SpaceX Secures Cursor Buyout Option in $60 Billion xAI Compute Deal

SpaceX在600亿美元xAI算力交易中获得Cursor收购期权 SpaceX获得了以600亿美元收购AI编程助手Cursor的期权，同时通过xAI的Colossus超级计算机为其提供海量算力。 本期节目报道了SpaceX与Cursor之间涉及xAI Colossus超级计算机的重大战略协议，并回顾了埃隆·马斯克在过去四年中不断扩张的科技与媒体生态体系。主持人分析了AI编程助手市场的近期格局变化、中国向一家太空数据中心初创公司注资84亿美元的举动，以及OpenAI推出的ChatGPT Images V2。最后，节目以轻松趣闻收尾，剖析了多起离奇事件，包括涉案14.1万美元的“熊爪行动”保险诈骗案、瑞典研究人员对摄入可卡因的三文鱼的研究，以及红龙虾餐厅（Red Lobster）的无限续大虾促销活动。 SpaceX 与 Cursor 达成重大战略协议，Cursor 将接入拥有百万张 H100 等效算力的 xAI Colossus 超级计算机，SpaceX 则获得以 600 亿美元收购 Cursor 的期权，并附带 100 亿美元的非稀释性反向分手费保障。 (0:00) 中国太空计算初创公司北京星际晨光（Orbital Chunguang）获得 12 家金融机构共计 84 亿美元的授信额度，计划在晨昏轨道部署太空数据中心基础设施，以规避地面的能源、水资源及监管限制。 (16:18) OpenAI 推出 ChatGPT Images V2，新增全景 360 度 VR 图像生成能力并显著降低商业产品摄影门槛，引发了关于 AI 评测基准演进及模型极端工况可靠性的讨论。 (20:43) 加州保险部门查处了名为“熊爪行动”的特大骗保案，三名嫌犯因身穿仿真熊皮道具服蓄意破坏劳斯莱斯等豪华车辆、累计骗取逾 14.1 万美元保险赔付款而被判刑入狱。 (28:59)

The program highlights major shifts across the artificial intelligence sector, headlined by SpaceX's strategic compute and acquisition agreement with Cursor, while analyzing global infrastructure bets, software consolidation, image generation benchmarks, and offbeat regulatory investigations. SpaceX–Cursor Strategic Deal and Musk's AI Ecosystem (0:00) SpaceX has agreed to provide AI developer Cursor with compute access to xAI's Colossus supercomputer—featuring 1 million H100-equivalent GPUs—in exchange for an option to acquire Cursor for $60 billion, secured by a $10 billion breakup fee. The transaction resolves major structural bottlenecks: it provides Cursor the capital-intensive compute needed to train models like Composer 2 against OpenAI and Anthropic, offers xAI an outlet for surplus compute following internal coding-model hurdles and co-founder departures, and grants the ecosystem an enterprise-friendly front under CEO Michael Truell. Analysts note that acquiring Cursor's estimated $2 billion in revenue at roughly 30x multiple is accretive for SpaceX, which is projected to trade publicly at over 100x revenue. The partnership marks the latest step in Elon Musk's multi-year ecosystem consolidation. Beginning with an unsolicited $44 billion bid for Twitter in April 2022 (closed in October 2022), Musk founded xAI in July 2023, launched Grok that November, open-sourced Grok-1 in March 2024, merged X into xAI in March 2025, and finalized SpaceX's acquisition of xAI in February 2026. Consolidation in the AI Coding Market (15:31) The AI coding landscape is experiencing sharp consolidation, leaving fewer independent foundation players. The hosts highlight Cognition founder Scott Wu—creator of Devin—as the 'last of the Mohicans' maintaining independence, while competing platforms like Windsurf and smaller ventures risk turning into uncompetitive 'zombie corps' or being absorbed by better-capitalized incumbents. The discussion also touches on the dramatic paper valuation gains accrued by early venture investors, noting humorous reports of incarcerated FTX founder Sam Bankman-Fried tracking massive secondary gains in portfolio holdings Anthropic and Cursor via prison-approved communications. China's $8.4 Billion Bet on Orbital Data Centers (16:18) According to reporting from Space News Inc., Chinese space startup Beijing Orbital Twilight Technology (Orbital Chunguang) secured an $8.4 billion credit line backed by 12 major financial institutions, including the Bank of China, following an April 20 pre-A1 funding round. Incubated by the Beijing Astro Future Institute of Space Technology alongside a 24-member consortium, the firm plans to construct space-based computing infrastructure in a dawn-dusk orbit at 700–800 km altitude by 2035. The initiative highlights divergent motivations for space-based compute: while American interest centers on bypassing protracted terrestrial permitting, land constraints, and cooling water limits, China's orbital push proceeds despite fewer terrestrial regulatory hurdles and abundant domestic hydroelectric power. ChatGPT Images V2 and the Reliability Frontier (20:43) OpenAI's ChatGPT Images V2 reflects rapid progress in visual fidelity alongside persistent edge-case vulnerabilities. While historical benchmark prompts—such as an astronaut on horseback—are now trivial, models still encounter difficulties with inverted physical logic, liquid limits, or analog clock hands. Citing observations from Sam Altman and Andrej Karpathy, the hosts discuss the '99% problem': reaching high initial fidelity occurs rapidly, but adding subsequent degrees of human-demanded reliability requires years of refinement. Simultaneously, high-grade visual models are democratizing commercial aesthetics. Early-stage hardware and consumer startups can now generate agency-level product photography instantly, eliminating early design moats. The technology has also expanded technically, producing 360-degree equirectangular panoramic environments directly deployable into virtual reality and stereoscopic workflows. Operation Bear Claw and Peripheral News (28:59) The episode reviews a bizarre Southern California insurance fraud case dubbed 'Operation Bear Claw.' Three individuals were arrested in November 2024 and sentenced to jail after claiming more than $141,000 in payouts by faking wild animal damage on luxury vehicles—including a 2010 Rolls-Royce Ghost and two Mercedes-Benzes—using a full-body bear suit and kitchen utensils. State investigators uncovered the plot after video footage revealed an unnatural human silhouette and uniform scratch marks. The show concludes with lighter news: a New York Times report on a Swedish scientific study finding that salmon exposed to cocaine swam nearly twice as fast with increased endurance, and Red Lobster's announcement that its limited-run 'Endless Shrimp' promotion will return around April 20. The programmatic review emphasizes how high-level capital structuring and massive compute scale are reshaping the enterprise AI frontier, even as foundational models struggle with long-tail reliability and real-world deployment frictions.

本期节目重点分析了SpaceX与Cursor达成的深度算力与收购期权协议及其对AI生态的重塑，并探讨了中国巨额授信的太空数据中心构想、OpenAI图像生成能力的演化与商业影响，以及荒诞的车险诈骗案与科学趣闻。 SpaceX与Cursor的算力并购机制 (0:00) SpaceX与AI编程助手Cursor达成了一项旨在打造前沿编程与知识工作AI的重大战略合作。根据协议，Cursor将直接获取xAI旗下Colossus超级计算机的算力资源，该集群拥有相当于100万块H100 GPU的计算规模。作为对价，SpaceX获得了以600亿美元收购Cursor的选择权，并设定了高达100亿美元的非稀释性分手费作为未达成收购时的兜底保障。这一架构为Cursor锁定了关键的模型预训练算力，避免其独自承担高达50亿美元的预训练资本开支；同时，这也为xAI消耗了闲置GPU算力，并弥补其在自研前沿编程模型上面临的阻碍。此外，Cursor由首席执行官Michael Truell建立的企业友好型品牌形象，为福特、通用电气和可口可乐等对Grok争议品牌存有顾虑的企业客户提供了缓冲。 分析人士指出，该交易具备显著的资本运作空间。市场预计SpaceX未来上市估值可能达到营收的100倍以上，而以约30倍市销率估值收购Cursor约20亿美元的预期营收，将实现约70倍的倍数扩张，从而大幅增厚SpaceX的股权价值。 马斯克AI垂直整合与编程助手格局 (3:55) 节目回顾了埃隆·马斯克过去四年构建AI与媒体垂直整合帝国的路径：从2022年4月出资440亿美元竞购推特并于当年10月接管，到2023年7月成立xAI挑战行业前沿实验室；随后xAI在2023年底推出Grok，并在2024年推进开源与多模态升级；2025年3月xAI以全股票交易吸收X平台，直至2026年2月SpaceX收购xAI，为本次与Cursor的深度合作铺平了道路。在此期间，身陷囹圄的FTX创始人Sam Bankman-Fried（SBF）通过美国联邦监狱管理局（BOP）合规通信渠道让外界代理发帖关注该交易，其早期对Anthropic（二级市场估值正逼近万亿美元）和Cursor的投资已形成庞大的账面浮盈。 与此同时，AI编程助手赛道正迎来残酷整合。节目引述行业观点将Cognition创始人、Devin开发者Scott Wu称为该领域独立创始团队中“最后的莫希干人”。随着Windsurf等竞品及其他小型团队面临被并购或沦为缺乏活力的“僵尸企业”，独立编程工具市场的格局正迅速向少数头部企业收缩。 中国布局84亿美元太空数据中心项目 (16:18) 据Space News报道，中国商业航天初创企业北京轨道春光科技有限公司（Orbital Chunguang）完成了Pre-A1轮融资，获得希慎资本、中信建投资本与中欧资本的支持，并从中国银行等12家主流金融机构获得了总额达84亿美元的战略授信额度。该公司由北京市科委支持的北京未来宇航空间技术研究院孵化，联合了24家机构，计划在700至800公里的晨昏轨道部署天基算力基础设施，规划于2035年前建成大规模天基计算网络。 节目分析指出，中美两国发展轨道算力的核心驱动力存在差异。美国支持天基算力的论点多集中于规避地面繁复的环境监管审查以及土地和水资源消耗；而在中国，由于地面基础设施审批效率高且具备三峡水电等充沛电力资源，将数据中心移入太空以化解地面散热和土地能耗瓶颈的构想展现出鲜明的发展模式对照。 图像模型演化瓶颈与商业视觉平民化 (20:43) 在探讨ChatGPT Images V2的功能突破时，节目回顾了图像基准测试从早期的“宇航员在月球骑马”，演化至测试模型逻辑理解的“马骑宇航员”，再到倒满至杯沿的红酒杯等物理边缘场景。针对模型生成带有两个时针的表盘等细节缺陷，节目引述Sam Altman与Andrej Karpathy的观点指出，将模型准确率提升到99%相对迅速，但要补齐后续每一个“9”的可靠性却需耗费数年，而人类用户对模型能力的评判标准也在持续水涨船高。 在商业应用端，ChatGPT Images V2与Midjourney大幅降低了高端产品摄影与品牌美学的门槛。过去硬件和消费品初创公司需要重金聘请专业摄影师或知名设计机构，如今即使是早期团队也能瞬间生成高质量产品图，甚至直接生成用于VR环境或立体视频流的360度等距柱状全景图。然而，这种技术普及也引发了审美同质化现象，促使追求差异化的品牌转向非标准化视觉风格。 假熊车险诈骗案与反常科学实验观察 (28:59) 节目还回顾了加利福尼亚州南部箭头湖发生的一起荒诞保险欺诈案。加利福尼亚州保险局在代号为“熊爪行动”的调查中查明，三名当地居民身穿橡胶与毛皮制作的全套熊服，钻入一辆2010年款劳斯莱斯古思特及两辆梅赛德斯-奔驰轿车内，使用厨房爪形工具划破内饰以伪造野熊袭击假象，成功骗取超过14.1万美元保险金。调查人员通过视频中异常平直规整的抓痕和人造破绽识破骗局，涉案嫌疑人已于2024年11月被捕并被判处监禁。 此外，节目提及了《纽约时报》报道的一项瑞典科研实验，研究人员给鲑鱼喂食可卡因后发现其游泳速度提高近一倍且耐力显著增强；最后，节目提到了红龙虾连锁餐馆计划于4月20日左右限时回归“无限续虾”促销活动的相关花絮。 本期节目展现出前沿AI竞争正从单一的技术演进转向算力重构、跨领域并购与企业级品牌包装；而技术平民化在打破商业壁垒的同时，也伴随着可靠性瓶颈、审美趋同以及光怪陆离的社会衍生现象。

## Background 本期节目于 2026 年 4 月 23 日由科技播客频道 TBPN 发布在 YouTube 平台上。该视频属于《Diet TBPN》精炼系列，由 John Coogan 和 Jordi Hays 共同主持，旨在用约 30 分钟的高效节奏为观众提炼当天科技直播节目中最精彩的商业洞察与行业重磅新闻。 ## TL;DR 节目核心聚焦于 SpaceX 与 AI 编程工具 Cursor 达成的震撼合作，深入拆解了涵盖 100 亿美元天价反向解约金、百亿级潜在收购要约以及 Colossus 超算接入的底层条款，并系统复盘了马斯克帝国近四年的资本整合与估值套利逻辑。随后，主持人探讨了中国商业航天初创企业获得的巨额信贷支持与天基轨道数据中心前景，并对比了中美两国的基建与监管差异。此外，节目还详细评测了 ChatGPT Images 2.0 在极端压力测试与复古视觉审美上的突破，并分享了加州“假熊破坏豪车”骗保案与“三文鱼吸食可卡因”等荒诞现实趣闻。 ## Atmosphere & style 全片充斥着硅谷圈内人标志性的快节奏与辛辣幽默，两位主持人互动默契、包袱频出，叙事风格兼具高度专业性与无厘头解构感。情绪曲线上，节目从前半段层层剖析数百亿级收并购财技的严密推演，逐渐滑向后半段围观“笨拙假熊划车”监控和“嗑药三文鱼游速翻倍”时的捧腹大笑。整场收听体验如同一场信息密度极高又充满机智吐槽的科技酒吧闲聊，既有硬核前沿视角的冲击，又自带强烈的娱乐与解压属性。 ## Bulletin - [00:54](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=54s) SpaceX 与 AI 编程工具 Cursor 达成深度合作与期权协议，Cursor 接入等效 100 万张 H100 的 Colossus 超算，协议包含 500 亿至 600 亿美元收购意向及 100 亿美元非稀释性反向解约金保障。 - [16:18](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=978s) 中国商业航天初创公司“北京辰光科技”完成 Pre-A1 轮融资，并获 12 家主流金融机构总计 84 亿元人民币授信额度，计划在 700 至 800 公里晨昏轨道布局天基数据中心。 - [20:43](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=1243s) 新一代图像生成模型（如 ChatGPT Images 2.0）已具备 360 度等距柱状全景图生成与复杂逻辑推理能力，正大幅抹平消费品与硬件初创企业的商业视觉设计壁垒。 - [28:59](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=1739s) 加州保险局破获代号“熊爪行动”的诈骗案，嫌疑人身着熊皮玩偶服并使用厨房抓肉爪具划伤劳斯莱斯等豪车内饰，企图骗取超 14.1 万美元保险理赔。 ## Key takeaways - [09:16](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=556s) SpaceX 与 Cursor 结盟旨在打造最强编程与知识工作 AI，Cursor 补足底层大模型预训练与算力资本短板，xAI 则借 Cursor 优秀开发者口碑扭转 Grok 偏离严肃办公场景的形象。 - [03:56](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=236s) 从 2022 年以 440 亿美元收购 Twitter，到成立 xAI、整合 Grok、xAI 与 X 合并，再到 2026 年 SpaceX 收购 xAI 并纳入 Cursor，马斯克完成了从算力、模型到分发与工具的 AI 生态闭环。 - [10:58](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=658s) SBF 在狱中通过官方批准的通讯渠道雇佣枪手代发推文维持社交网络热度，体现了现代服刑公众人物重塑公众好感度的公关手段。 - [10:58](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=658s) 根据投资人拆解，该笔交易本质上是 xAI 将闲置算力以成本价开放供 Cursor 冲击前沿代码模型，Cursor 获得了进可 600 亿美元被收购、退可拿 100 亿美元抵消 GPU 租金的无风险免费看涨期权。 - [13:46](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=826s) SpaceX 收购 Cursor 存在利用高市盈率收购低市盈率资产的估值倍数套利空间，但太空运力垄断所享有的高倍数能否直接套用于竞争激烈的 AI 编辑器赛道仍存疑。 - [13:46](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=826s) 面对代码编辑器赛道并购浪潮，Cognition（Devin）创始人 Scott Wu 依然选择独立坚守，成为吸收竞争对手离职工程师的“最后莫希干人”。 - [18:18](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=1098s) 美国对地面数据中心水资源与环保电力的严苛监管，倒逼科技企业将轨道数据中心作为规避地面政策阻力的可行替代路径。 - [18:18](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=1098s) 中国由科研机构牵头 24 家联合体规划 2035 年天基算力系统，但在三峡大坝等成熟水电地面基建支撑下，地面超算在现实可行性上依然更具即时优势。 - [20:43](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=1243s) AI 图像测试从“宇航员骑马”演进到“马骑宇航员”、倒满红酒及指定表盘指针等反向逻辑测试，暴露出预训练数据统计偏差与逻辑推理的最后 1% 攻坚难题。 - [24:32](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=1472s) 现代图像生成模型已具备真正的商业审美品味，促使未来品牌视觉在模板化极简主义与极度鲜明独特的主观风格之间走向两极分化。 - [26:43](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=1603s) 新版 ChatGPT 图像生成在科幻概念、360 度 VR 等距柱状全景图以及拟物复古 UI 排版（如 Windows XP、Mac OS X）上取得显著突破，逐步抹平与 Midjourney 的审美差距。 - [32:21](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=1941s) 瑞典科学家研究发现可卡因使三文鱼游速暴增近一倍且耐力提升，引发对学术研究经费流向与科研选题合理性的幽默反思。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=0s) 今日重磅看点与嘉宾阵容 节目开场抛出了一连串重磅科技新闻：SpaceX 与 AI 编程工具 Cursor 达成了极具震撼力的深度合作协议；ChatGPT Images 2.0 迎来新动向；OpenAI 动作频频；彭博社（Bloomberg）还爆出独家猛料，称一组未授权用户自 Claude Mythos 发布首日起就一直在违规调用该模型。两位主持人还预告了后续将登场的豪华嘉宾阵容，包括来自 Adobe、Build Forever、AngelList、Zankap、Vast Data 以及 Gradient 的业内人士。 在切入正题前，主持人迅速把焦点拉到了 SpaceX 与 Cursor 这笔非同寻常的交易上。这笔交易刚一提出便被认定“值得敲锣庆祝（gong-worthy）”——因为协议中不仅包含了 SpaceX 收购 Cursor 的期权，更附带了一笔高达 100 亿美元的天价反向解约费（breakup fee）。对于双方而言，这无疑是一场各取所需的双赢合作。 - **今日核心要闻概览**：SpaceX 联手 Cursor 达成期权与算力合作、ChatGPT Images 2.0 更新、OpenAI 动态追踪、彭博社曝光 Claude Mythos 未授权访问事件。 - **做客嘉宾名单**：Adobe、Build Forever、AngelList、Zankap、Vast Data 及 Gradient 的相关代表。 --- ### [00:54](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=54s) SpaceX 联手 Cursor：算力巨兽与顶尖编程智能体结盟 主持人在调侃“我们现在甚至不需要对齐事实，只要直接抛出观点就行”的段子之后，详细拆解了 SpaceX 与 Cursor 官方合作的真实细节。SpaceX 此次与 Cursor 携手，旨在“打造全球最强的编程与知识工作 AI（create the world's best coding and knowledge work AI）”。Cursor 旗下拥有具备前沿模型性能表现的自主编程智能体 Composer 2，而通过该笔交易，Cursor 将直接接入 SpaceX 拥有相当于 100 万张 H100 算力规模的超级计算机“巨像”（Colossus）。 ```mermaid flowchart TD accTitle: SpaceX 与 Cursor 合作架构与期权交易示意图 accDescr: 展示 SpaceX 算力资源赋能 Cursor 以及双方期权交易架构流程。 subgraph SpaceX_Infrastructure["SpaceX / xAI 算力基础设施"] Colossus["'Colossus' 超级计算机<br/>(100 万张 H100 等效算力)"] GigaFactory["AI 算力千兆工厂<br/>(Compute Gigafactory)"] end subgraph Cursor_Capability["Cursor 智能体能力"] Composer["Composer 2<br/>(前沿水准自主编程模型)"] Graphite["Graphite 代码审查生态<br/>(AI 时代的 Code Review)"] end subgraph Deal_Terms["交易架构核心条款"] Acquisition["500 亿 - 600 亿美元<br/>潜在整体收购要约"] Breakup["100 亿美元<br/>非稀释性反向解约金保障"] end Colossus -->|输送前沿超算资源| Composer GigaFactory -.-> Colossus Composer --> Deal_Terms Graphite -.-> Composer ``` 主持人特别强调，将“Colossus”称为超级计算机是完全准确的，尽管埃隆·马斯克（Elon Musk）和 xAI 团队还造过很多夸张的词汇，比如“算力千兆工厂（Compute Gigafactory）”。回溯整个脉络，这一切演变令人瞠目结舌：从最初马斯克一句轻描淡写的“我想买下 Twitter”，发展到现在，马斯克麾下已经串联起了庞大的超级计算机数据中心、顶尖编程 Agent、社交媒体平台，甚至还兼并了代码审查公司 Graphite，拼图已然完整。 现场还穿插了一段对正在狱中服刑的 FTX 创始人 SBF（Sam Bankman-Fried）的辛辣调侃：想象一下 SBF 每隔几天获得一次与外界通话的机会，激动地询问外面：“我的宝贝 Anthropic 现在估值怎么样了？”得到的答复是“先生，二级市场已经炒到了 8000 多亿美元，并且直奔 1 万亿美元市值去了”。第二天 SBF 又打电话问：“那我投的 Cursor 呢？”电话那头只能尴尬回应：“先生，您可能不太想听……他们刚达成了 500 亿至 600 亿美元的收购协议，就算最后不卖，他们也能白拿 100 亿美元的非稀释性解约补偿金。” - **超级算力底座**：Colossus 具备等效 100 万张 NVIDIA H100 的峰值计算能力，被埃隆·马斯克形象地定义为“算力千兆工厂”。 - **商业条款细节**：涉及 500 亿至 600 亿美元的收购意向，若交易未完成，Cursor 亦可获得 100 亿美元的无稀释纯现金补偿。 - **生态整合**：结合此前整合的 Graphite 代码审查工具，形成了从编码、审查到基础设施的闭环。 --- ### [03:56](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=236s) 回溯马斯克四年布局：从推特收购案到 xAI 的光速狂飙 为理解当下这笔交易的分量，主持人系统梳理了过去四年间马斯克帝国的重大演变。一切始于 2022 年 4 月 14 日马斯克主动发起对 Twitter 的收购要约。彼时恰逢美联储激进加息、零利率时代（ZIRP）彻底终结，全球科技股陷入暴跌，Snap、Pinterest 乃至 Meta 等公司股价几乎腰斩。Meta 当时每年在现实实验室（Reality Labs）上砸入上百亿美元押注未来，市场在贴现率从 3% 飙升至 6% 的折现现金流模型（DCF）下，被迫对这类远期收入大打折扣。而马斯克在暴跌前夕锁定了 440 亿美元的高价，随后在诉讼与推拉中险些破裂，最终摩根士丹利及一众 VC 机构“被硬着头皮拉上了船”。 然而戏剧性的是，“永远不要做空埃隆·马斯克”的定律再次应验：当年被迫入局的资方，后来陆续置换获得了 xAI 与 SpaceX 的股权，最终反而收益颇丰。2022 年 10 月 27 日，马斯克手捧水槽走进推特总部，完成了那场著名的“Let that sink in”交接。紧接着在 2023 年 7 月 12 日，距离 ChatGPT 问世仅过去约八九个月，马斯克正式宣布成立 xAI，全面对标 Google DeepMind、OpenAI 与 Anthropic。 进入产品落地期后，xAI 与 X 的整合推进极快： - **2023 年 11 月 3 日**：正式发布初代模型 Grok。 - **2023 年 12 月 7 日**：Grok 深度集成进 X 平台，面向 Premium+ 会员开放推送。 - **2024 年 3 月 17 日**：xAI 宣布开源 Grok-1 模型权重。 - **2024 年 3 月 26 日至 28 日**：将使用权限下放至常规 X Premium 用户，并迅速发布 Grok 1.5。 - **2024 年 4 月 12 日**：上线具备多模态理解能力的 Grok 1.5V。 - **2024 年下半年至 2025 年初**：先后迭代 Grok 2 与 Grok 2 mini，并最终向全体 X 平台免费用户全面开放；2025 年 2 月 19 日正式推出 Grok 3 测试版。 主持人感慨，马斯克入主 Twitter 后将公司彻底切换回了“初创公司战备状态”，以极快的敏捷节奏砍掉冗余、强推功能（甚至把前朝搁置的 Community Notes 迅速推上线）。到了 2025 年，xAI 已经彻底从单纯挂靠在 X 上的外挂工具，转变为在产品层面全面吸收、重塑整个社交平台底层体系的 AI 引擎。 ### [09:16](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=556s) 马斯克商业帝国的资本拼图与 Cursor 的入局逻辑 主持人首先梳理了一条极具马斯克特色的资本运作时间线：2025 年 3 月 28 日，xAI 通过全股票交易收购了 X（前 Twitter），将这家前沿 AI 初创企业与社交平台深度绑定；随后，X 紧锣密鼓地接连推出 Grok 4 fast 与 Grok 4.1 等更快速的模型更新；紧接着在 2026 年 2 月 2 日，SpaceX 出手收购了 xAI。整套组合拳脉络清晰：收购 Twitter、创立 xAI、将 Grok 整合进 X、合并 xAI 与 X，再由 SpaceX 全盘接管 xAI。如今，Cursor 正式加入这一超级舰队。 ```mermaid flowchart LR accTitle: 马斯克 AI 资产整合与 Cursor 加入路径 accDescr: 展示从 Twitter 收购到 SpaceX 整合 xAI 及 Cursor 的完整演变过程。 subgraph S1["前期整合"] A["收购 Twitter / X"] --> B["创立 xAI 并接入 Grok"] B --> C["2025.03: xAI 全股票收购 X"] end subgraph S2["巨头合体与最新布局"] C --> D["2026.02: SpaceX 收购 xAI"] D --> E["Cursor 正式并入生态"] end ``` 从战略角度来看，这项交易极为契合双方需求。Cursor 迫切需要海量的算力、资源和充沛资本来训练前沿级编程模型——他们此前从未完整做过从头预训练（Pre-training），而 xAI 团队对此则驾轻就熟。更为微妙的是品牌形象的互补：Grok 历经种种争议，甚至曾被称为“在工作场所被明令禁用的次数比正式采用的次数还多”；相比之下，Cursor 拥有极具美誉度的开发者口碑，其愿景也早已超出单纯的编程工具，旨在“打造全球最强编程与知识工作 AI”。两位主持人在此打趣道，现在硅谷每家公司都在造同一个“你下命令它干活的盒子”，主持人戏谑地反问对方是不是想让一家公司垄断一切的“共产主义者”，引来一片笑声。 - **关键历史节点**：2025 年 3 月 28 日 xAI 与 X 合并；2026 年 2 月 2 日 SpaceX 收购 xAI。 - **核心诉求对接**：Cursor 补足缺乏大模型预训练经验与底层算力资本的短板，xAI 则借由 Cursor 摆脱 Grok 偏离严肃办公场景的公众刻板印象。 --- ### [10:58](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=658s) 狱中发推的 SBF、公关幽灵与 Richard Woo 的交易内幕拆解 就在 SpaceX 官宣与 Cursor 达成合作的推文下方，竟赫然出现了正在服刑的 FTX 创始人 SBF（Sam Bankman-Fried）的回复互动，引发两位主持人对“囚犯为何能自由发推”的吐槽。SBF 在个人简介中专门写道，这是通过美国联邦监狱局（BOP）批准的常规电话和邮件委托他人代发的。主持人调侃称，虽然剥夺服刑人员与外界沟通可能有侵犯言论自由之嫌，但这类法律显然诞生于前互联网时代，如今在监狱里雇一个代笔枪手（Ghostwriter）在社交网络上维持热度、营造轻度幽默人设，正如伊丽莎白·霍姆斯（Elizabeth Holmes）此前所做的那样，简直成了重塑公众好感度的顶级公关武器。 随后，节目引述了投资人 Richard Woo 对该笔交易结构的深度拆解： 1. xAI 此前面临前沿编程模型训练遇阻、部分联合创始人离职的窘境，手里积压了大量闲置 GPU 算力； 2. Cursor 拿不出 50 亿美元去冒风险做一次预训练来硬刚 Codex 和 Claude； 3. xAI 向 Cursor 提出：以成本价开放所有算力供其冲击顶尖代码模型，前提是获得对其优先收购权； 4. Cursor 获得了极佳的免费看涨期权——若能训练出超越 Opus 的模型，便能以 600 亿美元高价被收购；即使没达到最高预期，也能获得 100 亿美元保障资金以抵消租用 GPU 的成本，堪称双赢。 主持人还盛赞了 Cursor 联合创始人兼 CEO Michael Truell（节目口误为 Michael Troughton / Michael Treu）。这位极其年轻的创业者身上带着典型的“Stripe 风格”与温和美学——在与 Patrick Collison 录制的那期咖啡馆漫谈播客中，其形象沉稳而治愈。比起马斯克与 Grok 卷入的种种政治风波与争议狂欢，Michael Truell 所展现的气质正是可口可乐、通用电气（GE）或福特这类传统美国企业巨头最放心接纳的企业形象，“美学与氛围感在商业世界真的极其重要”。 --- ### [13:46](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=826s) 估值倍数套利神话与 Scott Wu 坚守的“最后莫希干人” 节目随后探讨了由 Matt Slotnick 和 Julian 提出的“估值倍数套利”（Multiple Expansion）逻辑。Julian 分析指出，表面上看 SpaceX 以 30 倍营收倍数收购 Cursor 似乎存在溢价，但对于即将在公开市场以预计超过 100 倍营收估值 IPO 的 SpaceX 而言，这笔交易具有极强的每股增益效应（Accretive）。将 Cursor 约 20 亿美元营收套入 70 倍的估值差空间中，带来的市值放大效应惊人。不过主持人也警示道，这属于企业收并购中的经典财技把戏，聪明投资者会对不同业务的现金流区别定价——SpaceX 对太空发射运力的绝对垄断地位理应享有极高估值倍数，而置身于高度内卷、寡头竞争的 AI 代码编辑器赛道，能否维持同样的高倍数仍是巨大的未知数。 现场还算起了一笔爆笑的“荒谬年化算法”：Cursor 获得了 100 亿美元资金保障，主持人调侃这已经超越单纯的营收跑率（Run Rate），相当于锁定了“单月入账 100 亿美元”；如果按硅谷传统的乘 12 算法，Cursor 当月即可宣布拥有 1200 亿美元年化营收，一跃成为当月“全球规模最大的 AI 公司”。 与此形成鲜明对比的是 Cognition 创始人 Scott Wu。主持人展示了网友制作的恶搞梗图：面对伴随并购大潮逐步落幕的竞争者们（包括 Windsurf 与 Cursor 等），孤军奋战的 Scott Wu 仿佛《最后的莫希干人》，正摩拳擦掌准备接纳那些心怀不满而跳槽的工程师，在群雄割据的幽灵船队包围下坚守阵地。 --- ### [16:18](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=978s) 中国空天算力入局：北京微纳星空（晨光）与 84 亿巨额信贷 节目的后半段转向一则来自《Space News》的重磅动态：中国正全力推进轨道数据中心建设，一家名为“北京辰光科技”（Beijing Orbital Twilight Technology，即微纳星空关联布局公司）的商业航天初创企业，宣布完成了 Pre-A1 轮融资，并获得了由中国银行等 12 家主流金融机构联合提供的总计高达 84 亿元人民币的巨额战略授信额度。参与本轮股权投资的阵容堪称豪华，囊括了海松资本（Hysan Capital）、中信建投资本（CITIC Construction Investment Capital）以及国投美亚（Cathay Capital）等重量级机构。 主持人对中国初创企业这种充满宏大科幻色彩的命名风格惊叹不已，调侃称美国公司哪怕起 AnySphere 或 Cursor 这种名字，也完全比不上“轨道辰光”那般气势磅礴。但针对高达 84 亿元的信贷规模，两位主持人展开了困惑与思考： - 如果这家公司既不自己代工芯片制造，也不自建运载火箭，究竟为何需要如此庞大的重资产资金？难道是为了向华为等厂商集中大批量采购定制化算力芯片？ - 针对将算力中心搬上太空的商业逻辑，主持人分析道，支持者往往主张地面建设超算面临着极为严苛的环保与电力监管阻力。把数据中心送入低地轨道，似乎成了规避地面严密管制的终极解法；适逢“世界地球日”，两位主持人最后诙谐地向地球致意：“恭喜你，地球，你做得棒极了——因为我们再也不打算在地面上盖更多高耗能的数据中心了。” ### [18:18](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=1098s) 太空数据中心与中美监管：三峡大坝的“水权”调侃 主播们深入探讨了中美在建设前沿基础设施上的巨大差异。在美国，太空数据中心的商业逻辑在很大程度上是由繁重的监管壁垒催生的；相比之下，中国的监管开销要小得多，几乎是“想建就能建”。主播调侃道，美国社会对数据中心消耗水资源的恐慌甚至演变成了一种近乎荒谬的情绪——仿佛水拥有不可侵犯的“人权”，既不能流过水车也不能流经大坝发电，只能静静地待在原地。而在中国，三峡大坝本身就是通过水流来生产电力的，水在发电过程中既没有被毁灭也没有受到污染，只是自然流过，但在美国的舆论视角里，像胡佛水坝（Hoover Dam）这种规模的存在简直就是一场生态噩梦。 中国在轨道算力领域也有长远布局。中国商业航天项目（如星河动力/北京未来星空技术研究院等孵化项目）背后有北京市科委及中关村管委会的支持，并牵头了由24家机构组成的联合体。根据此前简报，尽管全球大规模地面数据中心扩张迅猛，但也面临占地过大、能源消耗激增以及大气散热受限等挑战。该项目计划利用700至800公里的晨昏轨道（Dawn-Dusk Orbit），目标是到2035年建成可支持天基计算的大规模太空数据中心。虽然其规划在2025至2026年进行首颗试验星发射，但目前进展似乎有些延宕。主播们笑着调侃，中国长征系列火箭运力不错，即便火箭回收降落技术还没做到极致，但资金充足的情况下也可以直接多发几颗火箭来解决，不过在三峡大坝旁边直接建超算其实更有现实可行性。 - **轨道方案**：规划位于 700–800 公里的晨昏轨道，旨在 2035 年前实现大规模太空算力支撑。 - **阶段目标**：第一阶段（原定 2025–2027 年）聚焦核心技术验证，但首颗试验卫星暂未见发射入轨。 - **牵头生态**：由北京未来航天相关研究院主导，联合了 24 家机构共同参与。 ```mermaid flowchart TD accTitle: Space Data Center Motivation Comparison accDescr: US regulatory push compared to China orbital computing plan subgraph US["美国视角 (US)"] U1["地面数据中心面临水资源争议与环保监管"] --> U2["倒逼科技企业探索太空数据中心路线"] end subgraph CN["中国视角 (China)"] C1["拥有三峡大坝等庞大水电与地面能源基建"] C2["北京相关科研机构牵头 24 家机构联合体"] C3["规划 700-800km 晨昏轨道天基算力 (展望 2035)"] C2 --> C3 end ``` --- ### [20:43](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=1243s) ChatGPT 与图像生成模型 2.0：从“马骑宇航员”到极限压力测试 话题切入新一代图像模型（如 ChatGPT Images 2.0、Midjourney 以及 Grok 等），主播们对社区中涌现出的离谱压力测试图像叹为观止。有人尝试重现早期的经典模因，也有像 Gabriel 这样的测试者专门把模型推向混乱的极限——天空中飞奔着奶牛、悬挂的马匹和跳床，充斥着荒诞感与超载细节。主播解释了这些评测基准的发展脉络：早在 DALL-E 时代，大家最惊艳的画面是“宇航员在月球上骑马”，因为现实中不可能拍到这张照片，Photoshop 合成难度也很高。 随着模型能力提升，基准测试演化成了更严苛的反向逻辑考察，例如 Ethan Mollick 总结的复合提示词挑战： - **逻辑颠倒测试**：“马骑宇航员”而非宇航员骑马。早期模型往往依靠互联网数据的先验概率自动纠错，画出来的永远还是人骑马；只有具备真正语意推理能力的模型才能反转这种关系。 - **训练偏差测试**：一杯“斟满的红酒”。现实照片中人们倒红酒通常只倒半杯或三分之二，因此模型很难画出一杯真正满到杯沿的红酒，除非它克服了预训练数据的统计偏见。 - **时钟指针识别**：让模型按指定时间绘制表盘。新模型画面中依然画出了两根时针，暴露了微小细节上的逻辑破绽。 看到时钟上多出来的时钟指针，主播们开始大笑着开玩笑：“快把球门挪走（Moving the goalposts）！”“它已经完成了 99%，但只要还有瑕疵，AI 从业者就得接着加班。”主播借此引用了 Andrej Karpathy 对自动驾驶的经典论述：做到 99% 或许只需一年，但为了增加小数点后每一个 9 都要付出指数级的时间与精力，因为人类在图像和智能上永远苛求完美，绝不会容忍表盘上有两根时针。 --- ### [24:32](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=1472s) Dwarkesh 模仿秀与产品设计降权：AI 是否真正拥有了“审美品味” 主播们注意到一张流传甚广的“Dwarkesh Patel 风格自拍照”，配文完美复制了硅谷新移民入职搬到旧金山后的标准打招呼模板：“嗨，我是 Dwarkesh，走遍全美现居旧金山，随时欢迎聊聊 AI 科学与历史……今年夏天来喝杯咖啡吧！”这一高度拟真的 AI 生成图再次证明新模型在日常社交氛围感还原上的恐怖精度。 两人重点讨论了 AI 图像质量的一个根本性转变——模型终于拥有了真正的“品味”（Taste）。以往在消费品（CPG）和硬件初创企业领域，投资人可以通过产品宣发图的水准来逆向评估创始人的综合能力：“即使创始人本人没有顶尖创意、融资也不多，但只要足够有冲劲（scrappy），就能找到优秀的独立摄影师完成高质量视觉呈现。”而现在，诸如 Andreessen Horowitz（a16z）投资的电动滑板车公司等团队，已经可以直接用 Midjourney 等模型批量生成质感相当于数万美元拍摄成本的商业摄影大片。 这种技术普及在带来便利的同时也抹平了早期的视觉护城河。主播预测，未来的设计风潮可能会出现两极分化：大量公司会陷入对苹果或 Linear 极简主义设计风格的模板化内卷；与此同时，真正能够脱颖而出的将是那些敢于尝试极具主观性、甚至古怪且难以预测的设计风格的品牌。一旦某种独树一帜的美学被品牌自身牢固占领，即便他人能用 AI 轻易复制渲染图，也无法动摇其品牌的核心心智。 ### [26:39](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=1599s) ChatGPT 图像生成进化与 Windows XP 怀旧审美浪潮 主持人们由早期设计机构 Red Antler 树立独特品牌美学引发大量拙劣模仿的往事切入，复盘了 AI 图像生成工具的风格演进。过去很长一段时间里，Midjourney 凭借类似 ArtStation 的概念艺术风、细腻的绘画质感以及硬核科幻构图，牢牢占据了独特的审美高地；相比之下，初代 ChatGPT 生成的图像更像平庸呆板的商业微利库存照片（Stock Photography）。如今新版 ChatGPT 图像模型展现出了更鲜明的审美主张与风格掌控力，科幻场景的表达也大幅强化，不过在未加入极端提示词的情况下，整体风格仍时常略显偏向写实摄影。 展示环节中，Ben Hilack 发布的一张全景图引来热烈讨论，画面配文调侃为“噩梦级飞草局”（Nightmare blunt rotation），其中甚至出现了 Sam Bankman-Fried（SBF）的身影。该模型目前已能够直接生成完整的等距柱状全景图（Equirectangular Image），用户不仅可以在 VR 设备中浏览，还能导入立体全景查看器。这一突破也随之引发了对下游工具链的思考：用户未来必然渴望让这些全景画面动起来，如何将其无缝转化为动态视频将是行业接下来的焦点。此外，该模型在前端排版布局和界面设计方面同样表现惊艳。节目随后展示了社区极具创意的“如果 Codex 诞生于 2005 年或 2012 年”的概念图，画面深度还原了早期 Mac OS X 的拟物水滴质感以及 Windows XP 的经典蓝绿草地界面。主持人们捧腹笑称，这种复古审美早已该全面复兴——一边畅饮激浪（Mountain Dew），一边打着《光环》（Halo），旁边屏幕上还跑着 XP 皮肤版的 Codex，“这绝对能治愈你的精神内耗”。 - **Midjourney vs. ChatGPT**：早期 ChatGPT 图像模型缺乏个性，偏向普通图库风格，新版本则在科幻、布局设计及审美风格化上大幅追赶 Midjourney。 - **全景生成能力**：支持生成 360 度等距柱状图（Equirectangular Image），为 VR 沉浸式场景展示及未来的全景视频动画生成铺平了道路。 - **复古 UI 怀旧热**：社区掀起将现代 AI 编码助手（如 Codex）套用 2000 年代初期设计语言的热潮，重现 Mac OS X 和 Windows XP 的经典视觉元素。 --- ### [28:59](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=1739s) “熊爪行动”：加州假熊套路骗保案 随后话题转向一桩令人啼笑皆非的现实离奇案件。加利福尼亚州三名居民因精心策划了一场“假熊袭击豪车”的连环诈骗案而被判刑入狱，涉案骗保金额超过 14.1 万美元。诈骗团伙不仅网购了全套熊类毛皮玩偶服，还特意使用类似熊爪造型的厨房抓肉金属工具，在多辆豪华车内部划出密密麻麻的刮痕，并先后向三家不同的保险公司索赔。加州保险局在收到一起疑点重重的理赔申请后正式立案，代号为“熊爪行动”（Operation Bear Claw）。 ```mermaid flowchart TD accTitle: 假熊骗保案案发流程 accDescr: 诈骗分子身着熊装破坏车内饰并被监控拍下，最终被保险调查局识破逮捕。 subgraph Fraud ["伪造现场"] A["嫌疑人套上熊皮毛偶服"] --> B["携带厨房抓肉爪具钻入车内"] B --> C["蓄意破坏劳斯莱斯与奔驰内饰"] end subgraph Claims ["提请理赔与破案"] C --> D["向三家保险公司索赔逾14.1万美元"] D --> E["提交车内监控录像作为证据"] E --> F["保险局启动'熊爪行动'识破假熊破绽"] end ``` 案件最初发生在以野生黑熊出没闻名的加州箭头湖（Lake Arrowhead）。嫌疑人声称一头野生熊擅自钻入了一辆 2010 款劳斯莱斯古思特（Rolls-Royce Ghost），并提交了车内监控视频和座椅皮饰严重被毁的照片。然而，视频画面极具滑稽色彩：画面中只见一个身形笨拙、披着毛皮的人在后排座椅翻爬并胡乱拍打仪表盘；更破绽百出的是，座椅上的划痕笔直规整，与真正野生猛兽造成的野性撕裂痕迹相去甚远，被调查人员戏称为“史上最弱不禁风的假熊”。官方经鉴定认定作案者根本不是什么“身手矫健的黑熊”，并在顺藤摸瓜中发现该团伙还用同一套熊皮道具对另外两辆梅赛德斯-奔驰轿车实施了同样的破坏索赔。面对这种异想天开的犯罪手法，主持人不由联想到自己朋友的真实生活经历：有次朋友去养山羊的熟人家里拜访，仅因车门未关，转眼间钻进车厢里乱嚼的反倒真是一只活生生的山羊。 - **涉案车辆**：一辆 2010 款劳斯莱斯古思特（Rolls-Royce Ghost）及两辆梅赛德斯-奔驰汽车。 - **涉案金额与结局**：涉案金额逾 141,000 美元，三人最终均被加州警方逮捕并获刑。 - **破案破绽**：厨房爪具留下的划痕过于平行规整，车内监控中“假熊”拉开车门爬行与挥爪动作毫无兽性特征。 --- ### [32:21](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=1941s) 嗑药的三文鱼与红龙虾“无尽大虾”的回归 荒诞故事并未就此结束，节目接着引用了《纽约时报》的一则离奇科学报道——《这些三文鱼吸食可卡因嗨了，但那还不是最疯狂的部分》。瑞典科学家在实验中尝试给三文鱼喂食可卡因，结果意外发现：摄入毒品后的三文鱼游动速度居然暴增了将近一倍，而且续航耐力显著提升。主持人与导播 Tyler 忍不住大肆吐槽学术界的研究经费分配机制：想象一下那些为攻克阿尔茨海默病等重大医学难题四处奔走却拿不到经费的顶尖学者，看到这篇论文时内心该有多么崩溃——“我辛辛苦苦研究绝症没有赞助，你们居然拿国家科研经费去喂鱼吸毒？”对于这项实验的影响，报道中专家的锐评同样耐人寻味：“虽然尚不清楚在毒品影响下游得更快更远对鱼类有何长期损伤，但这显然不是什么好事。” 节目尾声，话题跳跃至水下生物的终极归宿——著名海鲜连锁餐厅红龙虾（Red Lobster）宣布将于近期重新限时上线其标志性的招牌促销活动“无尽大虾”（Endless Shrimp）。网络名人 Luke Metro 甚至在推特上打趣称，有效利他主义者（EA）应该发起一场敌意收购拿下红龙虾，并强制将其改造成纯素食餐厅。主持人们围绕自己是否去过红龙虾展开闲聊，均表示从未踏足过这类餐厅，反倒偶尔会去澳美客牛排馆（Outback Steakhouse）。来自明尼苏达州的幕后团队成员 Ben 被追问打分时，原本想给出不痛不痒的 5 分中评，但在主持人“不许打中庸分”的逼问下最终给出了 4 分的“不及格”差评。最后，节目在轻松调侃的道别声中落下帷幕。 - **三文鱼毒理实验**：瑞典实验表明摄入可卡因的三文鱼游泳速度提升近两倍，且耐力持久，引发对科研资金用途的幽默调侃。 - **Red Lobster 商业动向**：红龙虾餐厅计划在 4 月下旬限时恢复让其声名大噪也曾带来财务重负的“无尽大虾”放开吃活动。 ## Notable quotes > [00:54](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=54s) 打造全球最强的编程与知识工作 AI — SpaceX 与 Cursor > [00:54](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=54s) 我们现在甚至不需要对齐事实，只要直接抛出观点就行 — 主持人 > [00:54](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=54s) 我的宝贝 Anthropic 现在估值怎么样了？ — Sam Bankman-Fried > [03:56](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=236s) 永远不要做空埃隆·马斯克 — 主持人 > [09:16](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=556s) 在工作场所被明令禁用的次数比正式采用的次数还多 — 主持人 > [10:58](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=658s) 美学与氛围感在商业世界真的极其重要 — 主持人 > [16:18](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=978s) 恭喜你，地球，你做得棒极了——因为我们再也不打算在地面上盖更多高耗能的数据中心了。 — 主持人 > [20:43](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=1243s) 快把球门挪走！它已经完成了 99%，但只要还有瑕疵，AI 从业者就得接着加班。 — 主持人 > [24:32](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=1472s) 嗨，我是 Dwarkesh，走遍全美现居旧金山，随时欢迎聊聊 AI 科学与历史……今年夏天来喝杯咖啡吧！ — Dwarkesh Patel 模拟提示词 > [26:39](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=1599s) 噩梦级飞草局 — Ben Hilack > [32:21](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=1941s) 这些三文鱼吸食可卡因嗨了，但那还不是最疯狂的部分 — 《纽约时报》 > [32:21](https://www.youtube.com/watch?v=ss3vs6kJVMM&t=1941s) 虽然尚不清楚在毒品影响下游得更快更远对鱼类有何长期损伤，但这显然不是什么好事。 — 专家 ## Mentioned SpaceX, Cursor, ChatGPT, ChatGPT Images 2.0, OpenAI, 彭博社 (Bloomberg), Claude Mythos, Adobe, Build Forever, AngelList, Zankap, Vast Data, Gradient, Composer 2, Colossus, NVIDIA H100, Graphite, xAI, Elon Musk, Sam Bankman-Fried (SBF), Anthropic, Twitter / X, Meta, Reality Labs, Morgan Stanley, Google DeepMind, Grok, Grok 1.5, Grok 1.5V, Grok 2, Grok 3, Grok 4 fast, Richard Woo, Michael Truell, Patrick Collison, Stripe, Matt Slotnick, Julian, Cognition, Scott Wu, Windsurf, Space News, 北京辰光科技 (微纳星空), 中国银行, 海松资本 (Hysan Capital), 中信建投资本 (CITIC Construction Investment Capital), 国投美亚 (Cathay Capital), 华为, 三峡大坝, 胡佛水坝, 星河动力, Midjourney, DALL-E, Ethan Mollick, Andrej Karpathy, Dwarkesh Patel, Andreessen Horowitz (a16z), Red Antler, Ben Hilack, Windows XP, Mac OS X, Codex, 加州保险局 (Operation Bear Claw), Rolls-Royce Ghost, Mercedes-Benz, 纽约时报 (The New York Times), Red Lobster, Luke Metro, Outback Steakhouse ## Video description 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. 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=ss3vs6kJVMM)

[0:54] SpaceX 与 Cursor: 打造全球最强的编程与知识工作 AI [0:54] 主持人: 我们现在甚至不需要对齐事实，只要直接抛出观点就行 [0:54] Sam Bankman-Fried: 我的宝贝 Anthropic 现在估值怎么样了？ [3:56] 主持人: 永远不要做空埃隆·马斯克 [9:16] 主持人: 在工作场所被明令禁用的次数比正式采用的次数还多 [10:58] 主持人: 美学与氛围感在商业世界真的极其重要 [16:18] 主持人: 恭喜你，地球，你做得棒极了——因为我们再也不打算在地面上盖更多高耗能的数据中心了。 [20:43] 主持人: 快把球门挪走！它已经完成了 99%，但只要还有瑕疵，AI 从业者就得接着加班。 [24:32] Dwarkesh Patel 模拟提示词: 嗨，我是 Dwarkesh，走遍全美现居旧金山，随时欢迎聊聊 AI 科学与历史……今年夏天来喝杯咖啡吧！ [26:39] Ben Hilack: 噩梦级飞草局 [32:21] 《纽约时报》: 这些三文鱼吸食可卡因嗨了，但那还不是最疯狂的部分 [32:21] 专家: 虽然尚不清楚在毒品影响下游得更快更远对鱼类有何长期损伤，但这显然不是什么好事。

Silicon Valley’s newest obsession [3:55] let that sink in [15:31] Patrick: last of the Mohicans [20:43] Ethan Mollick: a horse riding an astronaut [20:43] Dorkhast [20:43] just moved to SF / let's grab coffee [29:56] the weakest bear ever [32:21] These salmon got high on cocaine. That wasn't the craziest part. [32:21] While it's unclear if swimming faster and further while under the influence harms these fish, experts say it's probably not great. [33:33] No fives!

[0:00] I've a bunch of major stories today. [0:02] SpaceX and Cursor partnering up. [0:06] More news out of Images 2.0. A bunch of [0:09] news out of Open AI. And Mythos, a group [0:13] of unauthorized users have been using [0:15] Claude Mythos since the day it was [0:16] released. It's a big scoop in Bloomberg. [0:18] We'll go through there's a whole bunch [0:19] of timeline. We can also pull up the [0:21] lineup and take you through who's coming [0:23] on. We got Adobe, Build Forever, [0:25] AngelList, Zankap, Vast Data, Gradient. [0:29] We're going through all the news of the [0:31] day. Well, let's start with SpaceX and [0:34] Cursor who are teaming up in a very [0:37] interesting deal. Is this a gong [0:38] already? [0:39] >> This is gong-worthy because it's an [0:41] option to buy the company but a $10 [0:43] billion [0:45] breakup fee. [0:47] Yeah. Incredibly incredibly [0:49] >> Yeah. [0:49] >> I think I think it's a win-win. It's a [0:51] win-win. [0:51] >> Yeah. [0:51] >> It's a win-win. I think it makes sense [0:53] for [0:54] both. [0:54] >> Well, well, well, let's go through the [0:56] facts first. So, um [0:58] SpaceX partners fact. [1:00] >> He needs facts. [1:01] Just just immediate takes. We assume you [1:03] already know all the industry. [1:05] >> Post fact. [1:06] >> Do you know all the information? Do we [1:07] need to give you any information? We'll [1:08] see. But, let's let's run through it. [1:10] So, SpaceX partners with Cursor to quote [1:13] create the world's best coding and [1:15] knowledge work AI. The deal gets Cursor, [1:18] whose agentic coding model, Composer 2, [1:21] basically operates at frontier level [1:23] performance, access to compute from [1:25] SpaceX's million H100 equivalent [1:28] Colossus supercomputer. And that is the [1:30] correct term for Colossus. It is a [1:33] supercomputer. There were some other [1:34] terms that Elon was throwing out. [1:37] Fantastic terminology from the xAI team [1:40] over there. What was the other one? [1:42] Uh [1:43] It was like AI [1:44] >> Compute giga factory. [1:46] >> Compute giga factory. They're all [1:48] hilarious and and very good. I like all [1:50] of these. [1:50] >> Computer? Computer giga factory? [1:53] >> The xAI timeline, I was like I was [1:55] thinking about it and I was like, when [1:56] did this actually start? Like, I I cuz [1:59] it just kind of came out of nowhere and [2:01] just absolutely blew up. I wanted to [2:02] actually review and reset on like the [2:05] timeline of events here, because it's [2:07] gotten so crazy to the point where it's [2:09] like the thing that started as like Elon [2:11] sort of being like, "I want to buy [2:13] Twitter." is now like a Neo Lab with a [2:16] massive supercomputer data center and a [2:20] coding agent and code review for the age [2:21] of AI, because don't forget they own [2:23] Graphite now. [2:25] Or potentially well. [2:27] Uh and and and a social media app. [2:29] >> the space code review company. [2:33] >> So [2:33] >> I'm really I'm genuinely so [2:35] >> excited for Amara. [2:36] >> for the Graphite team. I'm thrilled for [2:39] the Cursor team. I'm thrilled for [2:42] >> for Scott Woo. [2:43] >> Yeah. Yeah, Scott Scott Woo is licking [2:45] his chops. He's like, "What are you [2:46] going to What are you going to leave out [2:47] in the records for [2:50] >> He really is going to have fun with it. [2:52] >> I also think we needed we need to take [2:54] one moment [2:55] >> Yeah. [2:55] >> and uh just send some thoughts and [2:57] prayers to SBF who [2:59] >> Oh, yeah. [2:59] >> Imagine I I don't know how this works in [3:03] in whatever prison he's in, but I [3:04] imagine you know, what hopefully every [3:06] few days he can go and like talk to the [3:09] outside world. [3:10] >> Yeah. [3:11] >> And you can imagine him going, you know, [3:14] making the call [3:15] >> Yeah. [3:15] >> and he says, you know, "Sir [3:17] >> What would the mark be? [3:18] >> How how how how's Anthropic doing? How's [3:20] my How's my baby doing?" And he's like, [3:23] "Sir, [3:24] >> Sir, it's traded [3:25] >> north of 800 in secondary markets and [3:27] it's and it's quite possibly going to [3:29] north of a trillion. And then he's like, [3:31] "Th- thank you." Hangs up. Comes back [3:34] the next day. [3:35] Uh "I I haven't checked in with Cursor [3:37] in a little bit. How are they doing?" [3:38] >> Sir [3:39] >> You may not want to hear this, but [3:41] >> There's a $50 billion acquisition on the [3:42] table. [3:43] >> They've they've agreed to sell for 60 [3:45] billion and if they don't sell, they're [3:47] going to get 10 billion [3:48] of non-diluted capital. [3:50] >> Absolutely wild. [3:51] >> Anyways. [3:52] >> What a what a what a wild wild turn of [3:54] events. [3:55] >> Uh anyway, let's go through the timeline [3:57] because it's it's interesting to revisit [3:59] the flurry of news that has come out of [4:02] Elon in Inc. over the past just 4 years [4:06] when this all started. April 14th, 2022, [4:10] it's April 22nd now. So, we're talking 4 [4:13] years to go from just proposing buying [4:17] Twitter. He made an unsolicited offer to [4:20] acquire the company. He closed that [4:21] acquisition after a bunch of back and [4:23] forth saying, "Hey, maybe I don't want [4:24] it." It was very clear that the stock [4:26] would have traded down significantly [4:28] from that $44 billion [4:29] because there this was when the interest [4:32] rate hike happened and the end of ZIRP [4:35] and basically all software companies [4:38] sold off significantly. And so, we saw [4:40] declines in Snap and Pinterest and [4:42] basically anything that was [4:44] even meta sold off like 40 50% post [4:48] Although, that was like an anomaly and [4:50] they built back up, but the question [4:52] about like, "Okay, well, Reality Labs is [4:55] a [4:56] what? A $10 billion tens of billions [4:58] dollar bet on on revenues that might [5:01] come in like 10 years." Like, it was so [5:03] far away and the revenue ramp on VR and [5:06] metaverse projects was so slow that the [5:09] market just had to discount those future [5:11] revenues even if they were still bullish [5:13] on the idea of meta winning the VR race, [5:15] winning the metaverse at some point. It [5:17] just wasn't going to happen anytime [5:18] soon. So, you got to discount that by 6% [5:20] instead of 3% or whatever your risk-free [5:22] rate is in your DCF. And so, everything [5:25] sold off and Elon had locked in that [5:27] pre-end of ZIRP price and there was a [5:30] big question about could he get out of [5:31] it? Were they going to twist his arm? [5:33] They there was a little bit of arm [5:34] twisting. The arm was a little bit [5:36] twisted, but Morgan Stanley came on [5:38] board and [5:39] >> bunch of other It was almost broken. [5:41] >> But, the people who get who were who's [5:42] arms were twisted, Morgan Stanley and [5:44] all the VC backers who came in, they [5:46] wound up with SpaceX stock. And so they [5:48] wound up doing very well cuz they wound [5:49] up with XAI stock and then SpaceX stock. [5:51] And so it's all looking like, you know, [5:53] as crazy it was it was never bet against [5:55] Elon. Like that was the thing was like [5:57] surely this is the time to bet against [5:58] Elon. He's buying Twitter for 44 billion [6:00] dollars. It's a crazy idea based on the [6:02] market and where things are, but [6:04] everything sort of panned out. So he [6:06] closed the Twitter acquisition on [6:07] October 27th, 2022. And he took control [6:10] of the platform at the end of that [6:11] month. That was sink day, of course. He [6:13] he comes in with a sink. [6:15] >> Something about the What was the joke [6:15] about the kitchen sink? [6:16] >> Let that sink in. [6:17] >> Oh, let that sink in. Okay, it wasn't it [6:19] wasn't like I'm doing the whole the [6:21] whole Isn't the whole kitchen sink [6:22] another phrase as well? [6:23] >> Yeah, like throw the whole kitchen sink [6:25] at it. [6:25] >> Okay, yeah. He did that too cuz he threw [6:27] everything together. Yes, good job. [6:30] Yeah, yeah. [6:31] So XAI came later. He publicly announced [6:34] XAI on July 12th, 2023. And that's only [6:38] nine months after ChatGPT, maybe eight [6:40] months after ChatGPT. [6:42] Google quick followed with Gemini and it [6:45] was like [6:46] it was a very fast following I think to [6:48] get to to XAI off the ground. [6:50] >> And that was like just just for context, [6:52] that was right around the time you first [6:53] publicly talked about the potential for [6:56] a tech live stream, right? [6:58] >> Oh, yes. Yes, I did. [6:59] >> 2023. [6:59] >> I think that was I think that was [7:00] actually very close to the date. Yeah, I [7:02] was making YouTube videos. I made a [7:04] whole YouTube video about the Elon [7:06] Twitter acquisition and I was sort of [7:08] trying to justify it as a desire for [7:10] free speech than anything else. Like you [7:12] shouldn't look at it in financial terms. [7:13] I think that still probably holds. [7:16] But there wound up being a whole bunch [7:17] of other knock-on effects for Elon's [7:18] strategy. He was, you know, coming out [7:21] on July 12th, 2023 saying I'm back in [7:24] the AI horse race. I'm competing [7:25] directly with the big labs. I'm going to [7:27] go up against DeepMind. Still, remember [7:29] that's why he founded OpenAI. He wanted [7:31] to push back against DeepMind. [7:33] I'm going going to go up against [7:34] Anthropic. Going to go up up against [7:36] OpenAI. And so the first major product [7:38] milestone came just a Just a months [7:40] later, November 3rd. that's when they [7:42] introduced Grok, November 3rd, 2023. So, [7:45] it's been almost 3 years. The first real [7:46] integration with Twitter happened on [7:48] December 7th, 2023 when Grok started [7:51] rolling out inside of X for premium plus [7:53] subscribers, which was a new tier. It [7:56] was great It's actually crazy how [7:58] >> Everyone got verified, huh? [7:59] >> Yeah. I mean, the the the that year on [8:02] Twitter was insane. Like, just product [8:04] feature, product feature, and they were [8:05] shipping stuff that had been clearly [8:07] developed beforehand. Like, I think the [8:09] community notes idea had been [8:11] workshopped and even like engineered in [8:13] the pre-Elon era, but he just got there [8:15] and was like, ship that tomorrow. And [8:17] like, there were a lot of things like [8:18] that that happened. And there were other [8:20] things that got pulled back and review [8:22] and like, there were other pieces of the [8:23] puzzle that were like not doing so well, [8:25] and he was just like, cut, cut, cut. And [8:27] so, he sort of like put it back in [8:28] startup mode, and it felt a lot more [8:29] agile, and it still does. Uh that was [8:31] the moment Grok finally became a part of [8:33] the X product experience. In early 2024, [8:36] xAI started shipping updates quickly. [8:38] March 17th, 2024, it open-sourced Grok [8:40] 1. March 26th, 2024, Grok access [8:44] expanded to X premium subscribers. Two [8:46] days later, they announced 1.5. Then on [8:49] April uh 12th, 2024, they announced [8:52] 1.5V, which added multimodal [8:53] capabilities. Uh then Grok 2, Grok 2 [8:56] mini. By the end of 2024, Grok was [8:58] available to all X users. And so, 2025, [9:02] uh xAI moved from being connected to X [9:04] to absorbing it at the at the product [9:06] level. On February uh 19th, 2025, xAI [9:10] announced Grok 3 beta. Then March 28th, [9:14] 2025. So, over a year ago, for some [9:16] reason this feels more more recent than [9:18] that, uh xAI acquired X in an all-stock [9:21] deal. So, March 28th, 2025, uh [9:24] effectively merging the AI company with [9:25] the social platform. Uh after that, X [9:28] kept releasing faster model updates. [9:30] They did Grok 4 fast, Grok 4.1. Then [9:32] came the SpaceX deal. That was February [9:34] 2nd, 2026. So, the clean sequence is [9:37] that Musk first proposed the Twitter [9:38] acquisition, then founded XAI, then [9:40] brought Grok into X, then merged XAI [9:42] with X, and then SpaceX bought XAI. And [9:45] so, uh now Cursor is joining the team. [9:48] >> It makes so much sense, right? Cursor [9:50] needs compute, they need the resources, [9:52] they need the capital to train a [9:54] frontier coding model. [9:56] >> never done a pre-train, believe, whereas [9:58] the XAI team has, right? Correct. [10:00] >> and the the big big thing is that the [10:03] Grok brand has, you know, been through [10:06] so much. There was uh an idea being [10:09] thrown around last year that it had been [10:10] banned in more workplaces than it had [10:13] been adopted, right? So, more people had [10:15] said like, "You cannot use this product [10:17] in the workplace than we're actually [10:19] using it in the workplace." Cursor's a [10:21] great brand, right? It's a It's a brand [10:23] that I think can probably expand outside [10:25] of outside of coding, right? They say in [10:28] the announcement [10:28] >> And you [10:29] >> best and [10:30] uh to create the world's best coding and [10:32] knowledge work AI. [10:33] >> Yeah. [10:33] >> So, we're at a point right now where [10:34] everyone is building the exact same [10:36] thing. We got everyone on Earth building [10:40] building a box that you can tell to do [10:43] things and it does things. [10:44] >> so you just want one company to do it [10:46] all? Communist? Are you a communist? [10:47] >> No, no, I don't. I think I think I think [10:49] the competition I think the competition [10:51] is great. [10:51] >> You want the government to have [10:52] >> It's bringing out the best in It's [10:54] bringing out the best in the [10:55] >> Knockout, [10:56] >> Cursor's always had a great brand. [10:58] >> Wait, he said Wait, SBF replied to [11:01] SpaceX's tweet. So, he's like, "That's [11:02] me." [11:03] He [11:04] Pull up Pull up the first post here and [11:06] then scroll down. The SpaceX and Cursor [11:08] are now working closely. Yeah, there you [11:10] go. Look, SBF is is is somehow on [11:13] Twitter. How can you be on Twitter from [11:15] jail? [11:16] >> All these people in jail on on X, I [11:18] don't understand it. [11:18] >> No, it says it in his bio. It says it in [11:20] his bio. He says, [11:22] uh "We can use BOP-approved phone calls, [11:25] emails to tell others what to post on [11:27] our socials." Right? I think it would be [11:30] uh free speech violation to say like, if [11:32] you are in jail, you cannot distribute [11:35] information in the outside world in any [11:37] capacity, but it feels like this was [11:39] maybe some legislation that was created [11:41] pre-internet because it seems like [11:44] having like a ghostwriter on your [11:45] account is a really good way to like [11:48] start, you know, kind of like [11:49] influencing public perception, right? I [11:51] think this is Elizabeth Holmes has had [11:54] enough kind of moments where it was like [11:57] she, you know, her proxy posted [11:59] something that was mildly kind of [12:02] entertaining and over time that just [12:04] wears on people and eventually people [12:06] are like, maybe maybe she isn't so bad [12:08] after all, right? [12:09] So. [12:10] >> Well, Richard Woo broke down the [12:11] structure of the deal. He says the [12:13] structure of the deal is pretty [12:14] interesting here. I think what's [12:15] happening is, one, xAI is having trouble [12:19] training a state-of-the-art coding [12:21] model, hence co-founder departures. They [12:24] might have a bunch of idle GPUs. Cursor [12:27] doesn't have capital to blow on a $5 [12:29] billion training run to compete with [12:31] Codex and Claude. [12:32] xAI three, xAI says to Cursor, [12:35] "Use all the GPUs you want at cost and [12:39] get to a state-of-the-art coding model [12:40] as long as we have the option to buy [12:42] you." [12:43] Four, Cursor also gets a free option. [12:45] Train a model better than Opus and get [12:47] bought out for 60 billion or get 10 [12:49] billion that pays for all the GPUs you [12:52] you rented. Win-win. And I was [12:54] reflecting on [12:56] Michael Troughton's, I mean, [12:58] fantastic entrepreneur, [13:00] uh, remarkably young, [13:02] but his aesthetic is is very much in [13:05] like the Stripe world, I imagine, [13:07] because I feel like the one really [13:09] cinematic podcast he's done is that one [13:11] with, was it Patrick Collison? Is that [13:13] the one I'm thinking of? And it's like [13:14] them at the coffee shop and it's so [13:16] welcoming and warming. It's like that is [13:18] welcome in every corporate entity in [13:20] America, right? Like and to your point [13:22] about like, you know, oh, like if your [13:24] corporation is like, oh, what's going on [13:26] with Elon and politics, blah, blah, [13:27] blah. Do we really want Grok running [13:29] around with Mecca and Annie and all this [13:31] crazy stuff? But, if it's just like [13:33] Michael Treu from Cursor, he's so [13:34] reassuring. Like, I if I'm like Coke or [13:37] GE or Ford, I'm like, yeah, of course, [13:40] like we love Cursor. That's great. Very, [13:42] very like aesthetics do matter, vibes do [13:44] matter, and I think makes a lot of [13:45] sense. [13:46] >> Matt Slotnick says he loves math. [13:50] >> Okay, what is the math here? [13:51] >> What he's talking about Julian. He's uh [13:53] So, Julian says at a 30x revenue [13:55] multiple, at first glance it appears [13:56] that SpaceX is overpaying for Cursor. [13:58] However, the deal is wildly accretive [14:00] for SpaceX, given it is expected to go [14:03] public at more than a 100x revenue. A [14:05] 70-turn multiple expansion on 2 billion [14:08] of revenue adds up. So, [14:10] >> Wow. [14:10] >> yeah. [14:11] >> This is a notorious thing in corporate [14:12] M&A, and uh and occasionally it's gone [14:15] poorly. Um but, there is like an old [14:18] adage that if you can acquire earnings [14:20] at a lower earnings multiple and [14:22] maintain your current multiple, that is [14:24] accretive to stock, and so there's a [14:26] whole philosophy around that. But, [14:28] uh the smart investor should price each [14:31] earning stream differently, and uh [14:33] having like a monopoly on launch [14:35] capacity should be a higher multiple [14:37] than a super competitive oligopolistic [14:40] coding market, if that's what this winds [14:42] up being. It's a tricky, tricky [14:44] situation. [14:44] >> Well, Brian, wow. Cursor just hit a $10 [14:46] billion run rate. [14:48] >> They did. It's guaranteed, right? [14:49] There's no way that they won't make $10 [14:50] billion this year. That's not even run [14:52] rate, that's just like it's locked in. [14:53] It's more than contracted. I don't know. [14:55] >> Yeah, and then and and it's very [14:57] meaningful because it is effectively an [15:00] exit in that it's, you know, by by [15:03] itself. They can you know, they they'll [15:06] be in a position to actually, you know, [15:08] reinvest that. That's like a a nice [15:10] little Neo Lab kind of series A, [15:12] basically. [15:13] >> Ken says no. You got to multiply that 10 [15:16] billion by 12, because in that month [15:18] where they get the $10 billion check [15:20] from SpaceX, it will be a $120 billion [15:22] run, right? They will be the largest AI [15:24] company in the world that month. [15:27] Uh and then they just have to figure out [15:29] uh they don't collect until end of year [15:31] if they're still independent. Otherwise, [15:32] it's more like 25 billion. Okay. [15:34] >> Patrick is quoting Scott Wu. [15:36] >> Oh yeah, yeah, last of the Mohicans. [15:37] Scott Wu. I love it because uh despite [15:40] this does not look [15:41] >> images too, without images too, we just [15:44] this this kind of adds that. I actually [15:46] think this might have been done the [15:47] old-fashioned way now. It's just a [15:48] banana. [15:49] >> This is dinner banana. [15:50] >> There we go. [15:51] >> It's going to cutting out faces now. [15:52] >> watermark. [15:53] >> He says, here we go again, last of the [15:55] Mohicans. Uh yeah, Scott Wu is the last [15:57] one standing in this particular [15:59] category and uh yeah, he's licking his [16:02] chops thinking about uh who might want [16:03] to hop over to uh Cognition Cursor with [16:07] uh Cognition Wind Surf, Devin Cursor, [16:09] Zombie Corp as the uh as the ghost ships [16:13] continue to line up on the shores of [16:15] Scott Wu's uh territory. [16:18] >> Uh some news from Space News Inc. [16:21] >> This is very interesting. [16:21] >> China backs orbital data center startup [16:23] with 8.4 billion in credit lines. [16:25] >> They're peeled. They're peeled. [16:26] >> Elon is interested. [16:27] >> 8.4 billion in credit lines, that is a [16:29] lot of money. A Beijing-based start [16:31] space startup has secured early-stage [16:33] funding. Early-stage funding to the tune [16:35] of 8.4 billion, what are we doing here? [16:38] Uh [16:39] and extensive credit backing is part of [16:40] a broader Chinese push towards [16:41] space-based computing infrastructure. [16:43] Beijing Orbital Twilight Technology [16:46] Company Limited. [16:49] We don't know how to name companies like [16:50] that in this country. [16:51] >> Seriously, this is the new meta. [16:52] >> Yeah. What is uh what is Cursor's real [16:55] name? Any Sphere? That's a that's a [16:57] pretty good name. I like I like Any [16:58] Sphere and I like Cursor, but doesn't [17:00] hold a candle to Beijing Orbital [17:02] Twilight Technology Company Limited, [17:04] also known as Orbital Chunguang, [17:07] announced the completion of a pre-A1 [17:10] funding round April 20th. The round saw [17:12] participation from venture and [17:14] industrial investors including Hysan [17:16] Capital, CITIC Construction Investment [17:19] Capital, Cathay Capital, they got [17:20] everybody. It's a murderers row over [17:22] there. At the same time, Orbital Estate [17:25] said that it has obtained strategic [17:28] credit lines totaling 8.4 billion from [17:32] 12 major financial institutions [17:33] including the Bank of China. Are they [17:35] going to build rockets with this? I mean [17:37] I don't I I don't understand why this [17:38] would be so capital intensive if they're [17:40] not going to fab the chips and they're [17:41] not going to launch the rockets. [17:42] >> Well, they're going to put a lot of GPUs [17:44] in space. [17:45] >> Cuz they're just going to buy a lot of [17:46] Huawei chips or something? Are they [17:47] going to go to Huawei and get some [17:48] special stuff? [17:49] >> mean this is pretty interesting because [17:50] I feel like, you know, generally the [17:52] main bull case for space data centers is [17:54] that the like regulatory environment in [17:55] the US is going to be so hard to build [17:57] data centers. [17:57] >> Yeah, yeah. [17:58] >> Like supercomputers, but yeah. [17:59] >> Normal latency. Oh, yeah. Yeah, [18:01] supercomputers. So, you got to send the [18:02] supercomputers to space because there's [18:04] less regulatory [18:05] >> data centers in space. We just we we if [18:07] it's on the ground it's a supercomputer. [18:09] >> Okay, yeah. That makes sense. [18:10] >> Because we don't want any more data [18:11] centers on Earth. It's Earth Day, by the [18:13] way. Congratulations, Earth. You've done [18:15] fantastically. Uh [18:17] >> But but [18:18] the whole thing like in China is like [18:19] they you know, you can just like build [18:20] things there, right? There's like very [18:21] little regulatory overhead. So, it's [18:23] like I I feel like the space data [18:24] centers makes a lot more sense for the [18:26] US. [18:26] >> Yeah. [18:27] >> If the if the bull case is regulatory, [18:29] which it seems to be, that's like [18:30] generally consensus, I think. [18:31] >> I was laughing about how you know how [18:33] there's all this this fear about like [18:34] data centers using water in America, but [18:36] over in China they have the Three Gorges [18:37] Dam, which generates electricity from [18:40] water. And it technically uses a lot of [18:42] water. The water's not destroyed or [18:44] anything. It just passes through the [18:45] dam, generates electricity. [18:46] >> They're using the entire ocean of water. [18:49] >> Exactly. The Hoover Dam, that's a [18:50] nightmare. If you just like there's one [18:52] frame where you don't want the water to [18:54] be destroyed or made unpotable, so you [18:56] can't drink it anymore. But there's [18:58] another one where you think like water [19:00] has individual rights and should not be [19:02] used at all. Like it should not be used, [19:04] it should not flow through a water [19:06] wheel, it should not flow through a dam [19:08] and generate electricity at all. It [19:10] should be left to its own devices, [19:12] still, and just chill. [19:13] Maybe. Anyway, Orbital has incubated was [19:17] is incubated by the Beijing Astro Future [19:18] Institute of Space Technology, which [19:20] itself is backed by Beijing's Municipal [19:22] Science and Technology Commission and [19:25] the Science Park Administration. The [19:26] institute leads a consortium of 24 [19:27] organizations. The rationale for the [19:29] constellation in November briefing. So, [19:32] large-scale data centers have expanded [19:33] rapidly worldwide, but further growth [19:35] faces faces major obstacles, including [19:38] heavy land use. Weird. Soaring energy [19:41] consumption. Weird. Limits on [19:43] atmospheric cooling. These things [19:44] usually don't apply in China, but maybe [19:46] they are skating where the puck is going [19:48] and maybe they're thinking that they [19:50] will need to change their direction. [19:52] They're planning a dawn-dusk orbit about [19:54] 700-800 km above the Earth, aiming to [19:56] achieve a large-scale space data center [19:59] to support space-based computing by [20:00] 2035. Wow, thinking in decades over [20:03] there. The initial phase, spanning 2025 [20:05] and 2027. Wait. Why are they talking [20:08] about last year? [20:09] Uh we'll focus on core technology [20:11] changes in a first computing [20:12] >> They're manipulating time, y'all. [20:14] >> The experimental satellite was slated [20:16] for launch in late 2025 or early 2026, [20:19] but it does not appear to have launched. [20:20] Uh they have Uh [20:23] they have like decent launch capacity [20:24] with the Long March rocket. I think that [20:26] they just are not amazing at landing it, [20:28] but uh you got more money, you can just [20:31] yolo more rockets up there, I guess. [20:32] Problem solved. Although uh I would be [20:34] surprised if there's as much pressure to [20:38] not build supercomputers next to the [20:41] Three Gorges Dam. Uh moving on. More [20:43] images out of ChatGPT images to Image [20:46] Gen 2 [20:48] images V2. These are crazy. Has anyone [20:51] done that Rune test of the of the the [20:53] the guy driving the car into the what [20:56] you you know, the whole thing where he [20:57] was like it it it it's AGI if it can [20:59] understand this meme. I guess you can't [21:01] because like that particular meme has [21:02] been uh saturated on the internet and so [21:05] it it no longer is a challenge. This [21:08] is a crazy image. I am a I am a tiny [21:10] man. [21:11] >> What is this? [21:13] My space. [21:14] >> My son was born. [21:16] >> This is truly never [21:18] never forget where you came from. Blink [21:20] 182. It really packed so much stuff in [21:23] here. It's almost too much. I have been [21:24] noticing Gabriel over there is pushing [21:26] the model into chaos and seeing what [21:28] happens. Cows are flying, horses are [21:30] flying, trampolines are flying. [21:32] >> The next one by Ethan Mollick has like [21:33] every image batch mark [21:35] >> Yes. You were you were asking about [21:37] this. You were you said can and explain [21:39] the history of this. [21:40] >> It's it's originally like when when like [21:42] Dolly first came out, the the like image [21:44] that everyone was like, "Oh, this is [21:45] crazy." It was an astronaut riding a [21:47] horse on the moon. [21:48] >> Yeah. [21:48] >> So then it's like, "Okay, that's like [21:49] very easy." And then it became [21:51] >> Well, it's very hard to Photoshop that. [21:53] It's very hard to go get an actual [21:54] picture. Because wait, if someone's just [21:56] like make a picture of a dog and it's [21:57] photo real, everyone's like, "Oh, who [21:59] cares? We have a picture of a dog." We [22:00] don't have any pictures of astronauts [22:02] riding horses on the moon. [22:04] >> people riding horses, right? And so like [22:06] >> It should be able to process [22:07] >> where it's like maybe that it's not [22:08] doesn't need to fully understand like [22:10] everything that's going on because there [22:12] there's a lot of like references it can [22:13] pull from. [22:13] >> Totally. [22:14] >> So then it became can um a horse ride an [22:17] astronaut? [22:17] >> Yes. [22:18] >> Right? It's like the reverse. And early [22:19] image models would always just do the [22:21] flip. They would just put the astronaut [22:22] on the horse. [22:22] >> Yep. Cuz that's more logical. Generally. [22:24] >> And so so there's a number of a number [22:26] of these things where like if you ask a [22:27] person to draw it out, it's like, "Okay, [22:28] you just think logically, like this is [22:29] how it would look." But there's just [22:30] like no reference images online. So then [22:32] another another one was like a full wine [22:34] glass. [22:34] >> Yeah. [22:35] >> So you think of a full wine glass is [22:36] like it's still only like technically, [22:37] you know, halfway 3/4 full. [22:39] >> understood why that one didn't work. I I [22:41] I saw the whole video for [22:42] >> There's no human images online of a full [22:44] wine glass being like [22:45] >> Oh, full to the brim because people [22:46] typically fill them halfway. And so in [22:48] the training data, okay, got it. Yeah. [22:50] >> Yeah, and then there was the, you know, [22:52] >> What is this image? [22:53] >> That's not the one. [22:54] >> This is not the one. [22:55] >> This is not the one. [22:56] >> not the one from the timeline. That [22:58] one's way more aggressive than [23:00] than the This is the one. This one looks [23:02] aesthetic. And it still looks weird. [23:04] >> only The only thing is would a horse's [23:06] belly really look like that? [23:08] >> It is sort of like a humanoid horse. [23:09] >> It kind of looks like a human that [23:13] >> But it does have the wine glass [23:14] completely filled. And and then explain [23:17] the [23:17] >> pulled up was was Grok, I think. [23:19] >> Okay. Explain the Explain the clock. [23:21] What's going on with the clock? [23:22] >> Yeah, that one's just another like [23:23] understanding thing where if you give it [23:25] a time, can it do the the clock like the [23:27] correct time? There's also one where you [23:29] do a bunch of clocks on different time [23:31] zones. [23:31] >> Yeah. [23:32] >> And [23:32] >> And it couldn't do the clock. It [23:34] couldn't do the clock. It has two hour [23:36] hands. Do you know what that means? [23:38] >> We're moving the goal posts. [23:39] >> Do it. [23:40] >> Moving the goal posts. [23:41] Goal posts. [23:43] We We all know that [23:44] >> Do it. [23:45] It's time. [23:46] It's time. [23:50] >> We got to We got to move the goal posts. [23:51] It's It's 99% of the way there, [23:55] but it's not [23:56] It's not complete yet. [23:57] It's not complete yet. [23:58] >> You guys have more work to do. [23:59] >> No, no. Sam was talking about this with [24:01] Andrej Karpathy. He was saying that like [24:03] you know, he thought he thought the job [24:04] was finished with like chat GPT images [24:07] because it was really good. And then [24:09] they worked a lot harder and they [24:10] realized that there was more to do. And [24:12] this is like the Karpathy thing about [24:13] like yeah, you get the self-driving cars [24:16] to 99% and then takes another year to [24:18] get add another nine and then takes [24:19] another year to add another nine and [24:21] then another nine and it just takes [24:22] forever to improve these things because [24:24] we demand perfection. We do not accept [24:27] two hands on a clock. Keep grinding, [24:29] folks. Keep grinding. Anyway, um there [24:32] are a lot of other fun posts. Uh Sam [24:34] analysis, I thought this was just a real [24:36] image of Dorkhastagging a selfie and [24:38] then it was recontextualized via a meme. [24:40] It is in fact an AI image. Um and it And [24:44] it [24:44] >> even look like the new model, though. [24:45] >> Hi, I'm Dorkhast. I grew up all over the [24:47] US and now SF based and always down to [24:49] nerd out about AI science and history. A [24:51] little about me. I host the Dorkhast [24:54] podcast, study at UT Austin, just [24:56] published a book on the history of AI AI [24:58] scaling. Let's grab coffee or do a fun [25:00] activity this summer. Because of course [25:02] this is a a meme template that is very [25:05] popular and has gone viral with a lot of [25:07] new people that have been hired and [25:09] moved to San Francisco. [25:10] >> the other thing I've noticed with [25:11] images, too, is that I think it has [25:14] fully the new model has actually has [25:17] taste. [25:17] >> I agree. [25:18] >> And it has fully democratized like [25:22] high-quality lifestyle and product [25:25] Midjourney and I'm shedding [25:28] a tear just a little bit [25:30] >> Yep. [25:30] >> because you used to be able to tell [25:33] somebody was working on a CPG company. [25:35] You could kind of clock their ability [25:36] >> Yep. [25:37] >> as a founder based on the quality of [25:39] their images, right? Cuz maybe they're [25:41] not super creative themselves, maybe [25:42] they haven't raised a bunch of money, [25:43] but if they're scrappy, they can meet [25:45] the right photographer, they can put [25:46] stuff together. And now it's like [25:47] everyone just has a great product [25:49] photography. [25:50] >> Yep. [25:51] >> And I saw this morning with there's that [25:54] Andreesen back company that does like [25:56] electric scooters. [25:57] >> Oh, yeah. [25:57] >> They were just like ripping out a bunch [25:59] of images that look like they spent tens [26:01] of thousands of dollars doing them, but [26:03] they're clearly like Midjourney's, too. [26:05] Which is cool. It's just like funny that [26:07] we've entered a world where [26:09] >> Yeah, I've seen I've seen some people I [26:10] I I think that the the end result will [26:13] be like you will see more opinionated [26:15] and creative and [26:17] and more people like like there will be [26:19] a collapsing around like everyone will [26:20] copy Apple or linear, but then on the [26:23] flip side you will see people that are [26:25] that are doing things that are really [26:27] unexpected and those things will be [26:29] copied, but if the if the brand can run [26:32] through and actually establish itself as [26:34] like oh, it has this unique aesthetic, [26:36] it won't matter that somebody can can [26:38] recreate it. Cuz plenty of people did [26:39] that with the Red Antler stuff. It was [26:40] like the like there were brands that [26:43] really own that and carve that out as [26:45] like their aesthetic and then there were [26:46] a bunch of copycats and no one really [26:47] liked those. [26:48] I was reflecting the fact that like for [26:50] a long time Midjourney felt like it had [26:51] a very unique aesthetic, like [26:53] uh like ArtStation and like painterly [26:56] and sci-fi really well. And then ChatGPT [26:59] images V1 and some of the other image [27:01] models just felt like stock photography. [27:03] And now I feel like I'm starting to see [27:05] more stuff out of images with ChatGPT [27:07] that feels more opinionated and has a [27:09] stronger aesthetic. And it can do more [27:12] of like the sci-fi stuff. Although, [27:14] sometimes it leans in a little bit too [27:15] much to the photorealism. I think you [27:17] got to get kind of crazy with the [27:18] prompts to actually get something that's [27:20] abstract. [27:20] >> Post from Ben Hilack on the topic. Says [27:23] "Nightmare blunt rotation." [27:26] >> Oh, yeah, it can do it can do 360 [27:28] images. It can do 360 images, which is [27:30] really cool. [27:31] >> Sam Bankman-Fried. [27:33] >> So, it generates a full equirectangular [27:36] image, which is something you could view [27:37] in VR. And then And then you can load [27:40] that into a [27:43] panoramic, you know, stereoscopic image [27:47] generator. The big question is like [27:48] people are going to want to animate [27:49] these. So, what what what's the [27:51] downstream tool chain for actually [27:52] turning this into video? Are we going [27:53] back into video at some point? We will [27:55] see. Anyway, people are having a lot of [27:57] fun with it. And [27:59] it should should help with like [28:00] front-end and and a lot of other stuff. [28:02] It seems like it's done very well on [28:05] on on on design and layouts and all this [28:07] stuff. Well, oh, I wanted to get your [28:08] take on this. Uh [28:10] Uh imagine if Codex existed in 2005. And [28:15] it's this very retro people are doing [28:16] imagine if Codex existed in 2012 and [28:18] it's the it's the Apple design cues from [28:21] the early Mac OS 10. And then [28:25] the Windows XP aesthetic. Do you think [28:28] the Windows XP aesthetic is overdue for [28:32] a comeback? This is in the middle of the [28:34] GPT [28:35] image to [28:38] feed right there. This aesthetic. [28:41] >> It's got to come back. This and a [28:42] Mountain Dew. [28:43] >> It would fix you. It would fix you. [28:45] >> This would fix me. No, uh you know, I [28:48] have some fond memories of Texas. [28:49] >> John John. Imagine playing Halo while [28:54] you just look over and this is just [28:56] running there. [28:57] >> I don't know. [28:57] >> You're just drinking a Mountain Dew. [28:59] >> There are a couple other important uh [29:01] warnings in here. One of them, I know [29:04] you had thought about potentially [29:07] putting on a bear costume and trying to [29:09] defraud your car insurer. And if you in [29:13] the audience were thinking about [29:14] dressing up like a bear and attacking [29:17] your car in order to collect a insurance [29:20] payout that's up to $141,000, [29:23] think again. Think again because humans, [29:26] this is in the New York Times, humans [29:27] who used a bear suit to defraud car [29:31] insurers were sentenced to jail. You [29:33] >> Good. [29:33] >> might be thinking, "I I'll just throw on [29:35] this bear suit and I'll attack my car [29:37] and then I'll call my insurer and I'll [29:39] collect a check and it'll be easy easy [29:41] money. Easy money get rich quick [29:42] scheme." [29:43] >> No. [29:43] >> No. [29:44] >> Three Southern California residents were [29:46] sentenced to jail after masterminding a [29:48] scheme in which they staged fake bear [29:50] attacks on their luxury cars, then [29:52] collecting more than $141,000 in [29:54] insurance payouts. To carry out the [29:56] attacks, the residents had a person in a [29:58] bear suit climb into the cars and use [30:00] claw-like kitchen essentials to leave [30:02] scratch marks. The Los Angeles residents [30:04] then filed claims to defraud three [30:06] different insurance companies. So, I [30:08] don't understand why they had to be in a [30:10] bear suit. Like, were they filming? Were [30:12] they filming? [30:12] >> Yes, there is security footage. It's in [30:14] the car. [30:15] >> Okay. [30:15] >> Okay, we're going [30:15] >> No way. [30:16] >> Yeah, yeah, we're going to pull up the [30:17] the security footage. [30:18] >> Yeah, on that Rolls-Royce Ghost. [30:20] >> Yeah, a 2010 Rolls-Royce Ghost. I know [30:23] someone who has one of those. I don't [30:26] think he's responsible for this, but uh [30:30] dressing up like a bear to get into the [30:33] car and rip everything and destroy it so [30:36] that you can file an insurance claim. [30:40] >> Hey, I'm just a bear. I'm going to just [30:42] get in this car. [30:43] >> The schemes are really out of control [30:46] these days. That is a wild, wild one. [30:48] The California Department of Insurance [30:50] began its investigation called Operation [30:52] Bear Claw [30:54] after an insurance company flagged a [30:55] suspicious claim about a about a bear [30:58] rifling through a 2010 Rolls-Royce Ghost [31:01] in Lake Arrowhead, California, where [31:02] there are in fact bears. So, it's not [31:04] that unreasonable. The defendants [31:06] claimed that a that the bear had damaged [31:09] the interior of their Rolls-Royce with [31:11] photos they submitted to the insurance [31:12] company showing scratch seats and doors. [31:15] Video footage submitted with the claim [31:16] and released by the department shows [31:18] what appears to be a bear crawling [31:19] around in the back seat of the [31:20] Rolls-Royce and swatting at the [31:21] dashboard. Investigators later [31:23] discovered fraudulent claims submitted [31:26] to two other insurance companies that [31:28] claimed a bear had damaged the same [31:29] interior [31:30] of two Mercedes-Benzes. [31:32] >> Just scratching everything. [31:33] >> That is like the most perfect scratch [31:35] line. Like it's so suspicious. Like it's [31:38] so clearly like I feel like a bear's [31:40] claws would be way more organic and like [31:42] it's like very suspicious. [31:43] >> doing more damage. That looks like the [31:45] weakest bear ever. [31:47] >> Yeah, because it's not a bear, it's a [31:48] human. The department concluded that the [31:50] culprit was not in fact a particularly [31:52] nimble bear. The defendants were [31:54] arrested in November of 2020. [31:56] >> So, I would I would my my first thought [31:58] is not super believable that the bear [32:00] walks up to the Rolls-Royce Ghost, opens [32:03] the door, and just goes around. [32:05] >> But, few weekends ago, I was hanging out [32:08] at [32:09] uh former guest of the show's house. He [32:11] has goats. Friend drives up in his car, [32:14] parks, and leaves the door open. Few [32:17] minutes later, what's in the car? [32:18] >> A bear. [32:18] >> A goat. [32:19] >> People do odd things with animals. Uh [32:22] people dress up as bears. Scientists [32:24] also apparently have been giving salmon [32:26] cocaine. I have no idea why, but uh the [32:31] New York Times has a post, "These salmon [32:33] got high on cocaine. That wasn't the [32:35] craziest part." Scientists in Sweden [32:38] made an unexpected discovery. I think [32:40] Tyler read this piece and can spoil the [32:43] ending because [32:44] >> What are they doing over there, Tyler? [32:45] >> unexpected [32:46] >> I think I think they tested like what [32:47] happens when you give the fish cocaine. [32:49] >> Okay. [32:49] >> And I think the answer was that they [32:50] swim much faster and they can swim for [32:52] longer. [32:53] >> Okay. [32:53] >> So [32:54] >> And that was unexpected? [32:55] >> Imagine being the scientist that like [32:57] didn't get a grant for their research. [32:59] >> Yeah. [32:59] >> You know, and they were working on [33:01] something. [33:01] >> I'm really close to breaking [33:02] Alzheimer's. [33:03] >> and they're reading this article and [33:04] they're like, wait, like I was going to [33:06] do Alzheimer's research and then these [33:09] scientists just wanted to give hard [33:11] drugs to fish. [33:12] >> Nearly twice as fast. [33:14] >> Nearly twice as fast. I guess that's [33:16] somewhat interesting, but this does feel [33:17] like something that could be one shotted [33:19] by LLM. [33:20] >> There's no way There's just no way you [33:21] could have [33:22] >> You had to prove it. [33:23] >> to prove it for science. [33:24] >> This is a good quote. While it's unclear [33:26] if swimming faster and further while [33:28] under the influences harm these fish, [33:30] experts say it's probably not great. [33:33] Speaking of [33:35] >> What else you got? [33:35] >> things under the sea, Red Lobster brings [33:38] back endless shrimp for a limited time [33:40] starting April 20-ish. [33:43] >> Luke Metro says this is an EA cause area [33:46] hostile takeover of Red Lobster. Make it [33:49] vegan. [33:50] >> Make it vegan. [33:51] >> Make it vegan. [33:53] >> No, they got to they got to make a play [33:54] to make it come back and this seems like [33:56] something that customers were clamoring [33:58] for. So I don't Have you ever been to a [34:00] Red Lobster? [34:01] >> Nope. [34:01] >> I don't think I've ever been to Red [34:02] Lobster. I've done the Outback [34:04] Steakhouse. [34:04] >> There are things in life that I don't [34:06] think I will ever do. [34:07] >> No. [34:07] >> And Red Lobster, I believe is one of [34:09] them. Ben, have you been to Red Lobster? [34:11] >> I've been there. [34:12] >> Yeah? [34:12] >> You What's your [34:14] >> They're big in Minnesota. One to 10. [34:16] >> One to [34:17] Yeah, one to 10. It's fine. I I think [34:18] they're all gone now. [34:19] >> What What What One to 10? What What are [34:20] you rating it? [34:21] >> Like a five. [34:21] >> Like a five? [34:22] >> before it was it was [34:23] >> No fives. No fives. You got to pick a [34:25] side. [34:26] >> Four. [34:27] >> Four. [34:27] >> Four. It's okay. [34:28] >> Subpar. [34:30] It's sub five, in fact. [34:31] >> Going flash back. [34:32] >> We love you. [34:33] >> Goodbye. [34:34] >> See you tomorrow.