Aschenbrenner Offloads Public Book to Citadel After July Slump

7月暴跌后，Aschenbrenner将公开市场持仓转让给Citadel Situational Awareness LP通过向Citadel进行场外投资组合转让，在月内迄今67%的暴跌中得以幸存，使Leopold Aschenbrenner保全了其Anthropic持股并避免被清盘。 在本期节目中，主持人John Coogan与特邀联合主持人Tyler (Ultradoo)探讨了Leopold Aschenbrenner对其对冲基金Situational Awareness LP遭遇严重回撤的回应；此前一份遭泄露的LP信件披露了该基金通过与Citadel进行场外大宗交易重组其公开市场股票持仓。主持人们分析了华尔街的掠夺性交易机制，将Aschenbrenner的处境与Archegos等以往重大市场爆仓事件进行对比，并探讨了他持有的包括Anthropic在内的高确信度非公开资产。此外，他们还梳理了大型科技公司（包括Microsoft、Apple、Amazon、Meta和Alphabet）最新的重磅季度财报，并重点介绍了OpenAI最新的模型降价与基准测试工具优化。 Leopold Aschenbrenner 透露其对冲基金 Situational Awareness LP 在 7 月遭遇 -67% 的净业绩暴跌，险些造成永久性资本损失，但得益于此前的高速增长，该基金年内（YTD）净回报率仍达 +80%。 (0:01) 尽管竞争对手恶意狙击做空其公开持仓并引发“挤兑”动态，但与当年 Archegos 崩盘不同，该基金通过有序交易完成了平仓，并未引发银行危机或系统性风险。 (3:22) 为消除杠杆与追缴保证金风险，Situational Awareness 放弃了出售 35 亿美元 Anthropic 私有股权的计划，转而将公开股票组合打包出售给 Citadel 并平掉所有空头。 (14:39) 最新财报显示，微软、亚马逊和 Alphabet 等巨头在 AI 基础设施上的庞大资本支出正直接转化为强劲的营收与利润增长，云业务加速扩张。 (20:28)

Following a steep July drawdown at Situational Awareness LP, Leopold Aschenbrenner offloaded his public portfolio to Citadel to eliminate leverage while preserving core private stakes. Meanwhile, Big Tech's latest quarterly earnings reflected soaring AI-driven cloud revenue despite selective market reactions, alongside ongoing price drops and benchmark adjustments at OpenAI. Situational Awareness Restructures After Severe July Losses (0:01) Situational Awareness LP addressed an estimated net month-to-date loss of -67% for July by executing a private block trade with Ken Griffin's Citadel. The fund, which had surged from an initial $250 million to a peak of roughly $45 billion in assets under management in under two years, closed all short positions and removed leverage from its public book. Despite the July decline, the fund remained up +80% year-to-date. In an internal letter to investors, Aschenbrenner acknowledged that the fund had risked permanent capital impairment, asserting that the remaining public equity and options positions are now fully paid for without margin risk. Leverage, Predatory Trading, and the Private Asset Strategy (3:22) Market commentary highlighted that concentrated, high-beta exposure with leverage ranging between 2x and 4x left Situational Awareness exposed when rival traders aggressively shorted names linked to the fund. While online critics drew comparisons to Archegos Capital's 2021 collapse, observers noted that Aschenbrenner's unwind involved an orderly private transaction that spared prime brokers systemic contagion. Situational Awareness initially agreed to sell a $3.5 billion private stake in Anthropic to a consortium led by Green Oaks and Sequoia, but canceled the sale on Thursday morning, opting instead to liquidate its public book to Citadel and maintain private investments in Anthropic, Thinking Machines, and FluidStack. Big Tech Earnings Highlight Surging AI Infrastructure Returns (20:28) Massive capital expenditure into AI hardware translated into widespread top- and bottom-line revenue gains across major technology firms. Microsoft posted fiscal Q4 revenue of $90 billion, driven by a 43% year-over-year jump in Azure revenue that contributed to a single-day market cap gain of $450 billion. Alphabet generated $119.8 billion with an 82% surge in Google Cloud revenue to nearly $24.8 billion, while Amazon recorded $200.6 billion in sales with AWS advancing 37% year-over-year to $42.4 billion. Conversely, Meta faced market pressure despite $60.8 billion in revenue, missing EPS targets ($6.18 against $7.22 expected) amid $31.1 billion in quarterly capex and $3.6 billion in legal and severance expenses. OpenAI Pricing Shifts and ARC-AGI Benchmark Harnessing (23:36) OpenAI cut pricing on lower-tier models, shifting the evaluation frontier from raw token cost toward total cost per completed task. In response to competitive scores from Anthropic's Claude 3.5 Opus on the ARC-AGI V3 benchmark, OpenAI evaluated its GPT-5.6 Thinking model. After identifying that initial testing harness constraints limited context retention, OpenAI found that configuring two API settings tripled the model's benchmark score while reducing required output tokens by sixfold, underscoring the critical impact of evaluation harnesses on testing complex reasoning models. While Situational Awareness avoided liquidation by de-leveraging its public holdings and holding onto private assets, frontier AI development continues to pose evaluation challenges as task complexity doubles every six months and capital spending faces heightened earnings scrutiny.

本期节目聚焦对冲基金 Situational Awareness 在遭遇7月巨额回撤后平仓去杠杆的过程，并审视了大型科技巨头在AI资本开支驱动下的最新财报业绩以及前沿模型的基准测试进展。 Situational Awareness 巨额回撤与重组 (0:01) 对冲基金 Situational Awareness LP 在7月遭遇估算达 -67% 的月度净回撤，接近永久性资本减损的边缘。然而，由于前期业绩翻倍暴涨，该基金年内净收益率（YTD）仍维持在 +80%。该基金在不到两年前以2.5亿美元起步，资产管理规模（AUM）一度攀升至近450亿美元的峰值。创始人 Leopold Aschenbrenner 在泄露的致投资人信中承认，核心多头头寸的剧烈波动以及多空价差的猛烈反转沉重打击了业绩，但他坚称AI底层技术面仍在加速推进。投资人 Shelto 甚至预测，在吸取极其昂贵的风险教训后，该基金将在本年代末超越 Citadel。 为了遏制危机并消除融资风险，基金将公开市场股票头寸通过大宗交易一次性出售给 Ken Griffin 旗下的 Citadel，以此彻底清除了所有借贷杠杆，并平仓了所有空头头寸。重组后的基金继续维持公募与私募混合架构，公募部分转为无保证金风险的全额持股与期权，同时完整保留了其高确信度的非公开股权资产，基金并未清盘或转换为纯一级市场基金。 华尔街博弈机制与所谓救助争议 (3:22) 当 Situational Awareness 的持仓开始下跌时，市场竞争者迅速觉察并主动做空其公开关联标的，引发了类似银行挤兑的恶性循环。科技投资人 Roy Driscoll 分析指出，虽然宏观层面的巨额资本开支逻辑依然成立，但 Aschenbrenner 在约12只高贝塔集中持仓上施加约4倍杠杆，导致投资组合毫无容错空间。网络舆论一度将其与爆仓100亿美元的 Archegos（Bill Hwang）相提并论，但主持人指出二者存在本质区别：Hwang 拥有超过5倍杠杆并波及主经纪商，而 Aschenbrenner 的清算是在场外与 Citadel 达成的有序竞价交易，并未引发任何系统性传染。 针对社交媒体上关于该基金是否获得不公平“救助”的争论，讨论明确反驳了这一观点。主持人认同 Joe Weisenthal 的分析，即此次交易是三方竞价后由愿意接盘的对手方完成的市场行为，而非政府或系统性救助。此外，该基金原计划向 Green Oaks 和红杉资本主导的财团出售价值35亿美元的 Anthropic 私募股份，但在周四上午撤回了交易，选择保留 Anthropic 股权，仅将二级市场资产打包出售给 Citadel。 科技巨头资本开支转化为强劲营收 (20:28) 在企业基本面方面，微软、苹果、亚马逊、Meta 与 Alphabet 最新披露的财报显示，AI基础设施的巨额资本开支正直接转化为显著的收入增长，尽管市场对不同公司的分化反应更加严苛。微软第四财季营收达900亿美元（同比增长18%），每股收益4.74美元，受 Azure 业务43%同比增速的驱动，其市值单日暴涨4500亿美元，创下美股单日市值增长最高纪录。Alphabet 季度营收录得1198亿美元，每股收益9.11美元，其中谷歌云业务同比激增82%至接近248亿美元。亚马逊营收增长20%至2006亿美元，AWS 同比增长37%达到424亿美元，推动股价大涨约15%。 相比之下，另外两家巨头的市场反应较为复杂。苹果季度营收达1094亿美元，每股收益2.02美元，估值虽曾短暂突破5万亿美元，但随后股价单日重挫近9.5%。Meta 营收实现608亿美元（同比增长28%），但每股收益6.18美元低于市场预期的7.22美元，加上其季度资本开支高达311亿美元以及计入36亿美元的法律与遣散费用，导致投资者情绪受挫并令股价承压。 模型推理基准与调用框架优化 (23:36) 前沿人工智能实验室正在成本曲线与高级推理基准上展开新一轮角逐。OpenAI 显著调降了包括 Luna Terastack 在内的轻量级模型价格，推动了成本效益帕累托前沿的位移。评估重点正从单一 Token 的名义费用转向解决特定任务的总成本，因为如果模型缺乏效率、需要消耗十倍的上下文窗口来解题，最终调用成本反而更高。 在 ARC-AGI V3 基准测试中，Anthropic 的 Claude 3.5 Opus 展现出优异成绩，促使 OpenAI 针对性调试其专注于复杂数学与推理的 GPT-5.6 Thinking 模型。针对 GPT-5.6 初期表现不佳的情况，OpenAI 查明原因出在评估框架（evaluation harness）限制了上下文记忆；在开启两项特定 API 设置后，该模型的基准跑分提升了两倍，同时输出 Token 需求减少了6倍。这表明正确的测试环境配置对于充分释放复杂推理模型的下游能力至关重要。 本期内容反映出前沿AI领域的双重现实：一方面，宏观基建投资与云服务需求正强力支撑大型科技公司的业绩爆发；另一方面，二级市场围绕AI标的高杠杆集中押注面临残酷流动性挑战，而前沿模型的实际应用效能亦高度依赖精细的框架配置与综合调用成本把控。

## Background 本期内容由科技播客频道 TBPN 于 2026 年 8 月 3 日发布在 YouTube 平台。作为精炼版的 Diet TBPN 节目，本期由 John Coogan 与特邀联合主持人 Tyler 共同搭档，重点聚焦当天硅谷与华尔街瞩目的焦点事件，涵盖利奥波德·阿申布伦纳（Leopold Aschenbrenner）的对冲基金危机、科技巨头最新财报分化以及 OpenAI 的降价与模型评测进展。 ## TL;DR 节目核心拆解了 Situational Awareness 基金创始人利奥波德在 7 月遭遇 67% 巨幅回撤后发给 LP 的坦白信，详细复盘其遭遇做空围剿、通过向 Citadel 打包转让公开股票资产实现去杠杆自救，并最终叫停出售 35 亿美元 Anthropic 核心股权的全过程。主持团队横向对比了 Archegos 比尔·黄的爆仓灾难，并探讨了业内关于这场交易究竟属于市场化资产甩卖还是机构救助的激烈争辩。最后，节目深度梳理了微软、苹果、亚马逊等大厂受 AI 资本开支影响而剧烈分化的财报表现，并解析了测试框架（Harness）对 OpenAI 5.6 Soul 等前沿模型在 ARC-AGI-3 推理基准成绩上的决定性影响。 ## Atmosphere & style 全片节奏极快且充满硅谷一线交易台的敏锐与戏谑，两位主持人配合默契，在严肃的金融工程拆解与辛辣的八卦调侃间自如穿梭。从将 Citadel 救场调侃为《沙丘》史诗名场面，到严肃推演 4 倍杠杆在流动性枯竭时的灭顶之灾，整场对话兼具硬核逻辑与戏剧张力。听众能直观感受到明星基金经理在生死边缘断臂求生后的如释重负，以及当前二级市场在审视巨额 AI 算力账单时既亢奋又极其神经质的复杂情绪。 ## Bulletin - [00:01](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1s) 利奥波德·阿申布伦纳向 LP 发信澄清危机：Situational Awareness 基金 7 月遭遇 -67% 巨幅回撤，但通过将公开股票资产打包出售给 Citadel 实现去杠杆脱困，年初至今净回报仍达 +80%，并成功保全 Anthropic 等一级市场核心持仓。 - [20:27](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1227s) 大科技巨头最新财报呈现分化，资本开支与云计算增速成核心博弈点：微软 Azure 增速逆势提速至 43% 推动股价暴涨 16%，市值单日飙升 4,500 亿美元；而苹果与 Meta 虽业绩亮眼但因巨额 CapEx 导致股价承压。 - [23:36](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1416s) OpenAI 大幅下调模型调用费率，其前沿数学模型 5.6 Soul 在 ARC-AGI-3 跑分异常被证实为评测 Harness 缺陷；修正两项 API 设置后，模型在 token 消耗减少 6 倍的同时得分暴增三倍。 ## Key takeaways - [00:01](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1s) 利奥波德·阿申布伦纳通过 LP 致信直面危机，承认 7 月遭遇极端行情与华尔街做空绞杀，单月预估净跌幅为 -67%，但 2026 年初至今净回报仍维持在 +80%。 - [02:46](https://www.youtube.com/watch?v=xa5MncRRgoU&t=166s) Situational Awareness 基金此前管理资产规模峰值触及 450 亿美元，因高 Beta 集中度持仓（仅 12 只股票）加杠杆，在 7 月多空收益差逆转与流动性枯竭中险些陷入穿仓。 - [04:16](https://www.youtube.com/watch?v=xa5MncRRgoU&t=256s) 基金在周三深夜果断自救，通过私下场外大宗交易将公开股票投资组合转让给 Citadel，彻底平仓空头并出清借贷杠杆，转为 100% 自有资金买断现货与期权模式。 - [11:28](https://www.youtube.com/watch?v=xa5MncRRgoU&t=688s) 本次危机与 Archegos 存在本质区别：比尔·黄通过高杠杆引发投行百亿美元亏损及瑞信倒闭，而 Situational Awareness 资产竞拍价高于清算线且主经纪商未受实质损害，不属于政府救助。 - [19:07](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1147s) Situational Awareness 在周四清晨果断撤回了原计划向 Green Oaks 和红杉以 35 亿美元出售 Anthropic 股权的协议，通过出售二级市场资产成功保全了核心一级资产。 - [13:53](https://www.youtube.com/watch?v=xa5MncRRgoU&t=833s) 超大规模云服务商（Hyperscalers）的资本支出并未放缓，利奥波德对算力与基础设施发展的产业判断依然成立，本次爆仓的核心原因在于投资组合风险管理与 4 倍杠杆构建缺陷。 - [20:27](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1227s) 二级市场对大科技财报的评估焦点从单纯营收利润转向资本开支（CapEx）与商业化转化，微软与亚马逊凭借云业务高增长股价大涨，苹果与 Meta 则遭遇剧烈回调。 - [23:36](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1416s) 模型评测中单 token 价格（Cost per token）已非唯一指标，实际解决任务的效率和单任务成本（Cost per task）更能决定工业级落地的最终账单。 - [23:36](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1416s) 测试框架（Harness）对释放大模型核心推理能力至关重要，缺乏记忆和特定上下文反馈机制会导致前沿模型下游评测得分大幅失真。 - [26:35](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1595s) 相比抽象的“完成人类 12 小时工作”等量化指标，ARC 谜题通过人类直观易懂而机器攻克艰难的反差，更易让大众直观感受到 AGI 的演进飞跃。 ## Detailed walkthrough ### [00:01](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1s) 联合主持人回归与利奥波德·阿申布伦纳的 LP 致信 节目一开始，特邀联合主持人 Tyler 坐在了热点席位上（接替暂时缺席的搭档）。主持人 John 随即抛出了当天的重磅话题：不仅 Tyler 回归了，知名投资人利奥波德·阿申布伦纳（Leopold Aschenbrenner）也宣告强势回归。利奥波德用行动表明：“除非你们把我从 Situational Awareness LP 硬拖出去，否则我绝不离场。”尽管他在这波市场暴跌中遭受重击，但他“虽跌倒却未出局”。TBPN 团队挖掘出了一封利奥波德发给 LP（有限合伙人）的正式信函，在圈内掀起轩然大波，彻底澄清了前一天笼罩在市场上的种种疑云与流言。 此前市场与内部团队都在激烈辩论：利奥波德将旗下公开股票投资组合打包卖给肯·格里芬（Ken Griffin）的城堡投资（Citadel），这究竟算不算被强平爆仓？基金是不是彻底完蛋了？网上甚至有人恶意嘲讽他是不是要去麦当劳打工了。主持人反驳了这种荒谬论调，直言利奥波德职业生涯还很漫长，许多人都在为他加油。利奥波德在信中开门见山，极为坦诚且直击要害：“这个月我们让大家失望了。我们距离无法挽回的永久性资本减损，比我们所能接受的底线还要近。尽管我们最终找到了保护基金和各位投资人的解决方案——即将公开股票资产转让给 Citadel 并通过特定结构获取流动性——但我们设立基金的初衷，是永远都不该让自己陷入这般境地。” 利奥波德在信中强调，“波动性是获取长期投资回报的代价”。过去两年里，Situational Awareness LP 创造了不可思议的业绩，早期回报曾高达 1000% 左右。其管理资产规模（AUM）从不到两年前创立时的 2.5 亿美元起步，一路暴增到数十亿、百亿，甚至峰值时达到约 450 亿美元。当时年仅二十多岁、初出茅庐的第一任对冲基金经理拿到 2.5 亿美元本就令人咋舌，而他却在极短时间内把规模做到了半个千亿美元级别。 - **基金规模轨迹**：从最初不到两年前募集的 2.5 亿美元，迅速膨胀至数亿美元、百亿美元，峰值触及 450 亿美元 AUM。 - **历史业绩**：在过去两年中曾实现近 1000% 的惊人涨幅。 - **核心危机应对**：通过将公开市场的股票头寸打包交易给 Citadel 换取流动性，化解了迫在眉睫的穿仓风险。 --- ### [02:46](https://www.youtube.com/watch?v=xa5MncRRgoU&t=166s) 危机复盘：遭遇血腥做空与对冲基金的“银行挤兑” 主持人和 Tyler 盛赞利奥波德信函的写作风格，认为其语言极为清晰、直白，毫不推诿回避，颇有保罗·格雷厄姆（Paul Graham）推崇的那种朴实有力的文风。利奥波德在信中立誓，必将吸取惨痛教训，并详细拆解了基金遭遇的困局：首先，整个投资组合在 7 月份遭遇了巨幅回撤，尤其是最后一周核心仓位的极端波动加剧了危机。众多 AI 核心标的单月腰斩甚至跌幅更甚，导致基金原本运行良好的多空收益差（Long/Short Spread）发生暴力反转。尽管这是极其罕见的极端市场月份，但他拒绝以此为借口，坚持以最高标准要求自己。 更为残酷的是第二点：随着这些极端行情的演进，市场上针对与 Situational Awareness 名字公开关联的标的，出现了越来越恶意的反向针对性交易。这印证了前一天马丁·什克雷利（Martin Shkreli）透露的内幕——华尔街闻到了“水里的血腥味”。当其他掠食者嗅到某家大基金陷入困境时，便会疯狂集中做空其重仓股，向下狠狠施压，企图将对方彻底推向深渊。这种无情的围剿机制在本质上等同于“针对对冲基金的银行挤兑”。由于做空打压和市场流动性瞬间枯竭，基金难以把仓位维持在风险参数以内。 ```mermaid flowchart TD accTitle: Situational Awareness LP Crisis Resolution accDescr: 展现 Situational Awareness 基金遭遇市场崩盘被华尔街做空狙击到最终自救重组的全过程。 subgraph Trigger ["第一阶段：市场反转与流动性枯竭"] A["7月AI概念股暴跌50%+"] --> B["多空收益差暴力反转"] B --> C["华尔街闻到血腥味<br>恶意针对性做空重仓股"] C --> D["流动性迅速干涸<br>演变为对冲基金版'银行挤兑'"] end subgraph Rescue ["第二阶段：果断断臂求生"] D --> E["周三夜间至周四清晨<br>达成大宗交易方案"] E --> F["将公开股票组合打包转让予 Citadel"] end subgraph Restructure ["第三阶段：去杠杆与重组新生"] F --> G["平掉全部做空仓位<br>彻底解除组合杠杆与融资依赖"] G --> H["保留全额付费的公开多头及期权"] H --> I["成功保全私有化投资资产<br>（如 Anthropic 核心股权）"] end ``` --- ### [04:16](https://www.youtube.com/watch?v=xa5MncRRgoU&t=256s) 断臂求生去杠杆：保留战火与第二幕的开启 面对生死关头，利奥波德在周三深夜至周四清晨采取了决定性行动：通过与 Citadel 的大宗交易出售了部分公开市场头寸，彻底清除了基金的所有杠杆，以杜绝进一步亏损。所有空头头寸全部平仓，彻底摆脱了对证券组合融资借贷的依赖。目前，基金公开簿记转为完全用自有资金买断（fully paid for）的长线现货股票与全款期权，零融资杠杆，亦无任何保证金爆仓风险。更关键的是，这一断臂求生举动成功保全了基金旗下的核心非公开一级市场资产（包括外界关注的 Anthropic 关键持仓）。 在信中，利奥波德用一整段极其铿锵有力的单句担起责任：“我对这些事件承担全部责任（I take full responsibility for these events）。”他坦然面对惨痛现实：当 AI 技术商业基本面持续加速、而 AI 概念股大幅杀跌时，深度押注该领域的基金单月暴跌在情理之中；但剧烈波动绝不应威胁到基金的生死存亡。他正式粉碎了外界的全部谣言：Situational Awareness 没有关停，没有彻底清算，也没有转型为纯一级市场基金，更不会变成其他大机构名下的一个交易台。他们将继续以一二级混合基金的形式运营，只是公开市场交易将严格遵循全款无杠杆模式，重新整顿风控团队以“留存火种，改日再战”。 这一波 AI 交易的极端波动不仅折磨了中小盘高波动的“瓶颈交易”概念股，连美股七巨头（Magnificent 7）的财报季也陷入空前震荡。例如微软（Microsoft）等万亿美元市值的超级巨头，单日振幅竟能高达 9%、10% 乃至 15%，仅因资本开支（Capex）与回报周期的一个微小预期变化就剧烈起伏。利奥波德公布了未经审计的 7 月净收益数据：单月亏损高达 **-67%**。这虽然看似触目惊心，但在经历如此巨震后，该基金 **年初至今的净收益率仍然高达 80%**，跑赢了同期几乎所有对冲基金。利奥波德全副身家基本都在该基金中，与 LP 深度绑定。主播与嘉宾一致赞叹这封信极其提振信心，堪称继他当年轰动硅谷的万字长文之后，写出的“第二份价值十亿美元的 PDF”，风投投资人 Shelto 甚至公开预言：Situational Awareness 未来终将变得比 Citadel 还要庞大。 - **风控调整**：彻底消除公开市场杠杆与证券质押融资，只持有 100% 自有资金买断的现货及期权，完全免疫追加保证金风险。 - **最新收益数据**：7 月单月预估净跌幅为 -67%，但 2026 年初至今（Year-to-Date）净回报率依然稳在 +80%。 - **创始人利益绑定**：利奥波德本人名下绝大部分净资产均投入于该基金中。 ### [08:53](https://www.youtube.com/watch?v=xa5MncRRgoU&t=533s) Citadel 与 Ken Griffin 的“救世主”姿态与交易室名场面 节目由关于利奥波德·阿申布伦纳（Leopold Aschenbrenner）的坚定支持言论切入——有人认为利奥波德对过去两年的预测比任何人都准，如今他在吸取了极其昂贵的风险教训之后，将变得无可匹敌。然而，肯·格里芬（Ken Griffin）和 Citadel 的介入却呈现出一种极为奇妙的戏剧张力。格里芬在整个交易中摆出的姿态不是趁火打劫的秃鹫，而是将自己塑造成“最后的救世主”与市场的定海神针，仿佛在向市场宣告：“只要我还在，我就绝不允许自己不是世界第一的对冲基金经理。” 主持人调侃着调出了所谓的 Citadel 交易室“泄露视频”，屏幕上赫然是电影《沙丘》（Dune）中极具压迫感与史诗感的浩大场面。两人边看边笑谈这究竟是《沙丘 1》还是《沙丘 2》的镜头，感叹画面的沉重与霸气。格里芬极力想传达的正是这种无所不能的救市英雄形象：“当你走投无路时，这就是你唯一想拨通电话的那个人。” ### [09:57](https://www.youtube.com/watch?v=xa5MncRRgoU&t=597s) 媒体库里的“绝版”利奥波德：从极简公关到突然曝光的私照 随着《华尔街日报》（The Wall Street Journal）和《纽约时报》（The New York Times）对 Situational Awareness 基金清算与 Citadel 接盘事件的大篇幅报道，公众视野中突然涌现出一批此前从未公开过的利奥波德高清照片。在此之前，互联网上几乎只有他在 Dwarkesh Patel 播客上的截图以及唯一一张早期证件照。最广为人知的一张他身着绿色西装的图片，甚至大概率是 AI 生成的假图。 主持人惊讶地发现，《纽约时报》在昨日首次正式报道利奥波德时，直接从其内部档案库中甩出了一张他站在玻璃幕墙后、神情凝重沉思的抓拍大片。这显然是早前特意拍摄却一直压着未发的特写。利奥波德在公关策略上一向极其低调克制，从不上主流媒体滥做专访或四处拍时尚肖像，然而各大主流媒体却似乎早已在其身上秘密布局。紧接着，社交网络上又流传出一张据称是他前往欧洲参加婚礼时、在异国街头扛着一面美国国旗的照片，引来主持人一番逗趣调侃：这场爆仓风波的最大教训就是“千万别出国度假”，一旦放假去趟欧洲，家里的大盘就天翻地覆了。 ### [11:28](https://www.youtube.com/watch?v=xa5MncRRgoU&t=688s) 交易员的“爆仓哲学”与 Archegos 比尔·黄的史诗级对比 知名能源交易巨头约翰·阿诺德（John Arnold）在社交平台上点评道，自己过去招聘交易员时的哲学是：最理想的历史爆仓次数是“一次”，绝非零次，因为只有真正经历过灭顶之灾的人才会敬畏风险。现场讨论起利奥波德早年在 FTX Future Fund 的履历究竟算不算爆仓，但大家一致认为那只是赠款与慈善基金，不能算作对冲基金经理的投资爆仓。 随后，主持人引述了推特用户 Rambo 对利奥波德与 Archegos 创始人比尔·黄（Bill Hwang）的爆笑对比，直言与“史上最强赌徒”比尔·黄相比，利奥波德简直温柔得像一只温顺的小绵羊： ```mermaid flowchart TD accTitle: Archegos与Situational Awareness爆仓机制对比 accDescr: 对比Bill Hwang高杠杆导致投行巨亏与Leopold通过Citadel场外平仓的风险差异 subgraph Archegos["Bill Hwang (Archegos)"] A1["个人资本 360 亿美元"] --> A2["5-6倍杠杆: 1600亿股票敞口"] A2 --> A3["2天内闪崩爆仓"] A3 --> A4["银行巨亏100亿美元 / 瑞信破产"] end subgraph Situational["Leopold (Situational Awareness)"] B1["重仓头部高Beta AI股"] --> B2["4倍杠杆组合承压"] B2 --> B3["与 Citadel 达成场外大宗平仓"] B3 --> B4["主经纪商与银行无损 / 基金存续"] end ``` - 比尔·黄用区区 2 亿美元个人本金滚到了 360 亿美元，且全部为个人资产。 - 每天开盘前，比尔·黄会在会议室里组织员工做祷告祈祷。 - 凭借 360 亿美元资本撬动了高达 1600 亿美元的股票敞口，杠杆率高达 5 到 6 倍。 - 最终在 48 小时内彻底崩盘，不仅直接重创并最终导致拥有百年声誉的顶级投资银行（瑞信等）蒙受逾 100 亿美元的直接损失，还引发了全球系统性震荡。 - 反观利奥波德的危机，最终只是与 Citadel 达成了一笔私下的场外大宗交易平仓，并未在主经纪商或大型银行间掀起流动性危机与系统性传染，基金本身依然健全存续。 ### [13:53](https://www.youtube.com/watch?v=xa5MncRRgoU&t=833s) 宏观趋势与投资组合构建：4 倍杠杆的代价与未受损的一级资产 分析师罗伊·德里斯科尔（Roy Driscoll）指出，从这场风波中根本无法推导出“AI 交易逻辑已死”的结论。利奥波德在 2024 年关于算力和大模型发展轨迹的判断完全正确；即便从刚刚公布的亚马逊最新财报来看，超大规模云服务商（Hyperscalers）的资本支出（Capex）依旧在疯狂扩张，并未放缓。然而，认清宏观趋势只完成了交易的一半，另一半关键在于投资组合的工程构建——用 4 倍杠杆去赌高 Beta 股票在风险管理上无疑是致命失误。 主持人进一步指出，Situational Awareness 此前的 13F 持仓报告显示其组合仅重仓了 12 只股票，集中度极高，几乎没有分散风险的空间。尽管利奥波德在写给投资人的信中表态已全面撤掉杠杆，未来几个季度是否会将持仓拓展到 100 只标的或大幅增加期权对冲，将是观察其是否真正吸取教训的关键。此外，他的一级市场持仓完好无损，依然握有 Anthropic、MatX 以及 Fluidstack 等顶尖非上市独角兽的股权，手握这些优质王牌，他未来随时有能力再募集 20 到 30 亿美元。 - **多头逻辑稳固**：巨头 Capex 持续无虞，利奥波德对 AI 算力与基础设施发展的产业研判依然成立。 - **组合缺陷致命**：缺乏对冲手段，12 只股票的高 Beta 集中度放大了 4 倍杠杆的脆弱性。 - **底牌依然雄厚**：一级市场核心资产完全避开了二级市场的强制去杠杆冲击。 ### [15:15](https://www.youtube.com/watch?v=xa5MncRRgoU&t=915s) “每个操盘手都要被追加保证金一次”？ 围绕利奥波德的遭遇，市场上充斥着典型的“高罂粟综合征”（Tall Poppy Syndrome，即人们乐见少年得志的金融神童跌落神坛）。但包括 Founders Fund 的 Delian Asparouhov 在内的诸多业内人士却表达了宽容，认为华尔街历史上所有青史留名的神级交易员早年都曾遭遇过重大爆仓，这是身为“球场上真正活跃玩家”必须交纳的入场学费。 不过主持人对这种普遍论调提出了反驳：沃伦·巴菲特（Warren Buffett）一生从未真正爆仓，肯·格里芬即便在 2008 年次贷危机期间遭受重创，也从未将 Citadel 彻底干碎。大家开玩笑称，如果真有这套“爆仓即勋章”的逻辑，那最高明的投机捷径莫过于先募集一个仅有 1 万美元的超微型对冲基金，加上极限杠杆故意亏掉 500 美元宣告爆仓，然后就能向投资人自豪地宣称自己已经是“渡劫重生（Post-fall）”的成熟操盘手，准备接盘数亿美元的正规基金。 节目的最后，焦点转向了彭博知名播客《Odd Lots》的主持人乔·维森塔尔（Joe Weisenthal）与特蕾西·阿洛韦（Tracy Alloway）之间的公开讨论。维森塔尔在社交平台上深度梳理了主经纪商的清算机制与风险传导，而阿洛韦则提出了一个直击灵魂的问题：如果这只是一次正常的亏损，为什么利奥波德必须被 Citadel “救助（bailed out）”？这究竟是一场有针对性的救助承接，还是一次不惜代价的强制平仓？双方各执一词，再次将事件的定性推向讨论高潮。 ### [17:39](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1059s) 爆仓还是救助？Situational Awareness 基金的风波与对决 主持人围绕对冲基金 Situational Awareness 的动向展开热烈讨论：这究竟是一场“爆仓”（blow up），还是一场“救助”（bailout）？如果是政府出资救助，那性质完全不同，但事实显然不是这样。彭博社 Odd Lots 播客主播 Tracy Alloway 和 Joe Weisenthal 在社交媒体上就此展开了激烈争论。Tracy 质问：“为什么不能直接让投机者破产认栽？这难道不算救助吗？为什么不能像 Cathie Wood 那样，忍受糟糕的一天然后继续熬下去？”然而现实情况是，该基金面临海量赎回，局势正在恶化失控。Joe 则反驳，基金只是遭遇了追加保证金通知（Margin Call），必须向券商追加保证金，而他们偿付的方式就是将持仓出售给其他交易对手，“这怎么能叫救助？这只是正常的资产出售而已”。 现在市场上似乎只要有人接盘出售资产，大家就盲目扣上“救助”的帽子。但事实上，根本没有任何政府介入，该基金也谈不上“大到不能倒”（Too Big to Fail）。真实的情况是：基金遭遇严重亏损后发起了一场拍卖，共有三家竞标方参与角逐，而且最终的中标价格高于清算线，因此基金并未被直接强制清盘，得以继续运作。Joe 在拉扯数个回合后无奈表示：“我想我们俩对这个词的定义可能完全不同。” ```mermaid flowchart TD accTitle: Situational Awareness 资产处置流程 accDescr: 展示基金面临追保后通过竞拍出售资产并保全 Anthropic 股权的过程 subgraph Crisis ["危机爆发"] A["巨额亏损与赎回风波"] --> B["面临 Margin Call 追保"] end subgraph Resolution ["资产竞拍与决策"] B --> C["发起资产拍卖<br/>(3 家机构竞标)"] C --> D["中标价高于清算线<br/>避免被强制清盘"] D --> E["公开股票组合出售给 Citadel"] D -.->|原计划出售但最终叫停| F["35 亿美元 Anthropic 股权"] end subgraph Outcome ["最终结果"] E --> G["大幅出清风险<br/>保住核心仓位重返战场"] F --> G end ``` - **赎回潮与追保风波**：基金并未触发政府救助或大到不能倒机制，而是遭遇常规追加保证金与流动性赎回压力。 - **三方竞标拍卖**：竞标出价成功高出破产清算线，使基金避免了彻底清盘的命运。 --- ### [19:07](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1147s) 35 亿美元 Anthropic 股权出售叫停与好莱坞剧本猜想 《华尔街日报》记者 Berber Jin 披露了一则引人注目的独家猛料：Situational Awareness 此前曾试图将其持有的价值 35 亿美元的 Anthropic 股权出售给由 Green Oaks 和红杉资本（Sequoia）领投的财团。由于市场对 Anthropic 股权需求旺盛，双方在周三深夜一度达成了协议。然而到了周四早晨，Situational Awareness 却突然反悔，撤回了交易。 他们之所以敢拒绝这份 35 亿美元的交易，是因为他们转而将自己的公开股票投资组合（Public Equity Book）打包卖给了 Citadel。这一举措大幅出清了风险敞口，让他们有底气保留这笔私有股权并对团队说：“嘿，我们保留这部分仓位，我们对它极为看好、极其乐观。”主持人们调侃，这出大戏未来是否会被写成畅销书或搬上大荧幕？毕竟连戏剧性更强的 FTX 崩盘至今都还没拍出真正的院线电影——迈克尔·刘易斯（Michael Lewis）虽然写了《Going Infinite》，但大部分内容是在暴雷前采写的，缺乏细致的崩盘日记。不过这次 Situational Awareness 恐怕很难拍成大片，因为故事没有以剧烈的大爆炸收场，而是以一句酷劲十足的“行了，我们重返战场”作为积极结局。 - **35 亿美元 Anthropic 股权交易逆转**：周三深夜与 Green Oaks 及红杉谈妥，周四早晨单方面叫停拒绝。 - **转向 Citadel 卖出二级股票**：通过向 Citadel 出售公开股票资产化解保证金危机，死守核心资产。 --- ### [20:27](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1227s) 巨头财报与 AI 资本开支：冰火两重天的市场分化 科技巨头（Big Tech）在 AI 上的大规模资本支出仍在催生炸裂的财务业绩，但二级市场投资者的态度已变得极其挑剔。在过去两周内，微软（Microsoft）、苹果（Apple）、亚马逊（Amazon）、Meta 以及 Alphabet 均发布了季度财报，绝大多数在营收和利润上都超出了华尔街预期。然而，仅看“是否超营收/每股收益（Topline/Bottomline Beat）”已经变得索然无味，决定股价单日暴涨 10% 还是暴跌 10% 的核心因素，变成了各家的资本开支（CapEx）指引以及关于“AI 扩散与商业化落地”的沟通口径。 各巨头的财报数据与市场反应呈现出剧烈分化： | 公司 | 季度核心业绩 | 关键亮点 / 焦点数据 | 市场与股价反馈 | |:--- |:--- |:--- |:--- | | **微软 (Microsoft)** | 财年 Q4 营收 900 亿美元（同比增 18%，超预期的 874 亿）；EPS 为 4.74 美元（超预期的 4.21 美元） | Azure 收入同比增速逆势加速至 43% | 单日暴涨 16%，市值单日飙升 4,500 亿美元，创下美股历史最大单日市值增幅；单月累计涨幅达 25% | | **苹果 (Apple)** | 季度营收 1,094 亿美元，EPS 为 2.02 美元，双双超出预期 | 市值一度短暂突破 5 万亿美元大关 | 随后遭遇市场无情抛售，当日重挫 9.47%（近 10%） | | **亚马逊 (Amazon)** | 季度总营收达 2,006 亿美元（同比增 20%） | AWS 云计算业务暴增 37%，达 424 亿美元 | 盘后及次日大涨约 13.76% 至 15% | | **Meta** | 季度营收 608.8 亿美元（同比增 28%，好于预期） | EPS 仅为 6.18 美元（低于分析师预期的 7.22 美元）；季度资本开支高达 311 亿美元，并计入 36 亿美元一次性法律与遣散费 | 投资者对膨胀的 CapEx 极为敏感，引发股价重挫走低 | | **Alphabet** | 季度营收 1,198.8 亿美元，EPS 达 9.11 美元，轻松击败预期 | 谷歌云（Google Cloud）营收同比飙升 82% 至近 248.8 亿美元 | 市场焦点仍死死盯在不断升级的 AI 算力基建成本上 | 关于 Meta 在 AI 上的开销是否真正具备产出效益、到底有多少资金花在 token 推理上、又有多少耗费在抢夺人才的员工薪酬上，主持人们表示后续将深入探讨。 --- ### [23:36](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1416s) OpenAI 降价潮与 ARC-AGI-3 跑分之谜：测试 Harness 的致命影响 在模型前沿能力与效率的赛道上，OpenAI 大幅下调了 Luna / Luna Terasol 的调用成本。从帕累托前沿曲线（Pareto Frontier）来看，其性价比甚至远超许多开源模型。主持人们指出，如今评估模型不能只看“单 token 价格（Cost per token）”，更关键的是“单任务成本（Cost per task）”与 token 生成效率——如果一个模型价格便宜一半，但解决同一个问题需要吐出 10 倍的 token，那么最终账单反而会更加昂贵。 随后讨论转向顶尖大模型在 ARC-AGI-3 基准测试中的激烈博弈。此前 Anthropic 的 Claude Opus 5 跑出了极其惊人的分数，而 OpenAI 推出的专为解决未解数学难题而生的 5.6 Soul 却在 ARC-AGI-3 上意外受挫。主持人提及同伴曾打入过 ARC-AGI-3 全球排名前十，甚至差点拿到“职业级”的答题资格。针对 5.6 Soul 跑分不佳的异常，OpenAI 展开了深入调查并发现了关键根源：评测所使用的测试框架（Harness）存在缺陷，未能允许模型调用并记住先前的解题经验。在开启了两项特定的 API 设置后，模型在输出 token 减少 6 倍（6x fewer tokens）的情况下，测试得分直接翻了整整三倍。 这一发现凸显出近一年半以来 AI 评测中的重大趋势：测试框架（Harness）对大模型能力的释放具有决定性作用。如果选用了错误的 Harness 或限制了模型的发挥机制，下游任务的表现就会大打折扣。人们过去总幻想着“上帝模型（God Model）”只需单一架构去预测下一个 token 就能完美搞定一切，但实际上的工程集成复杂得多。ARC-AGI-3 是一个向没有软件工程经验的普通人直观展示 AI 进步的优秀标杆；然而随着大模型能够解决的任务时间跨度每六个月翻一番，高难度的测试任务已经所剩无几，业界在高效量化最前沿 AI 能力上面临着严峻的挑战。 ### [26:35](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1595s) 从 12 小时工作认知到 ARC 谜题：如何真正让人“感受到 AGI” 对于许多普通人来说，要直观理解 AI 的能力演进其实并不容易。当听到评测中提到 AI 能够完成一个“原本需要人类花费 12 个小时的任务”时，大家的反应往往是一头雾水：“需要我花 12 个小时的任务到底长什么样？我脑子里根本没有概念。”毕竟每个人的工作形态大相径庭，这究竟是指从头撰写一份详尽的长篇深度研报，还是指连续开上 12 个小时的会议？这种抽象的工时定义很难让人脱口而出并形成切身感受。 相比之下，向人们展示 ARC（Abstraction and Reasoning Corpus，抽象与推理基准）谜题则直观得多。无论是 ARC v1 还是 ARC v2 的题目，人类一眼看过去往往会觉得“非常简单”；但只要你带他们回顾 AI 在这些题目上的攻坚演进历程，解释机器要找出那些隐藏极深的规律有多么困难，就能轻而易举地让对方真正“感受到 AGI”（feel the AGI）。主持团队感慨这种体验十分有趣，并在节目的尾声向观众道别，预告下周一再会。 - **工时衡量标准的模糊性**：普通人很难将“12 小时的单项任务”映射到具体的工作场景中，究竟是连续 12 小时的会议还是复杂的项目产出，缺乏统一具象的认知框架。 - **ARC 谜题的启蒙价值**：利用 ARC v1 与 ARC v2 这种人类极易理解、而 AI 却需攻克深层抽象规律的测试集，是向外界生动普及 AI 进展与展示 AGI 临近的最有效方式。 ## Notable quotes > [00:01](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1s) 除非你们把我从 Situational Awareness LP 硬拖出去，否则我绝不离场。 — 利奥波德·阿申布伦纳 > [00:01](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1s) 这个月我们让大家失望了。我们距离无法挽回的永久性资本减损，比我们所能接受的底线还要近。尽管我们最终找到了保护基金和各位投资人的解决方案——即将公开股票资产转让给 Citadel 并通过特定结构获取流动性——但我们设立基金的初衷，是永远都不该让自己陷入这般境地。 — 利奥波德·阿申布伦纳 > [00:01](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1s) 波动性是获取长期投资回报的代价 — 利奥波德·阿申布伦纳 > [04:16](https://www.youtube.com/watch?v=xa5MncRRgoU&t=256s) 我对这些事件承担全部责任（I take full responsibility for these events）。 — 利奥波德·阿申布伦纳 > [08:53](https://www.youtube.com/watch?v=xa5MncRRgoU&t=533s) 只要我还在，我就绝不允许自己不是世界第一的对冲基金经理。 — 肯·格里芬 > [08:53](https://www.youtube.com/watch?v=xa5MncRRgoU&t=533s) 当你走投无路时，这就是你唯一想拨通电话的那个人。 — John Coogan > [09:57](https://www.youtube.com/watch?v=xa5MncRRgoU&t=597s) 千万别出国度假 — John Coogan > [17:39](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1059s) 为什么不能直接让投机者破产认栽？这难道不算救助吗？为什么不能像 Cathie Wood 那样，忍受糟糕的一天然后继续熬下去？ — Tracy Alloway > [17:39](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1059s) 我想我们俩对这个词的定义可能完全不同。 — Joe Weisenthal > [19:07](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1147s) 嘿，我们保留这部分仓位，我们对它极为看好、极其乐观。 — 利奥波德·阿申布伦纳 > [19:07](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1147s) 行了，我们重返战场 — John Coogan > [26:35](https://www.youtube.com/watch?v=xa5MncRRgoU&t=1595s) 需要我花 12 个小时的任务到底长什么样？我脑子里根本没有概念。 — John Coogan ## Mentioned TBPN, John Coogan, Jordi Hays, Tyler, Leopold Aschenbrenner, Situational Awareness LP, Ken Griffin, Citadel, Martin Shkreli, Paul Graham, Anthropic, Shelto, The Wall Street Journal, The New York Times, Dwarkesh Patel, John Arnold, Bill Hwang, Archegos Capital Management, Credit Suisse, Roy Driscoll, MatX, Fluidstack, Founders Fund, Delian Asparouhov, Warren Buffett, Joe Weisenthal, Tracy Alloway, Odd Lots, Cathie Wood, Berber Jin, Green Oaks, Sequoia Capital, Michael Lewis, Going Infinite, Microsoft, Apple, Amazon, Meta, Alphabet, OpenAI, Claude Opus 5, Luna, Luna Terasol, 5.6 Soul, ARC-AGI-3, ARC v1, ARC v2 ## Video description Sign up for TBPN’s daily newsletter at TBPN.com Diet TBPN delivers the best of today’s TBPN episode in 30 minutes. TBPN is a live tech talk show hosted by John Coogan and Jordi Hays, streaming weekdays 11–2 PT on X and YouTube, with each episode posted to podcast platforms right after. Described by The New York Times as “Silicon Valley’s newest obsession,” the show has recently featured Mark Zuckerberg, Sam Altman, Mark Cuban, and Satya Nadella. TBPN is made possible by: Ramp - https://ramp.com Public - https://public.com Cisco - https://www.cisco.com Console - https://www.console.com CrowdStrike - https://www.crowdstrike.com Figma - https://www.figma.com MongoDB - https://www.mongodb.com NYSE - https://www.nyse.com Railway - https://railway.com Shopify - https://www.shopify.com/ Follow TBPN: https://TBPN.com https://x.com/tbpn https://open.spotify.com/show/2L6WMqY3GUPCGBD0dX6p00?si=674252d53acf4231 https://podcasts.apple.com/us/podcast/technology-brothers/id1772360235 https://www.youtube.com/@TBPNLive [Source](https://www.youtube.com/watch?v=xa5MncRRgoU)

[0:01] 利奥波德·阿申布伦纳: 除非你们把我从 Situational Awareness LP 硬拖出去，否则我绝不离场。 [0:01] 利奥波德·阿申布伦纳: 这个月我们让大家失望了。我们距离无法挽回的永久性资本减损，比我们所能接受的底线还要近。尽管我们最终找到了保护基金和各位投资人的解决方案——即将公开股票资产转让给 Citadel 并通过特定结构获取流动性——但我们设立基金的初衷，是永远都不该让自己陷入这般境地。 [0:01] 利奥波德·阿申布伦纳: 波动性是获取长期投资回报的代价 [4:16] 利奥波德·阿申布伦纳: 我对这些事件承担全部责任（I take full responsibility for these events）。 [8:53] 肯·格里芬: 只要我还在，我就绝不允许自己不是世界第一的对冲基金经理。 [8:53] John Coogan: 当你走投无路时，这就是你唯一想拨通电话的那个人。 [9:57] John Coogan: 千万别出国度假 [17:39] Tracy Alloway: 为什么不能直接让投机者破产认栽？这难道不算救助吗？为什么不能像 Cathie Wood 那样，忍受糟糕的一天然后继续熬下去？ [17:39] Joe Weisenthal: 我想我们俩对这个词的定义可能完全不同。 [19:07] 利奥波德·阿申布伦纳: 嘿，我们保留这部分仓位，我们对它极为看好、极其乐观。 [19:07] John Coogan: 行了，我们重返战场 [26:35] John Coogan: 需要我花 12 个小时的任务到底长什么样？我脑子里根本没有概念。

[0:01] Leopold Aschenbrenner: We came closer to permanent capital impairment than is acceptable to us. [0:01] Leopold Aschenbrenner: Our fund must always be structured such that we can take a loss and fight another day. [0:01] Leopold Aschenbrenner: Volatility is the price of long-term investment returns. [0:01] Leopold Aschenbrenner: These were very expensive scars, but I am dedicated to ensuring they will be invaluable lessons... [0:01] Shelto: Prediction: Situational Awareness will be bigger than Citadel by the end of the decade.

[0:01] We have a guest co-host today. Introduce [0:04] yourself for those who don't know. [0:05] >> He's out again. I'm back. [0:06] >> Yeah, we got Tyler in the Ultradoo in [0:08] the hot seat. You're back. You know who [0:10] else is back? Leopold Ashen Brener's [0:12] back. He says, "You're going to have to [0:13] drag me out of sulp out of situational [0:16] awareness LP cuz he's down but not out. [0:20] Little beat up." But he shared a letter [0:22] that's making the rounds thanks to some [0:24] intrepid reporters on the DBN team that [0:27] posted this. uh he sent an LP letter [0:29] that clarifies a lot of the questions [0:31] yesterday. I mean even internally we [0:33] were going back and forth on like okay [0:34] he sold a bunch of the portfolio to Ken [0:37] Griffin to Citadel. Does this count as a [0:39] liquidation? Does this count as blowing [0:41] up? And these are like sort of vague [0:43] terms like what what does it mean to [0:45] blow up? There was definitely a draw [0:46] down. The fund definitely was [0:47] underperforming that month. But what [0:50] does it mean? Is the fund gone forever? [0:52] Is he going to work at McDonald's as [0:54] some people were trying to make it seem [0:55] like it was happening? Obviously that's [0:56] not going to happen. and he's going to [0:57] have a long career. Lot lots of people [0:59] are rooting for him. I'm certainly [1:00] rooting for him. There are some facts in [1:02] this letter that we should read through. [1:03] So he writes uh this is Leopold [1:06] Ashenbrunner to the LPs of uh [1:08] situational awareness LP. We let you [1:10] down this month. We came closer to [1:12] permanent capital impairment than is [1:14] acceptable to us. While we ultimately [1:16] found a solution that protected the fund [1:18] and you as investors, that was the sale [1:20] of the of the public equity book to [1:23] Citadel. there was some other structure [1:25] going on to get liquidity. Um he said we [1:28] ultimately found a solution that [1:29] protected the fund and you as investors. [1:31] Our intention in running the fund is to [1:33] never find ourselves in such a position [1:36] in the first place. Volatility is the [1:38] price of long-term investment returns. [1:40] Uh over the past two years, we have [1:42] delivered outstanding results. That's [1:44] 100% true. It was up what a thousand% at [1:46] one point or something like that. [1:48] >> Yeah, something like that. I mean it got [1:49] up to what 45 is the number? 45 billion [1:52] AUM from an original raise less than two [1:55] years ago I believe of $250 million [1:58] which seemed crazy at the time like he's [2:00] he's a young he's a young first-time [2:02] hedge fund manager he's got $250 million [2:04] that's crazy then pretty soon it was [2:05] like oh he's got a couple billion that's [2:08] crazy then it was like he's got tens of [2:09] billions then he's got uh half a centy [2:12] billion um so he says over the past two [2:16] years we have delivered outstanding [2:18] results despite occasional sharp [2:20] pullbacks probably they're not the first [2:22] time. There's been other pullbacks in [2:23] the market and those have probably been [2:25] amplified but never gotten to this level [2:28] of of actually distressing the fund in [2:30] this way. He said, "But our fund must [2:32] always be structured such that we can [2:34] take a loss and fight another day." And [2:36] that's a recurring theme in this. The [2:38] writing in this letter is really good. [2:40] Very clear, very direct. Uh not being [2:43] dodgy, very upfront. I I love the way [2:46] it's written. [2:46] >> Almost kind of like there's that PG [2:48] advice to like write very clearly. Yeah. [2:50] I I think it was very kind of in that [2:52] line. [2:52] >> There's a lot of that in here. Yeah. So [2:53] he says, "I will make it my mission to [2:55] ensure that we learn the necessary [2:57] lessons from this experience." Here's [2:59] where things stand. One, the portfolio [3:00] experienced a significant draw down over [3:02] the course of July, which was [3:03] exacerbated by extreme moves in core [3:06] positions over the past week. Many AI [3:09] names drew down by half or more uh while [3:11] our positive long short spread reversed [3:13] violently. Um while we could say much [3:16] more about how unusual the month was, we [3:19] hold ourselves to a higher standard [3:20] irrespective of market conditions. Two, [3:22] as these moves proceeded, we started to [3:24] see increasingly adverse trading in [3:27] names publicly associated with us. So [3:30] this is the rumor that uh Martin Skrey [3:33] was talking about yesterday. This idea [3:35] that there's blood in the water, [3:37] >> you can kind of sniff out if someone's [3:38] hurting and then [3:39] >> Exactly. And then short, sell those [3:41] positions, sell those names. uh go put [3:44] some pressure on those downward pressure [3:45] to actually intentionally hurt that [3:48] fund. Uh it's a knockout dragout fight [3:50] there on Wall Street clearly. Um but [3:53] that's the game you're playing that [3:54] that's that's why you get paid the big [3:55] bucks if you can pull it off. So uh [3:57] these these dynamics are essentially [3:59] similar to a bank run. Uh crazy to put [4:02] that word in there. A lot of people [4:03] would be dodging that but very very [4:04] direct. I love it. Uh vulnerability be [4:07] getting more vulnerability. We worked to [4:09] keep the portfolio within our risk [4:10] parameters, but gradually this became [4:12] more difficult as positions rapidly [4:13] moved against us and market liquidity [4:15] dried up. On Wednesday night, Thursday [4:17] morning, we took decisive action to [4:19] protect LP Capital. We traded a portion [4:22] of our public portfolio in a block [4:24] transaction to remove all leverage from [4:26] the fund and prevent further losses. All [4:28] shorts were closed and reliance on [4:29] portfolio financing removed. We [4:31] currently manage a fully paid for public [4:33] book long stock and long fully paid for [4:36] options with no margin liquidity risk. [4:38] This restored stability and allowed us [4:41] to preserve our private positions. So [4:43] this feels like down but not out for [4:45] sure. Yeah. And he says I take full [4:47] responsibility for these events. That's [4:49] just the full paragraph. He just says I [4:51] take responsibility. No equivocating. [4:52] It's great. Uh to be clear, this should [4:55] rightly have been a very painful month [4:58] in terms of the performance of our fund. [5:00] When AI stocks draw down dramatically [5:01] while AI technical business fundamentals [5:03] are improving, you should expect our [5:06] fund to be down a lot. We embrace [5:08] volatility, but it should never [5:10] jeopardize the fund. The fund was not [5:12] shut down. It was not liquidated or [5:14] transformed into a privateonly fund. [5:17] This was something that a lot of people [5:18] were speculating on was is this going to [5:21] be private only? Are they only going to [5:22] have their private book? Is it just [5:23] going to be the anthropic position [5:24] that's going to be riding? Or is it just [5:26] going to be liquidated and they're just [5:28] going to return capital LPS and just [5:29] say, "Hey, we're going to start [5:31] completely fresh. Do something [5:32] completely different." Even even like an [5:34] like an aqua hire like the, you know, [5:36] situational awareness becomes like a a [5:38] desk at another fund. None of that's [5:40] happening. He's very clear about this. [5:41] Situational awareness is not shutting [5:43] down. It's not liquidating and it's not [5:45] transforming into a private only fund. [5:47] He says, "We are continuing to operate [5:48] as a hybrid public private fund as [5:51] before. However, we will manage our [5:53] public book on a fully paid for basis [5:54] while we draw the lessons from these [5:56] developments. Most importantly, we took [5:58] the steps that were necessary to fight [6:01] another day." I love it. A rallying cry [6:04] to both the LPs and the employees. I'm [6:05] sure in the coming weeks, I will focus [6:08] on putting in motion the necessary [6:10] changes across a portfolio management, [6:11] risk team, and vigilance applied across [6:13] the board to ensure a higher level of [6:15] resilience going forward. AI may [6:16] continue to intensify market volatility [6:18] for years to come. And that is something [6:20] that is so clear outside of the [6:23] situational awareness bottleneck trade [6:25] longtail low market cap high volatility [6:28] stocks. Like I have never seen the MAG7 [6:31] trading like this where across earnings. [6:32] We're going to get into this with [6:33] recapping uh Meta, Apple, Amazon, [6:36] Microsoft. [6:37] >> Yeah. I mean I think it the the stat was [6:39] Microsoft had the biggest day ever. [6:43] Yeah. So you're seeing you're seeing [6:45] trillion dollar companies move by 10% 9% [6:49] 15%. It's insane that anything can [6:53] happen at that scale. Uh and so uh [6:56] clearly there is going to be a lot of [6:59] volatility. Uh and I think he's right to [7:01] point out that it is based on the AI [7:03] trade. There's so much uncertainty about [7:05] one little number about how the capex is [7:08] going to trade back. you know the the [7:10] investors in these in these large [7:12] companies let alone the small ones are [7:14] moving the stocks significantly and that [7:16] makes his job all the harder. Uh he says [7:18] these were very expensive scars but I am [7:20] dedicated to ensuring they will be [7:22] invaluable lessons for our organization [7:25] and for myself as we move forward. My [7:27] core promise to you is that we will not [7:29] waste the opportunity to learn from [7:31] these events on the portfolio itself. We [7:34] are very optimistic about the current [7:35] investment opportunity set. Of course, I [7:38] mean, the thesis still holds. Uh, the [7:40] underlying fundamentals are accelerating [7:41] at the very same time that prices have [7:43] declined significantly. Uh, thank you [7:45] for your patience and your partnership. [7:46] I'm fully invested alongside you. [7:48] Virtually all of my capital is in the [7:49] fund, and I intend to work relentlessly [7:51] to demonstrate that the events of this [7:52] month have made me a wiser and stronger [7:54] investor. Uh, he says he's available for [7:56] calls. But he also says that uh the as [7:59] an interant update the uh current [8:01] unodudited estimate of net monthto-date [8:03] performance this is for all of July [8:05] basically negative 67% [8:08] sounds atrocious until you realize that [8:12] net year-to- date they're still up 80% [8:14] which is like better than any investment [8:17] fund uh ever. So uh people are [8:20] definitely you know maybe down but not [8:22] out that there's going to be a second [8:24] act here which I think everyone's very [8:25] excited for. A lot of people were [8:27] praying for his downfall. It's very [8:29] unfortunate to see. I [8:30] >> mean, this I think this was really good [8:32] letter. I mean, this is like instills so [8:34] much faith. Like, yeah, he's completely [8:35] levelheaded. He's not like freaking out [8:37] selling everything. [8:38] >> You're calling it another billion dollar [8:39] PDF. [8:40] >> Yeah, this could be the second billion [8:42] dollar PDF. [8:42] >> It might be. [8:44] I mean, Shelto agrees. Shelto had a [8:47] great position. What did he say? [8:49] >> Uh, he said, uh, prediction situation [8:51] awareness will be bigger than Citadel by [8:53] the end of the decade. Leopold has [8:54] predicted the last two years better than [8:56] anyone else. Now that he can combine [8:57] that with very expensive lessons in [8:59] risk, he will be unstoppable. He is my [9:01] full confidence. [9:02] >> This is such a wild [9:05] Kane Griffin sitting there being like, [9:06] >> "You got to ride with your boy." [9:07] >> Shelto like you're going to take a shot [9:09] at me like that, bro? Really? Really? [9:11] You're going to come for me like that? [9:13] Uh because I will die before I am not [9:15] the biggest hedge fund manager in the [9:17] world. Pull up a live view from the [9:18] Citadel trading floor because we got [9:21] some leaked video. [9:28] This is not the vibe that Martin was [9:32] articulating. Ken Griffin's wanting to [9:34] be framed as like the savior, the legend [9:36] of last resort. Positive force. [9:39] >> This is the guy you want to call. [9:40] >> I guess [9:41] >> this guy right here. [9:43] >> Dune. So good. Is this from Dune 1 or [9:45] Dune 2? [9:47] >> I think one I want to say Dune one. [9:51] >> Yeah. Beautiful. [9:56] >> It's very heavy though. [9:57] >> So I I think one of the the big stories [9:58] of the last, you know, few days is we've [10:00] seen all these new Leopold photos. We've [10:02] never seen these before. Brand new rare [10:04] Leopolds from Wall Street Journal. [10:06] >> This might be the biggest story of them [10:07] all. [10:08] >> Yeah. Cuz I mean for a while the only [10:10] image of Leupold was basically there was [10:12] like one head shot and then it was just [10:13] steals from Dark Cash podcast. Yeah. [10:15] >> And now we're just seeing all these new [10:16] ones. Like where did these come from? [10:18] There was one and then there was a photo [10:19] that was done I think for the Wall [10:21] Street Journal, but then the New York [10:23] Times writes up the whole story of the [10:25] of the situational awareness uh deal [10:27] with Citadel and they just drop a banger [10:31] new photo that they just had in the [10:32] archive that they could have leaked. [10:35] Let's pull it up. It's here. Uh it's [10:38] Leopold looking very uh pensive behind a [10:41] glass wall. This one's in the Wall [10:42] Street Journal today. This one's new, [10:43] too. Everyone's been clamoring for this [10:45] because the one that goes viral is him [10:46] in that green suit. This is the one. [10:48] >> I'm pretty sure that's AI, right? [10:49] >> That's AI. But this one is not. This is [10:51] from the New York Times. They went and [10:53] shot this and then never published [10:55] anything. Like the first time Leopold [10:57] was mentioned in the New York Times was [10:59] yesterday and they used this photo. And [11:01] so you have to wonder if they were like [11:03] working on a profile. [11:05] But Leopold's been so quiet with his [11:08] public relations strategy. He's not [11:10] talking to media, doing photo shoots, [11:11] doing profiles constantly. Like he [11:13] certainly could be doing more in [11:14] Bloomberg, at Forbes and Fortune. like [11:16] he could be doing a lot, but he's had a [11:18] very narrow strategy and I think it's [11:20] worked very well for him. Um, but it's [11:22] funny that somehow all the mainstream [11:24] media just has secret Leopold photos and [11:27] dropping on the timeline. Uh, John [11:29] Arnold is chiming in. He says, "My [11:31] philosophy when I used to hire traders [11:33] was that the optimal number of past [11:35] blowups was one." He's not saying zero. [11:38] >> Yeah. [11:38] >> He says you got to learn your lesson. [11:39] >> The question is, how does the FTX future [11:41] fund count? Does that count? [11:42] >> I don't think that counts at all. [11:43] >> That it definitely doesn't count as a [11:45] full blowup. wasn't a fund manager of [11:47] it, right? It wasn't it like for [11:48] donations and then FTX was just the one [11:50] that was funding. [11:51] >> Yeah, it was like he was doing stuff. [11:53] >> That seems Yeah, that seems completely [11:54] separate. Is Is the wedding photo AI or [11:57] is this real? And is he carrying an [11:59] American flag? Let's pull up this image. [12:01] >> I saw Yeah, this one I've never seen [12:02] before until yesterday as well. [12:03] >> I I mean, like if it's the wedding, like [12:05] the wedding's happening right now, this [12:06] wouldn't exist. But I I'm wondering if [12:08] this like leaked onto the timeline from [12:10] someone who was there. Also, this photo [12:12] hit like Wednesday, and I think the [12:14] wedding would be over the weekend. Um, [12:15] but it's cool that he's just rocking [12:17] carrying an American flag international. [12:20] Uh, yeah, the real lesson here is never [12:22] travel internationally because he takes [12:24] one day off, one weekend off to go to [12:26] Europe and everything go everything [12:28] blows up. No, of course this one I was [12:30] laughing at before the show. Rambo says [12:33] comparing Leopold Ashen Brener to Bill [12:35] Wang. Wang is Wang is the goat. Wang [12:38] Bill Wang from Arc Ghost. Uh, Wang is [12:40] the goat of degenerates and Leopold is a [12:42] sheep compared to him. Did you know that [12:44] Wang turned 200 million into 36 billion [12:47] and it was all personal capital? The guy [12:50] literally led prayer circles in the [12:52] conference room before trading days [12:53] started. He had $ 160 billion of stock [12:56] exposure on just 36 billion of capital. [12:59] It's like five or 6x lever. His blow up [13:03] happened in two days and he literally [13:05] caused the collapse of one of the most [13:07] prestigious investment banks. Banks lost [13:09] a total of 10 billion combined because [13:11] of his collapse. Leopold is nothing [13:13] compared to [13:16] so [13:16] >> get her numbers up. [13:18] >> Yeah. No. Uh yeah. I mean that that [13:20] that's the interesting thing here is [13:21] that like it is this sort of like [13:22] dramatic unwind, but at the end of the [13:25] day it is just like an overthe-c counter [13:27] transaction with Citadel for a block of [13:29] trades and block of equity positions. [13:31] >> And the fund is still around. I mean [13:32] they still seem to be like probably [13:34] going to be doing very well. [13:35] >> They'll be Yeah, they'll be okay. [13:36] Citadel will be okay. [13:37] >> Be bigger than Citadel pretty soon [13:39] >> any day now. Uh and and and importantly [13:42] uh all of the all of the prime brokers [13:45] the big banks like they were not [13:46] affected. There was not uh there was not [13:48] like a liquidity crisis that yeah that [13:50] uh a contagion effect did not take root. [13:53] Uh Roy Driscoll says uh there's nothing [13:54] to learn from the situational awareness [13:56] situation about the AI trade. Leo was [13:58] right in 2024 and based on the Amazon [14:01] results, he's still right today. Hypers [14:03] scale capex continues unabated. There's [14:05] obviously something to learn about risk [14:07] management. Forex leverage with high [14:08] beta stocks is a mistake in trading [14:10] stocks. Half the battle is getting the [14:12] trend right, but the other half is [14:14] nailing the portfolio construction. [14:15] Well, [14:15] >> yeah. I mean, this is what um Martin was [14:17] saying yesterday, right? Like the [14:18] underlying completely makes sense, but [14:20] like you get into these crazy psychology [14:22] things where it's just like, yeah, who's [14:23] really [14:24] >> everyone's focused on the leverage. It [14:25] does also seem like there were like [14:26] every time the 13F would drop, it would [14:28] be like 12 names, which is like not a [14:30] lot of diversification. So, I wonder [14:32] like right now the message from the [14:34] letter is we're not using leverage right [14:35] now. we're going to be learning the [14:37] lesson. Yeah, maybe the lesson is, hey, [14:39] 2x leverage or 3x or something like that [14:41] or four in certain scenarios with [14:43] smaller trades, not portfoliow wide or [14:46] something like that. But it will be [14:47] interesting to see if there's a [14:48] difference in uh if if the the lesson [14:50] that's learned when the next 13F drops [14:52] in a couple quarters, we see, oh wow, he [14:55] has like a hundred names or or there's, [14:59] you know, he's he's using more options [15:01] or less options or, you know, whatever. [15:03] However it changes, that will be [15:04] interesting to see for sure. So Leopold [15:06] still has Anthropic, Maddx, and Fluid [15:08] Stack, tier one private companies. He [15:10] can probably raise two to three billion [15:11] more. It's not over for him by any [15:14] means, says Zephyr. There's been this [15:16] vibe of like it's too good to be true. [15:18] >> Tall Puppy Syndrome, you know, he's he's [15:20] the AI getting away with Wonder Kid. [15:23] Yeah. You know, [15:24] >> yeah, people uh people hate to see a [15:26] young a young [15:28] >> a young goat, [15:28] >> a young young hedge fund manager run it [15:30] up crazy. Lexi Goozy says, "Uh, keep [15:33] people keep making fun of Leopold on the [15:35] timeline, but everyone one everyone [15:37] needs to get margin called once in their [15:38] life. Is this true?" [15:40] >> Delian had a similar take, right? He [15:42] said, um, uh, basically all the goats on [15:44] Wall Street have had some sort of blow [15:47] up earlier in their career part of the [15:48] game of being a live player on the [15:50] field. [15:50] >> So, [15:51] >> but yeah, is that true? I [15:53] >> I don't think that's actually true. I [15:55] don't think Warren Buffett ever blew up. [15:57] I don't even think Ken Griffin ever [15:58] really blew up. I think he had a really [16:01] bad year in 2008 during the financial uh [16:04] during the housing crisis, the financial [16:06] crisis. But early on, I think he got his [16:09] start sort of post.com and was doing [16:12] convertible debt trading and and never [16:14] really like the entity has always been [16:16] Citadel. There was no precursor to that. [16:18] Um but but it's a fair take that like [16:20] clearly people can blow can can build [16:23] back up after there's a [16:24] >> Yeah. I mean there there's a lot of [16:25] comparisons to PT, right? [16:27] And then they're saying, "Oh, this is [16:28] also kind of [16:30] >> uh lose take, right? Go into VC, then [16:32] you can kind of do the long only thing." [16:34] >> The real the real hack would be to uh [16:37] just raise the smallest hedge fund ever, [16:40] $10,000, lever it, blow it up, and be [16:42] like, "Wow, I'm postfall." [16:45] >> Oh, yeah. He's postfall now. [16:46] >> He's postfall now. And so, but but if [16:48] you do it with like such a small amount [16:50] of capital, but you can still be like, [16:52] "Oh man, I learned so much. That was [16:54] really crazy. Those were crazy crazy [16:55] times. I lost $500. [16:58] Ready? Ready for the real fund now? Uh, [17:01] no. Uh, Leo still made incredible [17:03] returns. His fund will do incredibly [17:05] well in the long term. Lots of people [17:06] coming out in support. Um, one person [17:09] that's not in support. [17:12] Joe Eisenthal is uh is going back and [17:14] forth with Tracy. That's hilarious. So [17:16] Joe has been uh live tweeting this. He's [17:18] he's been making a bunch of great points [17:20] and and just illuminating the deeper [17:22] level of like what's going on with prime [17:23] brokerages and all these different [17:25] aspects of what's going on. But Joe [17:27] started by sharing the Wall Street [17:29] Journal article that said that Citadel [17:30] buys situational awareness stock [17:31] portfolio after big losses in AI and [17:33] Tracy says why does he have to get [17:35] bailed out at all? And this is another [17:37] question like is this a liquidation? Is [17:39] this a blow up? Is this a bailout? Uh it [17:43] would be it would have been a very [17:45] different conversation if this had been [17:46] like a government bailout of of [17:48] situational awareness. That's not what [17:50] happened. But [17:51] >> Tracy says, "Why can't we just let the [17:53] speculators fail?" Joe Wisenthal says, [17:55] "Who says he's getting bailed out? He [17:57] entered into a transaction with a will [17:58] and counterparty." And Tracy says, [17:59] "Isn't that a bailout? Why not just keep [18:01] managing the fund? Why not beat Kathy [18:03] Woods and have a bad day and live to [18:04] tell another tale?" Uh except there were [18:06] probably too many redemptions, so it was [18:08] spiraling. Uh Joe says he got margin [18:10] called. And Tracy says, "So it is a [18:11] bailout. Just let it fail, but maybe it [18:14] was too big and could see the [18:15] contagion." Uh Joe says, "I don't get [18:17] what you're saying. Someone got gets [18:18] margin called and they have to pay the [18:20] broker." And the way they pay the back [18:22] the broker is selling off shares to some [18:24] other counterparty. How was that a [18:25] bailout? Like he's just selling. And [18:27] people associate every sale with a [18:28] bailout now, I guess. But that's not [18:30] what this is. This was not uh the [18:33] government stepping in. This is not [18:34] >> He was not too big to fail. [18:36] >> No, not at all. [18:36] >> I mean, some people were saying that he [18:37] could have been too big to fail going [18:38] in, but [18:39] >> Yeah. I it doesn't seem like that's what [18:41] happened. It seemed like there that [18:43] there were significant losses and then [18:44] they ran an auction and there were three [18:46] three parties bidding and they and and [18:49] the bids came in above above like you [18:53] know liquidation level. So the fund is [18:54] not liquidated. Yeah. And it and it [18:56] remains. Uh and so Joe after fighting [19:00] back and forth for several posts he says [19:02] I think we might have a different [19:02] definition of the term here. And I think [19:04] you do. I think you do. Um very fun. [19:07] This is an interesting scoop from Berber [19:09] Jyn over at the Wall Street Journal [19:10] related to this. Situational awareness [19:12] tried to sell a $ three and half billion [19:14] dollar stake in Anthropic to a group of [19:16] investors led by Green Oaks and Sequoia. [19:18] Obviously, there's a lot of demand for [19:19] the stock. Uh the parties reached a deal [19:21] late Wednesday, but then situational [19:23] awareness pulled out Thursday morning. [19:24] They turned it down. They [19:26] >> they turned it down. [19:26] >> They turned it down. Um [19:28] >> they went out for like 3.5 billion, [19:30] something like that. [19:31] >> Something like that. And they turned it [19:32] down just to grind for the public equity [19:35] book. uh they sold that to to to [19:37] Citadel. That obviously cleared a lot of [19:40] the risk out and they said, "Hey, let's [19:41] keep this position. We're extremely [19:43] excited about this. We're bullish." [19:44] Yeah. And so I don't know. Uh I think [19:47] will this be the subject of a book? Will [19:49] this be the subject of an actual movie? [19:51] Is it is it drama enough? Have we gotten [19:54] the FTX movie yet? Because that's way [19:56] more dramatic and I don't think that [19:57] ever happened. Yeah. [19:59] >> Um I haven't seen it. And then there was [20:00] Infinity M. Not Infinity Machine. There [20:02] was a Going Infinite. Is that the one? [20:04] That was the Michael Lewis book. [20:05] >> Michael Lewis book, but that was written [20:06] like before the blowup and so it was [20:08] like sort of uh it didn't really tell [20:10] the story like day by day. But like an [20:13] in the room fly on the wall [20:15] minute-by-minute account of this would [20:17] be interesting, but it's not that [20:19] dramatic because it doesn't end with an [20:21] explosion. It ends with like a okay, [20:22] we're back in the fight, which is cool. [20:24] I mean, it's maybe more positive [20:27] outcome. Big tech's AI spending is [20:29] continuing to produce blockbuster [20:30] financial results even as investors have [20:34] become increasingly selective about [20:35] which companies they're willing to [20:36] reward. Over the past two weeks, [20:38] Microsoft, Apple, Amazon, Meta, Alphabet [20:41] all reported quarterly earnings that [20:42] largely exceeded Wall Street [20:44] expectations. It's very boring when you [20:46] pull the did they beat on topline, did [20:48] they beat on bottom line. It's like [20:49] everyone beats and then the stock goes [20:51] down 10% or up 10% based on capex [20:54] forecasts and also just messaging around [20:56] AI diffusion and AI uptake. Microsoft [20:58] led the group with shares surging after [21:00] reporting fiscal fourth quarter revenue [21:02] of $90 billion, up 18% year-over-year [21:05] and ahead of the 87.4 billion that [21:09] analysts were expecting. Uh that was the [21:11] consensus estimate. Uh EPS came in at [21:14] 474 versus expectations of 421. So they [21:17] beat topline, beat bottom line. Uh Azure [21:19] revenue accelerated 43% over [21:22] year-over-year. Apple [21:23] >> So they gained 450 billion in one day. [21:26] 16% [21:27] >> 50 billion in one day. [21:29] >> That's for biggest one day market cap [21:31] gain for any US company. [21:32] >> Look at that. The god. [21:35] >> That's really really impressive. It's up [21:37] 25% over the month. Uh very impressive. [21:41] Apple also beat expectations reporting [21:43] 109.4 4 billion in quarterly revenue, [21:45] earnings per share of 202. Uh stock [21:47] briefly pushed the company market cap [21:49] cap uh above the $5 trillion mark, but [21:52] it has been uh absolutely tanking today [21:56] uh down what 10% today or something like [21:59] that. Let's see down 9.47%. [22:02] Uh last uh we'll go through two more. Uh [22:05] Amazon has also impressed investors with [22:06] revenue climbing 20% to 200 billion uh [22:09] 200.6. AWS growing 37% to 42.4 billion. [22:13] uh sending shares sharply higher in [22:15] after hours trading. Here's Amazon. We [22:17] can pull that up as well. Uh the market [22:20] is up 13.76% [22:23] and the day is looking pretty good, too. [22:26] Up 15% today. The market's reaction [22:28] wasn't usually universally positive. [22:30] Meta posted stronger than expected [22:32] revenue of 60.8 8 billion, up 28% [22:35] year-over-year, but earnings per share [22:37] fell 6.18 uh $6.18 [22:41] uh fell short of the $722 [22:44] analysts had expected. Investors focused [22:46] on the company's 31.1 billion in [22:48] quarterly capex along with 3.6 billion [22:50] in onetime legal and severance costs, [22:52] sending the stock sharply lower. Let's [22:55] see what Meta is doing. Down just a bit. [22:57] Um Alphabet, meanwhile, reported revenue [22:59] of 119.8 8 billion while earnings per [23:02] share of9 uh11 comfortably beating [23:05] expectations while Google cloud revenue [23:07] surged 82% year-over-year to nearly 24.8 [23:11] 8 billion. Even so, investors remained [23:13] focused on the escalating cost of AI [23:16] infrastructure as hyperscalers continue [23:18] pouring hundreds of billions of dollars [23:19] into new compute capacity. And here's, [23:22] uh, here's Google. So, uh, we can dig [23:24] into this more. There's a whole bunch of [23:26] deeper questions about what is the [23:27] actual efficacy of Meta spending on AI, [23:30] how much are they spending on tokens, [23:31] how much are they spending on headcount, [23:33] all these things matter, but we will dig [23:34] into it another time. take us through [23:36] what's going on with uh OpenAI uh [23:39] pushing the model frontier access across [23:41] efficiency. What happened? They dropped [23:43] the cost of Luna. [23:44] >> Yeah. Yeah. So there's there's Luna [23:46] Terasol. This is the cheapest model. [23:47] Yes. Massively reduced cost. You can see [23:49] on the kind of Yeah. Pluto curve. This [23:51] is like actually much cheaper than a lot [23:53] of like open source models because you [23:55] because we've been talking about this [23:56] recently. It's like [23:57] >> there's cost per per task. Not just like [24:00] can it do it and how much do like it [24:02] depends a lot how token efficient the [24:03] model is. [24:04] >> Yeah. Because there you could measure it [24:05] on on cost per token, but if if a [24:08] certain model takes 10 times the amount [24:09] of tokens and it's only half the cost, [24:11] you wind up spending more. Yeah. [24:13] >> Why is the paro frontier in this graph [24:15] flipped? I feel like the paro frontier [24:17] used to be this direction. Am I am I [24:20] hallucinating that? [24:22] >> Has it always been this way? You always [24:23] want to be on the left side. I thought [24:24] you wanted to be on the right side or [24:26] something like that. [24:27] >> Well, it depends on where you are in the [24:28] PRIO. [24:29] >> I I I suppose [24:30] >> I guess I see I think I see what you're [24:31] saying. [24:31] >> I suppose. Uh anyway, uh we also I don't [24:34] think we touched on this, but ArcGIV3, [24:36] the leading labs been going back and [24:38] forth. Uh Opus 5 put up a very very [24:41] impressive number. Then OpenAI fired [24:44] back with 5.6 soul uh used to solve open [24:48] problems in mathematics. So why was it [24:50] struggling with Arc AGI V3 uh which you [24:53] at one point were in the top 10, right? [24:56] >> Yeah. Yeah. I I was uh I was globally [24:58] ranked Arc AGI V3. I I don't think it's [25:01] still up, but um [25:03] >> Arc AGI V3 player that's up there. You [25:05] were like you were uh pro-Aam. [25:08] >> Yeah. Yeah, I would say [25:09] >> you didn't go pro. You turned it down. [25:11] >> I turned it down. [25:11] >> You had the opportunity to be an AR. [25:13] >> They were going to give me like 10 more [25:14] tasks, five more tasks, something like [25:15] that. [25:16] >> Something like that. Uh but apparently [25:18] OpenAI was able to investigate uh the [25:21] low score of 5.6 Soul on ArcGIV3. And [25:24] the harness was not letting it remember [25:26] what it had learned. We found that [25:28] enabling two API settings tripled our [25:30] scores with 6x fewer output tokens. So [25:33] very interesting uh to watch these. [25:36] >> This is fascinating. I mean we've seen [25:37] this a lot over the past like I don't [25:39] know year and a half almost where the [25:41] harness like really matters a lot. Yeah. [25:42] >> And if you have the wrong harness or [25:44] it's like limiting the model in some [25:45] way. I mean it can have like massive [25:47] totally [25:48] >> like effects on the the downstream task. [25:49] >> Yeah. People were not expecting this. is [25:51] definitely like the the the the model [25:53] the god model will be just one model and [25:55] you'll just ask it to predict the next [25:57] token and it'll just do it perfectly. Uh [25:59] there's a lot more that goes into the [26:00] integration here. I still think ARC AGI [26:02] V3, I mean, fantastic benchmark. Love [26:04] the team. Obviously, Mike's been on the [26:05] show multiple times, but uh it's also [26:07] just a great way to actually illustrate [26:10] AI progress to someone that maybe just [26:13] doesn't want to build software or hasn't [26:15] built software before and doesn't really [26:17] can't really feel that visual [26:18] viscerally. [26:21] I can't say viscerally, I don't know. Um [26:23] >> well especially uh you know the the [26:25] famous like time horizon task doubling [26:27] every six months like that's basically [26:29] like we can't actually measure the the [26:31] high end now. It's like too hard. We we [26:32] we don't have enough task to like [26:34] measure it efficiently basically. [26:35] >> And a lot of people are just like what a [26:37] task that takes me 12 hours like what is [26:39] that? I don't even know. It's hard to [26:41] think of that off the top of your head. [26:42] Like what does that mean? Like building [26:44] a whole report or something or like a [26:46] lot of people work in like various ways. [26:47] like yeah 12 hours of meetings is that [26:50] one task I don't know but uh if you show [26:53] someone the archi v1 puzzle and it's [26:55] very easy and v2 is very p very very [26:58] easy and then you and then you walk them [26:59] through the story of how AI has [27:01] progressed on this and how hidden the [27:04] answers are you can pretty easily help [27:06] someone feel the agi which is very very [27:08] uh very very fun and we'll see you on [27:11] Monday [27:11] >> see you goodbye [27:13] >> flashback