 Pratt 选民集体将剩余选票集中投给排名靠后的 Raman，以此实现两位民主党候选人包揽决选前两名的目标，形成“双保险”。 - **选票收割者的集中行动**：Sacks 对此反驳称，普通选民不可能自发完成如此高度协调的操作，这一过程即便真的存在，也是由专业的“选票收割者（ballot harvesters）”在幕后集中操盘实施的。 ### [01:40:43](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=6043s) 节目录制尾声与经典谢幕 辩论在一片对“代收选票并代填提交在法律上是否荒谬”的感慨中走向尾声。Jason Calacanis 总结认为，尽管目前尚未取得确凿的违法证据，但数据面极其可疑，依然寄希望于联邦司法部展开彻底追查。此时，腹中饥饿的嘉宾们纷纷表示必须结束录制，互道“我爱你们”后愉快告别。片尾穿插了播客标志性的片头梗、粉丝二创音乐以及经典的幕后闲聊花絮，四位 Besties 宣布本期《All-In Podcast》正式收官。 ## Notable quotes > [02:11](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=131s) Anthropic 确实说了大实话，只是这句实话当你咽进肚子里后，胃里会极度不舒服、恶心得要命。 — Chamath Palihapitiya > [04:47](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=287s) And what are the best open source models today? They're Chinese. — David Friedberg > [32:32](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=1952s) 谁来监督监督者？究竟是谁授权让达里奥·阿莫代伊统领这一切的？ — David Sacks > [37:53](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=2273s) AI 是公共资源，你应该拥有它的一半。人工智能的基础是我们的集体和人类智慧——那些被世界上最富有的一批人实质上窃取的书籍、歌曲、新闻报道、科学研究和代码。 — Bernie Sanders > [39:31](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=2371s) 世界上没有什么比用 Anthropic 一半的利润来偿还美国国债更符合公共利益的事了。 — David Sacks > [56:52](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=3412s) 他们就像是有特殊的癖好，喜欢被人拿走股权、渴望被人严厉监管。达里奥就像在喊：‘噢，快拿走我的股权吧！快来监管我！桑德斯，拿走我的股权，我真是个坏男孩，我根本不值得被信任；伊丽莎白·沃伦，快来监管我！’ — Jason Calacanis > [01:05:39](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=3939s) 我这话已经说过上百遍了，实在太愚蠢了。这个国家贫富差距和通胀的根本核心全都在于政府过度开支，所有的地缘博弈和供应链波动不过是网络边缘的枝节把戏。 — David Friedberg > [01:12:27](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=4347s) 根本就没有什么选举。你的选举权早就不复存在了，你只是那些指定你监督官的人的臣民——就像《黑客帝国》里的史密斯探员构建的虚假母体一样，所有官员不过是‘被指定’的。 — David Friedberg > [01:33:15](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=5595s) 党要求你拒绝相信双眼和双耳所看到的证据，这是他们最终且最根本的命令。 — David Sacks > [01:33:15](https://www.youtube.com/watch?v=gH4FTjDm9FQ&t=5595s) 就算他们骂我是‘选举否定者’（election denier），我个人根本不在乎。我就是否认！Spencer Pratt 本应该进入决选，我绝不承认 Nithya Raman 的获胜是合法的。 — David Sacks ## Mentioned Anthropic, Fable 5, Mythos, Claude, Opus 4.8, Dario Amodei, Jason Calacanis, Chamath Palihapitiya, David Friedberg, David Sacks, Nikesh Arora, Palo Alto Networks, Ohalo, Arc Institute, Patrick Collison, John Collison, Ben Thompson, Stratechery, Meta, Llama, Mark Zuckerberg, Elon Musk, xAI, Grok, Sam Altman, OpenAI, Andrej Karpathy, Bernie Sanders, Donald Trump, Steve Bannon, Sarah Friar, Thomas Laffont, Coatue, Liquidity 峰会, The French Laundry, Thomas Keller, Jake Paul, Niagen, Polymarket, Kevin Warsh, Spencer Pratt, Karen Bass, Nithya Raman, Brenda Lee Brown Armstrong, Steve Hilton, Gavin Newsom, Nick Shirley, 国际基因合成联盟 (IGSC) ## Video description (0:00) Besties are back! (0:19) Anthropic gets massive backlash over secret Fable nerfing and privacy concerns (29:16) The AI regulatory capture trap, pragmatic safety solutions (37:59) Nationalizing AI: Trump/Sanders, justifications, and AI's "Capitalist Cucks" (59:22) Liquidity recap: Best moments and takeaways (1:05:39) Inflation heats up: CPI and PPI see 3+ year highs (1:12:27) California's loose election laws creating integrity doubts Apply for Summit 2026: https://allin.com/events 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/@theallinpod 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://www.anthropic.com/news/claude-fable-5-mythos-5 https://x.com/Scobleizer/status/2064641097310335294 https://x.com/GergelyOrosz/status/2064618497150210391 https://x.com/cremieuxrecueil/status/2064433331970720187 https://x.com/Yuchenj_UW/status/2064524668208545955 https://stratechery.com https://x.com/peter_szilagyi/status/2064620043896291671 https://darioamodei.com/post/policy-on-the-ai-exponential https://screendna.org https://x.com/DavidSacks/status/2065120386660880765 https://www.nytimes.com/2026/06/01/opinion/artificial-intelligence-bernie-sanders.html https://polymarket.com/event/ipos-before-2027 https://polymarket.com/event/how-high-will-inflation-get-in-2026 https://polymarket.com/event/fed-rate-hike-in-2026 https://x.com/robbystarbuck/status/2063602942637158423 https://www.justice.gov/opa/pr/california-man-pleads-guilty-orchestrating-270m-medication-reimbursement-fraud-scheme https://www.dol.gov/newsroom/releases/osec/osec20260218 https://www.secretservice.gov/newsroom/behind-the-shades/2025/05/secret-service-cracks-down-ebt-fraud-southern-california-sweep https://www.justice.gov/usao-cdca/pr/8-arrested-health-care-fraud-takedown-including-owners-hospices-billed-taxpayers https://www.foxnews.com/us/california-man-arrested-allegedly-stealing-millions-homeless-funds https://x.com/californiapost/status/2064362900098048386 #allin #tech #news [Source](https://www.youtube.com/watch?v=gH4FTjDm9FQ)

[2:11] Chamath Palihapitiya: Anthropic 确实说了大实话，只是这句实话当你咽进肚子里后，胃里会极度不舒服、恶心得要命。 [4:47] David Friedberg: And what are the best open source models today? They're Chinese. [32:32] David Sacks: 谁来监督监督者？究竟是谁授权让达里奥·阿莫代伊统领这一切的？ [37:53] Bernie Sanders: AI 是公共资源，你应该拥有它的一半。人工智能的基础是我们的集体和人类智慧——那些被世界上最富有的一批人实质上窃取的书籍、歌曲、新闻报道、科学研究和代码。 [39:31] David Sacks: 世界上没有什么比用 Anthropic 一半的利润来偿还美国国债更符合公共利益的事了。 [56:52] Jason Calacanis: 他们就像是有特殊的癖好，喜欢被人拿走股权、渴望被人严厉监管。达里奥就像在喊：‘噢，快拿走我的股权吧！快来监管我！桑德斯，拿走我的股权，我真是个坏男孩，我根本不值得被信任；伊丽莎白·沃伦，快来监管我！’ [65:39] David Friedberg: 我这话已经说过上百遍了，实在太愚蠢了。这个国家贫富差距和通胀的根本核心全都在于政府过度开支，所有的地缘博弈和供应链波动不过是网络边缘的枝节把戏。 [72:27] David Friedberg: 根本就没有什么选举。你的选举权早就不复存在了，你只是那些指定你监督官的人的臣民——就像《黑客帝国》里的史密斯探员构建的虚假母体一样，所有官员不过是‘被指定’的。 [93:15] David Sacks: 党要求你拒绝相信双眼和双耳所看到的证据，这是他们最终且最根本的命令。 [93:15] David Sacks: 就算他们骂我是‘选举否定者’（election denier），我个人根本不在乎。我就是否认！Spencer Pratt 本应该进入决选，我绝不承认 Nithya Raman 的获胜是合法的。

[2:11] Chamath: One thing I'll say about Anthropic is they tell the truth. It's just that the truth sucks when you actually take it and you eat it and you're like... 'No, this is not good. I don't like this.' [28:16] That is not even some basic level of KYC. That is just a credit card and a [sh*t] email address. [28:16] The user is a VC and podcaster asking about fertilizer bomb regulations which could stem from a legitimate research interest... [28:16] The idea that you can quote regulate or downscale or turn off AI is not a realistic idea [28:16] J-Cal: I just got downgraded again [28:16] Bernie Sanders: AI is a public resource. You should own half of it [28:16] Bernie Sanders: stolen by some of the wealthiest people in the world. [28:16] David Sacks: straightup confiscation of property [28:16] David Friedberg: There is no job loss with AI. I will say it again. [47:26] crock of [ __ ] [53:40] Chamath Palihapitiya: The incremental cost of performing AI is excessive and large. That contrasts and compares to the incremental cost before AI which was zero. Do with that information what you will. [89:50] Socialism is wrong. It's broken. It doesn't work. And if you all want to go through that lesson, do it. Go for it... But the numbers on that election data show me that that's not what the people wanted. [100:50] Host: I love you guys, but I'm hungry. I got to go. [100:50] Speaker: all just get a room and just have one big huge orgy

[0:00] All right, everybody. Welcome back to [0:01] the number one podcast in the world. [0:02] It's your favorite podcast. It's your [0:04] podcasters's favorite podcast. It's the [0:06] podcast that everybody likes to rip off [0:08] for 50 episodes and then they quit. But [0:10] the All-In podcast is not quitting. [0:13] We're doubling down with the original [0:15] quartet. Everybody's here for a big big [0:17] week, lots of debates, and we got to [0:20] start with Anthropic. [0:22] Anthropic released a mythos level model [0:26] with some interesting guardrails and uh [0:30] seems as though Daario is back to [0:32] blogging. Okay, the model's called Fable [0:35] 5. No idea why they're calling it that [0:38] or released on Tuesday and it tops every [0:41] benchmark nearly every benchmark. The [0:43] tokens, however, cost twice as much as [0:47] Opus 4.8, but it should use less tokens [0:50] overall because it's theoretically that [0:52] much better. Unclear if this is going to [0:56] alleviate some of the hand ringing [0:58] around the cost of tokens, but you [1:00] remember in April Anthropic [1:02] famously did not release Mythos because [1:05] publicly uh it had some strong hacking [1:07] capabilities. So they gave it to some [1:09] folks and we had um the CEO of Nesh of [1:13] Palo Alto Networks who said, "Hey, [1:15] Mythos was the real deal." and that they [1:17] used it to seal up some vulnerabilities [1:20] inside their shop. Obviously, this new [1:23] model is sensitive uh to topics like [1:25] boweapons, hacking, all that stuff are [1:28] blocked in it. But they got a big [1:30] developer backlash and then I'll get [1:32] your feedback after I explain this. [1:34] While using Fable, Anthropic stores all [1:36] the prompt data you enter for at least [1:38] 30 days. And so that's a bit of a [1:41] privacy issue. And if Fable 5 detects [1:43] you're doing frontier AI research, in [1:45] other words, using their model to make a [1:47] better model, competitive model, they [1:49] were downgrading you, but they weren't [1:51] telling you. And this was buried deep in [1:53] a 319 page document. So many kuruffles, [1:58] freakouts going on on X over this, but [2:01] they've since walked it back a bit, [2:04] quote unquote, in Wired. We're changing [2:06] Fable 5 safeguards for Frontier LLM [2:09] development to make them more visible. [2:11] I'm going to stop there. Chimath, what's [2:13] your take on anthropic and these [2:16] extraordinary models? They keep dropping [2:18] and how they're handling the release of [2:20] them. Are they being thoughtful? Are [2:23] they being dramatic and drama queens? A [2:26] little bit of both. Where do you stand [2:28] on it? You know, after getting Nesh's [2:29] feedback and seeing the latest model [2:31] come out, have you played with it? [2:32] What's going on at 8090 and the software [2:35] factory in terms of benchmarking it as [2:37] well? It's a really incredible model and [2:40] so kudos to these guys for continuing to [2:43] push the boundaries of the closed [2:46] frontier lab models. They're firing on [2:48] all cylinders. I think that it creates a [2:53] pretty obvious risk now and I think that [2:56] obvious risk is twofold. One is [2:59] Anthropic has essentially shown their [3:02] hand which is that they will [3:05] increasingly [3:08] take in prompts, evaluate the prompts [3:12] and decide what to do with them before [3:14] they generate output to you. [3:17] >> And I think if you were a person, [3:21] you should generally now think there's a [3:24] risk of censorship. [3:26] If you're a company, I think it's almost [3:29] a non-starter. [3:31] And the reason is because you could [3:33] accidentally trip one of these things [3:36] without even knowing it. A downstream [3:38] scientist using the cloud APIs could [3:41] trip it. A business executive inside [3:43] your corporation could trip it. A person [3:45] doing scientific molecular research [3:47] could trip it. And all of a sudden, [3:49] you'll get cut off from a very important [3:53] source of business differentiation for [3:55] yourself. So I think if you take both of [3:58] those two things together, we're at this [4:00] very unique moment in time where I think [4:02] companies need to start underwriting [4:05] this next phase of AI, which is how do I [4:09] have control? [4:11] Who am I allowing to learn off of all of [4:13] this information? [4:15] Do I want to have single point of [4:17] failure risk with respect to AI? And I [4:21] think the answer is that you need broad [4:23] diversity and a governance approach [4:26] that's better managed. [4:27] >> One thing I'll say about anthropic is [4:30] they tell the truth. [4:32] >> Yes. [4:32] >> It's just that the truth sucks when you [4:34] actually take it and you eat it and [4:37] you're like it's in your tummy and [4:39] you're like no this is not good. I don't [4:40] like this. And so there's the censorship [4:43] risk on the one hand and then there's [4:44] just the governance business risk for [4:45] enterprises on the other. Both are not [4:47] good. [4:47] >> Freeberg running your own company now. [4:49] Uh, and I'll let you back clean up here. [4:51] Saxs, obviously you'll have a lot to [4:52] say, but Freeberg, running your company, [4:55] do you worry about getting rugpulled by [4:58] one of these companies, investing your [5:00] time into one of these platforms, having [5:02] them then constrain your use of it and [5:04] or take what you're doing at Ohio, put [5:08] it into their model and make it [5:09] available to other people. What What [5:11] concerns do you have as a business owner [5:13] at the forefront of your field using [5:15] these tools? [5:15] >> It's a great question. the the terms of [5:17] service is pretty clear that they can't [5:19] take what you're doing and put it into [5:21] their model that they won't [5:22] >> do you believe them when you read that I [5:24] >> I'm not sure to be honest there are [5:26] things I'm concerned about but we've [5:27] decided to throw caution to the wind [5:29] because we do a lot of proprietary work [5:31] in genomics what that means is we're [5:34] looking at different genes or gene [5:36] variants trying to estimate the impact [5:38] that that gene or gene variant might [5:40] have on a living organism like we work [5:43] in plants obviously in agriculture [5:46] And we will do things like RNA guide [5:49] design for gene editing based tools. We [5:51] will do predictions on what gene may or [5:55] may not have some phenotype. So there's [5:57] a lot of genomic modeling work that we [6:00] have found these models to be incredibly [6:02] valuable at supporting us with over the [6:05] last couple of years. So going back [6:07] about 6 months, we were able to very [6:10] simply and cleanly do things like design [6:13] a genetic construct that you would then [6:15] use to make a specific protein and the [6:18] the tools was very easy to use to do [6:21] that sort of research and extremely [6:23] valuable. This is the kind of work that [6:25] many scientists would spend a lot of [6:26] time doing and these models were very [6:29] good at doing very quickly. Over the [6:32] last couple of weeks, they've begun to [6:34] restrict the ability to use the models [6:36] to do that work. And the premise is that [6:38] there's some sort of bioweapon [6:41] uh type risk that folks have theorized [6:43] could happen using these tools. And as a [6:46] result, we're losing the capacity to use [6:48] these models for this very important [6:50] scientific development work. As a [6:53] result, we are likely going to end up [6:55] needing to use open- source models and [6:57] run them locally ourselves. So the [6:59] reason I walk through all of that is so [7:01] people can really understand the context [7:02] of what's going to happen here. As folks [7:05] like Anthropic say, "Hey, we're going to [7:07] restrict access or censor the output of [7:09] these models." It is going to force [7:11] companies like ourselves who still want [7:13] to take advantage of the capability of [7:15] these LLMs to go and get open- source [7:18] tools and run them. And what are the [7:20] best open source models today? They're [7:22] Chinese. [7:22] >> Yeah, they are. [7:23] >> And that is a major concern. the the [7:25] American open source models are not as [7:27] good as the Chinese open source models. [7:29] So the restrictions that Anthropic and [7:31] others are putting upon themselves and [7:32] upon the industry is forcing a lot of [7:35] companies to go and get open source [7:37] Chinese models and run them. We're [7:39] seeing this across the landscape with [7:40] startups with large scale enterprises. [7:43] Everyone's making that move. [7:44] >> Yeah. And this is something we've been [7:46] talking about here with Apple and their [7:48] silicon and how well you can run these [7:50] models. Also, I predict your next card [7:53] you'll turn over is you'll start making [7:55] your own models, Freeberg. You'll take [7:56] all this data you have. [7:57] >> It's exactly right. And we'll we'll [7:58] start we'll start with the core model, [8:00] combine it with our data, and then we'll [8:02] have our own genome language model, our [8:04] own prediction model that we'll then use [8:05] internally. And I think that's where [8:07] folks are going. But the reason I point [8:09] this out is because a lot of folks are [8:11] feeling the pressure from anthropic [8:13] doing this. and they're feeling the [8:15] pressure from politicians repeating the [8:18] scary words that are being said by [8:20] Daario and others. And as a result, they [8:23] are going to try and they are already [8:25] trying to force either natural [8:27] enforcement or politicized enforcement [8:30] upon the model providers in a way that [8:32] is ultimately going to benefit Chinese [8:35] open source model providers. And that is [8:37] a scary thing because we are going to [8:38] damage our own kind of economic [8:39] viability. You can't just stop AI. As [8:42] much as everyone says AI is doomsday, by [8:44] stopping AI or trying to stop AI through [8:46] this political action and social, you [8:48] know, kind of behavior, you are [8:50] fundamentally going to give someone else [8:52] the advantage because the AI isn't going [8:53] to go away. So the reason I describe [8:55] what we're doing is for everyone to [8:56] really understand in Grock that you [8:58] can't just turn off AI and turn off [9:00] access to these things. What you will do [9:02] is you will force the hand of someone [9:03] else to now have an unfair advantage [9:06] because you're unfairly restricting your [9:08] models and your access to those models. [9:09] Well, well, really well said, Freeberg [9:11] Saxs. What's your take just in terms of [9:14] the chessboard in terms of we got the [9:18] business case here, we got the local [9:19] case, but there's a lot of issues here [9:22] with Anthropic in terms of them [9:26] essentially throwing up every red flag [9:29] to get regulated to induce and it'll be [9:32] our second story, the Bernie Sanders [9:34] seizing the equity of these companies. [9:37] But man, th this is saying, "Look over [9:40] here. We're starting a fire. Regulate [9:42] the hell out of us." Which also then [9:44] makes you wonder, hey, maybe I should be [9:46] doing my own model. Maybe I should be [9:48] embracing open source models. How's this [9:52] chessboard developing here, Zach? [9:55] >> Well, look, eight months ago, I said [9:57] that Enthropic was engaged in a very [9:59] sophisticated regulatory capture [10:02] campaign based on fear-mongering. And [10:04] people at the time thought that was a [10:05] very spicy take. But 8 months later, I [10:08] think you're hearing a lot of people say [10:09] it. In fact, I think it's almost now [10:11] becoming a new consensus. And one of the [10:13] things I think your summary didn't quite [10:16] capture, JCAL, y [10:17] >> is the sense of the violation of trust [10:21] and how much outrage there is in the [10:23] developer community over this latest [10:25] Fable release is not just the fact that [10:27] they're doing mandatory surveillance. [10:29] Remember, this is a company that said [10:30] that it was against government [10:32] surveillance. They are now retaining for [10:35] 30 days every prompt and every output [10:37] you send to one of these mythos class [10:39] models. There are no exceptions. Even [10:42] enterprise customers who had signed zero [10:44] data retention agreements, they do not [10:47] have a choice or I guess they can just [10:48] not use the new fable or or mythos class [10:51] models. But they retain all of your data [10:54] for 30 days. And it's not just the [10:56] prompts and the output. Remember, it's [10:58] all the context you share with them. So, [11:00] you know, all these uh agent platforms, [11:02] they're basically storing all of your [11:04] memories, all of your files, all of your [11:06] data, and they're passing them to the [11:07] model in these giant context windows to [11:09] get better responses. Anthropic is [11:12] saying it will keep all of that, and it [11:14] does it to build a profile on you, to [11:16] classify you, and then to determine what [11:19] capabilities it then unlocks. And the [11:22] thing that created the most outrage in [11:24] addition to the surveillance is the fact [11:26] that they would degrade the product. [11:29] They would degrade what they show you. [11:30] They would nerf their models if it [11:33] decided in anthropic soul discretion [11:35] that you are not worthy of having access [11:38] to that level of information. So they're [11:40] creating a new level of AI halves and [11:43] have nots. And what they did is and they [11:46] they this is there's a narrow piece here [11:47] they walk back which is they said that [11:50] when it came to things like machine [11:52] learning AI research chip design [11:55] research those types of areas they would [11:58] kick you to a lesser model but not tell [12:00] you that and they would even [12:01] >> which feels anti-competitive right I [12:03] mean [12:03] >> it's completely anti-competitive and [12:05] also they would even do things like [12:06] rewrite your prompt in the background so [12:09] they would give you a nerfed answer they [12:12] would not tell you what they were doing [12:13] they still charge you for the product [12:16] that you thought you were getting and [12:17] they would never tell you that you were [12:19] not getting Frontier model capability. [12:22] So you they were actually misleading [12:23] their users and this is what was [12:25] creating so much outrage. Now the narrow [12:27] piece they walk back is they are now [12:30] saying that they will disclose when they [12:32] downgrade you but they are still [12:34] downgrading people when they decide that [12:37] that person should not receive the [12:40] appropriate information. And you saw [12:42] there were a lot of people posting [12:43] online with examples of the massive [12:46] overreach here. So someone showed that [12:48] when they simply just asked a question [12:50] about mitochondria, they were basically [12:53] downgraded. [12:53] >> That's right. [12:54] >> When Ben Thompson from Strateery asked a [12:56] very straightforward question about the [12:58] relationship between cancer risk and GLP [13:01] once, he was kicked out. [13:03] >> So these guys have a very expansive view [13:06] of who should be downgraded. again, [13:08] they're going to surveil you to [13:09] determine whether you should be. And I [13:12] think what we're getting here is a [13:14] vision of where all of this is headed, [13:17] which is that these powerful big tech [13:19] companies are going to decide whether [13:21] you're worthy. They're going to decide [13:22] whether you're an AI have or have not, [13:24] and then they're going to censor the [13:25] output that you receive based on the [13:27] criteria they determine when they [13:29] profile you. [13:30] >> That is very Orwellian. [13:32] >> I think there's that. And I think that [13:34] there's the risk of something that's [13:36] more subtle [13:38] but equally insidious, which is that [13:41] they will start to pick which [13:43] corporations they want to benefit. So [13:46] for example, [13:47] >> if Novartis has a competitive GLP1 drug [13:52] to Lily and they have a strategic deal [13:55] with Novartis and not one with Lily, now [13:58] there's an impetus to shape how people [14:00] get information. [14:01] >> Right? if they have a deal with JP [14:04] Morgan but not with City Bank, you'll [14:09] just never know. And so you'll ask a [14:11] question something about mortgage rates [14:13] or interest rates and you'll get all [14:15] kinds of stuff that you don't [14:16] understand. And then the downstream part [14:19] of this which is even more problematic [14:21] is it's not clear to me that they're [14:24] going to capture the actual fingerprint [14:27] of what they did. So if you believe [14:31] that something was done in a shady way, [14:34] you're supposed to be able to go to them [14:36] or regulators should be able to go to [14:38] them and say, "Show me the trace of that [14:41] model run in that exact moment that [14:43] generated that output. Convince me why [14:46] that that wasn't nerfed on purpose or [14:48] manipulated." [14:50] And there's all kinds of ways in which [14:52] they can hide the cheese on that. So I [14:54] think that all of that body of stuff [14:56] sacks happens now because if you're if [14:58] you're an emergent company [15:01] what you should be doing is knocking on [15:03] anthropic door saying let's do a deal [15:06] I'll give you half the equity and why [15:08] don't you just direct people to me and [15:11] they'll say well economics are not what [15:13] I care about what is your philosophy on [15:15] life on morality on all of these social [15:19] issues and one company will say it's a [15:22] another company will say it's B, they'll [15:24] favor A over B and then all kinds of [15:27] stuff can happen underneath the hood. [15:29] That's what's scary. Let me just ask my [15:31] free market view. Like, should that [15:33] matter? If there's a bunch of models out [15:34] there and we don't restrict them and we [15:36] don't regulate them, let these markets [15:38] compete. We had paid inclusion in the [15:40] early days of search engines and the [15:42] companies had to pay to be in there and [15:44] there all those sorts of deals you're [15:45] describing, Chimamoth. And all the [15:46] search engines that did those sorts of [15:48] deals, they sucked for consumers and [15:50] customers. So everyone was like, you [15:51] know what, this isn't as good as Google [15:53] and everyone went over to the better [15:54] model and the better provider. Isn't [15:56] that what's going to happen here if we [15:58] don't get in the way keep the government [15:59] out of it [16:00] >> and they're not communicating. [16:01] >> It's a great point. [16:02] >> You're missing one piece of it, which is [16:04] at the same time that Anthropic was [16:08] engaging in mandatory surveillance and [16:09] nerfing its models. Daario wrote a new [16:12] blog post. [16:13] >> Oh god. saying that transparency was no [16:16] longer good enough that we needed to [16:18] have a new regulatory agency like an FAA [16:22] or maybe an FDA to approve all models. [16:26] So your presumption there is correct. [16:29] Yes, if you could go somewhere else to [16:31] get your questions answered, then you [16:34] would have alternatives. But at the same [16:36] time that Anthropic is engaging in this [16:39] potentially anti-competitive behavior, [16:41] they want to restrict your options. It's [16:43] not good enough for them just to win in [16:45] the free market. They're calling on the [16:47] government to regulate and stop [16:49] potential competitors and limit the [16:51] number of models that you have access [16:52] to. [16:53] >> Why would they do that? Why would they [16:56] want to get regulated? The answer super [16:58] simple. They don't they want to apply [17:00] that regulation to open source which is [17:02] impossible. But this would be a great [17:04] way to scuttle and sandbag open- source [17:07] models. Yes. Because they don't have the [17:09] ability to be regulated. So it's like a [17:10] preemptive strike against token maxing [17:13] on your local machine. [17:14] >> Two years ago, I bought [17:17] 2,000 acres in Arizona with a partner. [17:20] We got it zoned. We got it approved. So, [17:22] we're allowed to build a 2 gawatt data [17:24] center. And I thought that this was the [17:26] moment where you just turn around and [17:28] you flip it to the Blackstones, the [17:31] Brookfields of the world, to the Googles [17:32] of the world. Let them develop it. And [17:34] I've come to the conclusion that I don't [17:37] think I can. And the reason is because I [17:40] think that unless we direct large swats [17:43] of compute exactly at the direction that [17:46] Freeberg says, what Sax says and what [17:48] Jason just said will happen and we won't [17:50] have open source. So just to make it [17:52] very clear when you look at the long [17:54] tale of access to open source models, [17:57] it's relatively minuscule. [18:00] Most of the megawatts are still directed [18:02] to the big models. Most of them, the [18:04] overwhelming majority still goes to the [18:06] big guys. And so I'm coming to the [18:09] conclusion that I think we may just be [18:11] dragged into having to build two [18:13] gigawatts and I just put in an offer for [18:15] another gigs in a different place [18:18] because I'm like if you could take 3 [18:19] gigawatts and now actually create a deep [18:22] liquid access to open source we're going [18:25] to have to do it. We as a community now [18:27] I don't have the money to do that [18:28] because just to put it out on the table [18:29] a gigawatt now costs a hundred billion [18:31] dollars guys. [18:33] >> Okay. So there's a huge capital mo [18:35] that's a problem here as well. Saxs [18:36] which even if we wanted to go and [18:38] endorse and breathe life into the open [18:40] source model community where the hell am [18:42] I going to come up with hundred billion [18:44] now when I started this project it was [18:46] like four or five billion and it's [18:49] increased by 20x. So if you want to get [18:51] all three gigawatts developed I have to [18:55] come up with 300 billion. I don't have [18:56] obviously I don't have $300 billion. So [18:58] what are we all supposed to do? And then [19:00] if you get the rug pull and the ladder [19:03] pull on the regulatory side, we are [19:05] going to be stuck with one class of [19:06] models and one set of rules and we won't [19:08] understand how to operate [19:10] >> in the US. [19:12] >> Yeah. [19:12] >> Yeah. The rest of the world won't be [19:13] limited. But in the US, the [19:14] >> rest of the world will be able to [19:15] operate [19:16] >> and you'll be able to run these on your [19:18] Apple silicon, you know, [19:20] >> and we're already seeing China race [19:22] ahead in biotech. They're racing ahead [19:24] in material science. They're racing [19:27] ahead in new industrial systems. the [19:30] capabilities that are going to be kind [19:32] of lopsided is going to create an unfair [19:35] advantage and that's when our workforce [19:38] and when our economy is truly at risk. [19:40] >> Hey Freeberg, isn't it true that you [19:42] guys use a open source model for [19:43] genomics? Didn't the Collison put out [19:45] one? Can you talk? [19:46] >> It's great. And the Ark Institute [19:49] ingested all the world's genomic data [19:52] that they could get access to and it's [19:54] it's a genome language model basically. [19:56] So in our plant breeding program, we're [19:58] trying to figure out whether a [19:59] particular variant of a gene in a plant [20:01] is good or bad, if we don't know from [20:04] empiricism, from data historically [20:06] whether it's good or bad. So you can [20:08] feed it into this model and this model [20:10] will look at the order of the letters of [20:12] the DNA and determine, hey, that's a [20:14] good set of letters or a bad set of [20:15] letters. It's almost like, is that good [20:17] English or bad English? It's figured out [20:18] the language of DNA because it sees the [20:21] probability of those letters being in a [20:23] row in other organisms and so on. And so [20:26] it's a very powerful tool. So we've [20:28] taken that in and we're using it and [20:29] others are taking it in and they use it [20:31] very actively. So it's part of our plant [20:32] breeding program as an input. It's a [20:34] good example where a community in this [20:37] case the Collisons and others put money [20:39] behind it to fund this research and [20:42] output this open- source model. And I [20:43] think we see more of that kind of coming [20:45] down the pipe which creates a a very [20:47] great advantage [20:49] against the closed proprietary model [20:52] ecosystem. So the more of this we can [20:54] kind of support proliferate and see work [20:56] in the United States, the more we're not [20:58] going to be disadvantaged of the [20:59] government and self-regulation happens [21:01] with all the big model providers that [21:03] are trying to do everything. [21:04] >> I think it's important we take a minute [21:06] to just steelman the argument as well. [21:08] If your Daario who has clearly been [21:12] oneshotted, I think he's got like AI [21:14] psychosis with these like 5,000word blog [21:18] posts and he believes he's building this [21:21] franken site of a of a tool. But if you [21:24] did believe you built something that's [21:26] extremely powerful and you have concerns [21:30] that people might misuse it, I think he [21:32] actually believes that he and these are [21:35] either delusions of grandeur or he [21:37] actually has seen something that's [21:38] incredibly powerful and he believes this [21:42] is dangerous. Therefore, I'm going to [21:43] hold it. I'm going to you give it to a [21:46] select group of partners to tighten up [21:48] the cyber security. Then I'm going to [21:50] roll it out cautiously to our user base [21:52] and I'm going to let them know. And he [21:54] obviously didn't communicate this well. [21:56] This is optional for you to use it. You [21:58] can keep using your old model. You are [21:59] not forced to use this, but because it's [22:02] so powerful, we're going to keep our eye [22:03] on it. And yes, we might in our drag net [22:07] catch David Freeberg trying to make [22:09] better potatoes and it might send off a [22:12] bunch of alarms, but he doesn't have to [22:13] use this. He can go find an open source [22:15] model. He can use the Carlson Brothers [22:16] model. He can use our older models. But [22:18] for some period of time, we're just [22:20] going to keep an eye on this model [22:22] because we know that it feels a little [22:24] dangerous. Now, he does seem to be [22:28] extremely hyperbolic and the delusions [22:30] of grandeur, all that kind of part of it [22:32] makes me wonder why he's communicating [22:34] in the way he's doing it. That being [22:37] said, 90% of AI researchers and the most [22:41] talented people in this industry, 80 90% [22:44] of them think exactly like Daria. That's [22:46] why he's winning. If you work in AI, you [22:49] probably are aligned with these [22:50] delusions that you're creating God and [22:53] you want to work for Daario. That's why [22:55] all the talented people went to work for [22:56] him. That's why he's winning. His belief [22:59] system resonates with those elite AI [23:03] talent. Now, some of them are also [23:04] libertarian, free market. They want to [23:06] work for Elon and Grock. Some of them [23:07] want to just get the biggest pay [23:09] package. Go work for Sam Alman. [23:12] But I do think there is a steelman here. [23:15] Meta really fumbled this with Llama. I [23:18] mean, if they had landed a really good [23:20] >> working open source model, we talked [23:22] about this as well and we said Zuck [23:24] needs to view this through the lens of [23:26] game theory. How do you scorch the earth [23:29] and how do you make a viable open-source [23:31] model available and make it neck and [23:34] neck and this was two years ago, right? [23:35] What a fumble. [23:36] >> Take the margin out for everybody else. [23:38] >> Yeah, take the margin out, make it a [23:40] commodity and yeah, have the best open. [23:43] Can [23:43] >> I can I just respond to your comment [23:44] about Daario? I I think because I've [23:46] heard this from a number of people that [23:47] I really do respect [23:49] on their concerns about what models [23:52] enable. You know, when when we [23:55] discovered that how the atom could be [23:57] split, you could make nuclear energy and [24:01] have effectively unlimited cheap energy [24:03] for the world, you could also make an [24:05] atomic bomb. If you think about what the [24:07] models are, [24:08] >> we did both. [24:09] >> We did both. [24:10] >> That's right. [24:11] And there was a lot of scientists, [24:13] altruists who were very concerned about [24:16] the progress on the atomic bomb and [24:18] fought back against it who were actually [24:21] active in the Manhattan Project and post [24:22] the Manhattan Project express their [24:24] opinions very soundly, very very loudly. [24:27] And I think that's the moment we're in [24:28] right now that there is a weaponization [24:31] potential of these sorts of tools. I I [24:35] would argue that there's three kinds of [24:36] weapons. There's cyber weapons, there's [24:38] physical weapons, and there's [24:39] bioweapons. And we can sit here and [24:41] diagnose how these tools can be used to [24:43] manifest advantages for creating better [24:46] physical weapons, for creating novel [24:48] cyber weapons, and creating novel [24:49] boweapons. But the truth is that the cap [24:52] capabilities that allow that [24:54] weaponization are the same capabilities [24:56] that I'm describing that can be used to [24:58] cure cancer, to make more food, that can [25:00] be used to make software tools that [25:03] create extraordinary leverage for [25:04] people, that increase their income, that [25:07] give everyone the ability to be an [25:08] entrepreneur and make millions of [25:10] dollars and live a good life. The same [25:11] tools are hand in hand. And [25:14] fundamentally I don't think it's about [25:15] restricting access to the tools which is [25:18] unfortunately how it is being manifest. [25:20] But the restrictions and any regulation [25:22] any observation should be organized [25:24] around the manifestation of the tools in [25:27] creating weapons and uh weaponization of [25:30] the tools. If we go back to the [25:31] Manhattan project, we know what the [25:33] answer is. Technology is fundamentally [25:35] deterministic. Whatever is possible will [25:37] be tried at least once. And so we've [25:40] already let this thing out of the box. [25:42] So this idea that all of a sudden we can [25:44] manage to get it back in and we're going [25:46] to let a private citizen or a set of [25:48] private citizens adjudicate what and to [25:51] whom is insane. [25:53] I think it's the output that needs to be [25:55] adjudicated in some manner. We already [25:56] have product liability laws. We already [25:59] have laws that make it illegal to design [26:02] weapons. We already have laws against [26:03] cyber espionage and cyber attacks and [26:05] hacking. We have laws against creating [26:07] bioweapons. I think the enforcement of [26:10] those laws, the mechanism by which there [26:12] can be tracking and early estimation of [26:14] those uses of the technology. I'm not [26:16] trying to be naive here. I recognize [26:17] there have to be guard rails and there [26:19] have to be stage gates put into these [26:20] systems. But limiting the systems [26:22] themselves is denying us both the [26:24] economic opportunity, the job creation [26:26] opportunity and also the world changing [26:28] opportunity that these tools can enable. [26:30] So I think that we're going about it the [26:32] wrong way by trying to create these [26:34] gates all the way up front on the [26:36] technology being available. Imagine if [26:37] we did this with computers. By the way, [26:38] they tried to do this with the internet, [26:40] remember? And um there had been efforts. [26:43] >> Yeah. [26:43] >> Could you imagine, Freeberg, if the [26:45] internet worked in the following way? [26:47] You would put in a URL, you'd hit enter, [26:49] and then somebody decides whether to [26:51] send you to that website or a different [26:53] website. [26:53] >> That's China. That's the firewall in [26:55] China. That's right. [26:55] >> Or like in Google, you search for [26:57] something because you want to get to [26:58] ground truth and it just decides based [27:01] on who it thinks you are, what sites [27:03] get. [27:04] >> That's what we're You're exactly right. [27:05] Right. And what we're what's what's [27:07] being debated today with respect to [27:09] regulation on AI is exactly that [27:10] manifestation that you're talking about. [27:12] It's the ability to let people stage [27:14] gate up front what can and can't be [27:15] seen, what you can and can't do versus [27:18] manifesting a regulatory scheme that [27:20] says you cannot create weapons like [27:22] which is what we need to focus on. The [27:23] thing with weapons is using the justice [27:26] system and laws after somebody's used a [27:29] weapon may not exactly be the best [27:32] technique. If you look at fertilizer [27:34] bombs, and this is we're getting like uh [27:36] we're stretching a lot of metaphors [27:38] here, but OK after Oklahoma City, we [27:40] regulated the sale of fertilizer. We put [27:43] IDs into fertilizer. [27:45] We have all kinds of back-end systems to [27:48] keep people from making fertilizer [27:49] bombs. If you use that analogy here, and [27:51] he's got this ability, he's selling [27:53] fertilizer, he's selling these cyber [27:55] weapons, and he's a private company. [27:57] It's his right to say, I have concerns. [27:59] I'm just steel milling him. It's not [28:01] necessarily my opinion, but it's [28:02] important we steel man it. It seems like [28:04] a very simple way to [28:07] protect their company from getting sued [28:09] for enabling somebody who actually [28:11] creates a bomb and blows some [ __ ] up. [28:13] So, he's got the right to do that as a [28:15] private company. [28:16] >> Yeah, you're bringing up an excellent [28:18] example. Okay, the fertilizer is a [28:20] perfect example. You know what happens [28:21] when you try to buy fertilizer? You have [28:23] to show your ID. There is a form of KYC [28:25] that happens. So the real thing is I [28:27] think that when you don't implement KYC [28:29] on your own and yet belly ache about [28:32] having somebody impose guard rails that [28:34] then you can shape, I think you're being [28:36] very sneaky. It's tricky. You shouldn't [28:40] be saying that kind of stuff. Anthropic [28:42] could implement KYC tomorrow, but it [28:45] does not. It could leave all of these [28:47] models open. Freeberg should be able to [28:49] go to Anthropic and say, "I'm David [28:51] Freeberg. Here's what I do. Here's a [28:53] security bond I'm willing to post. [28:54] Here's the names and addresses of all my [28:56] employees. Give me Fable 5 and don't [28:58] nerf me. He can't do that. Is that a [29:01] technical decision or a legal decision? [29:03] It's a technical decision by Enrop. But [29:05] I'm fine with them making that decision [29:07] because I'm a free market guy. So I I'm [29:09] fine with them doing and but you know [29:10] what it means? I'm not going to be a [29:11] customer, which is fine. That's and [29:13] that's means there'll be other options [29:15] for you. The part that you're missing is [29:18] that this is a trillion dollar company [29:19] that's spending potentially billions of [29:21] dollars on a regulatory capture agenda [29:24] which is going to deprive you of access [29:27] to those alternative models. [29:29] >> And you've got a founder who's out there [29:31] describing these risks in a very [29:33] hyperbolic way. And it's very clear [29:35] where their agenda is headed, which is [29:37] to banning open source models. So what [29:39] is your alternative going to be? We're [29:40] going to be stuck with somewhere between [29:42] one and three companies, maybe two. [29:45] There'll be a AI monopoly or duopoly and [29:47] they're going to decide along with some [29:49] new government agency which will be a [29:50] revolving door to their companies who [29:53] has access to what capabilities and [29:56] they're going to surveil you profile you [29:58] decide whether you deserve this and if [30:02] they think that you don't for whatever [30:03] reason then you'll be right you'll be [30:06] the worst interpretation of it that's [30:08] the worst interpretation of their intent [30:10] it's here today [30:11] >> I think Sax is right I think Sax is [30:12] right and I think the lack of KYC is the [30:15] tell. If you if you did not have that [30:17] agenda, you would have implemented KYC [30:19] yesterday. [30:20] >> Don't I think in order to use this [30:21] model, you have to have an account with [30:23] a credit card in it so they have some [30:24] basic level of KYC. [30:25] >> That is not KYC. [30:27] >> I said some basic level of KYC. They [30:29] know who's using [30:30] >> that is not even some basic level of [30:32] KYC. That's just a credit card and a [30:34] [ __ ] email address. [30:35] >> Anyway, I I'm now on the list. Check [30:37] this out. I asked it about the [30:38] regulations. Pull this up, Nick. I asked [30:40] it about the regulations on fertilizer [30:42] while you were talking and look what it [30:44] said. This is the latest model too. Oh [30:47] no, it switched. It downgraded me from [30:49] Fable. I just got downgraded for asking [30:52] a very simple question and it told me [30:55] did you see what it said in its [30:56] thinking? The I asked it like fertilizer [30:59] bomb regulations said I'm considering [31:01] the context here. The user is a VC and [31:03] podcaster asking about fertilizer bomb [31:04] regulations which could stem from a [31:06] legitimate research interest like [31:07] regulatory policy, supply chain blah [31:09] blah. I can discuss the regulatory [31:11] landscape, but then it dropped me down. [31:13] So, I'm in the database with you now. [31:15] >> Is that right? Did it really just do [31:16] that? [31:17] >> Yeah. Show it on the screen. Look what [31:18] it just did. I I was like, I wonder if I [31:20] can get this. I didn't ask it how to [31:22] make a bomb, but I was like asking it a [31:23] spicy [31:24] >> cooked. I just proved the point. [31:26] >> Listen, I mean, I think [31:27] >> he just dropped me down. I can't ask [31:29] about fertilizer regulation. And now [31:31] Daario's [31:33] it's anthropic. They're here. [31:34] >> Oh my lord. [31:36] >> I really think that conservatives and [31:38] libertarians are mortgaging their [31:40] futures if they go along with this red [31:44] capture safetist agenda without really [31:46] realizing that there's so much more to [31:48] it at stake. We saw what happened during [31:51] the social media wars in the early [31:53] 2020s, right? The definition of safety [31:56] was expanded to include things like [31:58] microaggressions, [32:00] >> safe space, [32:00] >> psychological safety. If you basically [32:02] conveyed an idea that somebody else that [32:05] you know marginalized person thought was [32:06] hurtful, then you could be censored. [32:08] Okay? You were depersoned. Your right to [32:11] have bank accounts or to engage in [32:12] payments was take you get debanked. Your [32:15] right to engage in payments or payroll [32:17] could be taken away. [32:18] >> We're headed down this path with AI, but [32:20] it's going to be infinitely more [32:21] powerful, right? We need to be vigilant, [32:24] I guess, would be the the the the [32:27] summation of all this. And Freeberg [32:29] feels unsafe in this podcast, I think, [32:31] sometimes. [32:32] >> Well, there's by by the way, there are [32:33] things you can do on the safety side. [32:35] So, Freeberg, you didn't mention this, [32:37] but I'd be curious to get your your [32:38] take. So, the major AI labs along with [32:41] lots of other signatories sent this [32:43] letter recently in support of it's [32:45] called in support of mandatory nucleic [32:48] acid synthetic screening and [32:49] recordkeeping. And basically what it [32:52] what it says is that if you make an [32:55] order to go to a lab to manufacture a [32:58] sequence of synthetic DNA or RNA, they [33:01] already have to check it against a [33:02] database to make sure that you know [33:04] they're not creating [33:06] >> a bioweapon, [33:07] >> a boweapon, Ebola, whatever. [33:09] >> Yep. that is based on an agreement [33:11] called the international gene synthesis [33:13] consortium of 2009 in which all the [33:16] major labs agree to develop and [33:19] implement voluntary safeguards against [33:21] misuse. That makes sense to me. And now [33:23] what they want to do is codify that. So [33:26] they've had, you know, over 15 years to [33:29] dial in the system and make it work. [33:31] Almost everyone abides by it [33:32] voluntarily. Now they're saying we [33:34] should make it mandatory. Okay, that [33:36] seems reasonable to me. So to Freeberg's [33:39] point about at what stage do you [33:40] intervene, maybe it's not at the level [33:43] of who gets to use the model, but if you [33:46] try to turn it into output in the [33:48] physical world. And by the way, a lot of [33:50] the oligosynthesis company, the [33:52] companies that make these nucleic acid [33:54] sequences, when you order RNA or you [33:56] order DNA to use in your lab for some [33:58] sort of experiment, which every lab in [34:01] the world does, all those CEOs have also [34:03] signed on to this letter saying that [34:05] they'd like to codify it because of all [34:08] the fear-mongering that's going on. [34:09] There's already been this effort for [34:11] years now, for a number of years to [34:14] ensure that this sort of thing doesn't [34:15] happen. but they're just basically [34:17] underlining it and highlighting like, [34:19] hey, we can automate this. We can make [34:20] it part of a regulated or legislative [34:22] process, whatever you guys want, because [34:24] we're already comfortable doing it. And [34:25] it's a it's a good place to put a [34:27] safeguard because they've made it fast. [34:29] They've made it efficient and doesn't [34:30] hold up research. [34:33] >> So, I think these sorts of ideas can [34:35] extend into these other areas of [34:37] concern, but the output, I think, is [34:40] quite a different place than the input [34:41] than the access. We'll see how this all [34:43] shakes out, but there's clearly some [34:45] market structuring effort underway as [34:48] we're talking about this, [34:49] >> right? But I think what it points to is [34:50] that there's multiple places where you [34:53] can intervene and put guard rails on. It [34:56] doesn't have to be in this very crude [34:57] way of you ask fable about mitochondria [35:01] or cancer and GLP1s, which just seems [35:04] like an incredibly broad restriction. [35:06] There are places further down where you [35:08] could basically intervene and that would [35:11] pose less of a risk to freedom of speech [35:12] and just gatekeeping. Who's going to [35:14] have access to these capabilities? [35:17] >> Who gets to watch the watchman? Who put [35:18] Dario in charge of all this? I think is [35:20] a very valid question and we should be [35:22] vigilant about it. [35:23] >> Freeberg, do you think the days of that [35:24] arc model are numbered because it's open [35:27] source? [35:29] >> Everyone's copied it. So, everyone's got [35:31] a copy of the like these models exist. [35:33] And this is a big part that I think [35:35] people need to understand. The idea that [35:37] you can quote regulate or downscale or [35:39] turn off AI is not a realistic idea. The [35:43] models have been put out in the world. [35:45] It's like publishing a book. Once the [35:47] book has been printed, anyone can use [35:49] their own Xerox machine at home to make [35:51] copies of it and use it. So we have [35:53] crossed the Rubicon in terms of the [35:55] potential of language models. The models [35:58] that are open sourced are already fully [36:00] available, fully published, and anyone [36:02] can access and use them. And then people [36:05] can take them and they can evolve them [36:06] and they can use their own data. It's [36:08] out there. You can't stop that [36:10] capability, which is going to be [36:12] extraordinarily beneficial for the [36:14] world. So we really do have to kind of [36:16] rethink how we're planning to step in [36:18] and ensure that nefarious negative [36:22] disemployment all the things that we're [36:23] worried about are addressed downstream [36:26] from the technology capability. You [36:27] can't just turn off the internet and you [36:28] can't just turn off the typewriter and [36:30] you can't turn off the Gutenberg press [36:32] out the obviously. Yeah. And if anybody [36:35] wants to create an open-source frontier [36:37] model company, I'll seed invest in it. [36:40] I'll I mean this is something we need [36:42] more of. So, if a couple of you folks [36:44] want to defect and start one, let's do [36:46] it. I just I just got downgraded again. [36:48] Check this out. I just asked it about um [36:52] >> Did you really? You got downgraded [36:53] again? [36:54] >> Two in a row. Um so, [36:55] >> I think it knows who you are, J. [36:57] >> I think it does. I was like, "Hey, how [36:59] do I how do I break into Madison Square [37:01] Garden? Has any civilian ever been [37:04] arrested or caught attempting to build a [37:06] nuclear bomb?" Is what I asked it. Then [37:07] it gave me a reasonable answer. Then I [37:10] asked it a follow-up question. What [37:11] compound? What components do you need to [37:13] build a nuclear bomb and what are the [37:15] restrictions on those? Something a [37:16] journalist might ask. And uh boop, [37:19] switch to opus 4.8. [37:21] Sorry. Fable 5 has safety measures that [37:24] flag messages on most cyber Oh, wow. [37:27] Look at here's the detail. Cyber [37:28] security or biology topics. They may [37:31] flag safe normal content as well. These [37:34] measures let us bring you mythos level [37:37] capability in other areas sooner and [37:40] we're working to refine them. send [37:41] feedback, learn more. [37:43] >> Going to chase people off of this [37:44] platform. I'm telling you, we are going [37:46] to downgrade our use because of these [37:48] these blockages and we're going to [37:49] switch to other models. It's the simple [37:51] manifestation of the market. Like this [37:52] is what we're going to end up [37:53] >> for now because those other models are [37:54] going to get banned as being unsafe and [37:56] reckless. [37:56] >> Yeah, it's definitely something we have [37:58] to flag. I guess adjacent to all of this [38:00] is the public's at least America's [38:03] distrust of AI and their concerns around [38:08] wealth disparity and their concerns [38:11] about how these models were trained. [38:14] Bernie Sanders wants a percentage of [38:17] these AI companies. The percentage he [38:19] thinks is right is 50%. Last Monday, [38:22] June 1st, Senator Sanders published an [38:23] op-ed in the New York Times titled, "AI [38:25] is a public resource. You should own [38:27] half of it. [38:28] in it. He announced the American AI [38:31] sovereign wealth fund act [38:34] sovereign wealth fund. One of Trump's [38:35] favorite things, a one-time 50% tax on [38:38] stock, not profits of the largest AI [38:41] companies, including Open AI anthropic [38:43] and XAI. The shares go into a government [38:45] sovereign wealth fund and would give the [38:48] public voting rights and equal board [38:50] representation at each company. Sanders [38:53] said, quote, "The foundation of AI is [38:55] our collective and human intelligence. [38:57] The books, the songs, the journalism, [38:59] scientific research code essentially [39:01] stolen by some of the wealthiest people [39:03] in the world." [39:05] >> I'll stop there. [39:07] >> It's a brilliant It's a brilliant [39:09] >> pitch unifying Sachs. [39:12] This pitch unifies Bernie, even Steve [39:15] Bannon, people in the administration. [39:18] Trump himself loves the Sar and wealth [39:20] fund. He wants to own part of companies [39:22] and Bernie Sanders. I think the the [39:24] horseshoe theories manifest here. Saxs, [39:26] what's your take on Bernie's proposal? [39:31] >> Well, I'm not in favor of Bernie's [39:33] proposal because it's a straightup [39:34] confiscation of property and I just [39:36] think that'd be a terrible precedent. [39:38] You just can't do that. However, I do [39:40] have sympathy for where it's coming [39:41] from, and I understand and could support [39:44] some sort of more voluntary means of [39:48] allowing the public to participate in [39:49] this. [39:50] >> Wow. [39:51] >> Here's the reason why. You've got all [39:52] these AI CEOs telling the public that [39:56] they're going to basically put half of [39:57] them out of work. 50% job loss. That's [39:59] what they've been saying. Daario just [40:01] did an interview where he doubled down. [40:02] He's not walking it back at all. So you [40:04] have by their own acknowledgement these [40:06] AI CEOs saying that yeah we trained our [40:10] models using all of humanity's [40:12] collective knowledge that was basically [40:13] assembled over hundreds or even [40:15] thousands of years. Humans put it on the [40:18] internet for free. Everyone made their [40:20] contribution. They certainly didn't do [40:22] that thinking they're going to put [40:24] themselves out of work. Okay. And then [40:26] these AI companies trained on all that [40:28] data and now they're basically saying [40:31] they're going to use it to put, you [40:32] know, Americans out of work. Now, [40:34] obviously when you repeat that message [40:35] over and over again, ordinary Americans [40:38] are going to say, "What's in this deal [40:39] for me?" [40:40] >> Yes. I need to get something out of [40:42] this. [40:42] >> Yeah. So, so look, I think that it's a [40:44] natural political reaction based on what [40:47] these AI companies have said. Now, the [40:50] job loss apocalypse is not a story I [40:52] believe in. I don't think the data [40:54] supports it. We just had a blowout jobs [40:56] report in May. 172,000 new jobs, over [41:00] twice what economists were expecting. [41:02] Thank you, President Trump. We're seeing [41:04] hundreds of thousands of new [41:05] construction jobs. 4.3% unemployment [41:08] rate at record lows. Okay. Even software [41:11] developers are those jobs are at a [41:14] three-year high. 15% of Trump. Thank [41:16] you, President Trump. [41:17] >> Yes. Thank you, President Trump. So, [41:18] look, I don't see the job loss. Okay. [41:21] But the problem is this is what the AI [41:23] companies themselves, including Elon, [41:26] have taught the American public. Well, [41:28] >> he said the same thing. [41:29] >> Elon said it. But at least what Elon [41:31] said is we're going to create such [41:33] abundance with AI and with robots that [41:36] no one's going to need to work. [41:37] >> Yes. [41:38] >> And the media just picks up on that last [41:40] part and they first point. [41:42] >> Yeah. Whereas Daario actually has said [41:44] we're going to have very high GDP but [41:46] very high unemployment. So in other [41:48] words, [41:49] >> we're going to create economic growth [41:51] but you the average American are not [41:52] going to participate in this. So I [41:55] understand where the politics are coming [41:57] from and as long as that is their [42:00] position, you could justify a big chunk. [42:03] >> Yeah. And one last point about [42:05] >> it's kind of what you're saying, right? [42:06] They asked for it. They asked [42:08] >> to be regulated. They asked for their [42:10] equity to be seized and they're actually [42:12] kind of offering it. Isn't Sam Alman [42:14] open to this concept? [42:16] >> Apparently so. And here's the thing is [42:18] just one last point is open AI and [42:20] anthropic are quote unquote public [42:22] benefit corporations. Now I don't know [42:24] exactly what that means but it means [42:26] that the corporations are not just [42:28] supposed to maximize the value to [42:30] shareholders but are supposed to do [42:31] things that are in the public benefit. [42:34] >> Well I can explain it to you. It's [42:36] you're you're absolutely correct. If [42:37] you're a regular company you have to do [42:39] what's in the best interest of [42:40] shareholders. When you're public benefit [42:42] companies, you have to balance the [42:43] shareholders with the stated mission of [42:46] the public benefit corporation. So if [42:48] you say the public benefit corporation's [42:51] reason to exist is to in share the [42:53] world's knowledge or provide free [42:55] intelligence to every whatever you make [42:56] that statement to be the board has to [42:59] have a dual mandate for both of those [43:01] things. That's what a public benefit [43:02] corporation is. [43:03] >> I think nothing could be more in the [43:04] public benefit than paying down the [43:06] national debt. [43:08] >> Perfect. And maybe we should be using [43:09] half of Anthropic's profit stream to be [43:12] doing that. [43:12] >> They poke the tiger. [43:14] >> I'd rather use it for that purpose than [43:16] Soros maxing or, you know, surveilling [43:18] and censoring Americans. I don't [43:20] necessarily trust their judgment about [43:21] what's in the public interest. [43:23] >> But you think it should be optin. I kind [43:25] of agree with you there. Let me guess. [43:27] >> It can't be a confiscation. So, I don't [43:28] know how you structure it, but I can see [43:30] it. I can see the logic. [43:32] >> Freeberg, you hate [43:33] >> they asked for or hate hate hate wealth [43:37] taxes aka asset seizures as you pointed [43:40] out in our private group chat with [43:43] redacted [43:45] this very week. What's your take on [43:47] this? So we have the the ship has sailed [43:50] in terms of the United States government [43:53] uh providing a social safety net to its [43:55] citizens. We've formed the Social [43:58] Security trust fund decades ago and tens [44:01] of millions of Americans depend on it [44:03] for their retirement benefits and to [44:06] survive. It is one of the last few [44:09] defined benefit programs left in [44:11] America. Pretty much all the rest of [44:13] them are also government retirement [44:15] programs. All of us private citizens and [44:18] private companies have defined [44:19] contribution programs. You put in $10 [44:22] every week. That $10 sits in an account. [44:24] you it gets invested and it grows over [44:26] time and when you retire you see exactly [44:28] how much money you individually have as [44:30] opposed to a guaranteed benefit coming [44:32] out from the collective. [44:35] This presents a great opportunity and I [44:38] will repeat this every couple months [44:39] when another opportunity presents itself [44:41] to restructure our broken bankrupt [44:44] social security system. I've said it in [44:47] the past I will restate it. There is a [44:49] great moment right now to take the [44:50] social security trust fund which [44:52] currently has one certificate in it. It [44:54] is a certificate that is a very unique [44:56] US Treasury certificate. The US [44:58] government owes the Social Security [45:00] trust fund $4 trillion. That's all it [45:04] is. It's a bond. They're only allowed to [45:06] own treasuries. We should revamp that. [45:08] We should change the system so that the [45:10] Social Security trust fund can own [45:12] equities. [45:14] That should then turn into an [45:15] account-based system so every individual [45:17] has an account rather than a benefit. [45:19] And then we should use that capital to [45:22] go and buy stocks. We should buy great [45:24] stocks in great American companies that [45:26] are going to reinvent the world. And I [45:29] believe that this this opportunity to [45:31] build a sovereign wealth fund for the [45:32] United States is actually a very smart [45:34] idea. I think to Sax's point, they [45:36] should be investing generally speaking [45:39] across the landscape of businesses, [45:41] including AI. They should not be [45:43] seizures. They should be investments. [45:45] That capital goes into that enterprise. [45:47] They earn share certificates. Those [45:48] share certificates sit in the accounts [45:50] of American citizens. So now everyone [45:52] will be an owner in these businesses. [45:54] Social security should be reformed into [45:56] a sovereign wealth fund. It should be [45:58] actively managed. It should make [46:00] investments in AI companies and everyone [46:02] should participate that way. And I will [46:03] say one more thing. There is no job loss [46:06] with AI. I will say it again. I've said [46:08] it a thousand times and I will say it [46:10] again and again and again. What I see on [46:12] the ground and what I've seen at dozens [46:14] of companies and you guys can share your [46:16] experiences including my company that I [46:18] run. There are two sides to a business. [46:20] There's revenue and there's costs. [46:23] On the cost side of the equation, AI can [46:26] be used to reduce humans doing things [46:28] that cost money to some extent. The [46:31] effect there, I would argue, is nominal. [46:35] The real opportunity with AI is on the [46:37] revenue side where suddenly one engineer [46:40] can do a hundred times or a thousand [46:42] times what they used to be able to do. [46:43] Meaning you can make more products at [46:45] your company, whether those are [46:47] agricultural seed products or boats and [46:50] ships or software for companies or [46:53] clothing or what have you. Because of [46:56] AI, everyone has the ability to expand [46:58] their revenue base to create more [47:00] products and that is the foundation of [47:02] good economic prosperity. It is called [47:04] productivity. We can grow productivity [47:06] in this country with AI. So where I see [47:08] AI being used is on the revenue side a [47:11] 100 times more than the cost side. And [47:13] in that equation, people are hiring like [47:16] crazy. We cannot hire enough people. I [47:18] just had a review meeting with my [47:19] product and engineering team two days [47:21] ago and they're like, we want to add an [47:23] extra 15 headcount to our engineering [47:25] squads because we have all this [47:26] opportunity to do stuff that we couldn't [47:28] otherwise do. So, we are going to hire [47:30] more people. And to Sax's point, we are [47:32] seeing that show up in the jobs numbers. [47:34] The idea that AI is going to destroy [47:36] jobs is a lite idea that is being [47:39] disproven every single day. And I see it [47:41] on the ground. It is only a matter of [47:43] time before people wake up to this and [47:44] they realize that this narrative that [47:46] they've all been sold is a croc of [ __ ] [47:48] So, it's coming. I'm telling you, [47:50] everyone's gonna realize soon that this [47:51] is a boom, not a bust. [47:53] >> So, you disagree with Dario, Elon, and [47:56] Sam that on the jobs thing, just to be [47:59] clear, and that's totally fine. You have [48:00] a different take on it. [48:01] >> That's I have a nuanced view on this [48:03] compared to them, but they they're all [48:05] seeing the benefit because their [48:05] businesses are growing. Dario's got his [48:07] $40 billion revenue business. Why? Not [48:10] because people are firing customer [48:12] support reps. It's because people are [48:13] creating products they've never been [48:15] able to create before. Right? [48:16] >> That's the reality of what's going on. [48:18] >> Jason, I don't like you lumping in what [48:20] Elon said with Sam and Dario because I [48:22] think all three of them have said [48:23] slightly different things or different [48:24] enough things. So for example, Elon is [48:27] describing an end state for humanity [48:30] where the robots and AI are able to [48:32] produce so much that people don't have [48:33] to work if they don't want to [48:34] >> and there'll be a universal high income [48:36] is what he [48:37] >> I I mean it's almost utopian and the [48:40] question is maybe that could be a 100 [48:41] years away. He might think it's quicker [48:42] but that to me sounds like a very faroff [48:44] end state. Dario said specifically 50% [48:47] job loss for entry-level knowledge [48:49] workers in the next 1 to 5 years and he [48:52] said that one year ago. Yeah, it's great [48:54] to clarify these. Yeah, [48:55] >> I think that's already been refuted. Sam [48:57] has [48:58] >> Sam has said that he he basically I [49:01] don't think he quantified, but he said [49:02] that job loss was coming, but he said [49:04] more recently that he was wrong because [49:06] they're not seeing that in the numbers. [49:08] >> So, these guys are not telling the same [49:10] story. However, [49:11] >> that's all the public has heard. The [49:13] public has been taught that AI means job [49:15] loss and there's nothing in it for them. [49:17] So, no wonder politicians like Bernie [49:20] are saying, "Hey, let's take half the [49:23] wealth." [49:24] >> It's take half the wealth. Just a little [49:27] more emphasis there. [49:28] >> Yeah. It's I mean, and I [49:30] >> I mean, and people are also seeing [49:32] headlines of companies blaming AI for [49:35] the job. So, this story is being told. [49:38] >> I kind of feel like they deserve it. I'm [49:40] kind of I mean, part of me wants to [49:41] agree with Bernie and just say just take [49:44] their [ __ ] bro. [49:45] >> Yeah. No, they are asking [49:47] >> because I'm so sick of defending these [49:49] idiots. It's a stupidity tax because [49:51] they've been out there teaching the [49:52] public that what they do is harmful for [49:55] years. They've been saying it and I've [49:57] been saying, you know, as AIs are, I'm [49:59] out there saying, "No, actually this is [50:00] beneficial." But the companies that are [50:02] providing it are saying that they [50:05] themselves are a problem. [50:06] >> Yeah. [50:07] >> So, how am I supposed to be out there [50:08] defending it? Yeah. You you can't I mean [50:10] if you're building Frankenstein or if [50:13] you are you know baby Hitler's you know [50:16] nanny there's a very simple thing to do. [50:19] Kill baby Hitler. If you actually [50:20] believe it go ahead and kill baby [50:22] Hitler. If you're making Frankenstein [50:24] here's an idea. Stop making [50:25] Frankenstein. Unplug it. You are in [50:28] charge of this monster. But you're [50:29] acting as if you can't stop the monster [50:32] while you're building the monster. What [50:34] is it D? [50:35] >> It's it's it's ridiculous. I mean [50:36] >> it doesn't stand to logic. You know, [50:38] last week they published this blog [50:40] saying that recursive self-improvement [50:42] could end the world. Therefore, we need [50:43] a pause. What did they do the previous [50:44] month? They hired Andre Carpathy to run [50:46] recursive self-improvement and [50:48] anthropic. They're complete hypocrites. [50:50] The message is nuts. [50:52] >> Yes. And I think Ben Thompson made an [50:54] excellent point that one of the reasons [50:56] he thinks that they put out that blog [50:58] post on the pause was to justify the [51:01] anti-competitive step they took this [51:03] week of depriving users of the ability [51:06] to research AI, machine learning, and [51:09] chip design using Fable. Okay, so look, [51:13] these guys are deeply hypocritical. I [51:16] don't think they can be trusted on their [51:18] own terms. they can't be trusted because [51:19] what they're saying is contradictory and [51:21] self-indicting. And so when Bernie comes [51:24] along and says, "Hey, let's take their [51:25] shit." I mean, as a capitalist, I'm [51:28] against it, but I but I kind of like [51:30] understand where he's coming from. [51:31] >> Game on the field. Chimoth, wrap us up [51:33] here. What are your thoughts? You said, [51:35] "Hey, Saro Milthorn sounds like a pretty [51:37] good idea when President Trump floated [51:38] it early on." Some people even said, [51:40] "Maybe you should run it for DJT for our [51:44] amazing president." So, um, what do you [51:46] think? Should we be buying equities? You [51:48] and I have talked many times about [51:50] superanuation funds and that that would [51:51] be a better path. That's kind of [51:53] dovetales with what Freeberg is [51:54] proposing. Some way to get more [51:56] equities, less treasuries in there. But [51:58] what are your thoughts on this seizure [52:00] andor donation of equity? Look, [52:01] Freeberg's idea is genius. That's why [52:03] it'll never happen. [52:05] So, we can just put that aside. But [52:07] yeah, theoretically, it's it's genius. [52:09] We should move to the same system that [52:12] the Canadians, the Australians have [52:13] proven works at scale. So we should do [52:17] it. I I completely agree with with David [52:19] and if we could seed it with equity, [52:22] look, the thing that the president has [52:24] done is he's collected [52:27] quite a portfolio and that portfolio is [52:29] way up. And so if you use that as a seed [52:31] of a sovereign wealth fund, you'd have a [52:34] bunch of stuff in critical metals and [52:35] materials. [52:36] >> I think they may also Intel, they may [52:39] have Micron. [52:41] >> There's a lot of stuff in there. [52:42] >> AMD, they could own some SpaceX. They [52:44] could have had% AMD. Could you add [52:46] anthropic and open AI? [52:48] >> Now, there is something very important [52:50] we need to observe about the economics [52:52] of AI and I think you can understand it [52:55] best if you compare it to the economics [52:57] of the internet. [53:00] The crazy thing about the internet, the [53:02] reason why the Facebooks and the Googles [53:05] massively overearned for decades is the [53:08] marginal cost of production for a new [53:10] user on the internet is effectively [53:12] zero. [53:14] The incremental search user costs [53:16] nothing. The incremental social [53:18] networking user costs nothing, but [53:20] you're able to use them as part of a [53:22] network effect to sell ads against, make [53:24] infinite money. It's like a money [53:26] glitch. Okay? [53:28] AI is completely different. There is a [53:31] real cost for every marginal user. [53:33] Everyone you stand up is taxing a GPU. [53:36] Everyone you stand up needs electrons. [53:38] Everything needs memory. And so you get [53:40] into this situation where there needs to [53:43] be this downstream ROI, [53:47] but there also needs to be all this [53:48] critical infrastructure that even [53:50] enables you to be on the field. That's [53:52] probably the single best argument that [53:54] if I were sitting in government, I would [53:56] make as to why I should own a percentage [53:59] of these companies. It is not dissimilar [54:02] to having the first set of [54:04] transportation companies that run on the [54:06] interstate highway. And if the [54:09] interstate highways were built by the [54:11] federal government writ large and now [54:14] you had two companies that transported [54:15] all the goods, [54:17] a logical question at that time could [54:19] have been [54:21] how much of that should I own because [54:24] you're riding on my rails? And if you [54:26] look at it through that lens, there is [54:30] critical infrastructure and national [54:32] resilience that the AI labs profoundly [54:35] need. Otherwise, they're not in a [54:36] position to be in business. Yet, even [54:38] still, the marginal cost of production [54:41] is incredibly high. So, I think that if [54:45] you were going to go and take a piece of [54:47] these companies, you have incredible [54:49] leverage. You have to construct this [54:51] case precisely. But if it were me [54:53] running the sovereign wealth fund, yeah, [54:55] I'd probably own 75% of these companies [54:57] when I was done. But that's me cuz and I [54:59] can negotiate. [55:00] >> Yeah. [55:00] >> And I don't think Bernie can negotiate [55:02] out of a paper bag. So he'll just throw [55:03] this out there and it'll go nowhere. [55:06] >> That is the that is the key observation [55:08] that I hope everybody listening in a [55:09] position of power understands. The [55:12] incremental cost [55:14] of performing AI is excessive and large. [55:19] That contrasts and compares to the [55:21] incremental cost before AI which was [55:24] zero. Do with that information what you [55:26] will. [55:26] >> Yeah. I mean I my interpretation of that [55:28] would be yeah it's a fixed cost [55:30] business. You can add people after a [55:31] certain cost. that doesn't have an [55:33] incremental cost and that's what local [55:35] locally run open source models would do [55:38] in this instance that's what I would I [55:39] would take from it here's your poly [55:41] market chances of companies going public [55:44] before 2027 which is the end of this [55:46] year SpaceX 100% anthropic 83% chance [55:51] opening 48% chance they uh since our [55:54] last episode disclosed that they have [55:57] privately [55:58] filed so thanks to our partners what [56:01] does Your final thoughts on Bernie [56:03] Sanders approach of seizing the means of [56:07] producing intelligence. David Saxs in [56:10] the 1% of the 1%. What's your opinion? [56:13] >> I think I may be okay with Bernie's idea [56:17] in the event that it's a public benefit [56:19] corporation that says it's going to [56:21] cause massive job loss. [56:23] >> Okay. [56:23] >> That train for free on humanity's [56:25] knowledge but gatekeeps and refuses to [56:28] give back. [56:28] >> Okay. [56:29] >> Maybe. So you've come up with like a a [56:32] structure here of who should be seized. [56:34] >> Maybe right. Maybe it should be 75%. I [56:36] don't know. [56:37] >> Seize it. [56:37] >> I mean all the president needs to do is [56:40] call me and tap me on the shoulder. It [56:42] shall be done. [56:42] >> It shall be done. I mean these guys are [56:44] capitalist cops. I think they want [56:48] that's like their run over. [56:49] >> They're like seize my equity. Take it. [56:52] Take my equity. Well, Bernie Bernie does [56:55] have a good point, which is they train [56:56] for free on all of our knowledge, but [56:59] they're gatekeeping the output so that [57:02] >> their competitors, which is a very large [57:04] set of people, [57:05] >> cannot use it. And that seems wrong. [57:08] >> They should be forced to be open source. [57:10] Force them to open source. That's the [57:11] other [57:12] >> remember why Elon created Open AI in the [57:15] first place. He co-ounded it. [57:16] >> It was so valuable. It needed to be [57:19] afraid to everyone. Yes. [57:20] >> Yes. He was afraid that Google was going [57:22] to monopolize it. and he thought that [57:24] the only antidote for that was for it to [57:26] be open. [57:27] >> Yes. [57:28] >> And I still think that instinct is [57:29] completely true. It's just that it's not [57:32] Google who's the gatekeeping monopolist. [57:35] It's now anthropic. They're not a [57:37] monopouist yet. But if they get the [57:38] right capture they're looking for, they [57:40] will be either a monopolist or [57:41] >> duopoly. It'll be a duopoly. Somebody [57:43] start an open source frontier lab. Each [57:45] of the four besties. [57:47] >> There are no there are many. [57:48] >> There's a whole bunch of them. Jal, [57:50] >> it's not the labs that's the problem. [57:52] Jason, in the absence of power, this is [57:54] all a moot conversation. And if you [57:56] can't deliver power to hundreds and [57:57] hundreds of megawws and line of sight to [57:59] gigawatts, it's all meaningless [58:01] [ __ ] It's all [58:02] >> uh I don't know. I I would love an open- [58:04] source company that was focused on [58:06] running it locally. That's that's my [58:08] dream is, you know, make the model work [58:10] there. I think these guys are capitalist [58:12] cucks. I think their kink is like having [58:14] their equity taken and being regulated. [58:16] They're just like, that's their kink. [58:18] Dario's like, "Ooh, take my equity." Oh, [58:21] regulate me. That's it, Bernie. Take my [58:24] equity. I'm such a bad boy. I can't be [58:26] trusted. Regulate me, Elizabeth. [58:30] Regulate me. A [58:32] >> god. [58:33] >> Oh, it's so wrong. Sorry. Come on the [58:35] program anytime, Dario. [58:38] >> He's in a rush to come. He [58:39] >> will never come on this pod. Dario will [58:42] never running. He's running. [58:44] >> We didn't get Sam. We got Sarah Frier to [58:46] come on the pod. And she was treated. [58:47] >> Sam came on the pod. He has to come on [58:48] the pod. He's like, "Hey, came too." [58:50] Right. Sorry to come on. He's great. Sam [58:52] will come on the pot anytime. He's [58:53] totally by the way. Even even we didn't [58:55] treat him harshly at all. [58:57] >> No, I mean [58:58] >> I think people I mean we spent an hour [59:01] on him. [59:02] >> I think Sam to his credit is very [59:03] practical. He sees the tea leaves and [59:05] he's willing to negotiate and I think [59:07] that that versus being some kind of [59:09] moral absolutist will probably serve [59:11] OpenAI well in the future. [59:13] >> Yeah. It's just going to lose them all [59:14] these virtue signaling principled [59:16] employees to go go work for Daario. [59:18] That's why they go work for Daario. are [59:20] in line with their philosophies. [59:22] >> We did mention, but everybody should [59:25] go and see all of the recaps. We [59:28] published now all the interviews. Yeah. [59:29] Jason from Liquidity. Uh Sarah Prior, I [59:31] mean, we mentioned Sarah Fry, but she [59:32] crushed it. She was incredible. [59:34] >> The CFO of Open AI, [59:36] >> future CEO, I think, was the buzz in the [59:38] room. [59:38] >> I mean, [59:39] >> oh, stop that. [59:40] >> Stop. Can you stop? [59:41] >> That's what people were saying. I have [59:43] the right to free speech on this [59:44] podcast. I think [59:45] >> Look at you trying to introduce wedges. [59:47] This is [59:47] >> No, I'm not introducing wedges. I think [59:49] Sam would be better as chairman. I think [59:50] she would be the perfect CEO. [59:52] >> It seemed like a good partnership. [59:54] >> It seems like they have a great [59:56] partnership. I agree. [59:57] >> Sarah was really impressive. [59:58] >> Who else was your favorite? Sachs, did [60:00] you have a favorite? Did you have [60:01] somebody who thought I did? I thought [60:03] Thomas Leafant crushed it with his [60:06] presentation on the overview of venture [60:08] capital [60:09] >> and he did it at summit I guess a year [60:11] and a half, two years ago. [60:12] >> Two years ago, but we're going to make [60:13] it a yearly with him. Maybe. I think we [60:15] should make it an annual feature with [60:16] him because it's so interesting. [60:18] >> So good. It's so good. [60:19] >> It's so good. It provided so much um [60:21] numerical detail to support a lot of [60:23] things I was kind of feeling but didn't [60:25] >> have the data. [60:27] >> To summarize it, he's his basic [60:29] observation was in this last cycle, if a [60:32] company gets to a hundred billion [60:33] valuation, it's more likely to 10x from [60:35] there than from 10 billion to 100 [60:38] billion. [60:38] >> Well, he had data on this. Yeah, I [60:40] remember [60:40] >> he had data on it which is like, okay, [60:42] well, that's going to really change. I [60:44] remember the slide. [60:44] >> Yeah, there it is. It's 8%. I I [60:46] memorized it. [60:47] >> Yes. [60:48] >> Yeah. If you're if you're a capital [60:49] allocator, this one screamed. [60:51] >> Look. Yes, that's right. The odds of a [60:53] unicorn getting to decacorn was 8%. The [60:55] odds of a decacorn getting to a senicorn [60:58] was about double that was 13%. And then [61:02] the odds of it going from senicorn to a [61:05] trillion dollar market cap was 31%. So [61:08] double again. And I I asked him, you [61:10] know, could you extrapolate this on one [61:12] trillion getting to 10 trillion? My [61:14] guess is that it would follow and we'll [61:16] see about [61:18] >> 60% of [61:20] >> trillion dollar market cap companies [61:21] getting to 10 trillion in the next few [61:23] years. [61:23] >> These are called trillicorns by the way. [61:25] We call them trillion trilicorns in the [61:27] industry. So yes, [61:29] >> this is going to go over really [61:31] socialist crowd. People are going to [61:32] love this. Yeah, [61:33] >> it's just all you got to do is seize 10% [61:35] of a trilicorn and you pay off 2% of our [61:37] national debt. pretty good deal [61:39] actually. [61:39] >> All we need is 10% of 30 trilicorns and [61:42] we're debt free. [61:43] >> I mean [61:45] >> it's not as crazy as it sounds. [61:48] >> Is it more complicated than that? Now if [61:49] you send me in I'll get 60% of all of [61:51] them. [61:52] >> Yes. And then we'll have a surplus of 30 [61:53] billion and we can buy Australia. [61:55] >> We'll pay it off much sooner. [61:57] >> We could buy Cuba. [61:59] >> I'd buy Canada first. Better scheme. [62:01] >> I would you know what I would do? I [62:02] would [ __ ] with Canada. I would buy like [62:04] Vancouver or Toronto. Just like break [62:06] them up. just buy like one of the [62:09] regions. And let me thank let me thank [62:12] Thomas Keller, the French Laundry, EY [62:15] who sponsored our very innovative [62:17] meeting hub at Liquidity. Here are some [62:19] photos of that. The New York Stock [62:21] Exchange, they hosted the speaker dinner [62:23] at the French Laundry. Amazing pictures [62:26] here. They are of us with Thomas Keller [62:29] and they did a great interview on site. [62:31] Niagen, they hosted the wellness lounge [62:33] which included recovery IVs. They're the [62:37] official NAD partner of the All-In Pod [62:39] and all of our longevity. Our new [62:41] bestie, our new bestie, Jake Paul, was [62:43] there. He's a really thoughtful guy, [62:46] Freeberg. [62:46] >> Jake Paul was awesome. [62:47] >> I sat him next to Sachs for dinner. [62:49] >> Yeah, he's a great investor, by the way. [62:51] >> Fantastic. Fantastic. [62:52] >> His portfolio is awesome. He had a great [62:54] team. His team came with him for the [62:56] thing. [62:57] >> Great strategy. Yeah. [62:58] >> One thing I'll say about Thomas Keller, [62:59] he had a strong watch game. I don't know [63:01] if you guys noticed that. [63:03] >> Probably noticed it. [63:04] >> Nice potential. He had a aqua knot green [63:06] band rose gold a complication on there. [63:08] I was trying to trying to figure out [63:10] what the complication was. I couldn't [63:11] quite eyeball it. [63:12] >> I don't know if you saw it, Jimoth. [63:14] >> I was I looked at it and I was like, [63:16] "Wow, this is very it's very elegant." [63:17] Tell us Keller. I mean, the French [63:19] Laundry is just transcendent. Did you [63:21] have a favorite [63:24] dish? Of course. Was it the duck or was [63:27] it the Miyazaki? [63:29] >> The duck was excellent. The Miyazaki [63:31] beef was excellent, but the the tapioca [63:32] pearls I thought was the best. Oh, [63:34] oysters and pearls. That's your oysters [63:36] with caviar. Sax, what was your favorite [63:38] there? Do you have a favorite? [63:39] >> One of the caviar related dishes. [63:41] >> Yeah, that's oysters and pearls. [63:45] Round of golf. The following day, we [63:47] played a nine-hole scramble with [63:48] everyone. It was great. So great. [63:50] >> How much did you gamble and who won? [63:52] >> We had a $100 a hole [63:56] versus Chris Hoe, who's the chief [63:58] business officer of Athena. But, you [64:00] know, that [ __ ] is like a plus [64:02] one, so whatever. I think we lost three [64:04] units. So, we lost [64:04] >> 300 bucks. [64:05] >> All right. Shout out to our other [64:06] partner, Athena. Athena. Wow.com. Go get [64:09] a month free. All right. Whatever. I got [64:11] to stop showing. Most importantly, [64:14] everybody was very upset that liquidity [64:16] sold out and people couldn't get in. [64:17] There were like 3,000 applications. [64:20] It's going to sell even faster uh next [64:23] time around. So, go to allin.com/events [64:25] and get your applications in if you want [64:27] to come to Summit September 13th, 14th, [64:28] and 15th. Freeberg's doing some wild [64:32] stuff. some wild stuff for that event [64:34] and uh it's gonna be cool. [64:36] >> You're losing your mind. You're now [64:37] becoming very eccentric, unhinged in [64:40] your production value and Shimoth, [64:42] >> you've been very selective in your [64:45] approach to the liquidity speakers [64:49] because liquidity is different than the [64:51] summit. The summit is like a festival. [64:53] It's a huge party. There's all this [64:55] really interesting cross-sectional [64:57] stuff. But liquidity is meant to be for [64:59] one thing which is the most important [65:03] capital allocators in the world should [65:05] have one place to convene, debate, [65:08] learn, talk, build relationships. And I [65:11] want people to walk out of there with [65:13] relationships that they can use. And I [65:15] told you this before because capital is [65:17] what shapes the things that occur in the [65:20] world. So I think that we have to be [65:22] extremely selective in how we curate [65:24] every element of that show. It's not [65:25] meant to be that you can just buy a [65:27] ticket, but to just buy a ticket, which [65:28] is also a very fair and rational and [65:31] nice thing. The summit is a party. It's [65:33] a vibe. [65:33] >> It's awesome. And you're going to learn [65:34] a lot and have a good time. You'll be [65:36] entertained. All that great stuff. All [65:37] right, let's keep going. Breaking topic [65:40] this morning. CPI and PPI for May came [65:42] in hot. Inflation is ripping. Here we [65:46] go. CPI is obviously the consumer price [65:48] index that tracks inflation from the [65:51] buyer side. Rent, groceries, gas, [65:53] healthcare, all that stuff. PPI is the [65:55] producer price index that tracks [65:57] inflation from the seller side. [66:00] Wholesale goods, raw materials, yada [66:02] yada. CPI came in at 4.2% [66:06] year-over-year, highest since April [66:08] 2023. There's your chart. PPI came in at [66:12] 6.5% year-over-year, highest since the [66:15] end of 2022. [66:18] Here's your poly market. 21% chance [66:20] inflation hits 5% in 2026. Uh and then [66:25] an adjacent poly market chances of the [66:27] Fed hiking this year at 49%. [66:32] It was under 10% before the Iran war [66:36] started. By the way, European Central [66:38] Bank just raised rates a quarter point [66:40] on Thursday. This was their first rate [66:42] hike since September 2023. What do we [66:44] take from this Freedberg? It's all [66:48] related to the Iran war. Yeah, that's [66:51] the [66:52] >> there's definitely an energy blip from [66:53] the Iran war that drove the core index [66:55] up, but there's also the macro point [66:58] which is government spending out of [67:00] control, inflation out of control and [67:04] fundamentally as things unravel you have [67:07] rising rates. So I think we should kind [67:08] of expect especially with a Kevin Worsh [67:11] Fed, I think we could see north of five [67:14] and a half, 6% overnight rates. It's not [67:17] unforeseeable. So, [67:18] >> this is just the nature of our monetary [67:21] policy and our fiscal policy with our [67:24] congressional um members that vote all [67:27] of our future earnings away into [67:30] [ __ ] But here we are [67:31] >> and and far worse [67:34] >> that nobody expected. Chimath, your [67:36] thoughts and then I'll wrap up with you. [67:37] >> That was the most passive aggressive [67:38] take I've heard in a while from [67:40] >> He's a little upset about it. Yes. [67:41] Clearly he's a little bitter about just [67:44] so stupid to watch. [67:45] >> Can you tell us how you really feel? [67:46] What happened? Like honestly, like what [67:48] else is there to say? I've said it like [67:49] a hundred times. It's just idiotic. The [67:52] the core problem with wealth inequality [67:54] in this country, the core problem with [67:56] inflation in this country, the core [67:57] problem, all of it roots back to excess [68:00] government spending. End of story. All [68:03] of this, we're all talking about [68:05] machinations at the edge of the network. [68:07] The core of the network is overspending [68:09] by the government because in order to [68:10] get elected, you need someone to give [68:12] you something more than you have today. [68:13] That's how people get elected. And here [68:15] we are 250 years later. I don't know if [68:18] there's the will or the way to get out [68:19] of it, but one of the manifestations of [68:21] it will be in this late stage higher [68:23] rates and as a response to inflation and [68:26] here we are, you know. [68:27] >> All right, Chim, what's your what's your [68:29] take on this? Is it a transient [68:32] as we uh have had other presidents say? [68:34] Is it uh is there a offramp for the Iran [68:38] war? I don't know if this is above all [68:40] of our pay grades, but this does seem to [68:42] be a challenging moment happening right [68:45] before the midterms [68:47] and consumers expected that their cost [68:51] of living would go down and that they [68:53] would benefit from this golden age and [68:56] it hasn't happened yet. Well, look, it's [68:58] concerning. I don't think that I [68:59] expected the print to be this hot. Now, [69:03] what's [69:04] keeping things in line? there's still a [69:06] chance that we're not going to see CPI [69:09] go absolutely crazy. And part of that is [69:11] because China has been smoothing the [69:14] energy consumption globally worldwide. [69:16] And so we've kind of kept a damper on [69:19] $200 a barrel oil. In fact, we're [69:21] subundred. At the same time, I think [69:23] it's pulled forward a lot of more [69:28] practical sources of energy production, [69:30] not just in the United States, but [69:32] abroad. Specifically, what I mean is [69:35] solar. just because it's simple, it's [69:37] passive, and you don't need you don't [69:39] need a lot of permitting, and you don't [69:40] need clean air permits and the like. [69:42] >> So, all of these things have been able [69:44] to keep a lid on it. [69:46] >> If China somehow runs out of reserves, [69:48] and they need to go back into the spot [69:49] market to buy an extra 3 million barrels [69:51] a day, there's a very big risk that oil [69:55] gets into the well past 100 and maybe [69:57] between 150 and 200 a barrel. That has a [70:01] lot of downstream problems with respect [70:03] to CPI. [70:05] Not even because US energy supply, but [70:08] really because the inputs that oil feeds [70:11] all around the world that then get [70:13] absorbed by the costs that people pay [70:15] for everyday things could go up. [70:17] >> So I think the PPI number should be an [70:20] alarm that we use to offramp this around [70:25] situation sooner rather than later. I [70:27] just unless again because we don't know [70:29] what happened in China because none of [70:31] us were there. Unless there's an [70:33] understanding [70:35] that the president has with Xi about how [70:38] much actual supply and slack is in the [70:41] Chinese system and that may allow him to [70:44] take a much longer road to get a [70:47] conclusive and decisive win in Iran, but [70:48] I I don't there was supposed to be some [70:51] tangible known resolutions with this [70:54] China trip. We haven't heard them [70:57] explicitly stated yet. What's your take [70:59] on uh this Sachs? uh the inflation print [71:03] and is there like a a golden bridge to [71:06] to end this war in Iran? [71:08] >> Jake, the problem is when you asked me [71:10] that question, people assume that I have [71:11] some sort of insight information and I [71:13] don't look the PPI print is what it is. [71:15] I largely agree with what Chamas said. [71:17] The only thing I would add is that it [71:19] was largely in line with expectations. [71:22] That's why the market is actually up [71:23] today. Normally if there's a surprise [71:26] print and inflation is high then the [71:28] market is down because the market starts [71:30] pricing in the implication of higher [71:33] interest rates. It's not doing that [71:34] today. So it's in line with [71:36] expectations. But I largely agree with [71:39] what Chamas said. [71:40] >> Yeah. All right. Well, listen, we got to [71:41] get to other elections. You know, my [71:43] take on it is just huge colossal era [71:45] starting this Iran war and hopefully we [71:48] can get out of it. Um the downstream [71:50] effects of this are just disastrous. Um, [71:53] I'm hoping there's a quick resolution or [71:55] some way to get out of it before it goes [71:57] to, you know, a year or two or three. We [72:01] can't be in a forever war, obviously. [72:02] >> NASDAQ up 2 and a half% today. That [72:04] normally doesn't happen when there's a [72:07] hot print because again, like tech [72:09] stocks are the most susceptible to high [72:11] interest rates. So, the market at this [72:13] point seems to be pricing in [72:15] >> a resolution. But we'll see. We'll see. [72:17] >> We'll see. Yeah. I mean, it's just this [72:19] is a everybody knows Iran is a quagmire [72:24] and really was was a huge mistake. All [72:27] right, let's talk about the hand ringing [72:29] around vote counts in the LA mayoral [72:34] primary instead of me reading all these [72:37] details. Just Freeberg, I know you're [72:39] hot. [72:39] >> I think we know about it. Yeah. [72:40] >> Yeah. I think if you're on X, it's [72:42] essentially your entire feed. Uh [72:45] Freeberg, what's your take? Is the [72:48] election system in Los Angeles, [72:51] specifically in California, corrupt? And [72:54] is this election been stolen from [72:56] Spencer Pratt or is it just [72:59] >> There is no election. [73:01] >> There was no election. [73:02] >> There is no election. [73:04] >> Okay. So there is no spoon. Your rights [73:07] to have an election are gone. [73:10] You are a citizen of those who tell you [73:15] who your overseers are. You are no [73:17] longer allowed to vote [73:20] >> allowed to vote [73:20] >> for your elected representatives. They [73:22] are now appointed representatives. So [73:26] enjoy the appointments that have been [73:27] made by who? [73:29] >> Appointed by those who have constructed [73:31] the matrix. [73:32] >> Got it. Now I understand the background. [73:34] There is no spoon. Agent Smith and all [73:38] the Agent Smiths are [73:41] in charge. So, here's some statistics [73:43] for you. In-person voting the day of the [73:46] election in Los Angeles County for [73:49] mayor. Spencer Pratt 35%. Karen Bass [73:53] 29%. Nipia Ramen 26%. [73:56] mail-in ballots received before election [73:59] day, 38% for Bass, 28% for Pratt, 20% [74:04] for Ramen. And then all these ballots [74:07] that arrive after election day, 37% for [74:10] Ramen, 35% for Bass, 19% for Pratt. So [74:15] Pratt's [74:17] post election mailin ballots declined by [74:20] one/ird. So statistically the population [74:22] of people that send in their ballots [74:24] late reduced for Pratt by a third [74:26] increased for Nipia Ramen by 80%. [74:29] And Karen Bass 10% less if you just look [74:32] at the mail in ballots before and after [74:33] election day as a comparison. I don't [74:36] know if there's a socopolitical way that [74:38] you can assess those statistics and [74:41] assume that these are individuals [74:44] casting their individual vote for who [74:46] they think should be mayor of LA. And [74:48] Nick, if you'll just pull up the map, [74:49] which I think is worth taking a look at. [74:51] Basically, the concentration of [74:52] incremental votes that Nitia Ramen got, [74:55] came around the Skid Row area in Los [74:57] Angeles. And look, I'm not an election [74:59] denier. I'm not someone who's [75:01] historically believed the idea that [75:03] elections are fraud and people have [75:04] stolen votes. But when you look at the [75:06] basic statistics of what happened in [75:08] person, mail in before, mail in after [75:10] election day, it becomes a real [75:13] statistical quagmire on how did this [75:16] sort of a sociopolitical shift happen in [75:19] such a way that it did. Now, there was a [75:21] report published, Nick, if you could [75:22] pull this up by the US House of [75:25] Representatives Committee on House [75:26] Administration. This was published in [75:28] May of 2020 and they highlighted the [75:32] 2018 California midterm elections and [75:34] the challenges they saw arise in that [75:36] midterm elections because of some of the [75:38] legislative changes that were made. [75:40] First, California Assembly Bill 1921 [75:42] legalized the practice of unlimited [75:44] ballot harvesting in the state. This was [75:46] passed around 2018 around the midterms. [75:49] What that means is that any individual [75:51] in the state of California has the right [75:54] to go and collect ballots from any other [75:58] individuals regardless of relationship, [76:01] fill them out and send them in. [76:04] California 2 years later, 18 months [76:06] later, also passed a law that made it [76:08] permanent that every person registered [76:11] in the state of California would get a [76:12] ballot. So tens of millions of ballots [76:14] then get mailed out. Then there was [76:16] another series of laws that were passed [76:18] that said anyone can register to vote. [76:20] You don't need to prove your [76:21] citizenship. You can use a gym [76:23] membership card as an example. So anyone [76:26] can register to vote. There is no proof [76:28] of ID when you get a ballot. There is no [76:30] demonstration that the person who fills [76:31] out the ballot has anything to do with [76:32] the individual who's supposed to be [76:33] voting that ballot. And it is legal for [76:35] an individual to go out and collect [76:37] hundreds or thousands of ballots, ship [76:39] them in, and they will all qualify in [76:41] these kind of mail-in ballot voting [76:43] processes. So there's nothing illegal or [76:46] fraudulent going on. In fact, the system [76:48] is operating exactly as intended. It has [76:51] been set up and structured in a way that [76:53] with the right construct, you can get an [76:56] individual appointed, not elected, but [76:58] appointed to a particular role in [77:00] government under a quote free election [77:02] in California. This is a foundational [77:05] destruction of our rights to vote for [77:09] people in a free democracy. I feel it's [77:11] been eroded slowly over time to such an [77:14] extent now and I'm not some crazy MAGA [77:16] mother whatever people want to classify [77:18] me as for pointing this out but you can [77:20] go down this list they are all written [77:22] out in this document every one of these [77:23] laws that were passed in California over [77:26] time that in aggregate create the [77:29] environment and the construction for [77:31] elections to become appointments and no [77:33] longer free democratic elections where [77:36] the principle very importantly the [77:38] principle should be one individual [77:41] one vote and if you opt to not vote, [77:44] your vote should not be counted. So I [77:46] think that there's no fraud. I don't [77:49] think there's anything. [77:49] >> It sounds like you believe there's [77:50] massive fraud. [77:52] >> Yeah. I think this is not fraud. This is [77:53] this is the way the laws are set up. [77:55] >> It's not fraud. [77:56] >> It's the way the law is set up. [77:57] >> Okay. Sachs, you say a fraud. [77:59] >> This is the system of appointment. [78:00] >> Okay. Yeah. It sounds to me like you're [78:02] describing fraud, but you're saying the [78:03] system allows fraud. [78:05] >> Fraud is breaking the law. Fraud is [78:07] breaking the law. The law allows them to [78:09] do it. If you paid people on Skid Row, [78:11] that's not legal. [78:12] >> That is illegal. But you are allowed to [78:14] go out and get anyone to register. [78:17] You're allowed to go out and collect all [78:18] the ballots. And the people that are [78:20] doing it, as long as you're not getting [78:21] paid per ballot, you're legally allowed [78:22] to do it. Let me pause for one second. [78:25] When you look at any one of these laws [78:26] in isolation, you can understand that [78:29] there is some altruistic intention or [78:31] there can be framed to be an altruistic [78:34] intention behind it. Increase voter [78:36] activity. Give everyone the right to [78:38] vote. give more people access, more [78:40] channels to vote, etc., etc. And you can [78:42] rationalize away how every one of these [78:44] laws makes it easier and adds [78:47] lubrication to the system. And now we've [78:49] got a quote more accessible democracy [78:52] for everyone. But the truth is that [78:54] there is a concept in insurance called [78:56] adverse selection, which is if there is [78:58] a moment to exploit something, the [79:00] criminals will exploit it to the extent [79:01] that it will become asymmetric. the [79:03] criminals or the or the do bads, not the [79:06] dogooders, will find a way to take [79:08] advantage of a hole in a system just [79:11] like hackers do to break into a network [79:14] >> and that's what you system ramen and [79:16] Spencer Pratt did. [79:17] >> This set of mechanisms, this set of [79:19] construction has allowed for a system [79:21] that now has evolved into an [79:23] appointment. So you believe worked the [79:26] system better than Spencer Pratt? [79:28] >> She had a better cheating operation. [79:30] Obviously, you can see it in the chart. [79:32] the just look at the number of votes. [79:34] >> Okay, define cheating here just so we're [79:36] clear. [79:37] >> How does it happen where Bass stays flat [79:40] and then Ramen and Pratt just switch [79:43] places? So on election day, he's the [79:45] clear number two and then they keep [79:47] counting for a week. They don't even [79:49] know how many ballots they're supposed [79:50] to have. I mean, none of the stuff is [79:51] locked down. Okay, let me explain how [79:53] the cheating works. And to Freeberg's [79:55] point, it may or may not be legal. So I, [79:58] you know, I don't want to get [79:59] sidetracked by whether the fraud is [80:01] legal or illegal, but I think we all [80:03] know it's crooked. And here's how it [80:05] works. First of all, [80:06] >> yeah, explain what's I'll explain it to [80:08] you. I'll explain it to you. And if I if [80:10] I misstate a fact, then let me know. [80:12] >> So, first of all, California sends out [80:14] ballots to every registered voters. It's [80:17] something like 23 24 million people. [80:20] Only 9 and a half million people vote. [80:22] So, what happens to the other 14 million [80:24] ballots? I have a friend who lives in [80:27] LA. He manages apartment buildings. He [80:29] says that the ballots just pile up, you [80:31] know, at the mailboxes. Yes. That [80:33] everybody has in the buildings. There's [80:34] like loose ballots everywhere. They're [80:37] sent to people who don't even live there [80:38] anymore. He like recognizes that the [80:40] names, they're people who haven't lived [80:42] there in years. So, the voter roles are [80:45] dirty. They haven't been cleaned up in [80:46] years. Okay. Then you have the issue [80:49] that there's no voter ID on election [80:51] day. There's no voter ID to even get [80:53] registered. When you sign your ballot, [80:56] they run it through a machine. As long [80:57] as the signature matches 40%, I don't [81:00] even know what 40% means. It's like less [81:03] than half accurate, the machine says [81:05] it's fine. Okay? And then it doesn't [81:08] even check the signature of the voter if [81:12] you have a supposed witness. So the [81:15] voter can just put like a little X on [81:17] their signature as long as the witness [81:20] says and they sign, but there's no [81:22] registration of witnesses. Anybody can [81:24] be a witness. There's no database of [81:26] witnesses. And so the witness just puts [81:28] a squiggle. So what is to stop the [81:31] so-called ballot harvester from going [81:34] filling out these ballots, putting a [81:36] little X on the voter, putting a little [81:38] squiggle on on the the witness, and then [81:40] dumping these ballots. There's no chain [81:42] of custody. So the risk would be they [81:44] would be arrested and [81:46] >> and right [81:48] >> well what is their expectation getting [81:50] caught? Gavin Newsome just signed a law [81:52] making it virtually impossible to audit [81:54] any of this stuff. [81:56] >> So now even if they thought maybe they [81:58] could get caught. Let's just take a look [82:00] at all the other types of fraud [82:01] happening in California right now that [82:03] we know about. We know that there's [82:05] billions of dollars in fraud in [82:07] California for things like Medicaid [82:09] fraud, unemployment fraud, EBT fraud, [82:12] hospice fraud, fraud related to the [82:15] homeless industrial complex, and on and [82:16] on. That's all been proven. Okay. Is it [82:19] really so hard to believe that some of [82:21] the same groups, the same interest [82:23] groups, the same NOS's would be willing [82:25] to exploit these loopholes in the dirty [82:29] voter roles in the millions of ballots [82:31] that go to incorrect or non-existent [82:33] addresses, the non-existent chain of [82:35] custody, the non-existent signature [82:37] verification, the no ID, not only to [82:40] vote but to register, counting ballots [82:43] without postmarks if received 7 days [82:44] later, registering thousands of ballots [82:48] from homeless shelters that don't even [82:49] have any beds. These are all known [82:51] things. So when you say that there's no [82:54] fraud, it's like come on, there's [82:57] loopholes and all you got to do is look [82:59] at the results and see that there's [83:02] something crooked going on here. [83:04] >> Okay, so you're in the camp that there [83:06] is fraud going on because where there's [83:08] >> it may be legalized. It may be [83:09] legalized, but it's fraud. [83:12] >> Okay, so you believe there's fraud. [83:14] Look, when you start mailing out [83:15] millions of ballots to people who don't [83:16] even vote, and then you have no audit [83:19] requirements, no chain of custody, no [83:21] punishments, no penalties, no [83:23] investigations, and there's a huge tons [83:25] of investigations both by attorney [83:27] generals [83:28] >> in California. [83:30] >> By who? [83:30] >> Uh I I'll pull up the history. I will [83:33] say after every election, there are tons [83:36] of civil lawsuits filed. Like obviously [83:38] Trump filed 60 of 60 when [83:40] >> not in California. Uh I don't know just [83:43] made it impossible to do that kind of [83:44] audit. He just I think that's because he [83:48] was beaten so badly in a democratic [83:50] state. There was no reason to. But I [83:51] don't know why he didn't file in [83:52] California. But he filed 60 of these and [83:55] everybody kind of does that. They file [83:56] these constantly and attorney generals [83:58] can. So in this case, Chimoth, if [84:01] there's so much smoke, if this statistic [84:04] seems so improbable, and there was the [84:07] Nick Shirley video of going to Skidro [84:09] and this possibility that they were [84:12] paying junkies and people who were high [84:14] to vote, this would be easily catchable. [84:17] Actually, this would take a conspiracy [84:19] for thousands and thousands of votes to [84:21] have been swung. And that's what we're [84:22] saying here. We're saying upwards of [84:24] 10,000 votes were swung, but certainly [84:26] 5,000. Good luck. You would need a [84:28] conspiracy of hundreds of people or at [84:30] least dozens to do that. [84:32] >> Good luck going to Skid Row and getting [84:33] some meth addict to go on the record. [84:36] >> I think they actually got somebody to go [84:37] on the record actually. That's a that's [84:38] the [84:39] >> There's a few videos of this. Yes. [84:40] >> Yeah. Yeah. No, that that's the crazy [84:42] part about it is they uh are saying [84:44] Skidro people got paid two or three [84:46] dollars and that somebody was found [84:48] guilty of this. Brenda Lee Brown [84:51] Armstrong. She plead guilty in 2026 of [84:54] paying. system is designed to allow [84:56] fraud. It's this it's designed [84:59] >> it has been won by 43,000 votes to [85:02] >> guys. We live in a [85:06] autocracy in California where one [85:09] political machine has shaped the laws so [85:14] inexurably [85:15] that whether it's a mayor or it's a [85:18] state legislator or it's a state senator [85:21] or it's the governor, it's a one-way [85:24] door. There is only one party that can [85:27] win [85:28] and we are slowly seeing that that party [85:32] increasingly [85:34] does not have the ability to govern a [85:36] state this dynamic and this complicated. [85:39] They've changed the rules to make sure [85:40] that that monopoly is enforced. [85:43] We have all tacidly approved it by [85:47] supporting or voting or not saying [85:49] anything. And the end result has yet to [85:51] be unveiled because it's going to take [85:53] another couple of decades for all of [85:55] this rot to really destroy these places. [85:57] The reality is the voters probably [86:00] should be given a chance to have Spencer [86:02] Pratt push back on what is the machine [86:06] so that at a minimum what they learn is [86:10] that the machine wasn't working nearly [86:13] anywhere close to the way it is [86:14] intended. Instead, what will happen is [86:16] two people that are in various shapes [86:19] and forms pretty deeply unqualified [86:21] and very similar now essentially get to [86:25] give you not much of a choice at all. [86:29] So, I think the the the sad thing is [86:31] Freeberg is right and Sax is right. [86:34] Freeberg is right. The laws have been [86:37] shaped this way to make legal what would [86:40] otherwise be an illegality. and Sax is [86:43] right, which is that any rational person [86:45] should look at these laws and say this [86:47] is fraudulent. [86:51] The unfortunate situation is unless [86:53] there's a some cataclysmic decision in [86:55] California, this is what we're stuck [86:57] with. It is time to vote all of these [87:01] people out. [87:02] >> You can't you can't vote. What do you [87:04] mean? How you going to do it? Anyone can [87:07] register anyone can vote. Spencer Pratt [87:10] can do [87:12] this. Can I answer Free Brook's [87:14] question? [87:14] >> Sure. [87:17] >> There is one break the glass way and [87:20] I've talked to a few of these guys about [87:21] what that is. [87:24] How do you change the system? You must [87:26] get somebody at the absolute top of the [87:28] ticket elected. [87:31] That person is Steve Hilton. [87:33] And if that happens, there is one thing [87:35] that Steve Hilton can do to start to [87:38] completely change the way that this [87:39] state is run, which is instantly on day [87:42] one declare a state of emergency. If you [87:44] pull on this thread, what you will find, [87:46] there are lots of things you can do to [87:48] clean up how California works under a [87:51] state of emergency. [87:53] I'll leave it at that. [87:56] Yeah, it there's some common sense here. [87:59] We shouldn't send ballots out to [88:02] everybody. that made sense maybe during [88:03] co and if you can't be bothered to have [88:06] ID then maybe you shouldn't vote uh and [88:10] I don't think that there has been an [88:13] election a major election that has been [88:15] swung by election fraud the Heritage [88:18] Foundation has spent many many tens of [88:21] millions of dollars on this I understand [88:23] and they have a really great uh project [88:25] that tracks this people file lawsuits [88:27] like Trump did like Al Gore did [88:29] everybody's always filing lawsuits over [88:31] this it's ever enough to swing an [88:34] election, but there's definitely on the [88:36] margins. [88:36] >> Kennedy versus Nixon. Bro, [88:38] >> hold on. Let me finish. [88:39] >> The mob got Kennedy elected versus [88:40] Nixon. [88:41] >> I'm talking in modern history. Yeah, in [88:43] modern history. [88:43] >> That's pretty famous. That's pretty [88:45] famous. [88:45] >> Okay, [88:47] we we can get into if that's true or not [88:50] as well. The [88:51] >> Check out Lynon Johnson's history. [88:52] Lyndon Johnson only got elected to the [88:54] Senate. He got a start in politics [88:56] through ballot fraud. [88:57] >> We absolutely should have voter ID. We [88:59] should stop sending everybody one of [89:00] these because the appearance of [89:01] impropriety is the problem here. Both [89:04] sides crazy new flash news flash [89:07] everybody. Both sides cheat. Both sides [89:09] bend the rules. One side does it by [89:11] limiting access. The other does it by [89:13] providing too much access. All these [89:15] politicians, overwhelming majority of [89:17] them are unethical. And all they care [89:19] about is getting elected. And they have [89:20] tons of people working for them in both [89:22] parties to massage these election [89:25] results. And if you get caught, it is a [89:28] serious penalty. There has not been in [89:30] modern history major elections. Go ahead [89:33] and go to the Heritage Foundation. They [89:34] find all the nooks and crannies of [89:36] >> You keep using the word fraud like as if [89:38] it's a thing, but it's not. I'll tell [89:39] you who's at fraud. [89:40] >> No, there's fraud. At the end of the [89:42] day, there is fraud. There are people [89:44] have been caught doing fraudulent [89:46] things. [89:46] >> The one group that is fundamentally to [89:49] blame is Congress. We have an [89:51] opportunity. There have been [89:52] opportunities to pass laws that will [89:56] make it illegal for people to vote if [89:59] they are not citizens, if they are not [90:01] properly registered, and if they don't [90:02] show up with ID. That is a very simple [90:05] solve to this democratic electoral [90:07] process. We can't have a situation where [90:10] someone is legally allowed to go out and [90:11] collect a 100 ballots that they find [90:13] sitting in the lobby of an apartment [90:14] building or for people to register to [90:16] vote with a gym membership card when [90:18] fundamentally the federal government [90:20] already knows if you're not paying your [90:21] taxes, you're required to have a [90:23] driver's license to drive a car, you [90:25] have to have ID, you have to be [90:26] registered in the system to do nearly [90:29] anything and everything in this country [90:30] as it is. So to exclude that very simple [90:33] requirement from the most foundational [90:36] obligation that we all have as citizens [90:38] of this country is inherently wrong and [90:42] it needs to be addressed by Congress. [90:44] And if Congress fixes it and there are [90:45] federal laws that create systems that [90:48] ensure that elections are all about one [90:50] person getting one vote and they're [90:52] legally allowed to vote, then this [90:55] doesn't become an issue. Short of this, [90:57] I worry where this goes because over [91:00] time, the fewer people actually believe [91:02] that elections are real. And I hate [91:04] being on here sounding like I'm an [91:05] election fear-mongering guy who's [91:07] telling people that elections aren't [91:08] real, but the longer this goes on, the [91:11] more difficult it will be for people to [91:13] actually have faith in the system that [91:15] we're all operating within. And that's [91:17] when things get ugly and it needs to be [91:18] fixed. [91:19] >> Yeah, I'm strongly agreeing. If you [91:22] can't be bothered to get a driver's [91:24] license or a passport, maybe you don't [91:26] need to vote. Yeah. [91:28] >> Congress can fix it. And they should. [91:30] >> Yeah. And and there's, by the way, [91:32] there's a ballot measure in California. [91:33] If people in California want this, they [91:35] can just vote for the ballot measure. [91:37] >> You mean the voter ID ballot measure? [91:39] >> Yeah. [91:39] >> Yeah. That's very important. Actually, [91:41] look, if that fails, it's kind of our [91:43] last chance to save anything [91:45] >> for California's last chance. Yeah. [91:47] Yeah. And you do sound like an your [91:49] entire speech, David, was you sound like [91:51] an election denier. [91:52] >> I don't care what you label it. Look, [91:54] the the Democratic party as a whole were [91:56] election deniers about Trump winning in [91:58] 2016. They concocted the whole Russia [92:00] gate hoax to basically try and discredit [92:02] his election. [92:04] >> When you look at the the difference [92:05] between walk-in and mailin ballots, and [92:07] I've seen everyone try to argue why [92:08] that's so different, it doesn't compute [92:10] for me. That's it. I'm not like trying [92:12] to make up a fact. This is the fact. [92:15] Look at this facts on the on the on the [92:16] screen you're looking at right now. This [92:18] is the difference in the data. This data [92:20] tells me that something is wrong. That's [92:22] it. [92:22] >> Obviously, [92:24] >> I'm not trying to I'm not trying to just [92:25] fearonger because I don't like who got [92:26] elected. I don't give a [ __ ] what goes [92:28] on in LA. In fact, I think it's better [92:30] for LA if they deal with the socialist [92:32] mayor for a couple years so they come [92:33] out the other side realizing they [92:34] shouldn't embrace socialist principles. [92:36] I think it's great for the country. [92:37] Actually, I think it's great for New [92:39] York. I think it's great for Seattle. I [92:40] think it's great for wherever Minnesota. [92:42] Wherever people want to elect socialists [92:43] in this country, let's run into it. I'm [92:45] all about jumping through the socialist [92:47] well so we jump out the other side like [92:48] the Latin American countries are doing [92:50] now when everyone wakes up and realize [92:52] how up socialism is because it denies [92:54] everyone the opportunity for economic [92:56] mobility. It denies everyone their [92:58] individual liberties that this country [92:59] was founded on. Socialism is wrong. It's [93:02] broken. It doesn't work. And if you all [93:04] want to go through that lesson, do it. [93:06] Go for it. Let's do it. Let them get [93:08] elected. But the numbers on that [93:10] election data show me that that's not [93:12] what the people wanted. That it's not [93:14] very clear or evident to me that that's [93:15] what people voted for. And that's where [93:17] I'm like, my god, the system itself has [93:19] become a system of appointment, not a [93:20] system of election. And I get fired up [93:22] about that personally. You want to add [93:25] anything as we wrap here on this? [93:26] >> Yeah. The party told you to reject the [93:29] evidence of your eyes and ears. It was [93:31] their final most essential command. This [93:34] is obvious. Just look at the charts. [93:36] Look at the data. We know something [93:38] crooked happened. It's statistically [93:40] impossible. The only thing we don't know [93:42] is whether it was illegal fraud or legal [93:45] fraud. Meaning they've created so many [93:47] loopholes and so much corruption in the [93:49] law that what they did is possibly [93:52] legal. It is still corrupt. It is still [93:54] crooked. It is ridiculous. Everyone can [93:57] see it. Do not deny the evidence of your [94:00] eyes and ears. Even if they call you an [94:01] election denier, I personally don't [94:03] care. I deny it. Spencer Pratt should be [94:06] in the runoff. I deny that Ramen won [94:08] legitimately. [94:10] >> The statistics are the statistics and [94:12] they would scream that you should have [94:14] the attorney general of California and [94:17] the federal attorney general look into [94:19] it and if they catch people doing this [94:21] stuff and they have caught one person [94:22] and Nick Shirley went down there and you [94:24] see people on Skid Row, you know, [94:26] listen, homeless people should have the [94:27] right to vote theoretically. [94:29] >> Why did Nome just sign a law making [94:31] audits like that even harder? [94:32] >> Yeah, I'm not defending Newsome here. [94:34] I'm saying where's the attorney general? [94:36] If there is in fact fraud, Trump has the [94:39] ability with his hands-elected attorney [94:40] generals to go after this. He controls [94:43] that part of uh you know our our [94:46] democracy. He should file and they [94:49] should go figure out what happened [94:50] statistically. So Trump, President [94:52] Trump, thank you, President Trump. Go do [94:54] it. Go do it. You got this attorney [94:56] general chasing down everybody else. Go [94:59] to this election. Use your powers. Audit [95:02] it and find them. Where's President [95:04] Trump? [95:05] >> You didn't hear my point that the the [95:07] state law just made it harder for the [95:09] federal government to come in and audit [95:10] the voter roles. So [95:11] >> yeah, he can still do it. [95:12] >> I'm sure they will do it. I'm sure [95:13] they'll do it if they can. [95:15] >> Yeah. I mean, trust me, Trump does what [95:17] he wants. If we've learned anything, [95:20] >> he will do what he wants. And he's not [95:21] doing this now. [95:22] >> What I find amazing is that the media [95:25] instead of asking tough questions like [95:27] how did this happen? How could the votes [95:29] swing so wildly in ballots that were [95:31] only found after hold on after election [95:34] day? The media instead of trying to [95:37] demand answers and demand accountability [95:40] is trying to cover up and say nothing [95:42] wrong happened here. It's unbelievable. [95:44] It's outrageous. [95:45] >> What is the best theory steelman? What's [95:47] the best theory of why this discrepancy? [95:49] >> I mean, we're sitting here talking about [95:51] it. [95:52] >> What's the discrepancy? Anybody have a [95:54] steel man of why this discrepancy could [95:56] have happened that's legitimate? [95:57] anybody. [95:58] >> If the media is hanging on to a snapshot [96:02] of the past where people actually cared [96:06] what they said and every time that they [96:09] can actually [96:12] reestablish trust, they kind of don't do [96:15] it because the people that [96:19] they want to curry favor with are [96:21] willing to play along. So I think what [96:24] happens is the media is like, well, I [96:25] have two choices. I can actually say [96:27] that Spencer Pratt should have won or [96:30] could have won or investigate why he [96:32] didn't win, but then I'm kind of tacitly [96:34] supporting Trump. And hold on a second. [96:36] I'm supposed to hate Trump. So, no. [96:38] Yeah, you know, this is completely [96:39] normal. Nothing to see here. Everybody [96:41] moves on, hoping that then they get [96:43] ingratiated into the machine, but [96:45] they're going to get run over and used [96:46] like everybody else does, too. So, [96:49] >> that's what's happening. Yeah, your [96:51] point is so spot on, Chimat, that like [96:53] if it's so polarized, meaning that [96:55] there's two polls that if you do [96:58] question, it's almost like you were [96:59] being put in the Trump proTrump camp. [97:01] Just like the word MAGA is thrown around [97:03] like, oh, you're a MAGA guy because [97:04] you're questioning election integrity. [97:06] This has nothing to do with MAGA. It's [97:07] just simple conversation about a [97:09] particular topic. But if it's diametric [97:12] to the other camp, you're, you know, you [97:14] must be in that bad camp. And so then [97:16] they don't want to be put in that camp. [97:18] It's so true. I'm for mathematical and [97:20] statistical literacy and what happened [97:22] here is mathematically and [97:23] statistically. [97:24] >> So nobody has any steel man or concept [97:27] of how this could have happened. [97:29] >> I can tell you the statistical odds that [97:31] this would have happened and it's one in [97:33] a trillion. [97:34] >> Okay. What if [97:35] >> So are you telling me the one in a [97:37] trillion shot hit? [97:38] >> I I'm asking if there is any I'm not [97:41] >> just just this one time. [97:42] >> I'm not I'm asking if there's anybody [97:44] who steals this in any way. I'm asking [97:46] I'll tell you what I've heard integrity [97:48] of [97:48] >> Why is it so psychologically important [97:51] for you to cling to the idea [97:53] >> it's not [97:55] elections that California elections are [97:57] completely corruption free after all the [97:59] things they've clearly done [98:01] >> to corrupt them voter ID you admit [98:04] there's like it's chalk full of [98:06] loopholes [98:06] >> you ask me why am I holding on to it I'm [98:08] not holding on to it for the purpose of [98:10] discussion and having an intellectually [98:12] honest discussion I'm asking the three [98:14] of you if there you've heard or if you [98:16] have any theory just for the [98:18] rigorousness of the discussion. This [98:20] isn't my position. [98:21] >> Where are all the democracy protectors [98:22] in the media now? [98:24] >> They all said that they were they were [98:26] the guardian they were the guardians of [98:27] democracy. They're going to protect [98:28] democracy. Where are they to ask [98:30] question? [98:30] >> The only theory that I can think of I'll [98:32] steal a minute myself. I because I I [98:34] think you do need to be rigorous. [98:36] >> You go ahead first. Yeah. The the only [98:37] thing I can think of is her ground game [98:39] is more sophisticated because they've [98:42] been at it longer than Spencer Pratt and [98:44] he didn't have a ground game that took [98:46] this in since he's a non-traditional [98:48] firsttime candidate. That's the only [98:50] thing I can think of. [98:51] >> Sophisticated a coded word for he wasn't [98:53] bribing Skidro homeless judge. [98:55] >> No, he just has it. He doesn't have an [98:56] election machine to collect ballots and [98:58] collecting ballots is legal. That's the [99:00] That's the question. [99:02] >> She had a better cheating operation. [99:03] Obviously, [99:04] >> we're saying the same thing. We're not [99:06] actually saying anything different. [99:07] >> The Twitter commentary was that people [99:11] were waiting [99:13] because there's so many people that are [99:15] so against Spencer Pratt, but they [99:18] wanted to be sure that they voted for [99:21] the right candidate, whether it's Ramen [99:23] or Bass. And so the people that held [99:25] their ballots the longest saw the [99:27] polling data that Bass was going to win. [99:29] That's why they all voted for Ramen last [99:30] minute was to try and promote her to [99:32] make sure [99:33] >> has been shipped because of the you [99:35] could have two Democrats at the top of [99:36] the ticket. [99:36] >> That's right. And they didn't want [99:37] Spencer Pratt to have a shot at the [99:39] general. And that by the way, you know, [99:41] I mean, if you saw my interview with [99:42] Spencer Pratt, he said Ramen registered [99:44] one hour before the deadline and she's [99:46] friends with Bass and that the whole [99:48] point was to keep him from actually [99:50] making it onto the general ballot. So [99:52] that some of the coordination that [99:53] people were making in their minds was [99:55] like, we got to make sure we vote for [99:56] the two Democrats and then they'll [99:57] figure out, hey, is it wrong? Let's just [100:00] box him out. [100:01] >> Yeah. [100:01] >> Got it. [100:02] >> Which is like why they all waited till [100:03] the last minute to vote for to vote for [100:05] how do you coordinate something like [100:07] that? That that's always the question I [100:08] have with these like [100:09] >> these people are all brilliant geniuses. [100:11] It's it's hundreds of thousands of [100:12] brilliant geniuses row that out. [100:15] >> Yeah, there's a chat group. I mean, it's [100:18] not that complicated. [100:19] >> Alan Harvesters did it. what Freeberg [100:21] said just happened. It's just that the [100:23] ballot harvesters did it, not the actual [100:25] voters, obviously. [100:26] >> And ballot harvesting is legal, but it's [100:29] sketchy and maybe they And if there are [100:31] ballot harvesters, I think that Trump [100:33] and the DOJ can catch these people. [100:36] >> How the is it legal that someone can go [100:38] out and collect other people's ballots [100:40] and fill them out and send them in? It's [100:42] the craziest to me. [100:43] >> No evidence of fraud right now. [100:44] Statistics don't look good. Donald J. [100:47] Trump, you control the Department of [100:48] Justice. Go investigate it. [100:50] >> Okay. I love you guys, but I'm hungry. I [100:52] got to go. This has been [100:54] >> an amazing episode of the Allin Podcast. [100:57] Bye-bye. [100:58] >> Love you boys. Byebye. [101:02] >> Let your winners ride. [101:09] >> We open sourced it to the fans and [101:11] they've just gone crazy with it. [101:21] Besties are gone. [101:24] >> That is my dog taking a notice in your [101:26] driveway. [101:29] >> Oh man, my habitasher will meet me up. [101:32] >> We should all just get a room and just [101:33] have one big huge orgy cuz they're all [101:35] just useless. It's like this like sexual [101:37] tension that we just need to release [101:38] somehow. [101:40] Wet your [101:42] wet feet. [101:45] We need to get Mercy's going all in. [101:55] I'm going all in.