OpenAI Astra Solves Math Open Problems as AI Solopreneurs Triple

OpenAI Astra攻克数学未解难题，AI一人企业数量激增至三倍 AI推理突破与自主智能体工作流正在推动单人运营企业急剧增长，这类企业在无需雇佣员工的情况下实现了数千万美元的营收规模。 主持人 John Coogan 和 Jordi Hays 深入探讨了近期围绕 YouTube 创作者 Hank Green 的网络舆论风波，此前观众指责他使用 ChatGPT 调研并撰写视频脚本。随后节目转向重大 AI 里程碑事件，包括 OpenAI 内部的 Astra 模型攻克了十个复杂的数学公开难题，并反映了人类数学家所感受到的存在主义焦虑。最后，主持人分析了在 AI 智能体驱动下、实现数百万美元营收的单人软件公司的迅速崛起，以及 Mark Zuckerberg 针对 Meta 全栈 AI 基础设施的长期战略。 知名科普博主 Hank Green 因在视频调研流程中使用 ChatGPT 遭遇网络声讨，部分批评者误指责其在镜头前直接照读 AI 生成提示词，引发关于科学传播中 AI 伦理的激烈讨论。 (0:00) OpenAI 研究员 Noam Brown 宣布其 Astra 模型家族的内部版本已成功攻克数学、量子复杂度和理论计算机科学领域的 10 项重大未解难题，引发研究界对其能否推动 AGI 的争议。 (2:30) 新西兰数学家 Kirwin Hampshire 发表文章《数学的至暗之夜》，生动反映出在 AI 逐步攻克纯数学难题背景下，人类数学家所经历的深层存在主义危机。 (9:50) Stripe 与商业登记数据显示单人百万美元级企业数量激增，创业者 Alex Broca 借助 AI Agent 自动化编程、客服与运营，独自支撑起拥有上万付费用户、年入千万美元的软件业务。 (15:00) 扎克伯格为 Meta 庞大的 AI 算力与基础设施支出辩护，强调自研全栈垂直整合才能避免受制于第三方平台风险，但华尔街对其资本支出及短期商业化变现仍存顾虑。 (30:00)

TBPN hosts John Coogan and Jordi Hays examine the widening cultural and economic friction caused by artificial intelligence, contrasting public backlash over creator Hank Green's research workflow with OpenAI's mathematical milestones, the surge of solo-run software businesses, and Meta's costly infrastructure bets. AI Backlash in Science Communication (0:00) Veteran educational creator Hank Green faced widespread social media criticism after using ChatGPT as a research aid for an episode of Ask Hank Anything covering the Bouba/Kiki effect. Critics pointed to his on-screen line, 'I appreciate the pushback,' claiming it was raw chat output read off a teleprompter, though Green clarified it was an edited response directed at his guest. While Green acknowledged using ChatGPT to locate papers, summarize PDFs, and convert units under an existing corporate AI policy, the intense scrutiny led him to reconsider his publishing pace. Coogan and Hays argue that demanding science communicators completely avoid AI creates a contradictory standard, particularly because modern scientific fields are deeply integrated with the technology. However, the hosts acknowledge legitimate viewer concerns highlighted in current debates, including diminished audience trust and documented instances of AI hallucinations corrupting academic literature. Frontier Reasoning Models and Academic Disruption (2:30) OpenAI researcher Noam Brown announced that an internal version of the upcoming Astra model family solved 10 open problems spanning mathematics, theoretical computer science, and quantum complexity. While AI skeptic Gary Marcus argued that solving formally verifiable benchmarks does not equate to solving open-world problems, AI researcher Daniel E.E. countered that critics continually move the goalposts for machine intelligence. The breakthrough highlights growing anxiety among academic researchers, exemplified by University of Auckland mathematician Kirwin Hampshire's viral essay, The Dark Night of Mathematics, which expressed grief over machine capabilities surpassing human mathematical intuition. The hosts note that as automated models tackle long-standing conjectures like P vs. NP or Navier-Stokes, human researchers will likely pivot toward translating abstract mathematical solutions into applied fields like chemistry and battery science. The Rise and Fragility of AI Solopreneurs (15:00) AI automation is facilitating a wave of highly profitable, single-person software enterprises. Stripe data reveals that the count of solo operators making over $1 million doubled between 2023 and 2025, while solo firms exceeding $10 million nearly tripled. A Wall Street Journal report cited 40-year-old founder Alex Broca, whose solo software operation supports 10,000 customers and a $10 million revenue trajectory using AI for customer service, coding, and refund operations. Additionally, a Harvard Business School study led by Rembrandt Koning found that AI-focused startups operate with 25% lower headcounts. Despite prominent successes such as Claire Vo scaling a solo tool to seven figures in profit, the hosts stress significant risks. High model usage costs can erode early capital, ventures remain vulnerable to imitation because the same tools are universally accessible, and many efforts fail quickly, as seen with former corporate executives attempting solo AI-assisted consulting. The hosts also question whether some autonomous business agents represent durable enterprises or merely interactive simulations. Meta's Full-Stack Bet Against Vendor Lock-In (30:00) During Meta's earnings call, Mark Zuckerberg defended the company's aggressive capital expenditures on proprietary AI models over third-party licensing strategies like Apple's. Reflecting on Meta's past infrastructure investments, such as building two dedicated data centers to power the Reels recommendation algorithm, Zuckerberg argued that full-stack vertical integration protects against regulatory vulnerabilities and platform lock-in from external model vendors. However, market analysts including Stratechary's Ben Thompson noted Wall Street's skepticism regarding massive infrastructure spending that lacks near-term monetization clarity. The hosts also point to internal friction within Meta regarding the substantial corporate resources directed to the Meta Super Intelligence unit relative to core revenue teams. Across media production, pure mathematics, and software startups, generative tools are compressing workflows and lowering operational headcounts. However, unresolved tensions remain around the defensibility of AI micro-businesses, public trust in AI-assisted education, and the financial sustainability of frontier infrastructure investments.

本期节目聚焦人工智能对内容创作、前沿科学以及商业组织形态的深层冲击：知名科普创作者 Hank Green 因使用 ChatGPT 进行视频调研遭遇舆论声讨，OpenAI 内部模型 Astra 在理论数学领域取得重大突破，同时自动化 AI 工具正在催生大批无需雇员的“单人千万美元企业”。 Hank Green 遭遇 AI 舆论反弹 (0:00) 资深 YouTube 科普创作者 Hank Green 近期因在系列节目《Ask Hank Anything》中使用 ChatGPT 检索论文、提取引文及换算单位，招致社交媒体用户的强烈批评。争议的导火索源于某位管理者发布的视频片段，指控 Green 照本宣科念出 ChatGPT 生成的剧本，批评者甚至将他在镜头前说的“感谢你的指正”（I appreciate the pushback）视为 AI 聊天的典型套话。Green 随后公开澄清，该表述是在回应现场嘉宾的意见，因镜头剪辑导致脱落语境，但他承认曾使用 AI 辅助查找资料并总结 PDF，未曾直接让模型撰写剧本。面对舆论压力与持续的生产力损耗，Green 不得不承诺调整工作流程，并表示可能削减内容更新频率。 节目主持人 John Coogan 指出，反 AI 情绪让科普创作者陷入难以调和的困境。当前数学、材料学与生物学等领域的顶尖进展均深度依赖 AI 赋能，若创作者因迎合受众情绪而刻意回避或为使用 AI 频频道歉，最终将丧失解读现代科学前沿的能力。事实上，Green 的媒体公司早前已发布过规范使用 AI 的透明守则，但仍无法抵挡网络舆论的攻击。支持者据此反讽，批评者的要求无异于强迫从业者在现代技术环境下继续退回原始的手工检索时代。 OpenAI 数学突破与学界焦虑 (2:30) OpenAI 研究员 Noam Brown 宣布，其内部版本模型 Astra 成功攻克了涵盖数学、量子复杂性以及理论计算机科学领域的 10 项重大未决开放问题（open problems）。这一里程碑在学术界引发了激烈分歧。AI 怀疑论者 Gary Marcus 发文称，除非 Astra 能够解决与形式化验证无关的现实世界重大问题，否则不足为奇；研究员 Daniel E.E. 则反驳认为，批评者对通用人工智能（AGI）的定义标准正被无限拔高。主持人指出，尽管形式化验证任务属于特定范畴，但在验证药物是否抗癌或检测工程精度等实际工业场景中具备极高的应用价值。 机器在前沿推理领域的飞跃加深了学界的心理危机。奥克兰大学数学家 Kirwin Hampshire 在 Astra 发布前夕撰写的文章《数学的至暗之夜》（The Dark Night of Mathematics）中，记录了人类学者目睹 AI 超越自身抽象思维能力时的存在主义绝望。主持人讨论认为，传统学术界长期依赖解决定理来体现价值，但伴随千年大奖难题（Millennium Prize problems）、P 与 NP 假说或 Navier-Stokes 方程逐步进入 AI 求解视线，人类科学家的重心或将加速从纯理论证明转向电池能量密度、材料化学及生物医药等落地应用领域。 AI 赋能单人千万美元企业崛起 (12:10) 《华尔街日报》报道的案例展示了生成式 AI 正在重构商业组织结构，使极小规模团队甚至单人公司得以承接以往需要完整部门支撑的业务。40 岁的创业者 Alex Broca（亦称 Ben Broca）在无任何雇员的情况下运营软件公司，借助 AI 自动化代理处理全流程的代码编写、Debug、客户支持、邮件往来、订阅管理和退款发放，在 7 个月内吸引了 10,000 名付费用户，年营收规模奔向 1000 万美元。支付机构 Stripe 的统计数据进一步印证了该趋势：2023 年至 2025 年期间，平台上午雇员、年营收突破 100 万美元的单人经营者数量翻倍，而营收超 1000 万美元的单人企业数量更是增长了近三倍。 经济数据亦反映出就业生态的结构性转变。过去一年中，美国信息技术领域的企业注册申请量攀升近 45%，位列所有行业之首，但申请者的外部雇佣意向却创下最大降幅。哈佛商学院副教授 Rembrandt Koning 对 50,000 家初创公司的研究显示，AI 驱动型企业的平均雇员数量减少了 25%。旧金山面向单人科技创始人的孵化器近期 10 个名额吸引了 4,500 人申请。不过，单人模式同样蕴含重大商业风险：Troy Johnson 凭借单人开发的信用卡积分优化应用月入数千美元，并称 AI 如同“人手一把的王者之剑”；但前 Verizon 资深员工 Samir Ahmad 尝试依托 AI 创办咨询公司，数月内便因无法维系而重返职场。这表明软件门槛降低的同时也削弱了产品防御壁垒，创始人的行业声誉与人脉网络依然是决定胜负的关键杠杆。 全栈基础设施争议与 Meta 战略抉择 (25:00) 围绕 AI 工具是真实业务载体还是沦为高风险“模拟游戏”的讨论，主持人和分析师延伸探讨了大型科技公司的垂直整合策略。部分用户在 Pulsia 等自动化代理平台上实时监控广告支出的行为，被戏称为付费围观软件运转的游戏体验；与此同时，Meta 在财报电话会议中阐述的战略路径展现了底层算力控制权的残酷性。分析师 Ben Thompson 指出，面对为何不效仿 Apple 借助外部模型合作的质疑，Mark Zuckerberg 坚持 Meta 必须自主研发前沿模型，认为依赖第三方或国外开源方案存在极高的商业锁定与监管变更风险。 Zuckerberg 强调，Meta 的本质是一家全栈工程公司，早年 Facebook 在基础网络与自建数据中心上的可靠性是击败对手的关键；例如 Reels 短视频推荐引擎就需要两座专用数据中心提供算力支撑，若无自研软硬件协同，便无法应对 TikTok 的实时竞争。然而，华尔街对 Meta 激进的资本支出（CapEx）仍持审慎态度。投资者不满于公司一方面为外部算力买单，另一方面向内部超级智能实验室（MSL）倾斜巨额预算，在短期变现路径尚不明朗的情况下，内部资源分配的张力正持续加剧。 本期节目展现了 AI 浪潮的双重图景：技术层面，模型在抽象推理与企业自动化运营上不断突破人类生产力边界；现实层面，公众对生成式内容的信任危机、学术界的存在主义焦虑，以及巨头在重资产基础设施投资与短期商业回报之间的冲突，仍是悬而未决的核心矛盾。

## Background 该视频由科技播客频道 TBPN 于 2026 年 8 月 6 日发布在 YouTube 平台。本期为《Diet TBPN》30 分钟精粹版栏目，由 John Coogan 与 Jordi Hays 共同主持，节目简介指出该节目致力于在半小时内浓缩当天对硅谷科技前沿、行业风云人物与前沿趋势的深度剖析。 ## TL;DR 节目首先聚焦知名科普创作者 Hank Green 因使用 ChatGPT 整理文献引发的网络声讨，并将其与 OpenAI 内部模型 Astra 攻克 10 项前沿数学与量子开放难题的学术突破进行对比，揭示公众情绪与科研进展之间的断层。随后，主播深入剖析了由 AI 工具赋能的“一人百万美元公司”浪潮，探讨其在创造惊人个人利润的同时对就业市场带来的挤压，以及创业者可能陷入“模拟经营游戏”般的自我陶醉悖论。最后，节目借财报分析拆解了扎克伯格为何拒绝苹果式的轻资产外采路线、坚决押注全栈自研与庞大数据中心算力的战略考量，并以保罗·格雷厄姆被迫读完“史上最烂之书”的幽默轶事轻松收尾。 ## Atmosphere & style 整期节目的基调犀利敏锐且极具硅谷极客式的幽默讽刺，两位主持人与团队成员之间的搭话默契、节奏轻快，充满了“氛围编码”与自嘲式的调侃。讨论在面对人类智慧圣殿受冲击时的深刻哲思（如数学家的存在主义危机与巨头万亿级算力豪赌）与针对行业泡沫的冷峻解构（如调侃单人创业宛如《安德的游戏》）之间自如穿梭，给听众带来一种既身处科技剧变风暴眼、又能跳出狂热保持清醒审视的畅快现场感。 ## Bulletin - [00:00](https://www.youtube.com/watch?v=XouOvPiDdMw&t=0s) 知名科普博主 Hank Green 因在研究中使用 ChatGPT 检索论文引发网络抵触并致歉，凸显出前沿技术普及与公众对 AI 信任危机之间的撕裂。 - [00:00](https://www.youtube.com/watch?v=XouOvPiDdMw&t=0s) OpenAI 内部模型 Astra 被曝攻克数学及量子计算领域 10 个开放难题，引发形式化验证与现实科研价值界限的激烈争论。 - [14:48](https://www.youtube.com/watch?v=XouOvPiDdMw&t=888s) 得益于 AI 编码与自动化工具，Stripe 上年入百万及千万美元的单人公司数量翻倍增长，但 IT 行业企业招工意愿创历史新低。 - [26:50](https://www.youtube.com/watch?v=XouOvPiDdMw&t=1610s) Meta 坚定押注全栈自研与模型主权拒绝走苹果式授权路线，但巨额资本开支与双重支出引发了资本市场与内部派系的焦虑。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=XouOvPiDdMw&t=0s) 科普顶流 Hank Green 因承认借助 ChatGPT 辅助论文整理与数据提取遭遇舆论围攻，暴露出普通网络大众与技术现实间的巨大认知断层。 - [00:00](https://www.youtube.com/watch?v=XouOvPiDdMw&t=0s) OpenAI 研究员 Noam Brown 透露内部模型 Astra 解决 10 个数学与理论计算机开放难题，但批评者质疑其仅限于形式化验证问题。 - [05:00](https://www.youtube.com/watch?v=XouOvPiDdMw&t=300s) Hank Green 视频中一句对嘉宾的口语回应被误认作念出 AI 提示词草稿，反映出反 AI 群体对创作者使用 AI 的高度敏感与捕风捉影。 - [05:00](https://www.youtube.com/watch?v=XouOvPiDdMw&t=300s) AI 提效带来的产能跃升把创作者推入难以停歇的内容跑步机，迫使成熟内容创作者反思是否需要主动减速降频。 - [10:00](https://www.youtube.com/watch?v=XouOvPiDdMw&t=600s) 在科学传播领域盲目抵制 AI 并不现实，关键在于保持事实核查与使用透明，以防 OCR 识别错误及 AI 幻觉污染文献体系。 - [11:58](https://www.youtube.com/watch?v=XouOvPiDdMw&t=718s) 面对 AI 在高深数学证明上的跨越，数学学者陷入“至暗之夜”的精神震颤，但基础数学范式的突破有望外溢革新材料科学与化学应用。 - [14:48](https://www.youtube.com/watch?v=XouOvPiDdMw&t=888s) 一人企业迎来爆发潮，部分创业者依托 AI 实现千万级年营收，但频繁调用高阶商业模型带来的 API 成本可能引发倒贴风险。 - [20:00](https://www.youtube.com/watch?v=XouOvPiDdMw&t=1200s) 信息技术行业新商业注册量年增近 45% 的同时用工意愿骤降，哈佛商学院研究显示 AI 初创公司平均减少了 25% 的员工编制。 - [24:45](https://www.youtube.com/watch?v=XouOvPiDdMw&t=1485s) 部分低门槛一人公司缺乏实质商业壁垒，生成式 AI 创业有时更像成年人获得心理满足的模拟经营“电子游戏”。 - [26:50](https://www.youtube.com/watch?v=XouOvPiDdMw&t=1610s) 扎克伯格坚决拒绝苹果式的轻资产第三方模型采购路线，认为开源权重无法替代最前沿智能，且存在被卡脖子的监管与断供风险。 - [30:00](https://www.youtube.com/watch?v=XouOvPiDdMw&t=1800s) Meta 依赖自研芯片与底层数据中心的全栈工程基因，以当年突击新建两座数据中心支撑 Reels 算法为例，强调唯有掌控底层才能打造差异化体验。 - [30:00](https://www.youtube.com/watch?v=XouOvPiDdMw&t=1800s) Meta 目前承受着外租算力高昂运营开支与下一代超级数据中心巨额 CapEx 的“双重买单”压力，引发华尔街与内部盈利部门的摩擦。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=XouOvPiDdMw&t=0s) 周末热议：Hank Green 遭遇 AI 舆论反弹与 OpenAI 的数学突破 节目伊始，主持人 Jordi Hays 和 John Coogan 盘点了刚刚过去的周末科技圈内外的热点。最引人注目的事件之一发生在科技边缘的创作者圈子——知名 YouTube 科普创作者、全媒体企业家 Hank Green 因为被曝在研究工作中使用了 ChatGPT，在社交媒体上遭遇了猛烈的舆论围攻。尽管全球有数以十亿计的用户日常使用 ChatGPT 进行资料检索，但当这位坐拥庞大受众的顶流科普教父承认借助 AI 时，依然引来了数千人的口诛笔伐与长篇指责。Hank Green 不得不出面澄清、道歉并承诺未来调整流程。这起事件荒谬却极具讽刺意味：作为一名科学与教育博主，Hank Green 所处的领域恰恰是当前受 AI 冲击最深、融合最紧密的前沿。 与公众对 AI 工具的抵触形成鲜明反差的是，AI 在底层科研领域正取得惊人的突破。OpenAI 研究员 Noam Brown（在社交媒体上以 polynomial 知名）在周末透露，OpenAI 下一代旗舰模型家族的内部版本 Astra 已经一口气攻克了数学、量子复杂性以及理论计算机科学领域的 10 个重大开放性难题。这项成就的突破程度让主持人都调侃称“甚至不值得在节目里费力拆解了”，团队成员 Tyler 甚至准备在演播室的“Ultra Dome”里跑个 5 公里来个胜利绕场。然而批评者依然不买账，著名 AI 批评家 Gary Marcus 泼冷水称：“等 Astra 解决了一个不依赖于形式化验证（formal verification）的重大真实世界开放问题，再叫醒我吧。”AI 研究学者 Daniel Eth 则反讽道，某些人给 AI 设立的“及格线目标（goalposts）简直已经移到另一个星球去了”。 - **Noam Brown 的爆料**：OpenAI 内部代号 Astra 的下一代模型家族成功解决了数学、理论计算机科学与量子复杂性领域的 10 个未解难题。 - **形式化验证与现实价值的争论**：尽管 AI 在形式化验证领域表现更突出，但现实中大量极具价值的成果（如药物是否攻克癌症、复杂调度推荐系统）本身也是高度可验证的。 ```mermaid flowchart TD accTitle: AI学术进展与公众抵触的认知反差 accDescr: 对比OpenAI在硬核科学突破上的进展与网络公众对博主使用AI检索的抵触情绪 subgraph TechAdvancement ["前沿科技进展"] A["OpenAI 内部模型 Astra"] --> B["攻破 10 个数学与量子计算开放难题"] B --> C{"批评声音 (Gary Marcus)"} C -->|"质疑"| D["仅限于形式化验证问题"] C -->|"反驳 (Daniel Eth)"| E["评判目标线被无限推向外太空"] end subgraph PublicBacklash ["公众舆论反应"] F["Hank Green 承认使用 ChatGPT 做检索"] --> G["社交网络指责：丧失人类灵魂/AI幻觉"] G --> H["博主陷入被动道歉与减产反思"] end ``` --- ### [02:30](https://www.youtube.com/watch?v=XouOvPiDdMw&t=150s) 顶流科普教父 Hank Green 与《Ask Hank Anything》节目风波 为了理清这场争议的来龙去脉，主持人深入介绍了 Hank Green 的行业地位。Hank 是 YouTube 的“元老级（OG）”创作者，早在 2007 年——也就是 YouTube 上线后不到两年——就开始在平台上深耕。在过去近二十年里，他不仅积累了极高的声誉与观众信任，还建立起庞大的媒体商业帝国：他和哥哥 John Green 共同运营着常青 Vlog 频道 *Vlogbrothers*、坐拥数千万学习者的顶级教育频道 *Crash Course*，更是创作者经济行业盛会 VidCon 的创办人。作为整个平台最懂 YouTube 氛围、最被信赖的科学传播者之一，他的内容一向以严谨和人格魅力著称。 引发此次争议的是他上周三发布的一期实验性节目《Ask Hank Anything》。这个节目的形式颇为特殊：Hank Green 邀请嘉宾做客，但与常规人物访谈不同，是嘉宾向 Hank 提出各种刁钻的科学问题并展开深度讨论。如果遇到现场无法即时解答或未提前准备的硬核知识点，Hank 会在录制结束后亲自查阅资料、补充研究，并将自己面对镜头的解答切片剪辑回正片当中。 在引发争议的这期视频中，他们讨论了著名的心理语言学现象“布巴/奇奇效应（Bouba/Kiki effect）”——研究者向受访者展示一个圆润如云朵的形状和一个尖锐多角的星形，绝大多数人都会本能地把“布巴（Bouba）”指派给圆润形状，把“奇奇（Kiki）”指派给尖锐形状，证明词汇的声音质感与人类的形状知觉存在深层普遍联系。Hank 讲述了这个科学实验的起源、研究人员的测试过程与样本数据。由于录制结束后他需要快速查证实验背景与详细数字，以便及时剪辑成片，他自然借助了现代检索工具，却未曾料到这会引来一场关于“学术不端”式的无端指控。 - **Hank Green 的影响力资产**：创立于 2007 年，旗下涵盖 *Vlogbrothers*、*Crash Course* 以及全球顶级创作者大会 *VidCon*。 - **《Ask Hank Anything》节目机制**：嘉宾现场提问，未即时解答的问题由 Hank 在后期查阅文献后录制单人补拍镜头切入。 - **布巴/奇奇效应（Bouba/Kiki effect）**：著名的跨文化心理学实验，展示人类语音发音与几何视觉感知之间的联觉映射。 --- ### [05:00](https://www.youtube.com/watch?v=XouOvPiDdMw&t=300s) “我理解你的反对意见”：被误读的 AI 提示词与科普界的 AI 撕裂 争议的导火索源于社交媒体上一位名叫 JoJo 的账号发布的切片控诉。该片段指责 Hank Green 连文案脚本都是直接用 ChatGPT 自动生成的，甚至“粗心到把 AI 的回复口吻读了出来”。控诉者死死咬住的所谓铁证，是 Hank 在一段单人补拍视频中说的一句过渡句：“我理解你的反对意见（I appreciate the pushback）”。批评者认为，这显然是 ChatGPT 在对话交互中安抚用户的标志性机械套话，Hank 是不小心把提示词反馈直接照本宣科读进了正片里。但事实完全是一场乌龙：Hank 这句话是对访谈中嘉宾此前提出的质疑进行的回应，只是因为后期剪辑将其处理成了 Hank 单独面对镜头的补录画面，脱离完整上下文单看显得稍显突兀，却被网络舆论无限放大，演变成“知名科普博主直接念 AI 提示词草稿”的耸动头条。 随着周末舆论持续发酵，Hank Green 不得不出面否认脚本由 AI 代笔的传闻，但坦诚承认自己确实把 ChatGPT 当作文献整理的研究助手：用它追溯原始论文出处、提取 PDF 中的核心实验数据、抓取引用原句、统一单位或变换数据格式。然而这一诚实的表态并未平息怒火，部分激进观众依然在社交媒体上连番声讨，声称 AI“充满幻觉”且“剥离了人类学习理解的核心灵魂”。这让 Hank 感到心力交瘁，甚至坦言自己目前对使用 AI 的状态并不满意——AI 提升了生产效率，促使他发布更多内容，进而把他推入了一个难以停歇的内容跑步机（treadmill），他甚至考虑未来通过减少内容产出以脱离这种恶性循环。 两位主持人指出，这一事件暴露出科技界认知与普通网络大众之间的巨大“断层（whiplash）”。AI 已经变成了一个高度撕裂的社会议题：一方面数十亿人在日常中享受着它的效率红利；另一方面，反 AI 情绪在网络社群中依然极度高涨。这种矛盾在科学教育领域尤为致命——现代科学与数学的每一次前沿突破，注定都将深度依赖 AI 工具。如果科学传播者因为畏惧舆论围攻而在内容中刻意回避 AI，最终将彻底丧失解释当代科学进展的能力。公众迫切需要建立起更加成熟客观的认知：只要最终内容严谨求真、使用边界公开透明，借助 AI 进行资料梳理本就是现代科研与内容创作理应接纳的效率工具。 - **被指认的“AI 口癖”**："I appreciate the pushback"（我理解你的反对意见），实为 Hank 针对嘉宾反驳的口语回应，因单机位补拍剪辑被误解为 AI 生成文本。 - **Hank Green 实际的 AI 使用场景**：论文溯源、下载并解析 PDF 报告、跨文献抓取引文、换算数据单位及格式整理，全程人工把关核心事实。 - **内容创作者的“跑步机效应”**：生产力工具带来的产能跃升反向绑定了更高的内容更新频率，迫使成熟创作者反思是否需要主动减速降频。 ### [10:00](https://www.youtube.com/watch?v=XouOvPiDdMw&t=600s) Hank Green 风波与 AI 辩论的众生相 关于科普博主 Hank Green 因使用 AI 检索论文而引发的网络声讨，网上有人整理了一张图表，对比了支持 AI 与反对 AI 两派阵营中各自“有力”和“站不住脚”的论点。支持派最扎实的论点在于：AI 确实是一个功能强大且高效的工具；而反对派最糟糕的说辞则是全盘否定，宣称“AI 在学术研究乃至所有领域一无是处”。然而，支持 AI 的阵营同样提出了十分荒唐的辩解——比如有人反驳网友说：“Hank 使用数据中心怎么了？你发这条评论不也是靠数据中心托管的吗？”这种反驳显然站不住脚，普通网络评论占用的机房资源，与大型生成式 AI 系统“成吨吞吐 token、把 GPU 烧得滚烫”的高能耗计算完全不可同日而语。 反对 AI 一方提出的最有价值的批评在于：在科学传播中使用 AI，确实在一定程度上削弱了公众的信任感。事实核查确实必不可少，现实中也已经出现了大量粗制滥造的 AI 短语、胡说八道的幻觉污染科研文献的真实案例。此前就有一个典型事件：某份被扫描成 PDF 的文献由于光学字符识别（OCR）排版错误，两栏文字在导入时重叠混淆，拼接出一个原本根本不存在的怪异词组，结果后续一大批科研论文竟盲目将其当成真实术语引用。主持人忍不住为 Hank Green 鸣不平：“他们竟然因为我心中的神（GOAT）用大语言模型检索做视频要读的论文，就把他给‘取消’了！难道在 2026 年，大家还指望他像个原始人一样用传统的谷歌搜索吗？”两人随即调侃，现在的 Google Search 甚至都不知道还能不能彻底关闭 AI 模式，或许只剩 DuckDuckGo 还没有全面被 AI 接管。 - **AI 支持派合理论点**：AI 是极具能力且高效的研究工具。 - **AI 支持派荒谬论点**：将运行大模型的能源消耗等同于普通网页服务器托管一条评论的能耗。 - **AI 反对派合理论点**：在科普与科学研究中不加节制地使用 AI 会不可避免地侵蚀公众对科学传播的信任。 - **OCR 幻觉致灾案例**：因文献双栏扫描混淆产生的生造词，竟被多篇后续科学论文当作真实概念引用。 --- ### [11:38](https://www.youtube.com/watch?v=XouOvPiDdMw&t=698s) 数学界的“至暗之夜”：当 AI 逼近人类智慧圣殿 在推特（X）上，Brian Lowe 分享了一篇由奥克兰大学数学家兼研究员 Kerwin Hampshire 撰写的感伤长文，该文发布于 Noam Brown 公布最新数学突破之前。面对 AI 在高等数学证明中的惊人跨越，Hampshire 陷入了巨大的精神震颤与无力感中。他在文章中写道：“我无能为力，你们或许也一样无能为力。我拿不出任何解决方案、政策建议，也发起不了条理清晰的行动倡议。我只想坦白、赤裸地表达我此刻在情感与精神层面遭遇的冲击。我渴望被看见，渴望像我一样的人能被看见。我需要这个全新数学范式的缔造者们，在赐予我们最后一击（coup de grace）之前，能直视我的双眸，承认我们共同的人性与灵魂。我最渴望的，是我们能真正懂得我们当下究竟在做什么。” 这一被称为“数学至暗之夜”的心声引发了主持人的深刻讨论。数学教育是否会就此消失？虽然许多未解的高深命题看似离实际应用十分遥远，但纯数学与应用科学是截然不同的学科。节目中 21 岁的主持人 Tyler 逗趣地给数学教授们支招：“文章里说数学家拿薪水是去证明定理的，但实际上大学教授的大部分精力不都在教课吗？真正一辈子能解出多少大定理的教授又有几个呢？”大家进一步探讨，科学发展的历史一直都是如此：当旧的假想和猜想被攻克，必然会有全新的未知问题涌现。如果 AI 真的解决了诸如 P versus NP、纳维-斯托克斯方程（Navier-Stokes）等千禧年大奖难题（Millennium Prize Problems），数学边界被彻底拓宽后，这项技术外溢到材料科学、化学以及生物学等现实领域，或将带来革命性改变——比如彻底攻克某种底层原理，使电动汽车的续航里程瞬间翻倍。 ```mermaid flowchart TD accTitle: AI 对数学及应用学科的冲击链路 accDescr: 展示 AI 从攻克数学猜想向材料、化学和生物等实际应用领域渗透的过程 subgraph CoreMath ["数学核心突破"] A["AI 系统 / 大模型推理"] --> B["千禧年大奖难题 (如 P vs NP / Navier-Stokes)"] B --> C["基础数学范式被改写"] end subgraph ScientificImpact ["下游科学外溢"] C --> D["材料科学革新"] C --> E["化学机理突破"] C --> F["生物医药演进"] end subgraph RealWorld ["现实产业变革"] D --> G["消费级产品跃迁 (如电池续航翻倍)"] E --> G F --> G end ``` --- ### [14:48](https://www.youtube.com/watch?v=XouOvPiDdMw&t=888s) 一人即独角兽：百万美元单人公司的爆发 随着 AI 编程与自动化工具的成熟，商业世界迎来了“一人赚百万”乃至“一人赚千万”的全新浪潮。《华尔街日报》报道指出，AI 工具让创业者孤身一人起步变得轻而易举，且许多公司在规模扩大后依然选择不招一名员工。节目提到了创业者 Ben Broca 的案例：他去年 12 月成立了一家向创业者提供 AI 工具的软件公司，如今已坐拥 10,000 名付费客户，今年年化营收有望冲上 1,000 万美元大关，而整个团队至今依然只有他一个人。这位 40 岁的创始人坐在加州索萨利托（Sausalito）阳光明媚的客厅里，完全依靠 AI 处理邮件、编写与调试代码、接待客户咨询、办理新订阅甚至处理退款。Broca 坦言自己享受独自决策的自由，并直言：“妥协会导致平庸的结果。” 根据支付巨头 Stripe 的统计数据，平台上年收入突破 100 万美元的“单人企业”（solo operators）已有数千家，这一群体规模在 2023 年至 2025 年间直接翻了一番；年收入突破 1,000 万美元的单人公司数量在同期更是激增了近两倍。传统观念中运营一定规模的企业必须配备团队，但现在 AI 成了全能合伙人。除了 Stripe 上的繁荣生态，知名程序员网红 Ben Awad 也曾做过趣味实验：只输入一句提示词“去给我赚钱”，挂上每月 200 美元的 AI 订阅（如 Fable、Sol 等系统）并配置好 API 密钥，让 AI 自主折腾一周，观察它究竟能赚回几美分还是几十美元。 美国人口普查局的数据从宏观层面印证了这一历史转向：信息技术行业的新企业申请量在过去一年激增近 45%，位列所有行业之首；然而，该行业中“明确表示计划招聘员工”的企业比例却创下了有统计以来的最惨烈暴跌。旧金山面向单人科技创始人孵化器的负责人 Julian Weisser 表示，创业的门槛从没有像今天这样低——其最新一期加速器仅开放 10 个入选名额，却引来了 4,500 名申请者，竞争激烈程度相较去年 5 月创办时飙升了近五倍。不过，单人依靠 AI 狂奔的背后也潜藏着不菲的隐形成本，Broca 就曾透露，由于调用 Anthropic 的 Claude 模型处理客户端复杂请求的账单过于高昂，他在许多客户身上甚至一度处于倒贴亏本的状态。 - **Stripe 单人企业数据暴涨**：年营收超 100 万美元的单人运营者在 2023 至 2025 年间翻倍；年营收超 1,000 万美元的单人企业数量增长近两倍。 - **信息产业新趋势**：信息部门新商业注册量年增近 45%，但有招工意愿的企业比例出现断崖式下跌。 - **单人孵化器极度火爆**：Julian Weisser 运营的 Solo 创始人加速器，仅提供 10 个席位却吸引了 4,500 人申请。 - **高昂的 AI 调用成本**：依靠 AI 自动化并非零成本，频繁调用顶级大模型（如 Anthropic 的 Claude）可能导致单客服务成本失控甚至倒贴亏损。 ### [20:00](https://www.youtube.com/watch?v=XouOvPiDdMw&t=1200s) 一人公司狂潮：从“石中剑”到劳动力市场的结构震荡 节目切入了一篇关于 AI 赋能“一人超级企业”（One-Person Businesses）的《华尔街日报》报道。一位此前依赖商用 API 的创始人，面对按调用量激增的账单，果断转向来自中国的免费开源大模型；他从投资人手中拿到了 3000 万美元融资，同时因为无需组建传统的庞大软件工程团队而省下了数百万美元薪资。但这种低门槛也带来了隐忧：既然你能靠 AI 快速起步，别人抄袭你也轻而易举。在佛罗里达州奥兰多独自运营 AI 公司的 Troy Johnston 对此深有体会，他在报道中留下了一句极具史诗感的话：“人人都握有宝剑，我们现在全都有拔出石中剑（Excalibur）的能力了。” 主播们饶有兴味地调侃起 Troy Johnston 酷爱亚瑟王隐喻的风格。今年 40 岁的 Troy 用 AI 编写了一款帮助用户最大化信用卡权益的应用——根据用户当前在餐厅消费还是日常购物，实时推荐该刷哪张卡。主播们指出，这类产品其实早已存在（比如 The Points Guy 围绕信用卡积分建立的内容帝国），甚至普通人开一个 ChatGPT 线程把卡片信息扔进去也能做到类似的事，但通过“氛围编码”（vibe coding）做出一款深度集成的轻量应用，依然让 Troy 在零雇员的情况下实现了每月数额可观的净利润，展现出惊人的单兵作战变现力。 这种“独行侠创始人”对宏观劳动力市场的影响正引发经济学界的激烈争论。哈佛商学院研究创业学的副教授 Rembrand Koning 提出一个发人深省的问题：“如果所有公司都在减少招聘，但全社会的初创企业数量变成了原本的四倍，整体员工规模会变成怎样？”他与人合著的一项针对 5 万家初创企业的研究发现，专注于 AI 的初创公司在运营时平均减少了 25% 的员工人数。主播们对此展开了讨论：这与 Ramp 的企业消费报告形成了微妙反差——Ramp 数据显示积极采用 AI 的公司往往招聘步伐更快，其背后可能是企业在压低基础运营人头的同时，将资源集中在更高增速的核心业务扩张上，同时求职大环境的疲软也倒逼许多人自立门户。 不过，AI 创业的结局并非皆是大获全胜。两位背景迥异的创业者展现了截然不同的现实切片： - **Samir Ahmad（39岁）**：在 Verizon 坚守近二十年后遭遇疫情后的职业倦怠，在社交媒体对 AI 的狂热吹捧下辞职，创立了一家单人咨询辅导公司。他将 AI 视为自己的“幕僚长和二把手”，用于制定商业计划和撰写营销文案。然而该业务在短短数月内便难以为继，他最终重返职场，入职了一家公用事业企业。 - **Claire Vo（41岁）**：身兼科技高管的她在 2023 年底纯靠把 ChatGPT 的代码“复制粘贴”，开发出一款面向金融与医疗客户、用于管理新产品设计与文档的工具（ChatPRD）。她将订阅费定在极具杀伤力的 1 美元/月。在经历了 9 个月的单打独斗后，她仅雇佣了一名工程师，如今用户突破 10 万人，今年有望录得七位数的纯利润。她的营销、销售和客服全部交由 AI 运转，但她清醒地表示：“我认为大家过高估计了 AI 有多轻松，而过低估计了我为此走到这一步所付出的努力与行业人脉。” --- ### [24:45](https://www.youtube.com/watch?v=XouOvPiDdMw&t=1485s) 创业还是打游戏？生成式 AI 的“安德的游戏”式悖论 话题转向了为一人公司提供全套自主运营后台的平台 Pulsia。主播们查看了 Pulsia 的公开数据看板，发现上面展示了各家单人公司实时投放在广告上的支出：截至录播当天上午，平台上所有公司合计支出的广告费才区区 373 美元。这引发了主播们的尖锐质疑：“我只想知道，到底有没有五家公司，从自己业务里赚到的钱能超过他们每个月付给 Pulsia 的服务费？” 这延伸出一个更深刻的哲学拷问：当前大部分人用 AI 搞的所谓“一人创业”，到底是真正跑通了商业模式，还是本质上只是一款让成年人沉迷的“视频游戏”？人们享受的可能仅仅是看着这台精巧的数字机器嗡嗡作响、自动运转的心理满足感，实际产出却微乎其微。主播将此与 Midjourney 和 Suno 的发展历程作了精彩类比： > 当 Midjourney 刚问世时，全世界都在惊呼“艺术家要失业了”，大家以为接下来纽约现代艺术博物馆（MoMA）会举办 AI 提示词画展、佳士得拍卖行的压轴大作会出自 Midjourney 创作者之手。但现实完全不是这样——人们并没有用它去创造真正的纯艺术，大众只是极度享受输入 Prompt、生成图像然后发给朋友炫耀的过程。这就像买了一把吉他坐在卧室里自弹自乐，和你成为一名巡演职业音乐家是两码事；玩 Suno 生成音乐也是同理。 主播顺着这个逻辑打趣道，这种局面甚至可能会演变成科幻小说《安德的游戏》（Ender's Game）式的黑色幽默：你以为自己只是在屏幕前无聊地玩一款 AI 创业模拟经营游戏，每天点点按钮、给虚拟员工派活，结果突然有人冲进来告诉你：“安德，恭喜你，你刚才不是在玩模拟器——你刚才把最后一批工作岗位成功外包到了海外，并且顺手通过杠杆收购垄断了全美整个暖通空调（HVAC）产业链！” --- ### [26:50](https://www.youtube.com/watch?v=XouOvPiDdMw&t=1610s) 扎克伯格的 AI 护城河：Meta 为何坚决不抄苹果的作业 节目的最后部分聚焦于科技分析师 Ben Thompson（Stratechery 创作者）对 Meta 最新财报电话会议的复盘。分析师们向马克·扎克伯格（Mark Zuckerberg）提出了一个直击灵魂的问题：Meta 为什么不能像苹果那样，采取“什么都不做，坐收渔翁之利”的轻资产合作策略？即直接与前沿 AI 实验室签约授权，需要图像模型就采购图像模型，需要文本就采购文本，需要写代码就去采购市面上最强的代码 Agent，何必烧几百亿美元自研全栈基础设施和大模型？ 扎克伯格对此给出了极其详尽且坚决的回应。他明确拒绝了苹果式的组装路线，核心逻辑分为三层： ```mermaid flowchart TD accTitle: Meta 的全栈 AI 战略逻辑 accDescr: 展示扎克伯格为何拒绝采用苹果式的第三方授权模式而坚持自研全栈技术 subgraph Risks["外采模型的风险"] R1["前沿差距：开源模型仍逊于闭源前沿"] R2["外部依赖：供应商变闭源或监管断供风险"] end subgraph Core["Meta 全栈工程底座"] C1["底层芯片与自建数据中心"] C2["自研底层系统与前沿智能模型"] end subgraph Advantage["最终业务壁垒"] A1["极致性能与成本优化"] A2["无法被替代的定制化原生体验"] end R1 --> Core R2 --> Core Core --> Advantage ``` 1. **开源权重无法替代前沿智能**：目前的开源或开放权重模型尚未达到绝对顶尖的最前沿水平，而 Meta 的核心业务体验必须依赖最顶级的前沿智能，因此 Meta 必须亲自驻扎在技术前沿。 2. **严重的外部依赖与监管断供风险**：依赖第三方（尤其是海外或竞争对手）存在巨大的合规与政策不确定性；更致命的是商业背刺——某些供应商前期以开源招揽生态，几年后突然闭源并开出天价授权费，Meta 绝不能将身家性命绑在随时可能卡脖子的外部技术上。 3. **全栈工程基因（Full-Stack DNA）**：扎克伯格回顾了 Facebook 早期的创业史。很多人只看到 Meta 表面上是一家做社交 App 和广告业务的公司，但它本质上是一家拥有自研芯片、自建超大规模数据中心和底层低级软件的全栈科技巨头。当年扎克伯格亲自编写了大量系统底层代码，Facebook 之所以能击败众多早期对手，最朴素的原因在于“它的系统真的能跑通且不崩溃”，而别的社交网站动辄瘫痪。扎克伯格强调，只有深入技术栈的最底层，才能创造出极致个性化、高度优化且其他拼凑型公司在物理上根本无法实现的端到端质感体验。 尽管主播们指出苹果的轻资产授权整合模式在当下看起来极其讨巧且成效显著，但扎克伯格展示出的这种坚决不妥协的底层硬核工程师视角，解释了为什么 Meta 愿意以近乎偏执的姿态在这场 AI 军备竞赛中押上所有筹码。 ### [30:00](https://www.youtube.com/watch?v=XouOvPiDdMw&t=1800s) Meta 的算力护城河与扎克伯格的多面“战局” 拥有自主构建模型的“模型主权”（model sovereignty），正成为未来科技技术栈中不可或缺的一环，这也是 Meta 坚定不移押注底层自研的核心原因。回顾早期竞争，许多对手之所以节奏落后、无法快速推陈出新，很大程度上就是因为缺乏全栈整合能力；而 Meta 凭借自上而下的垂直整合，能够以极其激进的节奏攻城略地。以 Instagram Reels 的诞生为例，外界总以为“不就是在界面上加个新按钮、放几段短视频”，但其背后支撑 30 亿用户分发的，是一个计算极度密集、需要存储吞吐海量数据的庞大推荐算法系统。为了扛住 Reels 算法的算力负荷，Meta 当年甚至硬生生突击新建了两座数据中心。如果缺少底层的硬核算力储备与基建控制权，Meta 绝不可能在如此短暂的时间窗口内拿出足以抗衡 TikTok 的产品。 这种全栈垂直整合虽然保障了个性化推荐与极致用户体验，却让华尔街投资者焦虑万分。科技分析师本·汤普森（Ben Thompson）将其尖锐地形容为“财务上的尾巴摇狗”，指出 Meta 目前正陷入“双重买单”（double paying）的尴尬泥潭：一方面为了眼前的模型研发和业务运作，不得不高价外租 AI 算力并重金挖角顶级研究员；另一方面又得为下一代超级数据中心掏出天文数字般的资本支出（CapEx）。更雪上加霜的是，在缺乏清晰明确的变现商业路径之前，市场对所谓“营收改善”的叙事已经逐渐失去耐心。不仅外部资本市场的信心受挫直接反映在股价上，Meta 内部也充满了摩擦与抱怨——其他承担营收压力的老牌部门眼睁睁看着算力与资源无限向战略先锋团队倾斜，内部火药味十足。面对内外交困的质疑，马克·扎克伯格（Mark Zuckerberg）却展现出极强的意志力，完全没有退缩妥协的打算，正独自打一场横跨资本市场、内部派系与前沿技术的“多线战争”。 ```mermaid flowchart TD accTitle: Meta算力垂直整合与双重支出困局 accDescr: 展示Meta通过自建算力支撑业务与当前面临的双重资本支出压力的对比结构 subgraph InfraStrategy ["垂直整合护城河（以 Reels 为例）"] DC["新建自研数据中心"] --> RecSys["计算密集型推荐系统"] RecSys --> Scale["支撑 30 亿级海量用户与抗衡 TikTok"] end subgraph FinancialPressure ["当前的“双重买单”压力"] Rent["高昂的外租 AI 算力与研究员薪酬"] CapEx["下一代数据中心庞大资本开支 (CapEx)"] Rent & CapEx --> Squeeze["资本市场质疑变现路径 & 内部派系摩擦"] end ``` - **Reels 背后的算力账本**：Meta 当年专门额外建造了两座数据中心，仅用来支撑 Reels 推荐算法的算力需求；没有这种底层基建的掌控力，根本无法在正常时间窗口内正面迎战 TikTok。 - **“双重买单”困境**：Meta 目前同时背负着租赁外部 AI 算力、招募顶尖研究员的高昂运营支出，以及自建未来数据中心的巨额 CapEx，短期内变现闭环尚不明朗。 - **内部矛盾爆发**：由于核心业务部门需要为高投入的 AI 研发实验室（如 MSL）提供资金弹药，内部围绕资源倾斜与待遇落差产生了明显的内耗与不满。 --- ### [32:15](https://www.youtube.com/watch?v=XouOvPiDdMw&t=1935s) Paul Graham 读到“史上最烂之书”与节目收尾 在严肃的商业探讨之余，主播们聊起了知名创业导师、Y Combinator 联合创始人保罗·格雷厄姆（Paul Graham）在社交媒体上分享的一则极其荒诞的个人趣事。保罗·格雷厄姆提到，自己之前买了一本书，内容糟糕透顶；他嫌弃到既不想把它留在自己的书架上，但出于某种爱书人的怪癖又狠不下心直接扔进垃圾桶，于是这本“烂书”就一直被随手搁在门边的桌子上积灰。 主播忍俊不禁地插话调侃道，自己正好认识能帮忙“解决”的人，“虽然我私下不认识他们，但他们绝对很乐意帮你把这书撕碎、塞进碎纸机然后一把火烧掉”。而保罗·格雷厄姆更倒霉的遭遇还在后头：某天早晨他急匆匆赶着出门赴约，顺手抓起门边的那本书塞进包里——这是“第一个致命错误”。随后他去餐馆吃早餐，手头没有任何其他阅读材料或消遣，只得被迫坐在那里，硬生生把这本“人生中读过最差劲的书”看了一整个早晨。主播笑称这情节简直就像情景喜剧《消愁》（*Curb Your Enthusiasm*）里拉里·戴维（Larry David）才会遭遇的经典桥段。在欢快的笑声中，主播向观众表达了感谢，提醒大家收看太平洋时间周一至周五上午 11 点的直播，节目可在 Apple Podcasts 与 Spotify 收听，并呼吁观众前往 TBPN.com 订阅每日新闻通讯。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=XouOvPiDdMw&t=0s) 等 Astra 解决了一个不依赖于形式化验证的重大真实世界开放问题，再叫醒我吧。 — Gary Marcus > [05:00](https://www.youtube.com/watch?v=XouOvPiDdMw&t=300s) 我理解你的反对意见（I appreciate the pushback） — Hank Green > [10:00](https://www.youtube.com/watch?v=XouOvPiDdMw&t=600s) 他们竟然因为我心中的神（GOAT）用大语言模型检索做视频要读的论文，就把他给‘取消’了！难道在 2026 年，大家还指望他像个原始人一样用传统的谷歌搜索吗？ — TBPN 主持人 > [11:38](https://www.youtube.com/watch?v=XouOvPiDdMw&t=698s) 我无能为力，你们或许也一样无能为力。我拿不出任何解决方案、政策建议，也发起不了条理清晰的行动倡议。我只想坦白、赤裸地表达我此刻在情感与精神层面遭遇的冲击。我渴望被看见，渴望像我一样的人能被看见。我需要这个全新数学范式的缔造者们，在赐予我们最后一击（coup de grace）之前，能直视我的双眸，承认我们共同的人性与灵魂。我最渴望的，是我们能真正懂得我们当下究竟在做什么。 — Kerwin Hampshire > [14:48](https://www.youtube.com/watch?v=XouOvPiDdMw&t=888s) 妥协会导致平庸的结果。 — Ben Broca > [20:00](https://www.youtube.com/watch?v=XouOvPiDdMw&t=1200s) 人人都握有宝剑，我们现在全都有拔出石中剑（Excalibur）的能力了。 — Troy Johnston > [20:00](https://www.youtube.com/watch?v=XouOvPiDdMw&t=1200s) 如果所有公司都在减少招聘，但全社会的初创企业数量变成了原本的四倍，整体员工规模会变成怎样？ — Rembrand Koning > [20:00](https://www.youtube.com/watch?v=XouOvPiDdMw&t=1200s) 我认为大家过高估计了 AI 有多轻松，而过低估计了我为此走到这一步所付出的努力与行业人脉。 — Claire Vo > [24:45](https://www.youtube.com/watch?v=XouOvPiDdMw&t=1485s) 我只想知道，到底有没有五家公司，从自己业务里赚到的钱能超过他们每个月付给 Pulsia 的服务费？ — TBPN 主持人 > [24:45](https://www.youtube.com/watch?v=XouOvPiDdMw&t=1485s) 安德，恭喜你，你刚才不是在玩模拟器——你刚才把最后一批工作岗位成功外包到了海外，并且顺手通过杠杆收购垄断了全美整个暖通空调（HVAC）产业链！ — TBPN 主持人 ## Mentioned TBPN, John Coogan, Jordi Hays, Hank Green, ChatGPT, OpenAI, Astra, Noam Brown, Gary Marcus, Daniel Eth, John Green, Vlogbrothers, Crash Course, VidCon, Ask Hank Anything, Google Search, DuckDuckGo, Brian Lowe, Kerwin Hampshire, Tyler, Ben Broca, Stripe, Ben Awad, Fable, Sol, Julian Weisser, Anthropic, Claude, Troy Johnston, The Points Guy, Rembrand Koning, Ramp, Samir Ahmad, Verizon, Claire Vo, ChatPRD, Pulsia, Midjourney, Suno, Mark Zuckerberg, Ben Thompson, Stratechery, Apple, Meta, Instagram Reels, TikTok, Paul Graham, Y Combinator, Larry David, Curb Your Enthusiasm ## 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=XouOvPiDdMw)

[0:00] Gary Marcus: 等 Astra 解决了一个不依赖于形式化验证的重大真实世界开放问题，再叫醒我吧。 [5:00] Hank Green: 我理解你的反对意见（I appreciate the pushback） [10:00] TBPN 主持人: 他们竟然因为我心中的神（GOAT）用大语言模型检索做视频要读的论文，就把他给‘取消’了！难道在 2026 年，大家还指望他像个原始人一样用传统的谷歌搜索吗？ [11:38] Kerwin Hampshire: 我无能为力，你们或许也一样无能为力。我拿不出任何解决方案、政策建议，也发起不了条理清晰的行动倡议。我只想坦白、赤裸地表达我此刻在情感与精神层面遭遇的冲击。我渴望被看见，渴望像我一样的人能被看见。我需要这个全新数学范式的缔造者们，在赐予我们最后一击（coup de grace）之前，能直视我的双眸，承认我们共同的人性与灵魂。我最渴望的，是我们能真正懂得我们当下究竟在做什么。 [14:48] Ben Broca: 妥协会导致平庸的结果。 [20:00] Troy Johnston: 人人都握有宝剑，我们现在全都有拔出石中剑（Excalibur）的能力了。 [20:00] Rembrand Koning: 如果所有公司都在减少招聘，但全社会的初创企业数量变成了原本的四倍，整体员工规模会变成怎样？ [20:00] Claire Vo: 我认为大家过高估计了 AI 有多轻松，而过低估计了我为此走到这一步所付出的努力与行业人脉。 [24:45] TBPN 主持人: 我只想知道，到底有没有五家公司，从自己业务里赚到的钱能超过他们每个月付给 Pulsia 的服务费？ [24:45] TBPN 主持人: 安德，恭喜你，你刚才不是在玩模拟器——你刚才把最后一批工作岗位成功外包到了海外，并且顺手通过杠杆收购垄断了全美整个暖通空调（HVAC）产业链！

[2:30] Gary Marcus: Wake me when Astra solves a significant open world problem that doesn't revolve around formal verification [2:30] Daniel E.E.: the goalposts are on a completely separate planet. [5:40] Hank Green: I appreciate the pushback. [7:50] use Google search like a caveman in big 2026. [9:50] Kirwin Hampshire: I have no prescriptions, policy recommendations, or coherent call to action. I just want to be honest and open about my emotional and spiritual response... I need the architects of our new mathematical paradigm to look me in the eyes and acknowledge our shared humanity and soul before they deliver the coup de grâce. [20:00] Troy Johnson: Everybody has the sword and we all have the ability to unsheathe Excalibur now. [30:00] Paul Graham: the worst book I've known

[0:00] Let's get into the show. What happened over the weekend? Did anything happen over the weekend, John? There were a couple things. The big debate that I was tracking, sort of outside of tech, but tech adjacent, was the cancellation of Hank Green, the YouTube creator. Not quite a cancellation, more just some backlash. Hard to always put a proper sizing on a mob when a mob comes after a creator but uh hank green the youtuber and really media entrepreneur uh he's grown a huge business which we can sort of go into uh is getting pilloried on social media over using chat gpt for research very controversial these days only a billion people do it but yeah it's the number one happening after but he's getting he's getting a lot of backlash from certain members of his audience i want to characterize the whole audience as being part of this but uh it's a It's a very silly situation. There's at least thousands of people that are really mad about it. It seems like that. It's always hard to tell. I mean, thousands of people that are liking a post about it. There's like maybe dozens of posts. I don't really know how to put a scale on these things. But Hank Green is definitely going through it having to sort of apologize or qualify or sort of you know state that he will adjust things in the future And it just sort of interesting to hear how he went through this process what he says is going to change and where the backlash is coming from because there's a lot of misunderstandings about it. And what's interesting is that he is a science and education creator, and science and education are potentially the most affected by AI right now. And so it's a real challenge to simultaneously say, I'm going to cover math. I'm going to cover science, but I'm not going to touch AI. That AI stuff's bad because, as we also saw over the weekend, AI is making a bunch of advancements on math. We've been seeing this for a while, but the latest version of the story comes from Gnome Brown, polynomial over at OpenAI. He says, an internal version of Astra, OpenAI's next major model family, solved 10 major open problems in mathematics, quantum complexity, and theoretical computer science. The achievements are so, so extreme at this point that i don't even think it's worth us trying to break them down like we did that with the uh so instead problem tyler is going to run a 5k here in the ultra dome to do a little victory lap for the research team but not everyone is impressed with it because gary marcus says wake me when astra solves a significant open world problem that doesn revolve around formal verification And of course Daniel E E says the goalposts are on a completely separate planet. It is a good point. Obviously, AI is doing better in formally verifiable tasks at the same time. Still impressive because there's a lot of things that are useful and verifiable. Like, did this drug cure your cancer or not? Or did this job get done or not. Like we've been using these recommender systems for lots of things. They're very valuable all over. But it is funny. The debate over is this AGI, is this ASI? Those terms will always be vague. Going back to Hank. Yes. So all he did was admit that he used some sort of AI tool for research. Yes. So I will take you a little bit more through it. So Hank Green, he's an OG YouTuber. He joined YouTube in 2007, I think less than two years after the platform actually launched. And he grew, he got a lot of views, but he also built a huge audience and created a real media company around it. So he has Vlogbrothers, like a vlog channel. Then he has Crash Course, which is a really, really huge educational channel He runs VidCon which is basically the premier conference around YouTube and the creator economy I been I think once or twice It a lot of fun And over the last 20 years he become one of the most trusted educational creators on the platform He's also just like, he gets the vibe of YouTube very well because he's been around it so long. Never really stepped back fully, but always been, you know, solid audience there. So last Wednesday, he published an episode of a show called Ask Hank Anything. And it's an interesting concept for a show. So he brings on a guest, but then instead of just doing the interview, tell me your life story, ask the guest a whole bunch of things, the guest brings questions for him about science or whatever. They have a big, long conversation. And if there's something in the show that he can't answer on the fly or he's not prepped for, he will go do the research and then get the actual answer and then cut that into the final episode. So you'll be watching them hang out. they'll talk about some odd thing. He was talking about this, have you heard this kiki and booba thing? There's like two words that, basically there's two shapes. One's like a fluffy cloud, the other's like spiky star, essentially. And if you ask people generally, which one would you assign the word kiki to and which one would you assign the word booba to, [5:00] People always pick Uba is the cloud and Kiki is the spiky one. And it's like the sound of the word has a shape to it, and this is just something in our language that shows up all over the place. So he's like telling the story of this, like, just somebody ran a science experiment. They, you know, put a bunch of people here. They pulled a bunch of people. They put together this result, and this is what happened. And so he needs to compile all of that quickly because you get off the show, you have the rest of your job, but then you have to go answer these questions and have all the information. And of course he uses all sorts of research tools, but he was accused specifically of using ChatGPT to write the script, which is interesting, because after the episode went up, manager JoJo posted a clip of him from the episode and accused him of using ChatGPT to write the script. The key line is Hank saying, quote, I appreciate the pushback. That's sort of an AI phrase, but that wasn't one of the really trigger AI phrases like you're absolutely right or it's not this, it's that. I appreciate the pushback is something that the AI models say occasionally, but you wouldn't think it would make it into a script. But that why people jumped on it They were like wow he was so careless that he left in a turn of phrase that was the model talking to him about I appreciate the pushback That not what happens He actually responding to the guest pushing back on him and about this concept And then he answers it. But he was just talking, but it feels out of place because he's talking to the camera at that point. Even though in the video, he's talking to the guest after the fact. The way it's edited is him direct to camera. So him saying, I appreciate the pushback to the camera. What is this? Beans? I don't know. But him saying that, I appreciate the pushback, feels a little weird when you just watch it, but it makes sense in the context of the longer video. So the headlines proliferated over the weekend to the tune of Hank Green accidentally reads AI prompt feedback left in his script. Hank has to deny this, but he goes on to admit that he does use ChatGPT for research. Dun, dun, dun. This did not land well. People don't like the idea of him using ChatGPT for research. And clearly it's just a small subset of his audience that actually takes the time to flame online about AI usage. But it is, there's still dozens of posts, maybe hundreds of posts about how AI cannot be used for research because it hallucinates or it removes some key human element of the process of learning something like that It very odd for anyone who used modern models because there a lot that AI can do well yet but pulling a bunch of links and quotes together from across the Internet is something that they're pretty good at. And it's definitely reliable for that. And so Hank clarified the script was not written by AI. He was just going on ChatGPT and saying, like, hey, where did this original research come from? Pull up the paper. download the PDF, crunch it all together for me, you know, pull some quotes from it, change this into a different format. I want it in this units instead of that units. Those types of questions. He still said that he has not been happy with how he's been using AI and may wind up publishing less as a result. He feels like he's on a little bit of a treadmill because he's more productive with AI, but then he posts more and then that's a feedback loop. And of course, like at this point in time, he's like built his career over 20 years. He has a very sustainable business. He probably doesn't need to be on as much of a treadmill as perhaps an early stage creator might be. It doesn't feel like it's total audience capture, but there is this interesting opportunity here that I was sort of just identifying. Like AI is clearly this wedge issue. Billions of people use AI and get value from it, but at the same time it deeply unpopular and there lots of people who like to post angrily online about how AI is bad for a variety of reasons but education and science in particular are going to be intertwined with AI for the foreseeable future. Like every advancement in science is going to be AI enabled, and so if you're a science educator and you constantly have to be dancing around AI and be like, oh yes, like they solved this math problem, but I don't like it because AI was used, well, you're going to wind up just not being able to talk about math or science or whatever's happening because you're constantly doing this dance around AI. And so that's fine. There's that audience that will love that. But there's also an opportunity for a new audience that's maybe a little bit more nuanced about this and maybe just, yeah, it's fine that you use that for doing research. As long as the script sounds good, I'm fine. Or as long as you are clear about your policy, which is odd because that's what he, he was always clear. He's just still got attacked and had to go on this defensive. I believe he has like a published policy around how him and his employees at his media company can use AI or do use AI or don't in various scenarios. But this was the first time I've seen like real backlash to just pulling that up on ChatGPT. A lot of people in tech, I think we're getting like sort of whiplash from watching. [10:00] All of these arguments pile up. Someone put together a cool chart here of the good arguments and the bad arguments from the pro-AI crowd and the anti-AI crowd. So an example of a good argument around this from the pro-AI crowd would be AI is a powerful and capable tool. And then like a bad argument from the anti-AI crowd would be AI is useless in research slash in general. But there were bad arguments that were put forth by pro-AI people. something like you use data centers, like Hank uses data centers, and it's like, yes, YouTube is hosted on a data center. Yeah, so are you to write this comment? Yes, but that's not the actual data center that's required to host an online comment is wildly different than a massive Gen AI system, like cooking tons of tokens and actually setting the GPUs on fire, right? And then a good argument, the best argument from the anti-AI crowd was said that AI usage in science communication reduces trust at least a little, which is interesting. I mean, yeah, you do have to check these things. And we do see tons of examples of people actually leaking, you know, AI phrases and weird AI like hallucinations into scientific research There was that example of there was some PDF that was scanned and there was a word in one column and a word in another column that got bled together when the document was imported. And then a whole bunch of scientific research started referencing this phrase that doesn't exist and just came from basically a hallucination or like a quirk of the optical character recognition. So anyway, they canceled my goat for using LLMs to search papers that he would need to read to make his videos. They want him to use Google search like a caveman in big 2026. The crazy thing is I don't – can you even turn off AI mode in Google now? I think you can. I think you can. You can use DuckDuckGo. I don't think that has AI yet. We'll see. Brian Lowe on X is sharing a heartbreaking essay by a mathematician last week for this most recent news drop. So this was before Noam Brown showed the recent breakthroughs by Astra. He said, there's nothing I can do. There may be nothing you can do. I have no prescriptions, policy recommendations, or coherent call to action. I just want to be honest and open about my emotional and spiritual response. I want to feel seen I want folks like me to feel seen I need the architects of our new mathematical paradigm to look me in the eyes and acknowledge our shared humanity and soul before they deliver the coup de grace I need most of all for us to understand what we are really doing. The Dark Knight of Mathematics, Kerwin Hampshire, mathematician and researcher from the University of Auckland, who recently authored the viral essay, studied mathematics. Interesting. It feels like I would be surprised if mathematical education goes away. It feels like a lot of these problems should be interesting to apply, but I understand that's a completely different discipline. It will be interesting to see what happens next because there are more advanced problems, the Millennium Prize problems, P versus NP, Navier Stokes, right? There are a number of problems that are still unsolved. What happens when they're all solved? Do we create new problems? Where do we go from there? Do we start applying them in different ways? What do you think, Tyler? Yeah, I mean, obviously, like. Any advice for mathematicians? I think, so in this article he says, like, mathematicians are paid to, like, solve theorems. 21-year-old podcaster has advice for math. Exactly. So he says like mathematicians are paid to solve theorems which like obviously I not in academia but like it seems like that it like kind of their job but also it like you in a university right It like to teach it Yeah I mean there plenty of there plenty of math professors that they'd sort of try and solve theorems, but also mostly teach and, you know, don't solve that many theorems. Yeah, and also like, presumably, if you can solve all these conjectures, like, there's going to be new questions that open up. This is like the entire history of all science, right? Yeah, it will be interesting to see the application of this stuff, because it's so abstract at this point, And it feels like it's very – everyone's saying, like, okay, based on this, like, this is going to, like, flood through material science and flood through chemistry and biology. And that would be awesome. Everyone would love, you know, oh, all of a sudden, like, the electric cars have twice as much range because we solved some fundamental thing. It'll be interesting to see where the new bottlenecks are. There, of course, will be always. These math professors hate AI for one simple trick. Just scale, scale, scale, I suppose. It's time. What? It's time to talk about the rise of one person, one dollar companies. Wall Street Journal is saying, AI tools make it easier for founders to get started alone and many stay that way as they grow. Ben Broca launched a company last December that offers AI tools to entrepreneurs. That name's familiar. We've had him on the show. [15:00] He added 10,000 paying customers and is on track to bring in 10 million in revenue this year. One thing he hasn't added, any other employees. The 40-year-old is part of a class of entrepreneurs who are launching and often running new companies on their own. Artificial intelligence tools answer Broca's emails, help write and debug code, field requests from customers, sign up new subscribers, and grant refunds when issues arise. Broca relishes his ability to make whatever decisions he wants on his own, often from his sun-drenched Sausalito, California living room. I think compromises make lukewarm results, he said. Once upon a time, running a business of a certain size required a team. AI is turning that assumption upside down, and more aspiring entrepreneurs are going it alone. An analysis by the payments company Stripe, Tyler, look up Stripe, shows there are thousands of solo operators on the company's platform that are generating over $1 million in revenue with their ranks doubling between 2023 and 2025. That's pretty crazy. So this is on Pulsia, right? No, no, no, no, no, no. Oh, on Stripe. Definitely not. Okay. This is just Stripe. Okay. The number of solo, I'm sure, I would be curious if Pulsia has any companies that do more than, you know. A thousand. A thousand dollars I mean you looking at the marketing Because to be honest like I actually do success for Pulsia is just making back even a dollar more than you spending on Pulsia Yeah. There's a YouTuber who's been demoing different AI systems, Fable and Sol and Kimmy, and saying, like, go make me money is basically the only prompt. And he lets it cook for like a week. and he'll be on like a $200 a month subscription, see if it can make six cents, see if it can make $10. And he's getting closer every time. And he, of course, has to do some things, set up API keys and do little things. But it's an interesting experiment. Ben Awad, you should go check it out. So the number of solo operators, according to Stripe, also crossing the $10 million threshold, nearly tripled in that same span. In the past, people without business contacts or particular savvy might not have known how to get their ideas off the ground. Now AI can be a built-in business partner. Yeah, how does Stripe know if you're a solo operator? Like if you're, because if you're a podcaster and you set up a Stripe account to accept money from advertisers, you could be having a million dollars move through there. But if you hire an editor or not, that doesn't necessarily show up in Stripe. So they must do some sort of polling and ask everyone Yeah I think in your account at some point you say how many employees do you have Okay and if you say yeah I just got one But I guess one question I have with the data is like what if you just set up your Stripe account? It's like, how many employees do you have? And then you wind up adding people and you don't go on update. Yeah, possible. Yeah, because they don't have the payroll. I don't know how they would have visibility into payroll, especially like individual employees. Yeah, yeah. They do have a sense for how many people obviously are like added to your account, but sometimes it's like, oh, true account. If you add your CPA. Yeah. You know, and that's a contractor, not an employee. Yeah. Yeah. AI's ability to handle various administrative tasks makes it potentially useful for launching solo businesses in many fields. But the technology's ability to handle key tasks in tech like coding make that field a particular hotspot. Among all industries, new business applications in the information sector. I've seen the biggest percentage increase nearly 45 percent over the past year. yeah. At the same time, the rate of information sector applicants saying they plan to hire workers has experienced the sharpest decline of any measured industry. This census data set doesn't track solo operated businesses, but the numbers broadly show in tech and beyond that applications are flat among businesses likely to hire workers, but generally rising elsewhere. Economists say that a strong sign that solo operators are in the upswing Wow yeah that chart is really up and to the left New business formation this is in the information sector in particular Tyler, update. Okay, so they basically calculate the number of solopreneurs based on how many people have, there's special plugins or platforms that are directly for the solopreneur. So they basically use that to get a proxy of the general percentage of solo people on Stripe. On Stripe. Oh, they have like a special flow for Solo Printer. Yes. Oh, cool. So they say that they're almost certainly underestimating the number. Hey, Julian Weiser, I know him, says the bar for getting started has never been lower, said Julian Weiser, who runs a San Francisco-based accelerator for Solo founders working in tech. The accelerator, which offers founders seed money and mentorship in exchange for an equity stake, It attracted 4,500 applicants for 10 slots made available in its most recent cycle, nearly five times the number it drew when it launched last May. Now, he's been growing this a lot, but that is staggering. A lot of people want to be solopreneurs. Going it alone with AI can still be surprisingly expensive. Broca said he was losing money on many customers' accounts while paying to access Anthropics Cloud to run his clients' requests that AI [20:00] company as well as others charged based on usage. He has since switched to free open source models from China. Broca says he has raised $30 million from investors and at the same time has saved millions in salary since he hasn't needed a team of software engineers. Another risk, if it's easy for one entrepreneur to launch an AI-assisted business, copying them can be easy too. This creates anxiety for founders like Troy Johnson, who runs an AI-assisted business alone in Orlando, of florida everybody has the sword and we all have the ability to unsheathe excalibur now johnston what a great quote for the journal i love it he's 40 he used an ai he used ai to code an app that helps people get the most out of credit card benefits huh it's interesting pick pick which card you want to use because you might have multiple cards one that's good for dining and you build an app for that there's been a few apps that do that the points guy had a whole blog around it, a media company around it still does. But interesting to sort of like, yeah, go and actually vibe code that. A lot of those things, it's like you could probably just use the models themselves for this, just have a thread that says, hey, these are the cards I have. Go pull all of the data When I about to buy something let me know But at the same time there might be some value for something new with a deeper integration somewhere The company makes around a month in profit with no employees and continuing to grow What a run for Troy Johnson. Great story. Who loves King Arthur related metaphors for business. What one person businesses will mean for the labor market remains to be seen. Polling has shown that Americans are worried that AI will replace jobs and top economists are wrestling with that possibility too. But AI is also creating lots of new jobs, and the go-it-alone entrepreneurs show the technology can both open doors and limit employment opportunities. If everyone's hiring less, but you get four times more firms, what does that do to headcount? Said Rembrandt Koenig, an associate professor at Harvard Business School who studies entrepreneurship. He co-authored a recent study that found that among 50,000 startups the researchers examined, those focused on AI tended to operate with 25% fewer employees. It's interesting because haven't we seen the ramp data that said that AI adopting companies were hiring faster, but maybe they still lower operational headcount but hiring faster because of hiring growth? There's like three different factors that are going on here sort of mixing all together. Koning the professor also believes a soft hiring environment that has left some people mired in log job searches has encouraged more to try their hand at launching businesses That makes sense Some founders cite different motives. It's a perfect storm of post-pandemic burnout and a re-evaluation of one's priorities and also booming AI and a sense of what's possible, says Samir Ahmad, 39, who lives in Brenningsville, PA. Two years ago, Ahmad decided to leave the corporate job he had worked at at Verizon for almost two decades to start a solo coaching and consulting business. He had been seeing social media posts touting the ease and virtues of AI, which he used to chart a business plan and help with marketing. It was like my chief of staff, second in command. The business ultimately petered out within months, though, and Ahmad is back to full-time corporate role with a utility company. For Claire Vo, 41, AI helped turn her passing impulse into a business. She was working full-time as a tech executive when she tapped AI in late 2023 to help code an app that would help manage documentation and design for new products with customers ranging from financial services to healthcare firms. I was copying and pasting from ChatGPT, said Vo, who lives in San Francisco. She put her app online for $1 a month. Wow, that is cheap. And within weeks I thought we didn know how to make apps that cheap anymore I mean it is a subscription at least not one time but she put it online for a dollar a month And within weeks people downloaded thousands of times Nearly three years later Vo company which she ran solo for nine months before hiring an engineer now has 100,000 users and is on track to make seven figures in profit this year. Wow, that's remarkable. At a dollar a month. That's crazy. AI handles the company's marketing, sales and customer support. Well, AI is a shortcut. Vo said her network and credibility in the industry were key. I think people over-index on how easy AI is and under-index on how much I did to get to this point, she said. She's still – Yeah, I just want to see five companies that make more money from their business than they give Pulsia every month. Yes, so Pulsia has some public dashboards for how much people are spending or something like that? Yeah, so right now you can see all the different things that the company is on Pulsia doing, or at least some of them. Right now, so far today, companies on Pulse C have spent $373. Is that today? On ads. Today? Yeah. Oh, it's still morning. It's still morning, so we're pacing. Okay. I don't know what time zone this is in. But yeah, the big question is like, is any of this stuff actually working? [25:00] Or is it more like kind of a video game effectively? Yeah. That people just enjoy like watching the machine hum, but there's not really much happening? I mean, that was the thing for Mid Journey and Suno, I think in many ways. Like Mid Journey, when it launched, people were like... Totally. I mean, no. Okay. Hear me out. Okay, I'll hear you out. Okay. When Mid Journey launched... Get the steel man. When Mid Journey launched, people were like, this is going to take artists' jobs. And it was like, okay, so if that plays out, then I'm going to go to the MoMA, and there's going to be a show for someone that just prompted Mid Journey. And the highest auction at Christie's is going to be some Mid Journey artist. And that's not really what happened. Like, people aren't using Mid Journey to make fine art, but people love Mid Journey. Like, they love the activity of going on Mid Journey and generating and prompting and getting an image back. And then maybe they send it to their friends, maybe they use it a little bit. But it's not exactly the same of like the process of becoming a fine artist. It's more like they're enjoying the process of just making. It more like just having a guitar that you just like to practice and noodle on versus like actually being a touring artist And so like that certainly my experience with Suno is it fun to try and make a song and then listen to it and then be like wrestling with the thing And it's possible that that could be the same activity for like, okay, I'm going to go build an online business, see if I can get something out. But it's not really like a job. It's more of like an entertainment product. I don't know. What do you think? Yeah, I mean, you could easily see it turning into like Ender's Game scenario where it's like a game and then it's like, oh, that was actually a real business you were starting, Ender. You offshored the last job. You sent off labor overseas. You rolled up the entire HVAC industry. That wasn't a simulation. Mark Zuckerberg was answering questions about his AI strategy on the latest Meta earnings call. and Ben Thompson wrote about what Meta's position is in AI, how they're grappling with a few things. There was a bunch of interesting points in this Stratechary update. One I wanted to call out was what Ben Thompson thought the best moment on the call was when an analyst asked him why the company can't just use other models. Like, why can't you just do the Apple thing? Do nothing, win. Like partner with one of the labs do some license agreement If you need an image model you get an image model If you need a text model you get a text model If you need to speed up your programmers or your engineers hire the best coding agent and negotiate with them, right? And here's how Mark Zuckerberg answered it. He said, I can take the open source question. Let's see. So basically the question is, do we think that because there are some open weight models that we can just rely on those? I mean, right now, the open source models are not as strong as the frontier models. Good point. So no is the basic answer. Meta needs to be on the frontier with their intelligence that they use, so they have to be there. There's also just always the perpetual both policy debate and question around other companies' actions and whether that's a thing a company like Meta can rely on. So if you're using Chinese open source and there's some regulatory risk, it seems like that's sort of what he's getting at is these things might not rely. They might not be available all the time. And then also some of these companies, they might be open source for a few years and then go close source and start charging you an arm and a leg. So you don't want to be in a place where you become super dependent and then all of a sudden get hurt once you super dependent on a particular product So he says and I think that very tricky So on both fronts we believe we going to be able to do better work and we think that there some risk in that reliance I don believe that the right thing to do So that felt like not a great answer to me in the sense that the Apple approach seems to be working so well. Then he goes on to explain some of the history of Meta, and it's very, very interesting. He says, I think that we're a company that if you look at Meta from take a step back on this a lot of people view the surface layer Of we build some social media apps and we have an ad business We are really a full stack technology company We had we build our own data centers our own infrastructure our own chips our own low-level software when we get when we got started He says my background in engineering. I wrote a lot of the systems code A lot of the reason why Facebook worked was because it actually it just worked which is a crazy thing to say based on the history, but he makes a really good point. He says, like, it literally worked when other social networks did not work fast and efficiently, and I think we just have the ability to build things that can be more personalized, more optimized, more efficient. Some qualitative experiences are just not even possible for others to build because we go all the way down the stack, and it just seems like it's a good thing. [30:00] pretty clear that having kind of sovereignty over building your own models is going to be an important part of that stack going forward, which is why it's important for Meta. And so that was very interesting that that was a differentiator in the early days, that certain other competitor sites would just be slower. They wouldn't be able to launch new features quickly, and by vertically integrating all up and down the tech stack, they were able to do things very aggressively. This is the Reels thing. They built two extra data centers to be able to do the Reels algorithm, because if they didn't have that compute capacity, they could not have launched a competitor to TikTok on any normal time frame because it was actually a compute-intensive project, not just a design. People see, oh, they just put a new button there, and there's some videos. But it's like behind those videos is a massive recommender system that is very computationally intensive and stores a lot of data, And you can't just spin that up for 3 billion users or however many billions of users they have on a dime if you don't have the infrastructure, if you're not actually vertically integrated. So there's a whole bunch of other things where in terms of personalization, understanding the user, delivering a really first-class experience, they sort of do need to bring it in-house. but investors are very upset about this. They're not very happy because Ben Thompson calls it the financial tail wagging the dog and says that they have to sort of double spend right now. He makes a good point about this. They're double paying. The company right now is basically double paying for infrastructure without a clear path to monetization. And to make matters worse, it's improving monetization story lost a bit of its luster. So they both renting AI compute paying a bunch of money for new researchers and then also spending all the CapEx for the next data center So all of that needs to come together in this moment to actually deliver And the investors are starting to ask all these questions about what the strategy is But Zuck's sticking with it. He's not backing down. But we will see. Yeah, tough position to be in when capital markets don't have a ton of faith. And you just see that in the stock price. The company broadly, right? There's a lot of infighting frustration around just how MSL is treated versus the rest of the company, which is paying for MSL. So Zuck is at war, fighting a war with multiple fronts. It is. It's a big war. Let's go to this Paul Graham post. He had such a wild experience. He bought a book. It was awful. Didn't want it on my shelves, but I couldn't throw it away. So it sat on a table near the door. I know someone that will rip it apart, feed it to a machine, and burn it. I know someone. I don't know them personally, but I know they would love to take this off your hands. Rushing to an appointment this morning, I grabbed it to read it, first mistake, then went to breakfast and had nothing else. So I spent the morning reading the worst book I've known. it's such a funny such a funny just like I don't know it feels like a Curb Your Enthusiasm episode or something like that anyway thank you so much for tuning in to TVPN today we'll see you tomorrow at 11am Pacific we love you Apple Podcasts Spotify it's been an honor sign up for our newsletter TVPN.com goodbye