Midjourney Unveils Medical Scanner as Noam Shazeer Joins OpenAI

Midjourney发布医疗扫描仪，Noam Shazeer加盟OpenAI Midjourney正利用其庞大的现金流，在不依赖风险投资的情况下独立打造高密度声学人体扫描仪。 主持人剖析了Midjourney通过“Midjourney Medical”向医疗健康领域的大胆扩张，分析了创始人David Holz从Leap Motion到AI图像、再到现在高通量超声人体扫描仪的转型历程。他们深入探讨了硬件规格、数据处理需求，以及在全球部署5万台扫描仪以在水疗般舒适的环境中普及常规健康体检的长期愿景。此外，他们还探讨了包括Noam Shazeer和Dean Ball加入OpenAI在内的行业重磅动态，并轻松闲聊了旧金山房产恶作剧及科技圈趣闻。 Midjourney 宣布成立新部门 Midjourney Medical，进军医疗硬件领域；创始人 David Holz 依托公司完全独立运作积累的充沛现金流，自主资助前沿人体扫描仪研发。 (1:01) 该超声扫描设备配备拥有 35.8 万个传感器的超高密度换能器阵列，具备皮米与飞米级亚原子分辨率，能以 17 GB/s 的速率流式传输数据，并由 2 petaflops 算力集群处理高达 806 TB 数据完成半毫米精度的 3D 重建。 (12:28) Midjourney 设想打造由 5 万台扫描仪构成的全球网络，每月提供 10 亿次类似水疗体验的无感常规健康扫描，并计划未来探索利用高强度聚焦超声开展无创手术与细胞干预。 (14:53) 重磅 AI 人才动态频出：Transformer 奠基论文作者之一、前 Google DeepMind 高管 Noam Shazeer 携手 AI 政策专家 Dean Ball 正式加盟 OpenAI。 (19:53)

TBPN reports that Midjourney is expanding into medical hardware with an advanced acoustic body scanner, funded independently through its generative software cash flows, alongside key talent acquisitions by OpenAI including Noam Shazeer and Dean Ball. Midjourney Medical's Acoustic Hardware Architecture (12:28) Midjourney Medical unveiled an acoustic full-body scanner engineered around a 70-centimeter ring housing 40 integrated systems and 358,000 ultrasonic transducers. The apparatus operates within a fluid-filled tank where acoustic waves travel at 1,481 meters per second, traversing the ring in 480 microseconds. Dynamic controllers execute 100 million adjustments per second while firing up to 1,000 pulse bursts per second, allowing acoustic energy to dissipate through tissue before subsequent waves trigger. The system's acoustic sensors are designed to resolve reflections down to subatomic picometer and femtometer scales, outputting raw wave telemetry at 17 gigabytes per second. Capturing a single anatomical cross-section moves more than 40 gigabytes of data through the system to map biological tissue across 25 perspective angles. A standard 60-second procedure yields several hundred slices as a patient passes along a stationary platform. To convert this telemetry into a 3D volumetric map with 0.5-millimeter spatial resolution, the architecture relies on an on-site compute cluster of 21 servers delivering two petaflops of processing capacity to ingest up to 806 terabytes of raw data. Clinic Alternatives and High-Intensity Therapeutic Potential (14:53) Midjourney aims to bypass traditional clinical diagnostics by deploying scanners in spa-like wellness facilities reminiscent of saunas, removing the need for on-site physicians during scans. Initial deployment plans center on a flagship San Francisco location equipped with roughly 10 operational units, serving as a template for a long-term goal of 50,000 autonomous scanners conducting one billion scans per month globally. Beyond diagnostic imaging, the hosts highlighted emerging discourse surrounding high-intensity focused ultrasound. Observers suggest that dense acoustic transducer arrays capable of high-resolution biological sensing could theoretically invert their function from 'reading' tissue to 'writing' it, utilizing targeted acoustic energy to ablate tumors, dissolve pathological tissue, or stimulate cellular responses without invasive surgery. Bootstrapped Capital and Founder-Driven Hardware R&D (4:55) Midjourney founder David Holz previously developed deep-learning hand-tracking interfaces at Leap Motion, which twice rejected acquisition bids from Apple before struggling with consumer scale and supply chain friction. Applying those lessons, Holz structured Midjourney without venture capital, relying on an active Discord-based community where four-image prompt grids generated massive human-preference training flywheels. Sustained profitability and major licensing agreements—including a significant deal with Meta—provided Midjourney with the capital reserves needed to finance capital-intensive hardware development internally. Because Holz retains majority ownership, the company avoids external venture capital pressures, enabling long-cycle hardware development paired with Holz's hands-on product design, 3D visualization, and aesthetic direction. Strategic Talent Acquisitions at OpenAI (19:53) In frontier AI developments, researcher Noam Shazeer departed Google DeepMind—where he served as Vice President of Engineering and Gemini co-lead—to join OpenAI. Shazeer is widely recognized for co-authoring the foundational 2017 Transformer paper 'Attention Is All You Need', pioneering sparse Mixture-of-Experts (MoE) architectures, and co-founding Character.ai. OpenAI simultaneously hired governance specialist and frequent TBPN commentator Dean Ball into its policy division. Together, these moves signal an aggressive consolidation of top-tier model architecture experience and domestic regulatory strategy within OpenAI as competitive pressure across frontier labs intensifies. While Midjourney's software cash flows provide rare operational independence to pursue medical hardware, the company's leap from generative graphics to clinical-grade acoustic scanning faces steep technical and logistical hurdles in scaling toward a 50,000-device global fleet.

TBPN播客节目重点剖析了Midjourney依托充沛现金流进军医疗硬件领域并发布声学全身扫描仪的战略布局，同时探讨了Noam Shazeer与Dean Ball加入OpenAI等关键行业人才变动。 从Leap Motion到Discord：David Holz的独立造血路径 (1:01) 主持人回顾了Midjourney创始人David Holz的技术演进背景。Holz曾拒绝NASA博士项目创办Leap Motion，在2010年代初期便利用反向传播深度神经网络进行手势实时空间追踪，期间两度拒绝苹果收购，但受限于当时VR硬件供应链与生态未成熟，未能实现大众消费级规模。此后Holz创立Midjourney，打破传统硅谷风投模式，完全依靠自筹资金且不设传统网页或移动端应用，直接依托Discord多人公共服务器运营。Discord的共享环境让用户实时借鉴复杂提示词，而默认生成四张候选图供用户放大的机制，为Midjourney构建了无与伦比的人类偏好数据闭环。 由于几乎全资控股并获得巨额商业授权收益（包括与Meta的大规模交易），Midjourney拥有极为丰沛的现金流。这使Holz能够摆脱外部VC在18个月募资周期和季度KPI上的束缚，将资产负债表之外的利润直接投入前沿硬件研发，催生出全资打造的Midjourney Medical分支。 亚原子精度与每秒17GB数据流的声学硬件架构 (12:28) Midjourney Medical推出了一款旨在革新传统成像的超声波人体扫描仪。其硬件核心为一个直径70厘米的圆环，由40个集成子系统组成，容纳了多达358,000个超声波换能器。系统以每秒1亿次的采样速率进行动态微调，换能器轮流发射最高每秒1,000次的脉冲波阵列，声波在充液水箱中的传播速度为1,481米/秒，横穿整个水箱耗时480微秒（约两千分之一秒），并在下一个脉冲发射前在组织内自然耗散。据介绍，该阵列的传感器分辨率突破了微米与纳米尺度，能够捕捉皮米乃至亚原子飞米级别的声学微动反射，在人体回波过程中实现每秒17 GB的高速原始数据流输出。 在计算与三维重建方面，系统利用多探头从25个不同结构角度发射声波，单个人体切片的数据处理量就超过40 GB。患者在平台上匀速穿过固定环形舱的60秒内，机器可采集数百个切片。后台由21台服务器构成的计算集群提供2 PFLOPS算力，累计处理多达806 TB原始数据，将最终的三维人体映射精度推至0.5毫米级别。Holz更亲自动手使用Blender制作3D技术演示并谱写背景音乐，试图打破传统医疗设备冰冷刻板的形象。 5万台水疗式网络愿景与非侵入超声治疗潜力 (14:53) Midjourney对该扫描仪的长期规划极具扩张性，目标是构建一支拥有50,000台设备的全球机队，实现每月10亿次扫描，理论上可覆盖全人类的定期筛查需求。在服务体验设计上，Midjourney计划打破医院传统诊室的压迫感，首站选址旧金山设立包含约10台设备的旗舰店，将体检流程包装成类似出入水疗中心、蒸汽房或桑拿室的放松体验，且全自动化过程无需医生在场操作。 节目进一步探讨了高强度超声技术从生物结构“读取”跨越至“写入”的医学前景。除了高精度诊断成像，聚焦超声波在理论上具备无需开刀即可无创修改生物组织的能力，例如定向粉碎体内肿瘤和癌细胞。针对外界评价这种跨界是无VC约束的特权，主持人辩称顶级创始人如Elon Musk和Sam Altman在VC体系内同样拥有探索激进项目的自由，但Midjourney在巨头环伺下依靠单一独立实体取得商业成功并跨界硬件，依然构成了极具冲击力的叙事突破。 顶级AI人才重组：Noam Shazeer与Dean Ball加盟OpenAI (19:53) 在硬件议题之外，播客讨论了近期AI行业的重要人事地震。传奇人工智能科学家Noam Shazeer离开Google DeepMind的工程副总裁兼Gemini联合负责人职位，正式加入OpenAI。Shazeer的技术履历极为亮眼，不仅是开创现代大语言模型基石的Transformer论文《Attention Is All You Need》的共同作者，也是T5模型、Switch Transformer的核心作者、稀疏混合专家（MoE）架构的先驱，并曾联合创办character.ai。该变动被业界普遍视为年度最具分量的人才迁移之一。 紧随Shazeer之后，知名AI政策学者Dean Ball也宣布加入OpenAI政策团队。节目还梳理了湾区近期的技术文化动态，包括创作者Riley Walls买下旧金山公共街道并由办公软件公司Notion以4万美元竞得冠名权更名为“Notion Way”，以及政商领袖互动轶事与SpaceX上市后南加州科技生态的持续扩张。 本期节目展现了从生成式AI软件到尖端医疗硬件的罕见跨界范例，突显出独立资本在探索高技术风险项目时的自由度。然而，Midjourney声学扫描仪在临床医疗有效性、FDA等监管审批落地以及50,000台全球设备网络的工程可行性上，仍有待未来真实数据与医疗场景的进一步检验。

## Background 本期节目由科技对谈频道 TBPN 于 2026 年 6 月 23 日发布在 YouTube 平台。视频简介显示，这是其精简版节目“Diet TBPN”，旨在用 30 分钟浓缩每日科技直播脱口秀的精髓，该节目由 John Coogan 与 Jordi Hays 共同主持，被《纽约时报》誉为“硅谷的新宠”，长期聚焦前沿科技脉搏并曾专访多位硅谷重磅领袖。 ## TL;DR 节目核心聚焦 Midjourney 宣布进军医疗领域并推出革命性的超声波全身断层扫描仪，深入剖析了创始人 David Holz 从 Leap Motion 硬件泥潭到自筹资金打造商业奇迹的传奇历程与技术愿景。随后，节目解读了 Transformer 核心作者 Noam Shazeer 与政策专家 Dean Ball 携手加盟 OpenAI 所引发的 AI 领域人才地震，并深入对比了自筹资金与传统 VC 融资模式在战略自由度上的巨大鸿沟。最后，主持人们还分享了旧金山“Notion Way”街道拍卖、特朗普在 G7 现场向 Sam Altman 请教升降椅用法的搞笑轶事，以及关于社交媒体黑粉算法与法拉利首款电动车的辛辣吐槽。 ## Atmosphere & style 整期节目洋溢着快节奏、高信息密度的硅谷极客圈内人氛围，谈话风格犀利随性又充满网络迷因式的幽默调侃。主持人们在深入剖析飞米级微动感知与数十 PFLOPS 算力等硬核技术参数的同时，极擅长用平实生动的比喻拆解前沿科技，并随时穿插行业八卦与大厂槽点。听众仿佛置身于一场硅谷顶级创业者与研究者的深夜酒吧畅聊，在捧腹大笑之余又能精准捕捉到 AI 与硬件浪潮最鲜活的脉动。 ## Bulletin - [12:30](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=750s) Midjourney 正式公布医疗硬件部门 Midjourney Medical，推出基于 35.8 万个超声传感器的全身三维断层扫描仪，能在 60 秒内以 0.5 毫米分辨率完成全身三维成像。 - [16:31](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=991s) Midjourney 计划将首批扫描仪部署在水疗中心（Spa）而非医院，未来更有望利用超声波的精准聚焦能量从“只读”健康筛查升级为无创切除病灶的“读写”治疗能力。 - [19:54](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=1194s) Transformer 核心合著者、稀疏 MoE 架构先驱 Noam Shazeer 宣布离开 Google DeepMind 加盟 OpenAI；政策专家 Dean Ball 亦同期宣布入职 OpenAI。 - [22:06](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=1326s) 旧金山一条无名街道的公共命名权以约 4 万美元竞拍成交，由办公软件 Notion 拿下并正式命名为“Notion Way”。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=0s) Midjourney 进军医疗设备并非困境下的业务转型，而是创始人 David Holz 依托充足自生现金流展开的雄心勃勃的硬件扩张计划。 - [03:23](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=203s) David Holz 早年曾创办手势追踪硬件公司 Leap Motion，在深度学习未成显学时就应用了神经网络，曾两度在最后一刻拒绝苹果的收购要约。 - [04:59](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=299s) Midjourney 凭借零 VC 投资、长期无独立前端/APP、寄生 Discord 的反常规模式，通过多人协作网络效应与提示词共享获得了惊人的商业成功与高额利润。 - [07:19](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=439s) 单次生成四张低分辨率候选图的交互设计，不仅为 Midjourney 节约了推理算力，更为其构建起用户偏好反馈数据飞轮，奠定了其自我造血的技术与商业底气。 - [17:58](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=1078s) 自筹资金（Bootstrap）模式让 Midjourney 彻底摆脱了风投基金每 18 个月融资一次的指标死循环，拥有充裕自由度去进行科幻级硬件的豪赌。 - [12:30](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=750s) Midjourney Medical 扫描仪由 40 组系统组合成 70 厘米圆环，集成 35.8 万个高频换能器芯片，感知灵敏度达到皮米乃至飞米量级，数据采集速率达 17 GB/s。 - [12:30](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=750s) 系统后台配备 21 台专用服务器（2 PFLOPS 算力），在 60 秒扫描中消化 806 TB 原始数据，可自动辨析 25 种人体组织结构，达成 0.5 毫米级别的内部组织重构。 - [16:31](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=991s) 团队设定了全球部署 5 万台扫描仪、实现每月 10 亿次扫描的终极目标，并计划通过水疗中心等放松环境革新传统医院冰冷压抑的体检体验。 - [17:58](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=1078s) 该硬件的超高声波精度未来有望用于无创外科治疗，如同极紫外光光刻般通过精准聚焦声波直接在体内震碎病灶细胞，实现从“只读”到“写入”的跨越。 - [19:54](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=1194s) Noam Shazeer 从 Google DeepMind 跳槽至 OpenAI，作为奠定大模型基础的顶级技术架构师，其流动被业内视为年度最轰动的 AI 人才地震。 - [19:54](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=1194s) 政策学者 Dean Ball 同期加入 OpenAI，表明该公司正持续加大在国家安全、宏观政策与行业治理维度的资源倾斜。 - [22:06](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=1326s) 旧金山无名小街被成功竞拍为“Notion Way”，成为一场避免了投机币圈广告侵蚀、兼具城市公共美感与品牌创意的经典营销范例。 - [23:35](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=1415s) 特朗普在 G7 现场直接询问 Sam Altman 如何调节座椅高低的插曲，生动体现出顶级前沿技术与现实应用扩散之间的巨大断层以及科技慢速起飞的常态。 - [25:03](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=1503s) 在当今社交平台算法机制下，黑粉的负面攻击同样会为人物推高互动量与曝光度，使争议与流量深度绑定。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=0s) 节目开场：破次元壁的嘉宾阵容与 Midjourney 的“非典型转型” 节目开场气氛热烈，主持人们首先抛出了今天科技圈的重磅炸弹：Midjourney 宣布进军医疗设备领域。但这并非一次仓惶的业务掉头，而是一次业务扩张，被视作创始人 David Holz 极其耀眼的创业“第二幕”乃至“第三幕”。在深入探讨这款医疗扫描设备之前，两位主持人激动地宣布了当天的直播嘉宾阵容，尤其是 Derek Thompson 与网红拳手 Jake Paul 破天荒地登上同一档播客节目，这让大家惊呼“很多人原本认为这绝不可能发生，他们完全属于两个不同的世界，但今天我们把他们聚在了一起”。当天的阵容还包括 Arm 首席执行官兼软银董事会成员 Rene Haas，以及著名播客主持人 David Senra，可谓群星云集。 随后话题迅速切回正题——Midjourney 正式宣布成立全新业务部门“Midjourney Medical”。团队成员 Tyler 甚至亲临了发布会现场，带来了一手情报。主持人们指出，要理解为什么一家顶级 AI 图像生成公司会突然做起医疗硬件，就必须追溯创始人 David Holz 那段极其传奇且硬核的个人历史。 ### [01:18](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=78s) David Holz 的早年经历与 Leap Motion 的技术萌芽 David Holz 在佛罗里达长大，父亲是一名牙医。节目中主持人们还不忘拿演播室团队里同样是牙医之子的同事开玩笑，调侃他们一口好牙是所谓的“裙带牙齿”（neppo teeth）。David 从小就沉迷电子游戏，但他不是单纯地玩，而是很早就开始动手破解游戏、钻研计算机技术。他拥有极高的数学天赋，甚至一度攻读数学博士学位。但他最终选择放弃博士学位，前往 NASA 工作。 离开 NASA 后，David Holz 踏上了创业征途，创立了 Leap Motion。那是一次非常典型的风投支持型创业，Founders Fund 也参与了投资，但整个过程充满艰辛。Leap Motion 最初的产品是将精细的手势追踪技术引入 VR 领域。尽管当时 Leap Motion 的产品与如今庞大的医疗扫描设备看似天差地别，但其底层关于传感探测、空间建图和算法处理的核心逻辑却一脉相承。早在 2010 年代初，David Holz 就在 Leap Motion 中应用了深度神经网络和反向传播混合模型，在深度学习尚未成为显学的年代，其技术视野就已经极其超前。 ```mermaid flowchart LR accTitle: David Holz 创业演进图 accDescr: 展示 David Holz 从早期学术背景、Leap Motion 硬件手势识别到创立 Midjourney 及其医疗扩展的历程。 subgraph EarlyStage ["早期探索"] A["数学背景 / NASA 研究"] --> B["创立 Leap Motion"] end subgraph LeapMotionEra ["手势与传感交互"] B --> C["突破性深度学习/手势追踪"] C --> D["硬件供应链与市场瓶颈"] end subgraph MidjourneyEra ["生成式 AI 与新赛道"] D --> E["创立 Midjourney"] E --> F["Discord 社区与数据飞轮"] F --> G["扩展硬件：Midjourney Medical"] end ``` ### [03:23](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=203s) 硬件泥潭与两次告吹的苹果收购案 Leap Motion 诞生的时代，整个开发者社区都被 Oculus 激发的 VR 浪潮所点燃，然而做硬件业务却异常艰难。当时硬件供应链远未成熟，制造成本高昂，而且还要面对巨头的竞争挤压——Meta 收购了 Oculus，苹果也在暗中重金投入。其间，苹果曾先后两次试图收购 Leap Motion，甚至一度到了极其临近签约的第 11 个小时。有报道透露，苹果当时甚至已经为所有 Leap Motion 员工印好了入职欢迎手册，准备当天直接办手续，但交易最终却在最后一刻泡汤。 据传是 David Holz 本人拒绝了苹果的收购要约。他在深入接触后认为苹果“过于大企业官僚化”，整体氛围不符合团队预期，进去后很难获得成功。收购告吹后，公司经历了艰难的裁员阵痛，但他们依然坚持为极客和开发者制造设备。那个看起来像一条口香糖大小的 Leap Motion 感应器，虽然可以在电脑前隔空切《水果忍者》，但它始终停留在开发者套件阶段，从未能像 Oura Ring 或 Whoop 手环那样实现真正的大众消费级普及。在苦苦支撑近十年后，David Holz 终于得以抽身，开始构思他的下一个项目。 ### [04:59](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=299s) 离经叛道的 Midjourney：无前端、全依靠 Discord 的多人协作神话 在开启第二幕后，David Holz 抓住了 AI 图像生成的浪潮创立了 Midjourney，并取得了巨大的病毒式传播成功。令人惊叹的是，Midjourney 彻底颠覆了硅谷传统的创业法则：没有拿风险投资，最初几年没有专门的前端、没有独立网站、没有独立的移动 App，所有的交互居然全部放在一个 Discord 服务器里。他们把这个服务器一路扩张到数百万成员，成为 Discord 历史上规模最大的社区之一，以至于迫使 Discord 官方不得不专门为其升级后台系统架构。 这种寄生在 Discord 上的模式带来了意料之外的杀手级优势：第一天就天然拥有了覆盖全平台的高质量移动体验，更重要的是它创造了一种不可替代的“多人协作社交体验”。正如 David Holz 曾指出的观点：如果你只给用户一个空白的搜索框，告诉他们“这是一台能生成任何图像的神奇 AI 机器”，大多数人只会茫然地输入一个词——“狗”，然后得到一张普通的狗的照片，这跟去 Google 搜图毫无区别；AI 真正的魅力在于生成现实中不存在的东西，比如“宇航员在月球上骑马”，或者在自己的家乡开 F-16 战斗机，这些在以往需要庞大的 CGI 预算或 Photoshop 大师耗费心力才能合成的画面，如今只需一条提示词。在 Discord 频道里，所有人都能实时看到别人精心设计的复杂提示词，偷师那些多层嵌套的风格参考代码（`--sref`）和构图参数，然后立刻“二创混剪”成属于自己的版本。 - **反传统商业打法**：无外部 VC 投资、长期无独立前端/APP、完全依赖 Discord 承载核心产品。 - **多人协作网络效应**：公开频道的提示词交互让新用户能够无门槛借鉴高阶玩家的 Prompt 结构和样式参数。 ### [07:19](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=439s) 四图生成的数据飞轮与进军医疗的雄厚底气 Midjourney 在产品体验上的另一大开创性设计是单次生成四张低分辨率候选图的交互流程。系统不会一开始就直接给用户输出一张高清大图，而是先生成四张预览图，让用户选择最符合预期氛围、风格和构图的那一张进行放大或迭代。这种机制不仅优化了算力消耗，更为 Midjourney 带来了极其宝贵的高质量偏好数据反馈，从而构建起飞轮效应。相比之下，当时开源且支持本地运行的 Stable Diffusion 虽然效果出色，但由于运行在用户本地，无法像 Midjourney 这样在中心化服务端沉淀下庞大的用户反馈闭环。 时至今日，据外界所知 Midjourney 依然没有接受任何风险投资，公司的大部分甚至全部股权仍由 David Holz 掌控。凭借惊人的利润率和充沛的现金流，团队完全有资本去自主资助其他看似疯狂的探索项目。而事实上，打造一款医疗硬件设备的构想，David Holz 已经在内部酝酿并谈论了多年，如今这一宏大愿景终于浮出水面。 ### [08:22](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=502s) Midjourney 的硬件野心与 Midjourney Medical 宣传片 主播们首先探讨了 Midjourney 令人咋舌的财务底气。虽然官方从未公布过确切的财务数据，但保守估计其营收已达数亿美元，甚至“可能达到数十亿美元”。这种无需依赖任何外部风投资金的充沛现金流，让 Midjourney 拥有媲美顶级重度融资私有公司的资产负债表，足以放手尝试极具野心的“大力挥棒”（big swings）。原本大家最期待的其实是一款 VR 头显——在苹果发布 Apple Vision Pro 之后，创始人 David Holz 甚至曾公开发表评论表示“看来我们不得不自己动手做一款了”。事实上，Holz 此前放出了大约八张硬件预热图，这意味着还有一系列神秘硬件正在酝酿中，而今天率先揭开面纱的正是其涉足医疗领域的重磅项目：Midjourney Medical。 随后，直播间播放了 Midjourney Medical 发布的官方宣传视频。主播们一边开着玩笑确认画面里的山羊是他们自己直播流里的元素而不是 Midjourney 官方素材，一边迫不及待地拆解视频传达的视觉风格与背后深意。 --- ### [09:20](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=560s) 线下发布会体验与 David Holz 的“品味”和“气场” 谈及线下发布会的实际体验，亲临现场的主创透露，整个展厅延续了宣传片中温暖而昏黄的灯光色调。由于真实的扫描仪体积过于庞大无法直接搬入室内，现场展示了还原其巨大圆环管道的模拟装置，并提供了一台较小的原型机供人将手伸入体验。现场甚至有人通过低分辨率的局部手部扫描，当场查出自己因打拳击而留下的手腕隐患。主播们随即开启了硅谷式幽默：“其实写个 if 语句就行——如果住在旧金山，那么判定为腕管综合征。” 随后，讨论深入到了 Midjourney 的审美内核。主播指出，每当外界争论“AI 生成的内容到底算不算艺术”时，他始终将 Midjourney 视为一场属于 David Holz 个人的艺术展：“作为提示词输入者的你，仅仅是漫步在‘David 博物馆’里的参观者；真正的主创和策展人是 David Holz 及其团队。”Midjourney 的图像之所以始终带着高级的质感，是因为它并非平庸的图库生成器，而是被注入了极度鲜明的个人品味。就连这部充满科幻色彩的官方宣传片，也是 David Holz 本人亲自举着摄像机拍摄完成的。 在医疗行业层面，这种跨界带来了巨大的冲击。传统医疗器械行业长期给人一种冰冷死板、缺乏审美与生命力的感觉，主播直言“医疗器械领域毫无气场（Aura）与动势（Motion），而 David Holz 正在给这个行业注入灵魂”。同时，这也给所有创业者上了一堂生动的公开课——“永远不要对初创公司陷入绝望（never black pill on startups）”。Midjourney 不拿 VC 资金，也绝不考虑出售公司，它的入局不仅不会抹杀一切，反而逼迫现有的十多家生物医疗器械公司紧急行动起来，去和更多创业公司展开合作或收购。 --- ### [12:30](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=750s) 技术深度剖析：超声波全身扫描仪的硬核参数 随着第二段技术深度视频的播放，更多硬核硬件参数浮出水面。整部技术动画由 David Holz 亲自利用 3D 工具渲染，并配上了极其惊艳的背景音乐。视频没有旁白配音，完全依靠滚动的技术说明，全方位展示了其革命性的超声波全身断层扫描架构。 ```mermaid flowchart TD accTitle: Midjourney Medical 超声扫描与三维重建流程 accDescr: 展示从超声波传感器阵列发射、信号采集到服务器群AI重建三维断层图像的完整架构。 subgraph Hardware["硬件发射与感知环"] A["40组系统组合环<br/>(直径70cm / 35.8万传感器)"] B["换能器高频受控发射<br/>(1亿次/秒控制率 / 声速1481m/s)"] C["皮米/飞米级微动感知<br/>(480微秒穿透水箱)"] A --> B B --> C end subgraph DataPipeline["数据采集与传输"] D["实时吞吐采集<br/>(17 GB/s 流量)"] E["单层断面原始声波数据<br/>(超过 40 GB)"] C --> D D --> E end subgraph ComputeCluster["算力集群与AI三维重建"] F["21台计算服务器<br/>(2 PFLOPS 算力 / 806 TB 数据)"] G["AI解构识别<br/>(解析25种生物组织结构)"] H["全身3D高清映射<br/>(60秒扫描 / 分辨率达0.5mm)"] E --> F F --> G G --> H end ``` 整个扫描系统的核心是一个由 40 组系统紧密构建、直径 70 厘米的圆环阵列，内嵌高达 35.8 万个超声波传感器芯片。传感器中的换能器兼具扬声器与麦克风功能，其控制响应速率可达每秒 1 亿次，以结构化声波形式向人体发射。视频中给出了一个生动的比喻：“想象一个由 9,000 名歌唱家组成的合唱团，同时有 9,000 名听众在全神贯注地倾听。” - **极速微观感知**：声波在介质中以每秒 1,481 米的速度传播，穿透整个水箱仅需 480 微秒（约两千分之一秒）。各换能器轮流发射，并在声波消散期间以高达每秒 1,000 次的频率连续触发；传感器灵敏度不仅突破了微米和纳米级，达到皮米（picometer）量级，数百个传感器协同工作时甚至能逼近亚原子尺度的飞米（femtometer）级别。 - **庞大的吞吐与算力支撑**：整套系统以每秒 17 GB 的速率捕获全身声波回响，仅看清身体的一个横切面就需要处理超过 40 GB 的数据。 - **机械构造与扫描流程**：受检者站立在升降平台上缓慢下降穿过固定的圆环，在 60 秒的扫描周期内系统获取数百个横截面断层。 - **三维重构与组织识别**：后台配备了 21 台专用服务器，输出 2 PFLOPS 的强大算力，可消化高达 806 TB 的原始数据。利用 AI 技术能够自动辨析识别出 25 种不同的体内生物器官与组织构造，最终生成分辨率高达 0.5 毫米的内部组织 3D 映射图。 - **全球规模愿景**：如果十几台此类设备满负荷运转即可覆盖整个城市的体检需求；Midjourney 设立的终极目标是在全球部署 50,000 台扫描仪，实现每月 10 亿次扫描，理论上足以覆盖全人类的常规筛查。 --- ### [16:31](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=991s) 体验革命：从冷冰冰的诊所走向“水疗中心”与非侵入式治疗 针对 Midjourney 计划将首批扫描仪部署在“水疗中心（Spa）”而非传统医院的决定，主创们展开了分析。首个旗舰空间选址旧金山，预计将安装大约 10 台扫描仪，未来某些大型场所可能部署数百台，小型场所则配置一两台。David Holz 之所以选择水疗形态，是因为传统医院和诊所的体验冷酷压抑，而这种全自动化的扫描过程既快速又无需医生在场守护，受检者在享受类似桑拿或蒸汽室的放松氛围中即可完成深度健康体检。主播调侃道：“他们甚至应该在圆环里装上一圈自动注射针头，扫描完成后根据个人体质自动注射胜肽、激素或运动表现增强剂（PEDs）。” 除了作为强大的“读取（Read）”工具，这套硬件的性能甚至被评价为过于超标（overpowered）。由于其对声波能量的精准控制力，理论上它完全能够延伸至“写入（Write）”能力——即无需任何开刀切口的非侵入式外科手术。就像极紫外光（EUV）光刻通过高精度聚焦来雕刻微观结构一样，超声波如果以极度聚焦的方式精准汇聚于病灶，能够直接在体内将癌细胞或肿瘤震碎气化。这预示着该系统未来不仅能用于全面体检，还有望演化为下一代无创声学治疗平台。 ### [17:58](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=1078s) 自筹资金 vs 风投模式：Midjourney 的超声波野心 对话承接上一部分关于硬件与医疗设备的畅想。主持人打趣说，有了这类设备甚至可以顺势转向芯片制造领域，毕竟芯片更赚钱。随后，主持人引述了 X（Twitter）平台上的好友账号 Anabology 的观点：一旦研发出全身超声波扫描仪，显而易见的下一步就是利用超声波对人体进行干预治疗——包括精准消除组织、促使细胞分裂、重编程细胞，实现从单纯的“只读”（Read）升级为“可读可写”（Read + Write）。该账号还公开招聘有志于攻克这类科幻级医疗器械“写端”能力的工程师。 这引发了关于初创企业自由度与融资模式的热烈讨论。网络舆论普遍认为，Midjourney 涉足超声波扫描仪，是“自筹资金（Bootstrap）”公司所独享自由的绝佳范例——这种需要承担全部责任且带有一点“玩世不恭（devil-may-care）”态度的豪赌，通常是拿风投资金（VC-backed）的公司绝不敢尝试的。不过主持人也提出反例：像 Elon Musk 的众多项目，或是 Sam Altman 运作的方式，都是在充分利用 VC 资本游戏——为新想法组建新团队、独立拿钱在外部孵化，从而获得去造登月“质量加速器（mass driver）”等疯狂事物的资本与特权。 然而，两位主持人依然承认，绝大部分普通创始人确实被深锁在 VC 的莫比乌斯环中：每隔 18 个月就必须融一次资、疲于奔命去达成核心 KPI、为了生存打得头破血流，既没有喘息的时间也没有充裕的现金流。正因如此，Midjourney 能够在外界普遍视作同质化严重、红海厮杀的生成式图像市场中（不仅要正面对抗 Google 和 OpenAI 等巨头），不拿外部风投，仅凭自身造血就成长为一台盈利机器，打磨出极其忠诚的用户社群与坚实的商业模式，这完全打破了主流创业叙事，令人赞叹。 ```mermaid flowchart TD accTitle: 创业公司模式与战略自由度对比 accDescr: 对比了VC风投模式的死循环与自筹资金模式在探索科幻医疗等激进业务上的战略自由度 subgraph VC_Loop ["VC 融资模式循环"] V1["每隔 18 个月融资"] --> V2["紧盯单一核心 KPI"] V2 --> V3["缺乏充裕自由现金流"] V3 --> V1 end subgraph Bootstrap_Path ["Midjourney 自筹资金模式"] B1["强劲自身造血与现金流"] --> B2["摆脱资方指标束缚"] B2 --> B3["激进跨界：超声波全身扫描"] B3 --> B4["终极愿景：细胞与组织读写干预"] end ``` --- ### [19:54](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=1194s) AI 人才地震：Noam Shazeer 与 Dean Ball 加盟 OpenAI 话题随即转向了轰动整个科技圈的人事变动：传奇 AI 研究员 Noam Shazeer 宣布离开 Google DeepMind，正式转投 OpenAI。主持人发出好奇的惊叹：“Noam 究竟看到了什么？”Noam 在其离职声明中客气而真诚地表示，这是一个艰难的决定，他为在 Google 团队共同取得的成就深感自豪，同时也对加入 OpenAI 卓越团队的新旅程充满期待。 现场主持人开玩笑地提到了 X 网友 Yaxine 的调侃——对方一本正经地科普 Noam 的身高有 6 英尺 4 英寸（约 1.93 米），主播打趣道这简直是在“附会塑造我们的技术英雄”。但抛开玩笑，Noam Shazeer 的技术资历无可挑剔：他是奠定现代大模型基础的 Transformer 论文《Attention Is All You Need》的核心作者之一，主导了 T5 和 Switch Transformer 等里程碑式论文，更是稀疏专家混合（Sparse MoE）架构的开创者。此前他担任 Google DeepMind 工程副总裁及 Gemini 联合负责人，他的出走被业内广泛誉为“年度最重磅的 AI 人才流动”，引发了公众对 Google 内部现状的诸多揣测。 紧接着第二天，政策专家 Dean Ball 也宣布加入 OpenAI。与深耕算法的 Noam 完全不同，Dean 是一位在 AI 政策与治理领域深耕、见解平衡客观的专家。主持人提到 Dean 多次做客节目，他之所以对行业内各家公司都直言不讳地提出批评，核心动力源于他极度在乎美国作为国家整体在 AI 浪潮中能否“把路走对”。两位顶尖人物的连续加盟让全网瞩目，甚至连财经名嘴 Jim Cramer 都在凌晨三点发帖感叹“顶级 AI 思想家跳槽”，主持人戏谑 Cramer 把 OpenAI 直接简称为“AI”是省掉了“Open”，毕竟如今都是闭源模型，“直接改叫 AI 公司反而更纯粹”。 - **Noam Shazeer 技术功勋**：Transformer 论文合著者、T5 与 Switch Transformer 核心作者、稀疏 MoE（Mixture of Experts）先驱。曾名列 TBPN 在去年 AI 人才大战高峰期发布的“Metis List”榜首。 - **Dean Ball 的定位**：长期关注 AI 国家战略、公共政策的平衡派专家，其核心使命是推动美国在 AI 竞赛与治理中保持正确方向。 --- ### [22:06](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=1326s) 4 万美元拍下旧金山街道：“Notion Way”的诞生 随后，节目聊到了旧金山当地一则充满网络恶搞精神但结局极佳的趣闻：Riley Wells 完成了一项不可思议的举动——他拍下了旧金山一条无名小街的冠名权并组织线上竞拍，最终协作软件公司 Notion 胜出，这条街被正式命名为“Notion Way”（Notion 之路）。甚至连旧金山市长丹尼尔·卢里（Daniel Lurie）都给出了官方祝福，路面上也已经刷上了崭新的街道名称。 主持人对此赞不绝口，认为由 Notion 拍得简直是最好的结果。因为把这类公共命名权抛到互联网上公开拍卖极具风险，一不小心就可能落入投机炒作的加密货币公司之手，变成令人厌烦的商业牛皮癣。而“Notion Way”听起来宛如一条顺畅而富有哲理的正常街名，即使住在这条街的拐角处也不会觉得突兀反感。主持人还回忆道，当 Riley 最初带着这个想法找他们聊时，他们就非常看好，甚至曾承诺如果最后流拍无人出价，节目组愿意出资一两万美元作为“最后兜底买家（buyer of last resort）”。最终该项目以约 4 万美元的价格成功交割并落地执行，堪称一次教科书级别的营销行为艺术。 --- ### [23:35](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=1415s) 达沃斯与 G7 轶事：从 Aidan Gomez 到川普问椅 谈及当代顶级 AI 研究者，主持人们顺带提到了 Cohere 联合创始人 Aidan Gomez（同样是 Transformer 论文的合著者，被社区亲切地尊称为“Aidan Goatmez”）。主持人聊到 Aidan 在社交媒体上自嘲“在工作中意外变重要了”，并回忆他此前去参加播客节目《20VC》时居然穿着一身实验摇滚乐队 Death Grips 的 T 恤，风格极为硬核，主持人当场向他发出邀请，希望能请这位顶级摇滚乐迷兼技术领袖做客节目。 紧接着，主播爆料了近期政商峰会上的一个搞笑花絮：唐纳德·特朗普（Donald Trump）在 G7 会议现场遇到了同场的 Sam Altman，结果特朗普竟然当面询问 Sam 该如何把自己的升降椅调高调低。主持人调侃称这绝对让 Sam 瞬间内伤——你倾尽了十余年心血、砸进去了数百亿美元，打造出一个本该无所不知的终极机器；只要拿出手机给椅子拍张照片，ChatGPT 就能借助多模态与自主 Agent 瞬间去官网抓取该座椅的说明书、剖析机械结构，提供“升降椅超级智能”。然而眼前的总统完全无视这些黑科技的前提假设，宁可直接抓着本体去当人工客服。主播笑称：“这就是为什么技术扩散（Diffusion）总是如此缓慢，为什么我们还处于‘慢速起飞（slow takeoff）’阶段。” --- ### [25:03](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=1503s) 黑粉算法与法拉利电动车风波 在节目的后半段，话题延伸到了公众人物面对争议与“黑粉（Haters）”的心态调整。一位对谈嘉宾直言，世界上任何做出一番大事业的人都不可能没有黑粉，这是伴随成名第一天起就注定的现实。他回忆自己当年最初在学校走红时，食堂里的同学立刻就开始冷嘲热讽，以至于他“根本不记得没有黑粉的生活是什么样”。 他从流量本质上解构了黑粉的数学逻辑：“好事传千里，坏事传得更快。”尽管黑客会在网上恶评攻击，但吃瓜群众到头来往往根本记不住具体说了什么，唯一留在脑海中的只有你的名字和面孔。更关键的是，黑粉的每一次攻击、每一次互动，都在毫无保留地为社交平台的算法推波助澜。如果有 1 万个死忠粉发声，加上 1 万个黑粉在评论区对骂，那就是 2 万人在讨论你，点击量、观看时长和热门趋势随之翻倍。因此，所有站在世界之巅的人，往往同时也是被恨得最深的人。 见此氛围，主持人突然抛出了一个“挖坑式”的问题，询问嘉宾对法拉利最新推出的纯电动车 Ferrari Luce 有何看法、是否打算买一台。主持人善意地在旁边提醒：“你可得想清楚再回答，如果你以后还想从法拉利买到像 F80 这样的旗舰光环跑车（halo cars），这个回答可是永久记录在案的。”面对这一灵魂拷问，嘉宾立刻领会其中利害，果断给出了极其滑头且求生欲拉满的回答：“我的政治性回答是：无可奉告（No comment）。”随后又忍不住补刀了一句“那玩意儿就是坨烂货”，随即大笑着连续重复“无可奉告”，在欢快的自嘲与调侃中结束了本段对话。 ### [26:59](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=1619s) 节目尾声与告别：祝贺新基金与南湾未来展望 节目在轻松愉快的氛围中进入尾声，主持人和嘉宾互致谢意。主持团队首先对嘉宾的来访表达了由衷的感谢，并热烈祝贺其刚刚成立的新基金，同时嘱咐嘉宾在前往埃尔塞贡多（El Segundo）时替他们向那里的同行与朋友转达问候。席间大家谈及加州南湾（South Bay）蓬勃发展的硬科技与国防科技生态，主持人们调侃道，随着未来可能的“SpaceX IPO”临近，大家注定会在南湾投入大把时间，那里的动态与投资热度只会变得“越来越活跃”。 在最后的收尾环节，主持人以标志性的幽默台词“Flashbang out”作为信号切出节目，热情呼吁收听节目的观众与听众前往 Apple Podcasts 和 Spotify 给节目打下五星好评，并前往 TBPN.com 订阅每日更新的电子新闻通讯，感谢大家的收看与收听后正式向观众道别。 - **祝贺新基金成立**：为主持人与嘉宾之间的合作致意，并向埃尔塞贡多（El Segundo）团队送去问候。 - **南湾科技热潮**：展望加州南湾地区（South Bay）未来的发展机遇，特别提及在 SpaceX IPO 之后当地创投生态预计将迎来更多动作与活力。 - **互动与订阅指引**：号召听众在 Apple Podcasts 和 Spotify 留下五星好评，并访问 TBPN.com 订阅每日通讯。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=0s) 很多人原本认为这绝不可能发生，他们完全属于两个不同的世界，但今天我们把他们聚在了一起 > [08:22](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=502s) 看来我们不得不自己动手做一款了 — David Holz > [09:20](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=560s) 其实写个 if 语句就行——如果住在旧金山，那么判定为腕管综合征 > [09:20](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=560s) 作为提示词输入者的你，仅仅是漫步在‘David 博物馆’里的参观者；真正的主创和策展人是 David Holz 及其团队 > [09:20](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=560s) 医疗器械领域毫无气场（Aura）与动势（Motion），而 David Holz 正在给这个行业注入灵魂 > [09:20](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=560s) 永远不要对初创公司陷入绝望（never black pill on startups） > [12:30](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=750s) 想象一个由 9,000 名歌唱家组成的合唱团，同时有 9,000 名听众在全神贯注地倾听 > [19:54](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=1194s) Noam 究竟看到了什么？ > [19:54](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=1194s) 直接改叫 AI 公司反而更纯粹 > [23:35](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=1415s) 这就是为什么技术扩散（Diffusion）总是如此缓慢，为什么我们还处于‘慢速起飞（slow takeoff）’阶段 > [25:03](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=1503s) 根本不记得没有黑粉的生活是什么样 > [25:03](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=1503s) 好事传千里，坏事传得更快 > [25:03](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=1503s) 我的政治性回答是：无可奉告（No comment） > [25:03](https://www.youtube.com/watch?v=8rYW2eYQfvQ&t=1503s) 那玩意儿就是坨烂货 ## Mentioned TBPN, John Coogan, Jordi Hays, Midjourney, David Holz, Derek Thompson, Jake Paul, Rene Haas, David Senra, Arm, SoftBank, Midjourney Medical, NASA, Leap Motion, Founders Fund, Apple, Oculus, Meta, Discord, Google, Stable Diffusion, Apple Vision Pro, Anabology, Elon Musk, Sam Altman, OpenAI, Noam Shazeer, Google DeepMind, Dean Ball, Jim Cramer, Riley Wells, Notion, Notion Way, Daniel Lurie, Aidan Gomez, Cohere, Death Grips, Donald Trump, ChatGPT, Ferrari Luce, Ferrari F80, SpaceX ## 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=8rYW2eYQfvQ)

[0:00] 很多人原本认为这绝不可能发生，他们完全属于两个不同的世界，但今天我们把他们聚在了一起 [8:22] David Holz: 看来我们不得不自己动手做一款了 [9:20] 其实写个 if 语句就行——如果住在旧金山，那么判定为腕管综合征 [9:20] 作为提示词输入者的你，仅仅是漫步在‘David 博物馆’里的参观者；真正的主创和策展人是 David Holz 及其团队 [9:20] 医疗器械领域毫无气场（Aura）与动势（Motion），而 David Holz 正在给这个行业注入灵魂 [9:20] 永远不要对初创公司陷入绝望（never black pill on startups） [12:30] 想象一个由 9,000 名歌唱家组成的合唱团，同时有 9,000 名听众在全神贯注地倾听 [19:54] Noam 究竟看到了什么？ [19:54] 直接改叫 AI 公司反而更纯粹 [23:35] 这就是为什么技术扩散（Diffusion）总是如此缓慢，为什么我们还处于‘慢速起飞（slow takeoff）’阶段 [25:03] 根本不记得没有黑粉的生活是什么样 [25:03] 好事传千里，坏事传得更快 [25:03] 我的政治性回答是：无可奉告（No comment） [25:03] 那玩意儿就是坨烂货

[17:18] Anabology (quoted on X): The obvious next step once you have a full body ultrasound scanner is to use the ultrasound to do useful things to the body. Delete tissues, make cells divide, reprogram cells, read plus write.

[0:00] Midjourney. [0:01] >> [applause] [0:02] >> Why Midjourney pivoted to medical [0:04] machines? It's not a pivot, it's an [0:05] expansion. It's a It's a second act, a [0:08] third act for David Holtz, [0:10] fantastic entrepreneur, extremely [0:12] inspiring entrepreneur. Maybe we should [0:14] give some background on [0:15] David Holtz. [0:16] >> we talk about the show today first? [0:17] >> Yeah, please. [0:18] >> Because this is the first time that [0:21] Derek Thompson and Jake Paul have been [0:24] on the same podcast. And I want to just [0:26] take I want to take a moment [0:28] >> [applause] [0:29] >> to just appreciate [cheering] that [0:30] because a lot of people said it would [0:32] never happen. They said it was [0:33] impossible, that they were just in [0:35] >> all over the place today. [0:36] >> two different worlds, but we're bringing [0:38] them together. Derek Thompson's kicking [0:41] off our guest lineup today at 11:30. [0:42] >> Yeah. [0:44] >> You don't You don't need an introduction [0:46] there. We have Rene Haas from Arm CEO, [0:50] also on the board of SoftBank. [0:52] >> Very good. [0:53] >> And then we'll be closing it out with [0:55] David Senra and Jake Paul at the end of [0:58] the show. [0:58] >> in live in the TV been after. [1:01] Uh, anyway, yesterday Midjourney [1:03] announced a new division of the company, [1:05] an expansion, not a pivot, [1:08] uh, called Midjourney Medical. [1:10] Uh, we'll watch the video. We'll go [1:11] through the Midjourney scanner. Tyler [1:13] was actually at the launch event, has [1:15] some news to break it down, but uh, [1:17] first let's kick it off with the history [1:18] of David Holtz because it's fascinating. [1:20] Grew up in Florida, son of a dentist. [1:22] He's He's not the dog walker. He's on [1:24] team dentist over there, son of a [1:26] dentist. [1:27] >> We have a son of a dentist on our team. [1:30] >> And played video games growing up. [1:32] >> Dentist camp. Dentist son camp. [1:34] >> Dentist son camp. [1:35] >> Look at those teeth. [1:36] >> Oh, wow. [1:38] >> Wow. [1:39] >> teeth? [1:39] >> Neppo teeth. [1:40] >> He's got neppo teeth. [1:41] >> go back. Go back. Cut back. [1:43] >> Look at that. [1:43] >> Big smile. Big smile. There we go. [1:46] >> He's got the neppo teeth. That's crazy. [1:48] That's crazy. [1:49] >> That's crazy. [1:49] >> Wow. [1:50] >> [laughter] [1:50] >> Your Your father really saved his back [1:52] for you. [1:53] >> base with teeth wise. It's crazy over [1:56] there. [1:57] Uh, played video games as a kid, [1:58] obviously, but starts hacking them, [1:59] learns a little bit about computers, [2:01] super intelligent. Uh PhD in math track, [2:05] I don't think he graduated finished the [2:06] PhD, but I mean, you know, at that level [2:09] of [2:09] >> He turned it down. [2:10] >> engineering. He turned [2:11] >> He turned down the PhD to go work at [2:12] NASA, I think. He was at NASA, uh [2:15] eventually branches out into [2:16] entrepreneurship, starts Leap Motion. Um [2:18] pretty traditional venture-backed [2:20] startup, raised money, I think Founders [2:21] Fund uh was in at least one of the [2:23] rounds, but it was a grind by all [2:26] accounts. And so, the first broad [2:28] product bought brought hand tracking to [2:30] VR. Uh we should pull up a video of the [2:33] original Leap Motion because uh there's [2:35] some cool demos, and you can see how it [2:36] worked. Uh and it's not actually that [2:39] crazy to think about what we're going to [2:42] talk about today with the Midjourney [2:43] medical scanner if you know the history [2:45] of Leap Motion and some of the some of [2:47] the technology that was employed there. [2:49] Yes, this is a different different use [2:51] case, different product, very much [2:54] bigger, but in terms of sensing, [2:56] detecting, mapping, using uh algorithms, [2:59] yeah. [2:59] >> Yeah, and so uh when David was talking [3:01] about like the backstory, [3:02] >> Yeah. [3:02] >> uh he was kind of going through like [3:04] some of the the technicalities of how [3:05] the Leap Motion worked. It was [3:07] incredible cuz he was like, I mean, this [3:08] is early 2010s, and he's using like [3:11] these deep neural nets. He said it was a [3:12] mixture of backprop model. It was like [3:14] incredibly far ahead of the curve just [3:17] on like deep learning. And then, it's [3:18] just like on this kind of like weird [3:20] like side project. It's like doing hand [3:22] motion. [3:23] >> Yeah, everyone thought VR was going to [3:24] take off. There was so much energy [3:25] around Oculus in the dev community, but [3:28] it was a tough It was a It was a very [3:30] tough business. A lot of hardware [3:32] challenges, much less mature supply [3:34] chain, everything's expensive, and uh [3:36] there was all these complex market [3:37] dynamics with other big companies. Meta [3:39] bought Oculus, Apple was investing [3:41] heavily. Eventually, Apple actually [3:43] tried to buy David's first company, uh [3:46] Leap Motion. Um but the deal fell [3:48] through at the 11th hour. I I it [3:50] actually happened twice. There was a [3:51] reporting that like Apple had printed [3:53] welcome packets to all of the Leap [3:55] Motion employees being like, "You're [3:57] getting on boarded today." And then the [3:58] deal fell through, which is crazy. You [4:00] don't see that normally. [4:01] >> David turned it down. [4:02] >> He turned it down. There are There is [4:04] some crazy reporting about like getting [4:05] in and basically being like, "Ah, this [4:07] is too corporate." Or like, "This is not [4:09] the vibe. We're not going to be [4:10] successful here." It was really rough [4:11] for the company. They had to do layoffs, [4:12] uh but they kept building and they sold [4:14] devices to hackers. Yeah, look at this. [4:15] The The Leap Motion. This is Yeah, play [4:17] Fruit Ninja. Now, a lot of this stuff [4:19] can be done with cameras and hand [4:21] tracking. And the hand tracking on the [4:23] modern VR device is pretty good just [4:24] from the headset. Hand tracking in Apple [4:27] Vision Pro, my favorite device, is uh [4:29] fantastic. But at the time you needed a [4:31] third-party device, so you had Can Can [4:34] you scroll the video down to show the [4:35] actual device a little bit? [4:37] Yeah, this little like It looks like a [4:39] stick of gum. Very small, very neat, uh [4:41] very cool, but ultimately uh the very [4:44] much a dev kit. Never Never got to, you [4:46] know, mass consumer adoption like an [4:48] Oura Ring or a Whoop band or anything [4:50] like that. It was not must-have for, you [4:52] know, the average gamer, the average [4:54] consumer, the average technologist. Uh [4:55] after about a decade, David was able to [4:57] take a step back and start hacking on [4:58] the next thing. AI image generation, you [5:00] know, Midjourney, obviously huge viral [5:02] success. Broke a lot of the traditional [5:04] rules. So, no venture funding, no front [5:06] end, no website, no no app. Like, really [5:10] they they would I mean, now they have a [5:12] front end and and stuff, but uh years [5:14] went by where it was just a Discord. [5:16] Everything was in a in a Discord and [5:18] they scaled that Discord to millions of [5:20] members, so huge that the company [5:22] Discord, I think it was the biggest [5:24] Midjourney Discord was the biggest [5:25] Discord for a while. They had to scale [5:27] their systems at that company, but [5:28] you're doing all this like off-balance [5:29] sheet R&D basically. And on day one, you [5:32] just get a fantastic mobile app for [5:35] every device. [5:36] >> experience. [5:36] >> Social experience, that was the second [5:37] thing. So, yeah. [5:38] >> Yeah, because there was so much This was [5:40] the era like prompt engineering era, [5:43] right? Where they had their favorite [5:44] prompts. [5:46] >> It really helped kind of, uh, cultivate [5:48] >> multiplayer on day one. [5:50] >> Yeah. [5:50] >> And so, David has the quote, I I think [5:52] it's from him saying like, "If you show [5:54] someone a blank box and you tell them, [5:56] 'Hey, this is a magical AI machine that [5:58] can generate you a picture of [5:59] anything.'" People will just type in [6:01] dog. And they'll get a picture of a dog, [6:03] and they would get the same picture if [6:05] they just went to Google Images. That's [6:07] not what AI is actually interesting at. [6:09] It's the It's the uh astronaut riding a [6:11] horse on the moon. The thing that that [6:12] picture doesn't exist already. Uh so, [6:14] it's you [6:16] flying an F-16 in your hometown and like [6:19] personalizing and doing something that [6:21] would typically take a very large CGI [6:24] budget or or some sort of uh you know, [6:26] Photoshop master to actually whip up or [6:29] collage together. Uh you could do that [6:30] with just one prompt. And so, the [6:32] multiplayer nature of Midjourney, when I [6:34] jump into the Midjourney Discord, the [6:35] first thing I see is your crazy prompt. [6:37] You're 12 levels deep prompt injecting, [6:40] thinking of different keywords, doing [6:42] different S refs for style references, [6:44] doing all these different tweaks, and [6:46] then I can just see, "Oh, he He used [6:48] this model and he specified these [6:50] dimensions and he's getting really good [6:51] results." So, I'll take that, but I I [6:53] don't want to make it about I I want to [6:55] make my images about a dog. You're a [6:56] mosquito guy, so you're going to make [6:57] them about mosquitoes, [6:58] >> Yeah. [6:59] >> right? Um and and you can take my prompt [7:01] and just sort of remix it. And so, [7:03] >> Some of the mosquito propaganda prompts [7:06] I had back in those days were crazy. [7:08] >> Yeah. [7:08] >> I say I never ended up sharing them. [7:09] They were simply too good. [7:11] >> Yeah, you Well, you lost the you lost [7:12] the war and the mosquito's going away. [7:14] Whether you like it or not. [7:15] >> Google is is focused entirely on killing [7:18] mosquitoes, apparently. Uh by generating [7:19] four candidate images for each prompt, [7:21] Midjourney got a bunch of really [7:22] valuable data on what a correct image [7:25] looked like for a specific prompt. This [7:26] was another sort of Midjourney [7:28] innovation. I don't know if they were [7:29] the first ones to do this, but it was [7:31] really key to the Midjourney flow. It [7:32] would gen you Instead of generating you [7:34] an HD image, it would generate you four [7:36] low-res images, and you'd be like, "Ah, [7:38] that's that has the right vibe. That has [7:40] the right style. That has the right [7:42] layout. And then you would get that [7:43] feedback and be able to uh to iterate. [7:46] At the same time, Stable Diffusion had [7:48] some really good results, but it ran [7:49] locally, it was open source, and so they [7:52] weren't having the data flywheel like [7:53] Midjourney was. And so there were a lot [7:55] of really positive things there. Now [7:57] every company has a data flywheel and [7:59] can can iterate on that, but it was an [8:01] important innovation. Uh the company [8:03] still that we know of hasn't take [8:04] venture funding, so I think David Holz [8:06] owns most if not all of the company. Um [8:09] and after the big meta deal to devour [8:11] vibes, [8:12] um [8:13] there's clearly enough cash flow to fund [8:15] other projects. David's actually been [8:16] talking about this for years, this idea [8:18] of a medical uh device. [8:20] >> the meta deal I don't think any [8:22] specifics were ever released, but I [8:24] think it's safe to assume it was in the [8:26] hundreds of millions of dollars. [8:29] >> trillions potentially. [8:30] >> Potentially, potentially. But but enough [8:32] to be able to take some really really [8:35] really big swings [8:37] and be able to have a balance sheet like [8:39] a heavily funded [8:40] >> Yeah. [8:41] >> private company. [8:41] >> Yeah. So [8:43] it's really cool. I was I was hoping for [8:44] a VR headset, honestly. [8:45] >> [laughter] [8:46] >> I really wanted a VR headset. [8:47] >> I mean, this is just one of the hardware [8:49] projects, right? He says like I think [8:50] there were like eight little uh you [8:52] know, images, yeah, teaser things. So so [8:54] there's there's a bunch more hardware. [8:56] >> Because after the Apple Vision Pro [8:57] launch, I think he put out some sort of [8:59] review being like we're going to have to [9:01] do it ourselves. [9:02] >> [laughter] [9:02] >> Which I love. Anyway, let's play let's [9:04] play the Midjourney launch video for [9:06] Midjourney Medical and see uh what's [9:09] going on here in this two-minute video. [9:12] The goat is still on screen. The goat is [9:14] in our stream, not in the Midjourney [9:16] video. [9:20] Did the aesthetics of this video carry [9:22] through to the actual event experience? [9:25] I only saw a few photos, but did did the [9:27] event have a nice uh vibe to it? [9:30] >> Yeah, it was very cool. They had a [9:31] similar like warm lighting everywhere. [9:33] They had like a kind of fake versions of [9:36] of the real machine. It's like way too [9:37] big obviously to move around. But they [9:39] had the kind of big tubes that like [9:40] resembled it. Then they had a smaller [9:42] version you could put your hand in. I [9:43] did it and my hand. [9:47] >> Are you healthy? Is your hand healthy? [9:48] We need you coding at all times [9:50] prompting. [9:52] >> Uh I think it I it's hard cuz that was [9:54] like a a small like demo version, so the [9:56] resolution was lower. But uh someone [9:58] else said that uh [10:00] that they figured out that they had some [10:01] like wrist thing because of boxing. [10:03] >> tunnel. [10:04] >> Yeah, they they found some issue with [10:05] their [10:06] >> You can just use an if statement for [10:07] that. [10:07] >> time? [10:08] >> Yeah. [10:08] >> If if San Franciscan [10:11] then yes, carpal tunnel. [10:16] Very cool vibes. [10:18] I've always said this about Midjourney [10:19] that I think when you like people debate [10:22] is AI art art or whatever, I always [10:24] think of Midjourney as a David Holz art [10:28] experience. And I feel like you as the [10:30] prompter are merely the the [10:35] uh the museum goer in the museum of [10:37] David and he is the one who is creating [10:41] and the Midjourney team [10:42] is the curator. They are the artists and [10:45] you are experiencing the art through the [10:47] prompts, but the decisions and the taste [10:50] comes from David. And so are Midjourney [10:52] images tasteful? I think they have [10:53] taste. I think they have David's taste. [10:55] And then there's the opportunity for a [10:57] user to put a little bit more on top, [10:59] but by default, I think the reason [11:01] Midjourney has been successful is [11:02] because it's so opinionated and it's not [11:04] trying to just be uh stock photo [11:06] generator. [11:07] >> Yeah, even this video was fully filmed [11:09] by David. Like he was like the one [11:11] holding the camera. [11:11] >> That's cool. [11:12] >> Yeah. [11:12] >> Love this guy. [11:14] It's amazing. [11:14] >> It makes me a little bit emotional. It's [11:17] so cool. [11:17] >> Yeah. [11:18] >> It makes me feel like we're finally [11:20] living in the future. No one No one [11:22] making this kind of device has ever been [11:25] as cool as David Holz. [11:27] >> Sure. [11:27] >> Straight up. [11:28] >> Uh let's move [11:29] >> to I'm willing to state that as a fact. [11:31] It's just true. Like [11:33] >> [clears throat] [11:33] >> this is not the first medical device [11:36] >> Yes. But the medical device industry [11:38] does not have Aura. [11:40] >> Yeah. [11:40] >> They have no motion. [11:41] >> He's bringing motion [11:43] and Aura [11:44] >> to the medical device [11:45] >> field. [11:45] >> category. [11:46] >> Yes. [11:46] >> He's inventing He's inventing medical [11:48] devices. [11:49] >> Uh I think a lot of people are going to [11:50] want to get these scans [11:51] >> Yeah. [11:52] >> because it is a beautiful [11:54] representation, futuristic [11:56] representation of the human form. It's [11:58] personalized. [11:59] >> Also just never black pill on startups. [12:00] Like he could be so successful and he [12:02] doesn't have any venture money and he's [12:03] never going to sell this business. And [12:05] so you have like 12 biomedical device [12:08] companies that are like we have to we [12:10] cannot get steamrolled by David Holtz. [12:12] We got to work with this company. We got [12:13] to buy from that the YC company. Like [12:15] the competitive dynamics of the market [12:17] are always more complicated than just a [12:19] single winner comes and steamrolls all [12:21] the startups and all the incumbents. [12:22] That's not how things ever play out. [12:24] Life is more complicated. The economy is [12:25] more complicated complicated. So never [12:28] black pill. Let's play the deep [12:30] technical dive. The technical deep dive. [12:35] Did he make this music too, I wonder? [12:36] >> Yeah. And he also made this in [12:38] I think Blender, he said. [12:40] >> Woah. [12:40] >> Like all the animations are all made by [12:41] him. [12:42] >> [laughter] [12:43] >> By hand, though? [12:44] >> He's the one who yeah. [12:45] >> That way I [12:46] >> I guess he's using Blender. For I mean [12:48] for this it's a great use case for [12:49] Blender cuz you're cloning [12:51] you know personal [12:51] >> I don't know if it was actually Blender. [12:52] It was some I mean [12:54] >> CGI stack. [12:56] >> Incredible music selection. [13:01] >> Oh yeah, [music] there there there's [13:02] there's words we should be reading. [13:03] There's no voiceover. [13:05] But we can kind of give you the summary [13:07] here. [13:08] >> [music] [13:08] >> Um [13:10] The the overviews We'll start with [13:12] something called transducer. We're going [13:14] through Transducers act as speakers and [13:16] microphones. Imagine a choir of 9,000 [13:19] singers and an audience of 9,000 people [13:21] all listening carefully. [music] [13:24] The transducers are controllable at a [13:25] rate of 100 million times per second. [13:29] Together, they fire in a pattern that [13:30] sends out structured waves. Those are [13:33] the structured waves you're seeing. [13:35] The waves travel out at a speed of 1,481 [13:39] m per second. We construct a ring of 40 [13:41] of these systems [13:43] 70 cm in diameter. [13:47] The ring consists of 358,000 [13:51] ultrasonic sensors. The chips take turns [13:53] sending out waves. [13:56] As the waves have a chance to dissipate, [13:58] we fire the next one. [14:02] One by one, they fire at a rate of up to [14:04] 1,000 times per second. [14:09] You're vibing out. The waves travel [14:10] across the tank in 480 microseconds, [14:14] about 1/2000 of a second. Hundreds of [14:17] thousands of transducers listen. Each [14:19] sensor resolves motions smaller than the [14:22] width of an atom, not micrometers or [14:25] nanometers, but picometers. Working [14:27] together, hundreds of sensors can even [14:29] push into the subatomic femtometer [14:32] range. [14:34] Finer than the scale of atoms, a scale [14:36] where we don't even have anything you've [14:38] heard of. [14:40] The system captures data at a leisurely [14:43] 17 GB a second. [14:46] GBs of sonic reverberations flow around [14:48] your body. [14:53] And from vibrations, they form images [14:57] with this device. Each probe sees [14:59] different angles of your body. [15:04] We combined hundreds of waves and [15:06] thousands of sub images to get the final [15:09] product. Over Over 40 GBs of data moves [15:12] through the system [15:13] uh to just see one slice of your body. [15:17] We analyze the images, making out [15:18] organs, structures, and tissues. Of [15:20] course, all good uses of AI. [15:26] There you go. [15:28] Not going to read all that. Here we show [15:29] a slice of what we can see with 25 [15:31] different biological structures. [15:33] Look at this. [15:35] We do this again and again and again as [15:37] you move through the ring. [15:40] So, are you moving through the ring or [15:41] is the ring moving around you? [15:43] >> No, no, you're moving down. But, the [15:44] ring stays. [15:46] That's why you saw in the video a little [15:47] platform [15:48] >> That's interesting. Over a period of 60 [15:50] seconds, the goal is to obtain several [15:51] hundred slices of your body. [15:56] Reconstructing across 21 servers [15:59] with two petaflops of compute power. [16:03] And up to 806 terabytes of raw data. The [16:06] slices form a 3D map with the ability to [16:08] resolve internal tissue details as small [16:10] as half a millimeter. Wow. [16:13] >> Leminah Leap says, "Still can't believe [16:14] they missed the med journey opportunity [16:17] here." [16:19] >> Less than a dozen of these systems [16:20] operating at full speed can do a full [16:21] body scan of every Their goal is a fleet [16:24] of five of 50,000 of these scanners [16:26] capable of a billion scans a month. [16:29] That's basically everyone on Earth. [16:30] >> Tyler, what's the logic of putting these [16:32] integrating these into spas? Do they [16:35] want to make standalone spas or they [16:37] want to make a first one in San [16:38] Francisco that's like a flagship and [16:40] then I can't imagine all 50,000 scanners [16:43] would come with a full spa, but [16:47] >> Yeah, I think he he said the plan was [16:49] you start with the big one and I think [16:50] there's going to be like 10 or so in the [16:52] first SF location and then in some [16:55] places there'll be like hundreds of [16:56] these, some places there's like one or [16:57] two. But, I yeah, I think that the spa [16:59] is like mostly because at least he he [17:01] rationalized it because like oh, you [17:02] know, going to the clinic or whatever, [17:03] it's like not a good experience. [17:05] >> Yeah. [17:05] >> This should be like very quick. [17:06] >> Yeah. [17:07] >> And it should feel much more like going [17:08] to a you know, sauna or steam room or [17:10] whatever than than going to the doctor. [17:12] Cuz you you can imagine like you you [17:13] don't really need a doctor they're [17:15] during it, right? Yeah. It's like [17:16] autonomous. [17:17] >> They should create a bunch of needles in [17:19] the in the ring that come in and just [17:22] kind of connect with you, deliver [17:25] whatever peptides, hormones, PEDs that [17:28] it deems necessary. [17:29] >> Potentially. [17:30] >> Well, I mean, he he also talked about [17:31] this like uh [17:33] this like system is like pretty [17:35] overpowered, he said. So, you could like [17:37] theoretically instead of just like [17:38] basically reading like what's going on [17:39] in your body, you can like write stuff [17:41] too, right? So, you can like do [17:42] non-invasive surgeries whatever by like [17:44] beaming light [17:45] >> Yeah, yeah, we have we have folks who [17:46] have done that where it's like it [17:48] pulverizes the cancer or the tumor with [17:50] uh ultrasonic waves. [17:51] >> Yeah, like think about like UV [17:53] >> Yep. [17:54] >> uh lithography or something. You you [17:55] have if you like [17:56] you know, position the light in very [17:57] certain ways, you can like actually make [17:58] a change too, not just like read stuff. [18:00] >> potentially we could pivot this into [18:01] more chip fabrication capacity. [18:03] >> That would be the goal. Yeah, yeah. [18:05] >> Just Yeah, sorry. We can make way more [18:08] money just making more chips. [18:10] >> Yeah, one of my [18:11] >> [clears throat] [18:11] >> We're going to go to the data center. [18:12] >> Friends over on X, Anabology, great [18:15] account, says, "The obvious next step [18:17] once you have a full body ultrasound [18:18] scanner is to use the ultrasound to do [18:20] useful things to the body. [18:22] Delete tissues, make cells divide, [18:24] reprogram cells, read plus write. If [18:26] you're an engineer interested in solving [18:28] the right side of sci-fi medical [18:30] devices, we're building this and [18:31] hiring." [18:32] >> Mhm. Well, let's go through some [18:34] more of the reactions. Midjourney [18:36] launching an ultrasound scanner is such [18:38] a clear example of freedom enjoyed by [18:40] bootstrap companies that VC-backed [18:41] companies would never have. Only time [18:43] will tell if it's the right bet, but [18:45] such bold bets require ownership and a [18:47] kind of devil-may-care attitude. Uh [18:49] that's somewhat true, but uh the counter [18:51] to all that is like the Elon Musk [18:53] projects uh and and a variety of uh [18:56] founders. [18:57] >> And SPACs, but also, you know, [19:00] like Sam Altman has done this too where [19:01] it's like, "Okay, there's a new idea, [19:03] get a new team, fund it, and it's on the [19:05] side." Like there are there are a [19:08] variety of uh entrepreneurs that are [19:10] playing the VC game, marshaling capital, [19:12] and then have the permission to work on [19:14] crazy stuff, like a mass driver on the [19:16] moon, etc. [19:18] But, in general, I agree with this point [19:21] that a lot of founders get locked in a [19:23] VC loop of you got to raise money every [19:26] 18 months, you got to hit the core KPI, [19:27] you don't have the time or cash flow cuz [19:29] you're win win win fight fight fight. [19:32] And it's very very interesting in such a [19:33] narrative violation that Mid Journey has [19:35] become such a machine financially in, [19:39] you know, people would say it's the [19:40] commodity market. People would say it's [19:41] a highly competitive market. You're [19:42] going up against Google. You're with [19:44] Nano Banana. You're going up against [19:45] OpenAI. And yet, Mid Journey has carved [19:47] out a fantastic community, a fantastic [19:49] user base, a fantastic business model, [19:51] and it's really testament to something. [19:53] It's just awesome. I don't know what [19:54] else to say. What did Noam Shazeer see? [19:58] >> [laughter] [19:59] >> What did he What did he see? Because [20:00] Noam Shazeer, the legendary AI [20:03] researcher formerly of DeepMind, has [20:05] joined OpenAI. Very exciting news for [20:08] Noam and OpenAI. He says, "I'm excited [20:11] to share that I'll be joining OpenAI and [20:13] look forward to working with the [20:14] exceptional team there. It was a [20:16] difficult decision to move on. I'm [20:17] incredibly proud of the amazing team at [20:19] Google and everything we've built [20:20] together. It has been an honor and a [20:22] pleasure to work with all of you." [20:25] Everyone is very very excited about [20:26] this. And Yaxine shares interesting [20:28] details here. He says, "Noam Shazeer [20:30] 6'4". [20:31] >> [laughter] [20:32] >> I don't [snorts] know [20:33] >> Just getting right to the important [20:35] stuff. [20:35] >> But, I like that we're apocryphally [20:37] building up our our heroes. [20:39] >> Yeah. Remember the Metis list. [20:40] >> Number one on the No, number one on the [20:42] Metis list. [20:42] >> the Metis list which we released summer [20:44] of last year during the height of the [20:46] talent wars. [20:48] >> Co-author of the Transformer, T5 Switch [20:50] Transformer papers, one of the pioneers [20:52] of sparse MoE models, leaving his VP of [20:55] engineering Gemini co-lead role at [20:57] Google DeepMind to join OpenAI. [21:00] Lissan al Ghayeb says, "This is likely [21:02] the most significant AI talent move of [21:04] the year. It makes you wonder what's [21:06] going on at Google. Only mashed by the [21:09] next day Dean Ball joins Open AI. [21:12] Tyler, you've been a Ball now for a long [21:14] time. We've had him on the show multiple [21:16] times. [21:17] We love Dean and he's been [21:19] in the complete opposite camp. He's not [21:22] an AI researcher. He's policy, a lot of [21:24] experience there and had a really even [21:27] keel, I think, on a lot of the analyses [21:29] of should something be [21:30] >> I think the main thing is he really [21:31] cares about getting this right as a [21:34] country as a country. [21:35] >> Totally. [21:35] >> Right? And he's been critical of almost [21:38] every company [21:39] >> Yeah. [21:39] >> in the space, but only again because he [21:42] cares. What does Jim Cramer think about [21:44] these moves? [21:45] >> Yeah. [21:45] >> He posted at 3:00 a.m. this morning, [21:48] Noam Shazeer, top AI thinker, goes to AI [21:51] from Google. Big win for AI. [21:54] >> [laughter] [21:55] >> And [21:56] >> Drop the O. Drop the open. [21:58] Just AI. [21:59] >> It's cleaner. [22:00] >> It's cleaner. [22:02] It's closed source models anyway. Just [22:03] call it AI company, I guess. [22:06] In other news, Riley Walls did it. He [22:09] did the unthinkable. He bought a street [22:11] in San Francisco and auctioned it off [22:13] and Notion bought it and now the Notion [22:15] Way is officially born. I'm so I'm so [22:19] happy that Notion won this. It could [22:21] have been [22:22] you know, some crypto thing. It could [22:24] have gone so odd. You know, it's like [22:27] high risk. Whenever you turn something [22:29] loose like this on the internet, it can [22:31] go a bunch of different ways. Notion Way [22:33] is not something that I think would be [22:34] annoyed. I I I personally wouldn't be [22:37] annoyed if like a street corner in my [22:39] neighborhood was Notion Way. It is an [22:41] ad, but it's it's an it's a it's a term. [22:44] I would love it. [22:45] >> you lived there. [22:46] >> But but it's just a term that feels like [22:48] it could just be a street name, not not [22:51] something that just feels totally like [22:52] an ad like with a dot com in there or [22:54] something. So fantastic [22:56] >> Yeah, I remember I remember when [22:58] perfect execution from from Riley Walls. [23:00] I remember talking with him about this [23:02] idea for the first time. Absolutely [23:03] loved it. We We told him like [23:05] >> They painted the street, too. [23:06] >> We told him that we'd backstop it. [23:08] >> Yeah. [23:08] >> Basically, like if no one bid on it, [23:10] >> Yeah. [23:11] >> then then we would we would cover it. [23:12] >> for like a thousand bucks or something [23:14] like that." [23:15] >> think it was like I think I think [23:16] >> or what what what whatever the cost was. [23:18] >> it was like 20s 20-something. [23:20] >> were like, "Well, we'll we'll backstop [23:21] it. We'll be like the the buyer of last [23:23] resort if you need that." [23:24] >> [laughter] [23:25] >> And [23:26] what he did was 40K. Very, very [23:28] interesting. Um [23:30] >> They got the the blessing of the mayor. [23:32] >> Yeah. Yeah. [23:32] >> Daniel Lurie. [23:33] >> That's very exciting. Very exciting. [23:35] >> Aidan Gomez accidentally became [23:36] important at work. [23:37] >> I love Aidan Gomez. Can we get him on [23:39] the show? [23:39] >> life. We have to have enough Death Grips [23:41] fans on the show. We got to have Aidan [23:43] Gomez on. Really funny [23:47] thing to post from [23:48] >> Also on the Transformer paper, by the [23:50] way, Aidan Gomez. They call him Aidan [23:52] Goatmez for a reason. He's He's quoted. [23:55] No, I love him. He's fantastic. And he [23:58] did a podcast episode with 20VC wearing [24:01] a Death Grips t-shirt, and that was [24:03] incredible. Um Anyway, bunch of bunch of [24:05] other funny things happening at the G7. [24:07] Apparently, Donald Trump had to ask Sam [24:09] Altman how to adjust his chair up and [24:12] down, how to use the chair. And that's [24:14] got to hurt because you're Sam Altman, [24:16] you've spent billions, tens of billions [24:18] of dollars, years of your life, a decade [24:21] of your life building a machine that can [24:23] answer this exact question. [24:25] >> question. [24:26] >> You take a picture of the chair, it [24:27] tells you how exactly to operate it. [24:29] >> So, did you [24:30] >> Yeah, right in his face. [24:31] >> Right in his face. [24:32] >> Just completely reject the premise. [24:33] >> Yeah. Yeah. And And you know you know on [24:36] the tip of things like, "Actually, like [24:37] this is a great opportunity to use [24:38] ChatGPT. You can use images, upload a [24:40] photo of that chair. It'll It can go and [24:43] agentically pull the actual diagrams, [24:47] the the manual from that chair. It can [24:49] go to that website. It's chair [24:50] superintelligence. [24:52] >> Mr. President. [24:53] >> President is just not interested. He's [24:55] just like, "I'm going with a human on [24:57] this one. And that's why diffusion takes [24:59] so long. That's why That's why we're in [25:01] the slow takeoff. Anyway. [laughter] [25:03] >> You know, I don't think anyone who does [25:06] good things in this world or big things [25:08] in this world doesn't have haters. I [25:11] mean, it's the it's the day one story [25:13] and [25:14] I honestly [25:16] you know, since day one when I first [25:18] went viral, the video I was talking [25:20] about like I instant [25:22] >> Cafeteria. Right out of the cafeteria. [25:24] >> people in my school started hating. So [25:26] like I don't even [25:27] >> [laughter] [25:27] >> I don't even remember life without [25:29] haters. [25:30] >> Yeah, your your mom, you know, took that [25:32] video when you were born. I'm sure some [25:34] of the nurses were like, yeah, this kid [25:36] >> They'll [ __ ] this guy. [25:37] >> [laughter] [25:38] >> Yeah, yeah. [25:39] >> No, you have to have it and, you know, [25:42] good news travels fast, bad news travels [25:44] faster and the the haters actually will [25:46] talk about you more and say bad things. [25:49] But at the end of the day, the people [25:51] don't really remember what was said. [25:53] They just remember your name and your [25:56] face. Um and so you could do with that [25:58] what you want and, you know, it's really [26:01] they're adding to the algorithm at the [26:03] end of the day. So it's really just math [26:05] and if you just have fans that are [26:08] saying things, you know, let's say [26:09] that's 10,000 people, but add 10,000 [26:13] haters in there and now 20,000 people [26:14] are talking about you and it just adds [26:16] to clicks, views, talk, trending. [26:20] Uh so that's the way I've always looked [26:23] at it and the biggest and best people in [26:26] the world all are also the the most [26:28] hated. [26:29] >> Ferrari Luce has a lot of haters. Will [26:31] you be getting one? [26:32] >> Say it again. [26:33] >> The Ferrari Luce, the new electric [26:35] vehicle from Ferrari. Are you in the [26:37] market? [26:37] >> You got to be careful cuz if you want [26:39] any of those halo cars, you know, this [26:41] is going to be permanent. [26:42] >> permanent. [26:42] >> I know you might want to say I know you [26:44] might want a F80 at some point. You got [26:46] to be really careful. I just want you to [26:47] be really careful with what [26:49] >> This is a political answer. No comment. [26:52] [laughter] [26:52] >> Okay. [26:54] >> That shit's ass. No comment. No comment. [26:56] No comment. [laughter] [26:57] >> Oh well. [26:59] That was a lot of fun. Well, thank you [27:00] so much for coming on today. [27:01] >> Yeah, congrats on the new fund. Have fun [27:03] down in El Segundo. Hopefully give our [27:05] best to everyone. [27:06] >> It was great. [27:07] >> It the South Bay, I'm sure you guys are [27:09] going to be spending a lot of time there [27:11] post SpaceX IPO. It's just going to be [27:13] more and more action. [27:14] >> Yeah. [27:14] >> Flashbang [clears throat] [27:15] out. Leave us five stars on Apple [27:17] Podcasts and Spotify. Sign up for our [27:18] newsletter at tbpm.com. And thank you [27:20] for tuning in today. Goodbye.