China's GLM-5.2 Model Matches US AI Safety and Defense Capabilities

中国GLM-5.2模型比肩美国AI安全与防御能力 中国开源模型作为通缩力量削弱高利润服务业垄断，对美国AI市场的主导地位构成威胁。 主持人结合George Hotz和John Luttig的观点，分析了GLM-5.2对全球AI竞赛的影响，并探讨了中国开放权重模型如何挑战美国闭源模型的主导地位。讨论涉及模型蒸馏与Token经济学等技术议题，随后转向Meta的Gemini使用上限、内部安全政策以及非侵入性大脑解码研究。此外，他们还回顾了2006年雅虎收购Facebook失败的著名案例，审视了内存芯片成本上涨对AI利润空间的挤压，并报道了Chamath Palihapitiya为890筹集1.35亿美元A轮融资的消息。 中国初创公司智谱AI发布开源权重模型 GLM-5.2，在识别软件漏洞方面的表现已能匹敌 Anthropic 的 Claude Opus 4.8 等美国顶尖闭源模型。 (0:02) AI 模型的评估重心正从“每 Token 成本”转向“每任务成本”，市场也加速分化为高成本的前沿模型与廉价高效的开源小模型两大阵营。 (4:20) 因云算力基础设施出现瓶颈，Google 限制了 Meta 对 Gemini 模型的调用量；同时 Meta 内部也限制员工使用 Claude 和 Codex，以防发生意料之外的底层模型蒸馏。 (19:04) 高带宽内存（HBM）及 DRAM 产能严重紧缺推高芯片价格，Micron 等存储厂商大幅提价攫取巨额收益，正进一步挤压 AI 产业整体利润并波及消费电子市场。 (26:31)

TBPN explores how the rise of competitive open-weight models from China is challenging assumptions about proprietary AI dominance, while physical compute bottlenecks and skyrocketing memory chip prices transfer industry profits upstream. Open-Weight Models Challenge American Frontier Labs (0:02) Zhipu AI's release of the open-weight GLM-5.2 model has challenged previous industry assumptions that open-source artificial intelligence would lag behind well-capitalized closed-source laboratories. Security researchers at Semgrep reported that GLM-5.2 matches Anthropic's Claude Opus 4.8 in identifying software vulnerabilities, placing it among the top ten most utilized models on OpenRouter. The emergence of capable open models has reignited debate over open versus closed development paths. In 2024, John Luttig argued that exponential capital expenditure would starve open-source providers of returns, while George Hotz has more recently claimed that China utilizes open-source AI as a deflationary lever targeting the high-margin United States service economy. Model Distillation, Task Economics, and Security Buffers (4:20) The economic evaluation of AI is shifting from cost-per-token toward cost-per-task, as token-hungry open models can incur higher overall costs on complex reasoning. Simultaneously, developers observe a bifurcation: expensive proprietary models are reserved for high-stakes cybersecurity and coding agents, while smaller open-source models handle high-volume automation such as receipt OCR. Technical questions also persist around distillation, as Chinese models frequently train on outputs from American systems like Claude and GPT. While distillation lifts benchmark scores, developers note it can reduce generalization on open-ended tasks. On national security, although Dario Amodei warned in 2023 of catastrophic biosecurity and cyber risks from open weights within two to three years, defense firms like CrowdStrike and Palo Alto Networks currently rely on frontier systems like Mythos and GPT-5.5 Cyber to patch vulnerabilities ahead of open exploits. Compute Constraints and Distillation Policies Within Big Tech (19:04) Infrastructure bottlenecks and proprietary data protection are shaping operational decisions across major tech companies. The Financial Times reported that Google capped Meta’s access to Gemini models around March due to broad compute infrastructure constraints. Internally, The Information reported that Meta restricted employee usage of rival systems, including Anthropic's Claude and OpenAI's Codex, to prevent accidental distillation into Meta's proprietary architectures. Outside software infrastructure, Meta published research in Nature detailing its Brain-to-QWERTY V2 project, which uses magnetoencephalography to decode raw neural signals directly into full semantic sentences. Memory Chip Bottlenecks Extract AI Industry Profits (26:31) High-bandwidth memory (HBM) and DRAM shortages have concentrated profitability in hardware manufacturing rather than foundation model development. Micron Technology, Samsung Electronics, and SK Hynix have capitalized on supply constraints, with Micron raising DRAM prices by more than 60% and NAND flash by over 80% in the quarter ending May 28, despite single-digit shipment growth. These price surges have pushed Micron's market capitalization past $1 trillion and spilled over into consumer electronics, contributing to a 15% price increase on Apple MacBooks and a tripling of retail RAM costs over 12 months. With AI labs subsidizing compute to maintain market share, input costs increasingly compound operating losses across the software ecosystem. Tech History, Restructuring, and Venture Syndicates (23:50) Corporate restructurings and venture investments highlighted broader commercial realignments. Comcast announced the separation of its connectivity infrastructure from NBCUniversal and Sky entertainment units. Silicon Valley history underscored the stakes of dealmaking through Wired's retrospective on Yahoo CEO Terry Semel's 2006 bid to purchase Facebook; after an initial $1 billion verbal agreement, Semel lowered the offer to $800 million following a stock drop, prompting Mark Zuckerberg to exit the transaction permanently. In current venture funding, Chamath Palihapitiya raised a $135 million Series A round for 890, backed by Salesforce Ventures, Craft Ventures, Launch, and The Production Board. The hosts leave open whether open-source AI can sustain its capability parity without direct access to frontier models for distillation, and whether surging memory hardware costs will eventually force foundation model providers to abandon loss-leading pricing models.

随着中国开源权重模型在安全评估与编程基准上迅速逼近美国顶级闭源模型，人工智能正演化为融合地缘经济与技术路线的深层博弈；与此同时，上游存储芯片厂商凭借产能瓶颈大幅提价，正加速转移并侵蚀模型开发生态的商业利润。 开源闭源技术路线与地缘博弈升级 (0:02) 智谱AI（Zhipu AI）于6月13日发布的开源权重模型GLM-5.2引发业界广泛关注。安全机构Semgrep测试显示，GLM-5.2在查找软件漏洞方面已能匹敌Anthropic的Claude Opus 4.8，在特定指令微调下甚至追平Mythos；聚合平台OpenRouter的数据显示该模型已跻身平台调用量前十。这一进展对此前美国科技界普遍持有的“高额资本开支将使闭源前沿实验室拉大差距”的预期构成了事实冲击。针对其技术路径，行业普遍存在关于蒸馏技术的争议，即新兴模型通过调用OpenAI或Anthropic接口生成的合成数据进行训练，甚至催生出通过VPN和多账号采集API输出的灰产网络；但随着大量公开网页和开源代码库普遍包含LLM生成文本，蒸馏的界限日渐模糊。开发者反馈表明，过度依赖蒸馏的模型虽在跑分和代码上表现优异，但在复杂开放场景的泛化表现上依然不及大规模基础训练的原生前沿模型。 在宏观战略层面，投资人John Luttig曾在2024年5月撰文论证基础模型本质是闭源游戏的逻辑，指出指数级扩大的资本开支将剥夺开源提供商的投资回报；然而George Hotz在2026年提出相反观点，认为免费且强大的开源模型是中国针对美国高利润服务型经济实施的“通缩型武器”，旨在冲击依赖智力溢价的商业模式与垄断壁垒，这使得华盛顿试图通过出口管制与安全审查来限制开源发展的手段陷入被动。 模型分层落地、安全防御与监管摩擦 (10:33) 在企业实际落地中，模型采购正呈现显著的两极分化。在代码智能体构建、核心网络攻防等具有灾难性风险的高价值场景中，企业倾向于不惜代价采购能力居顶的昂贵前沿模型；而在高吞吐量的点状自动化任务中（如金融服务公司Ramp对海量消费小票进行自动化OCR识别），机构则优先采用DeepSeek Flash等轻量、高速且廉价的开源模型。业界评估重心正加速从“每Token成本”转向“每任务成本”，因为某些单Token定价低廉的模型在应对复杂多步骤推理时极度消耗Token，整体任务成本反而更高。 安全风险与政策监管的博弈同步激化。针对Anthropic首席执行官Dario Amodei在2023年国会听证中预警的“高级开源模型将在两至三年内对生物与网络安全构成严重威胁”，当前网络安全防御主要依赖关键的时间差：CrowdStrike和Palo Alto Networks等头部安全厂商提前数月部署GPT-5.5 Cyber及Mythos对基础设施进行查漏补缺，使白帽防御者在黑帽群体利用开源攻击工具前掌握防御先机。同时，如果政府监管部门强行推行严格的客户背景调查（KYC）或设置长达数月的准入审批，往往会导致急于落地的科技企业转而采用唾手可得的开源模型，从而使行政审批形同虚设。 算力瓶颈制约与存储芯片垄断红利 (19:04) 尽管大型科技公司投入数千亿美元扩充资本支出，算力基础设施的绝对瓶颈依然在各环节显现。《金融时报》披露，Google Cloud由于自身计算基础设施吃紧，在3月前后不得不限制Meta对其Gemini模型的算力配额，直接导致Meta内部多个AI项目延期；与此同时，《The Information》指出Meta内部下达指令限制员工在特定研发部门调用Claude与Codex，防止外部专有数据意外蒸馏进Meta自身的基础架构中。在此背景下，Meta在《Nature》上发布了非侵入式脑电解码项目Brain-to-QWERTY V2，展示了将脑磁图信号实时转译为语义文本的研究进展。 在硬件供应链端，半导体上游环节正在实质性攫取AI产业链的利润。美光科技（Micron）、三星电子和SK海力士因垄断高带宽内存（HBM）和高端DRAM产能，掌握了极高的定价权。截至5月28日的财季报告中，美光在出货量仅录得个位数增幅的情况下将DRAM售价提高超过60%，NAND闪存价格涨幅更逾80%，推动其市值迈入万亿美元门槛。硬件成本的通胀正向终端传导：苹果将MacBook售价提高逾15%，消费级组装电脑的内存价格在一年内暴涨两倍并反超CPU单价。随着前沿AI实验室普遍采取牺牲短期盈利以争夺市场份额的策略，存储芯片的溢价正在持续加剧整个模型开发阵营的运营性亏损。 历史并购启示与产业资本版图重组 (23:50) 节目通过回顾硅谷经典商业历史反思了资产定价偏差。《Wired》披露的2006年雅虎收购Facebook案显示，当时年仅22岁、创业18个月的Mark Zuckerberg曾口头同意以10亿美元将Facebook出售给雅虎；然而因雅虎财报不佳及广告平台延期导致其股价单日暴跌22%，雅虎首席执行官Terry Semel趁机将收购报价削减20%至8亿美元。这一举动促使Zuckerberg彻底终止协议，并在两个月后雅虎试图恢复10亿美元原价时坚决拒绝，最终使雅虎错失了移动互联网时代最关键的社交平台资产。 在当下的资本运作中，跨行业重组与投资并购依然活跃。传统电信巨头Comcast宣布分拆旗下媒体业务，将包含NBCUniversal与Sky的电影、流媒体及主题公园板块与核心电信网络资产进行结构性剥离。在风险投资领域，Chamath Palihapitiya为其创办的企业890敲定了1.35亿美元的A轮融资，参投方汇聚了Salesforce Ventures、Craft Ventures、Launch以及David Friedberg旗下的The Production Board，促成了《All-In Podcast》四位联合主持背后投资机构在单一创业项目上的集中注资。 本期讨论表明，AI产业的竞争重心正从单纯的前沿模型基准比拼，扩散为开源地缘博弈、算力与存储硬件瓶颈对开发者利润的实质性挤压，以及企业在合规与实际交付成本之间的动态权衡。核心未决问题在于：当上游硬件供应商持续吸附产业链绝大部分流动资金时，前沿AI实验室在缺乏稳定商业回报模型的情况下还能维持多久的高强度现金消耗。

## Background 本期内容由科技商业脱口秀频道 TBPN（由 John Coogan 和 Jordi Hays 主持）于 2026 年 7 月 23 日发布在 YouTube 平台上。这期 Diet TBPN 浓缩精选了当期直播节目的核心亮点，重点探讨开源与闭源大模型博弈、存储芯片巨头蚕食 AI 行业利润、谷歌与 Meta 的算力争端以及康卡斯特重组等前沿科技和商业大事件。 ## TL;DR 节目首先从智谱 AI 的开源权重模型 GLM-5.2 登上《华尔街日报》头条切入，深入剖析了中美在开源与闭源前沿的安全对抗、单任务成本陷阱以及当前模型生态向“顶级旗舰”与“极速轻量”两极分化的格局。随后，主播们探讨了美光、三星与 SK 海力士凭借内存与闪存涨价狂揽数百亿美元并推升美光迈入万亿美元市值、反向蚕食下游 AI 模型利润的行业异象。此外，节目还评述了谷歌对 Meta 限制 Gemini 算力配额、Meta 脑电实时解码新技术、2006 年雅虎临阵砍价错失 Facebook 的历史悬案，以及康卡斯特重组分拆背后线下体验经济与数字博彩的残酷竞争。 ## Atmosphere & style 整期节目的氛围兼具硅谷内线茶话会的辛辣幽默与宏观商业观察的冷峻敏锐，主持人和嘉宾在硬核的技术拆解与不修边幅的插科打诨间自如切换。从调侃 GLM 像“黑市补剂”、嘲讽 Meta 读脑设备“占了半个足球场”，到复盘雅虎高管自作聪明的历史败局，演播室内的笑声与犀利吐槽不断。整体节奏紧凑、信息密度极高，让听众在轻松随性的闲聊体感中迅速洞悉数百亿美元资本流转背后的残酷商业博弈。 ## Bulletin - [00:02](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=2s) 智谱 AI 开源权重模型 GLM-5.2 登上《华尔街日报》头条，其在网络安全漏洞挖掘上的表现追平甚至超越 Claude Opus 4.8，并迅速冲入 OpenRouter 调用量前十。 - [19:00](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=1140s) 谷歌因内部计算容量严重紧张，对 Meta 采购 Gemini 模型的算力额度设定严格上限，导致 Meta 内部多个关键 AI 项目推进受阻延期。 - [27:57](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=1677s) 美光、三星与 SK 海力士垄断高速存储供给，DRAM 与 NAND 价格暴涨逾 60%-80%，美光市值突破 1 万亿美元，正如同“欧佩克”般吞噬下游 AI 产业链的巨额利润。 - [15:45](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=945s) AI 模型市场正加速两极分化为“高价极致智商的前沿闭源模型”与“极速超低单价的微型开源模型”，中间层（Medium-sized）开源模型在推理性价比与绝对性能上面临尴尬的定位危机。 ## Key takeaways - [00:02](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=2s) 智谱 GLM-5.2 的开源权重打破了商业 API 黑盒调用模式，免监管与可离线修改的特性既能赋能深度定制，也给白宫政策制定与网络安全防御带来了严峻挑战。 - [03:13](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=193s) 网络安全评测显示 GLM-5.2 具备强大的漏洞挖掘能力，但 Epoch AI 等跟踪数据表明，闭源前沿与开源追赶阵营之间的代际时间差自 2023 年以来仍基本保持恒定。 - [06:32](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=392s) 评估模型不能仅看 API 单价，许多模型由于长思考链导致单任务消耗极高（Token-hungry），叠加无经济可行性的“测试期算力缩放”刷榜，实际工程落地必须以单任务综合成本为准。 - [06:32](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=392s) Meta 曾试图通过开源将对手护城河商品化，但在严苛的资本开支与前沿推理壁垒面前，如今自身也被迫斥巨资采购外部闭源前沿模型，昔日开源颠覆闭源的剧本并未成真。 - [09:32](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=572s) 乔治·霍茨指出，中国开源顶尖生产力工具正在对高度依赖服务业的美国经济施加通缩压力，而在多极化竞争下，华盛顿试图通过监管俘获来垄断 AI 技术已成徒劳。 - [10:26](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=626s) 闭源前沿模型比开源公开模型拥有一段关键的时间窗口，CrowdStrike 等安全巨头率先接入前沿闭源模型加固底层系统，抢在开源武器平民化之前完成漏洞修补。 - [12:26](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=746s) 开源模型在代码基准上表现抢眼但在创意写作等领域缺乏泛化韵味，原因在于依赖代理高频抓取闭源输出与互联网合成数据污染，导致过度拟合跑分任务。 - [15:45](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=945s) 模型市场呈现两极分化：企业在代码 Agent 等容错率为零的关键场景宁肯承受天价调用前沿闭源，在批量高频任务中则选择廉价轻量模型，中端模型处境日益逼仄。 - [18:16](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=1096s) 政府依赖行政审核与 KYC 限制模型使用的监管逻辑严重滞后，漫长的数月审批流程往往会被开源社区迅速推出的性能持平替代品瞬间瓦解。 - [19:00](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=1140s) 谷歌限制 Meta 调用 Gemini 计算容量，不仅反映出全行业不计成本抢夺算力导致的供给告急，也折射出大厂对竞争对手利用大规模 API 调用反向蒸馏核心模型的深层戒备。 - [21:35](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=1295s) Meta 在《Nature》公布的 Brain-to-QWERTY V2 实现了脑电信号到完整语义句子的端到端实时流式解码，但所依赖的脑磁图（MEG）仪器庞大无比，离日常穿戴化尚有巨大距离。 - [23:50](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=1430s) 2006 年雅虎 CEO 塞梅尔因股价大跌在最后一刻将 Facebook 的收购报价从 10 亿压至 8 亿美元，激怒扎克伯格导致交易流产，彻底改写了现代社交网络的权力版图。 - [27:57](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=1677s) 下游基础模型厂商由于拉新竞争无法轻易提高订阅价格，而上游内存单价翻倍暴涨，使得 AI 淘金热中的实际真金白银现金流正全面向底层存储三巨头转移。 - [29:25](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=1765s) 康卡斯特分拆业务折射出传统重资产主题乐园面临的长折旧周期与高资本壁垒，当今休闲娱乐消费正分化为线上廉价数字内容（及数字博彩）与极少数高溢价现场体验的杠铃形态。 ## Detailed walkthrough ### [00:02](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=2s) 《华尔街日报》头条关注中国开源模型 GLM-5.2 与安全能力突破 当人工智能新闻登上《华尔街日报》头条、进入国际新闻版块而不再局限于商业金融专栏时，你就知道华盛顿和整个政界正在猛然惊醒。头版配图虽然是一场热浪，但字号最大的头条标题赫然写着：“随着安全模型取得突破，中国重置中美 AI 竞赛格局”（China resets the AI race with the United States as security models mark gains）。主持人感叹，原本以为开源与闭源的争论早该尘埃落定——要么是技术前沿彻底崩塌、模型走向完全的平庸商品化（commoditization），要么是开源阵营被甩得看不见尾灯；然而现实却如同钟摆，不断在“一切都完了”与“我们彻底杀回来了”之间反复横跳。 这波舆论震动的焦点是智谱 AI（Zhipu AI，亦称 Z.ai）于 6 月 13 日正式发布的开源权重模型 GLM-5.2。尽管距离发布过去了几周时间，但它在基准测试中的强势表现和开发者社区的好评，终于彻底引爆了主流政经媒体。《华尔街日报》援引安全研究人员的评估指出，该模型在挖掘网络安全漏洞（finding security bugs）的能力上已经能够匹敌美国最新顶尖模型，这不仅可能重新洗牌全球技术竞争，更直接给正在重修 AI 政策的白宫施加了巨大压力。与 OpenAI 或 Anthropic 必须通过商业 API 调用的黑盒模式截然不同，GLM-5.2 是完全开放权重的（open weight）。只要你拥有足够的算力和电力，就能在私有硬件上直接部署、修改并完全不受监管地离线运行。这种特质对于追求绝对自主掌控的机构是福音，但对于游走在暗处的黑客而言，同样意味着毫无约束的攻击武器——主持人和嘉宾戏谑地将这种状态命名为“无羁智能”（Unfettered Intelligence），甚至调侃这名字非常适合拿去给硅谷新成立的 AI 实验室当门牌。 根据聚合了 400 多个模型访问入口的平台 OpenRouter 的统计，GLM-5.2 发布后迅速跻身该平台使用量排名前十的模型之列。主持人大赞 OpenRouter 创始人 Alex Atallah 商业嗅觉敏锐，不亲自下场卷基准测试，而是将平台打造成所有开发者接入 AI 生态的第一道大门，堪称极其成功的“卖水人”商业模式。 - **头条发布细节**：智谱 AI 于 6 月 13 日正式发布开源权重模型 GLM-5.2。 - **开源权重特征**：免 API 准入限制、支持本地部署与任意代码修改，既赋能深度定制，也带来了不可控的安全威胁隐患。 - **市场采纳热度**：在涵盖超 400 个模型的 OpenRouter 路由平台上，GLM-5.2 迅速冲进前十最常用模型之列。 --- ### [03:13](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=193s) 基准测试之争：闭源先锋与开源追赶的 Elo 曲线 根据网络安全公司 Semgrep 的基准测试结果，GLM-5.2 在特定测试中击败了 Anthropic 于 5 月份发布的旗舰模型 Claude Opus 4.8；在给予进一步指令提示后，Opus 4.8 与 GLM-5.2 在漏洞挖掘能力上甚至能双双追平顶尖水准。在此之前，美国 AI 圈内广泛流传着一种乐观论调：认为开源 AI 相对闭源前沿的发展速度正在放缓；凭借庞大的资本市场、顶尖的数据中心基础设施和顶级科研人才，美国闭源前沿实验室将维持更陡峭的增长斜率，从而形成难以逾越的指数级复利壁垒。聊天室里观众 Ryan 开玩笑称，GLM-5.2 这名字听起来实在太像某种地下黑市兜售的多肽补剂（gray market peptide），引来演播室一阵欢笑。 随后节目切入了一张反映全球模型综合能力的评估图表。该图表由人工智能标准与创新中心（Center for AI Standards and Innovation，CAISI）基于 5 月初的模型评测体系构建，通过对多项基准测试加权折算出模型的综合 Elo 天梯分，串联起从 GPT-4o、o1、o3-mini、Opus 4.6 一路延伸至 GPT-5 系列的技术演进直线；图中的红线则代表中国模型的发展轨迹。嘉宾 Tyler 指出，由于 CAISI 采用的部分测试集并不对外公开，因此很难直接将后来发布的模型精确嵌入该榜单。不过从公开指标来看，GLM-5.2 显然从中国模型的历史趋势线上大幅向上跳跃了一大步。Tyler 同时补充，类似 Epoch AI 等独立研究机构的追踪则显示，自 2023 年以来，闭源与开源模型之间的实际能力代差整体上保持着一种相对稳定的延迟，并未如一些悲观论调所言被无限拉大。 ```mermaid flowchart TD accTitle: AI 模型演进与评测分歧图 accDescr: 展示中美闭源前沿与开源追赶在能力评估与成本结构上的对比分析 subgraph Evaluation ["CAISI Elo 基准测试趋势"] A["美国闭源梯队<br>(GPT-5系列 / Opus 4.8)"] -->|线性推进| B["前沿能力壁垒"] C["中国开源梯队<br>(历史红线趋势)"] -->|GLM-5.2 突破| D["大幅跳升追平特定场景"] end subgraph RealityGap ["现实落地维度衡量"] E["Token 价格<br>(单价极低)"] --> G{"实际任务效率评估"} F["思考 Token 消耗<br>(Token Hungry)"] --> G G --> H["单任务综合成本<br>(Cost per Task)"] end ``` - **安全评测对比**：网络安全机构 Semgrep 证实，GLM-5.2 在安全测试中压制了 5 月面世的 Claude Opus 4.8，具备极高的漏洞发现命中率。 - **机构趋势观察**：Epoch AI 的长期跟踪数据显示，闭源顶尖模型与开源模型的代际差距自 2023 年以来始终维持在相对固定的时间窗口。 --- ### [06:32](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=392s) Token 单价与任务成本之辩，以及闭源护城河的终极推演 在深入探讨模型能力时，讨论重心转向了“单 Token 成本”（Cost per token）与“单任务成本”（Cost per task）的本质区别。Tyler 和主持人指出，许多新模型虽然宣称单价大幅跳水，但在实际执行中极度“吞 Token”（Token hungry），完成一次高质量推理需要输出冗长的思考过程。此外，业内存在大量依靠“测试期算力缩放”（test-time scaling laws）刷榜的现象——只要为一个问题硬砸上百万美元的算力，所有模型都能给出惊艳答案，但这在企业级生产环境或哪怕是黑客实际攻击中都毫无经济可行性。GLM-5.2 尽管在单 Token 计费上远低于闭源旗舰，但若按最终完成一个具体任务的综合消耗来算，优势往往会被其庞大的 Token 吞吐量部分稀释。 节目进一步重温了投资人 John Luttig 于 2024 年 5 月（即 DeepSeek 震撼行业之前）撰写的一篇著名长文，该文曾预言“基础模型的未来终属于闭源”。Luttig 当时提出了三个核心支柱： 1. **闭源数据飞轮**（closed-source data flywheels）； 2. **模型预训练呈指数级攀升的 CapEx 资本开支强度**； 3. **开源生态持续走低的 ROI**。 Luttig 认为，开源模型虽然可以在需要小型化、高度定制化和企业内部私有工作流中立足，但真正统治价值分配与技术前沿的必然是重资本支撑的闭源巨兽；大厂开源底层模型本质上只是一种免费的市场营销手段，或是为了“将自身业务的互补品商品化”（commoditize your complements）。 这套理论恰恰精准对应了当年马克·扎克伯格（Mark Zuckerberg）在 Meta 大力推行 Llama 系列的战略逻辑：彼时 Meta 试图通过开源树立顶尖 AI 人才聚集地的技术形象，即便短期无法直接从模型服务中获取飞涨的 ARR（年度经常性收入），也能通过压低全行业底座成本来反哺自身应用，甚至期望将其他闭源实验室的护城河彻底抹平。然而时间来到 2026 年，现实却充满讽刺——最新行业报道披露，Meta 反而在 Gemini 等外部闭源前沿模型上投入了巨额采购预算，当年设想的“借开源反杀闭源”并未完全按剧本上演。与此同时，自 2025 年初 DeepSeek 崛起后，整个开源竞技场的博弈性质彻底蜕变，演变成了一场关乎全球地缘政治与产业博弈的复杂大戏。 - **成本结构认知升级**：单纯比较 API 单价容易产生误导，Token 消耗量与推理效率决定的“单任务总成本”才是衡量模型落地的关键指标。 - **测试期算力陷阱**：在推理性任务中通过无限堆砌计算资源虽能拔高基准分数，但对商业企业级落地而言缺乏实际工程价值。 - **闭源资本论验证与反转**：Meta 等巨头原期望通过开源压制闭源利润空间，但在巨额算力竞赛和推理能力跃迁面前，自身最终仍不得不掏出巨资采购最顶尖的闭源模型。 ### [09:32](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=572s) 乔治·霍茨的通缩论与开源 AI 的中美地缘博弈 节目首先探讨了知名黑客与创业者乔治·霍茨（George Hotz，网名 geohot）近期发表的一篇观点敏锐的博文——《AI 将带来巨大的通缩》（AI will be massively deflationary）。霍茨在文章中剖析了为何中国在开源 AI 领域的投资收益可能远大于美国企业。他指出，中国开发者之所以愿意把耗费可观资源训练出的先进模型直接免费开源给全球，背后其实有着深刻的宏观经济逻辑：美国是一个高度依赖服务业的经济体，而这类开源免费的顶级生产力工具，正在以极快的速度对服务业的传统价值进行通缩打击。在霍茨的推演中，如果免费的 AI 能直接击垮或极度压缩服务业的利润空间，这对主要依靠实体制造的中国经济而言无疑是一大优势。 霍茨还进一步直言，华盛顿政客和某些大公司试图通过监管俘获（regulatory capture）来维持垄断的想法完全是痴人说梦。当今的世界早已不再是单极格局，世界上没有任何单一势力或政府能够真正垄断 AI 技术的演进。因此，他犀利地把目前华盛顿特区围绕 AI 监管、版权和垄断的各种游说与博弈，讽刺为“泰坦尼克号上重新摆放甲板躺椅”——在历史性的大浪潮面前，这些政治操弄既荒谬又无关痛痒。 ### [10:26](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=626s) 达里奥·阿莫代伊的预言与闭源防御的“时间窗口” 主播们随后将话题引向开源大模型对美国安全带来的现实冲击。近期社交网络上疯狂重新流传着一段 Anthropic 首席执行官达里奥·阿莫代伊（Dario Amodei）在 2023 年国会作证的视频剪辑。虽然很多自媒体把它包装成阿莫代伊的最新发言，但那实际上是三年前的发言。然而，他在当时针对行业前沿时间线的预测展现出了惊人的前瞻性——阿莫代伊当时警告称，两到三年之内（也就是如今的 2025 至 2026 年），前沿模型将带来严峻的生物安全与网络安全风险。如果这类能力毫无防备地随着开源模型的无节制扩张而扩散，且缺乏相对应的反制能力，人类社会将步入一个极度危险的境地。 所幸的是，防御体系并没有停滞不前。主播们透露，与 CrowdStrike、Palo Alto Networks 等网络安全巨头的 CEO 交流发现，这些安全厂商早在数月前便已经接入了 Mythos 以及 GPT-5.5 Cyber 等顶级闭源防御模型，用以加固关键系统以抵御基于 LLM 的自动化渗透攻击。闭源前沿模型与开源公开模型之间存在的“代差”，恰好为白帽黑客创造了一段至关重要的时间窗口，让他们能在黑帽黑客利用开源模型批量挖掘漏洞之前完成补丁修复。然而，随着开源追赶速度的加快，这种安全时间差并没有显著拉大，美国网络安全与生物安全产业必须在能力被全面平民化、商品化之前，持续调整自身的防御姿态。 ```mermaid flowchart LR accTitle: 前沿闭源安全加固与开源风险时间差 accDescr: 展示闭源前沿模型如何先行为安全防御争取时间窗口，随后开源追赶普及风险。 subgraph Defense ["防御侧优先期（闭源窗口）"] A["前沿专有模型<br/>(Mythos / GPT-5.5 Cyber)"] --> B["顶尖安全企业<br/>(CrowdStrike / Palo Alto)"] B --> C["白帽黑客率先修复<br/>底层与系统级漏洞"] end subgraph Threat ["扩散与博弈期"] D["开源模型能力追赶<br/>(高危能力平民化)"] --> E["黑帽黑客尝试攻击"] C -.->|"建立加固防线<br/>抢占时间差"| E end ``` ### [12:26](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=746s) 评测 GLM-5.2 与开源模型的“蒸馏”争议 随后，主播泰勒（Tyler）分享了他对智谱 AI 最新开源的 GLM-5.2 模型的上手评测与行业反响。坊间针对该模型最激烈的质疑依然是那句老生常谈的“它肯定是蒸馏（distilled）出来的”。长久以来，西方科技圈有一种根深蒂固的声音，认为许多开源模型（尤其是来自中国的模型）之所以表现优异，完全是因为偷偷拿美国闭源实验室的模型输出来做蒸馏微调。此前 Anthropic 甚至公开指责阿里巴巴等实验室存在蒸馏其模型的行为。 主播们也谈到了这背后高度专业化的黑灰产运作：灰色网络服务商会将大量个人账号、企业订阅和 API 密钥打包隐藏在 VPN 之后，搭建起一个极其庞大、能以极高频率并发抓取闭源大模型输出的代理前端，专门用于采集高质量合成训练数据。主持人调侃道，如果这些团队真的只是直接抓取数据来训练，“难道在按下开始训练键之前，大家就不知道先全选‘查找替换’一下原模型实验室的名字吗？” 不过，现实中的数据污染远比直接调用 API 复杂得多。如今公共互联网、GitHub 开源仓库以及大量亚马逊 Kindle 电子书，已经被各类大模型的输出彻底淹没。当一个训练团队爬取公开代码库时，这些代码很可能原本就是由闭源模型生成的。模型在训练中不可避免地会继承某些独特的生成痕迹（例如某些高频特定句式“不是……而是……”）。 这种基于蒸馏或合成数据训练出的模型往往存在一个致命软肋：泛化能力较差。它们在基准测试（Benchmarks）上可以刷出极其惊人的高分，甚至往往是“不经意间的刷榜”，但一到实际复杂场景中，就会暴露出缺乏超大模型那种无法言喻的“底层韵味”（je ne sais quoi）。在实际使用反馈中，GLM-5.2 在纯代码任务上广受好评，表现非常出色，但在创意写作等更难被量化跑分的领域则略显逊色。当被问及该模型是否能通过某些敏感历史事件（如 Tiananmen Square 测试）时，泰勒指出该模型毕竟是完全开源的，社区开发者即使遇到内置限制，也可以非常轻松地通过本地微调直接绕过。 ### [15:45](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=945s) 模型市场的两极分化与中端模型的尴尬处境 围绕 API 调用成本与企业预算，两位主播深入剖析了当今 AI 模型市场残酷的两极分化现状。由于前沿模型在边际智商提升上的算力消耗呈指数级暴增，调用顶级模型的账单正在拖垮许多公司的资产负债表。这种极高的成本门槛直接催生了当前模型市场的两大主流阵营： 1. **绝对顶级的闭源前沿模型（Frontier Models）**：主打高风险、高价值场景。例如自主编码智能体（Coding Agents）或生死攸关的网络安全对抗，在这些场景下一旦出错代价惨重，企业宁愿支付天价 Token 费用也必须追求最高智商。 2. **极小、极快、极便宜的轻量级模型（Small/Fast Models）**：以 DeepSeek Flash 等为代表，专门用来解决单一、高频、重复的特定任务。 而在 OpenRouter 等聚合平台上，调用量长期霸榜的几乎全都是这些极便宜的小型开源模型。主持人举例打趣道，就像 Ramp 每天要处理海量的商业发票报销，系统每一张小票都要过一遍大模型，但“难道我需要一个通用超级智能（AGI）来告诉我刚刚在咖啡店花了 10 美元吗？不需要，它只需要做好基本的 OCR 信息抽取就行了”。 ```mermaid flowchart TD accTitle: AI 模型市场的两极分化生态 accDescr: 说明顶尖高价前沿模型与小快轻量模型主导市场，导致中端模型定位尴尬。 Market["AI 模型需求市场"] --> ClassA["第一极：前沿旗舰模型<br/>(高智力 / 高单价 / 闭源为主)"] Market --> ClassB["第二极：微型轻量模型<br/>(超低成本 / 极速 / 开源为主)"] Market -.-> MidClass["失落的中间层 (Mid-tier)<br/>定位模糊 / 商业化艰难"] ClassA --> UseA["关键代码 Agents、核心网络安全加固<br/>『不惜代价，容错率为零』"] ClassB --> UseB["海量小票 OCR、特定重复子任务、多智能体路由<br/>『OpenRouter 流量主力，主打性价比』"] ``` 这也让处在中间层（Medium-sized）的开源模型陷入了尴尬的定位危机：论极致能力，它比不上不惜成本堆叠的前沿模型；论推理性价比，它又不如轻量模型便宜迅捷。或许目前这类中端模型最大的受众，是不愿承担闭源高昂 Token 成本、又渴望使用个人编码 Agent 的极客与业余爱好者。 主播最后总结道，尽管存在种种争议，GLM 等开源阵营展现出的绝对实力绝不能被轻易忽视。“闭源壁垒正在无限扩大、无需担忧开源追赶”的论调是完全站不住脚的。开源模型的飞速迭代正极大压缩闭源实验室的先发红利期，每一个全新前沿模型的商业化变现窗口，正在变得前所未有的狭窄和充满挑战。 ### [18:16](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=1096s) 政府监管滞后与AI模型许可的现实悖论 主持人探讨了政府试图通过“了解你的客户”（KYC）机制或专属企业审批来严格限制模型使用的潜在漏洞。这种监管逻辑在飞速迭代的AI行业中显得格外笨拙：如果一家公司为了合法使用尖端前沿模型（比如假设中的“GPT-7”）而提交申请，政府部门往往需要耗费数月进行审核流程；然而就在等待批复的短短三个月里，开源界或竞争对手可能已经推出了性能几乎持平的替代品（例如“GLM-6”）。 当企业已经能够轻松获取性能“足够接近”的公开模型时，原本严密的行政审批壁垒瞬间失去了实际意义。这种技术演进速度与官僚决策流程之间的严重脱节，给政府正在拟定的一整套AI监管战略制造了巨大障碍，决策者不得不重新思考如何应对技术飞速扩散带来的政策失效。 ### [19:00](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=1140s) 谷歌对Meta限制Gemini算力供应与模型蒸馏之忧 《金融时报》（Financial Times）爆出猛料，谷歌由于自身计算容量遭遇严峻挑战，已对Meta使用其Gemini大模型的额度设定了严格上限。随着全行业对先进模型的计算需求爆发式增长，算力已成为极度稀缺的战略资源。消息人士透露，Meta曾在三月份试图向其竞争对手谷歌购买超出预期的庞大Gemini容量，但谷歌明确告知无法满足这一采购规模，直接导致Meta内部数个AI研发项目的推进受阻与延期。主持人在节目中打趣道，报道配图里居然有一辆印着Gemini标志的自行车，令人忍俊不禁。 两位主持人针对这一事件展开了深入推测。一方面，科技巨头普遍陷入了“刷Token狂潮”（token maxing），各家都在不计成本地疯狂砸钱抢占算力，谷歌因此在算力售罄的情况下优先保全自身核心业务完全在情理之中，这也直接印证了Google Cloud在财报中所展现的强劲上行势头；另一方面，这也引出了业内对“模型蒸馏”（distillation）的敏感担忧。虽然此前有传言称Meta由于担心知识产权和蒸馏争议而严禁内部员工在某些核心业务中使用Claude和Codex，但主持人们好奇谷歌是否也存有戒心，担心竞争对手利用大规模调用Gemini来提炼并训练自家的底层模型。 - 谷歌在三月告知Meta无法提供其申请的全部Gemini计算容量，导致Meta部分内部AI项目被迫推迟。 - 面对算力限制与降本增效的考量，Meta已在内部要求员工更加节制、高效地使用AI Token。 - 谷歌的多家客户均受到了不同程度的配额限制，但Meta由于采购需求极其庞大，受到的冲击最为显著。 ### [21:35](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=1295s) Meta发布脑电解码器：从字符到语义的“读脑”技术与巨型设备调侃 Meta在科研领域公布了一项名为“Brain-to-QWERTY V2”的突破性进展，相关成果发表于《自然》（Nature）期刊。这项非侵入式脑电信号解码技术构建了一个端到端的实时流水线，能够直接从人类大脑的原始电信号中实时解码出完整的句子内容。与此前局限于字符级预测的V1版本相比，V2版本跨越到了对词汇和抽象语义的直接解码，大幅提升了大脑信号转文字的沟通准确度。 ```mermaid flowchart LR accTitle: Brain-to-QWERTY 技术演进 accDescr: 展示从V1字符级解码向V2全链路语义解码的技术跨越 subgraph Input["信号输入"] A["原始脑电信号<br>(MEG设备采集)"] end subgraph Pipeline["解码架构演进"] B["V1 代架构<br>字符级预测"] C["V2 代架构<br>端到端实时流式解码"] end subgraph Output["语义输出"] D["单一字符拼写"] E["实时完整词汇与语义<br>(通顺句子)"] end A --> B A --> C B --> D C --> E ``` 这一前沿技术立刻引来了主持人的犀利幽默：过去大众总怀疑Instagram在偷偷监听麦克风，现在终于有了全新的阴谋论——“Meta可能正在直接监听你的思想”。然而，当屏幕上展示出该项研究所使用的脑磁图（MEG）仪器时，现场爆发出一阵大笑。这台所谓的“非侵入式”仪器体积硕大无朋，甚至需要专门的屏蔽房间，画面滑到一半都看不到全貌，看起来“足足占了半个美式足球场”。两人调侃道，虽然设备不需要切开头盖骨植入电极，但普通人绝对不可能把它当成日常可穿戴设备出门，不过能在实验室里坐下来、亲眼看到脑海中的想法直接跃然屏幕上，依然堪称极度震撼的演示。 - 话题延伸至知名风投家Rob Toews此前在《福布斯》上发表的五大AI预言，虽然其关于Anthropic规模扩大、台积电面临更激烈竞争等判断十分理性，但他预测“到2030年心灵感应将变得司空见惯”，被主持人认为过于激进。 ### [23:50](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=1430s) 雅虎错失Facebook：Terry Semel临阵砍价造成的历史性败局 节目翻开了硅谷历史上最惨烈的一段商业收购悬案：2006年雅虎时任CEO特里·塞梅尔（Terry Semel）如何搞砸了对Facebook的并购。当时创办仅约18个月的马克·扎克伯格（Mark Zuckerberg）面临着各方重压，终于在2006年7月原则上口头同意以10亿美元全现金的价格将Facebook卖给雅虎。当时这在战略上堪称绝配——雅虎拥有数亿庞大用户群却做不好社交网络，而Facebook拥有极具粘性的酷炫产品却急需大众流量。 然而，命运开了一个残酷的玩笑。就在扎克伯格点头后几天，雅虎因核心广告平台研发受阻发布了糟糕的盈利预警，股价在一夜之间暴跌22%。面对突如其来的压力，塞梅尔做出了一个灾难性的决定：他利用扎克伯格经验尚浅的劣势，单方面将收购报价直接砍掉了20%，削减至8亿美元。早已得知塞梅尔习惯在最后一刻压价的扎克伯格感受到了极大的侮辱，毅然掀桌离席。尽管两个月后塞梅尔反悔并重新开出最初的10亿美元全价，但扎克伯格已经借机说服了董事会和高管团队，坚信雅虎绝非严肃可靠的合作伙伴，Facebook独自发展将远超这个估值。小扎由此一战成名，成为《连线》（Wired）杂志笔下那个“狂妄拒绝10亿美元的年轻人”。 主持人们感叹历史的蝴蝶效应：如果当年雅虎真的吞并了Facebook，凭借雅虎日后的平庸官僚文化，Facebook很可能丧失对顶级人才的吸引力，甚至根本没有魄力和资本在日后豪赌收购Instagram与WhatsApp，从而彻底改写当今整个社交帝国的格局。 ### [26:31](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=1591s) 内存芯片巨头吞噬AI利润：美光、三星与SK海力士的“石油垄断” 科技行业正在上演一场波澜壮阔的财富大转移：AI开发商及下游产业链产生的巨额利润，正以惊人的速度流向最上游的内存芯片制造商。商业与财经各大媒体的头条都在探讨这一现象。美光科技（Micron Technology）最新季度的爆炸性利润增长让其股东欢欣鼓舞，但这份繁荣的代价却是所有抢购高速内存芯片的AI模型公司承担的高昂成本。 美光与韩国的三星电子（Samsung Electronics）、SK海力士（SK Hynix）构成了事实上的垄断三角，它们在当今AI生态中的地位，就像是欧佩克石油巨头之于航空业——供应着飞行动力所必需、短期内绝无替代可能的刚性命脉。由于AI大模型训练与推理所需的高带宽内存（HBM）制造难度极高、产能极其紧张，且新建半导体晶圆厂往往需要耗费数年时间，数据中心激增的算力部署瞬间推高了芯片单价。在截至5月28日的财季中，美光在出货量仅录得低个位数微幅增长的情况下，将DRAM芯片的价格较前三个月暴拉了超过60%。这种级别的利润转移在现代商业史上极度罕见，整个AI淘金热中真正的现金流，正在全面向底层存储巨头的口袋汇集。 ### [27:57](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=1677s) 内存芯片暴涨蚕食 AI 利润与美光突破万亿市值 数据中心使用的 NAND 闪存价格飙升超过 80%，这一轮供应链震荡正深远地改变着科技行业的成本结构。通常情况下，内存成本并不起眼，但仅在最近一个季度，美光（Micron）的客户就多支付了 180 亿美元；在过去一年里，内存价格足足翻了两番。这种涨价潮也蔓延到了人工智能之外的消费电子领域：苹果公司上周将 MacBook 的售价上调了 15% 以上。主持人分享了一件贴近生活的轶事——他一年前在亚马逊上组装一台超静音电脑（“我极度讨厌风扇噪音”）时购买的内存条，如今价格翻了整整两倍，“现在的价格甚至比 CPU 还要贵”。对于一个习惯了产品逐年降价的半导体细分领域而言，这无疑是一场翻天覆地的逆转。 消费类电子产品可以通过提高终端售价将成本转嫁给消费者，从而像高油价抑制汽油消费一样自然抑制芯片需求。然而，AI 公司却无法简单地将上涨成本转嫁给用户，因为它们正试图通过烧钱补贴来掩盖供应瓶颈。目前人工智能领域的核心痛点在于：终端口用户支付的订阅费根本无法覆盖庞大的基础设施算力成本，各大基础模型厂商正承受着巨额亏损。当前所有产品的定价策略依然是为了争夺拉新，而非盈利。原材料输入成本的暴涨引发了一场极其棘手的两难困境：要么吞下更惨烈的亏损，要么提高价格直接劝退潜在用户。与下游模型厂商的窘境截然相反，美光的股价一路狂飙，正式跻身“1万亿美元市值俱乐部”。主持人随口打趣提问：“泰勒，好奇问一下，欧洲现在有几家市值达万亿美元的公司？”得到的回答干脆利落：“我猜是零。”“答对了。” - **NAND 闪存涨幅**：数据中心闪存价格单季度上涨超过 80%，内存整体价格在一年内暴涨 4 倍。 - **美光季度吸金**：美光客户单季度为采购内存额外支出了 180 亿美元，助推美光市值突破 1 万亿美元。 - **消费电子转嫁成本**：苹果 MacBook 提价超过 15%，零售端装机内存条价格翻了三倍并反超 CPU 单价。 - **AI 商业模型困局**：模型厂商仍在补贴用户以换取市场扩张，内存涨价使下游厂商陷入“加剧亏损”或“涨价丢用户”的两难。 --- ### [29:25](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=1765s) 康卡斯特分拆业务与线下主题乐园的残酷经济学 康卡斯特（Comcast）计划对其业务进行重大重组，将 NBCUniversal 和 Sky 的宽带连接业务与旗下的电影、流媒体以及主题乐园运营彻底拆分。针对环球影城（Universal Studios）的剥离，主持人抛出了一个脑洞大开的创业假想：“难道就没有机会以现代技术栈（Modern Tech Stack）打造一个全新的主题乐园，做‘主题乐园界的 Anduril’吗？”然而，这个想法立刻遭到了现实商业逻辑的无情反驳。构建现代科技主题乐园的资本开销极高，不仅所有设备都要适配严苛的安全与工程标准，且过山车等大型游乐设施的折旧周期往往长达 20 年。正如对谈者所指出的，在迪士尼从事主题乐园开发的朋友都深知这门生意的艰难，“这绝对是一门极其残酷的生意”。 更有力的挑战在于娱乐形态的变迁。过去的主题乐园诞生于没有海量数字内容的时代，而如今用户在手机上就能随时消费满足每一种小众偏好的无尽内容。不过，市场上目前呈现出耐人寻味的“杠铃策略”（Barbell Strategy）：一方面，线下真实体验（IRL Experiences）正享有前所未有的溢价定价权——即便人们在 TikTok 上能免费刷到纽约尼克斯队的精彩集锦，依然有人愿意掏 5,000 美元买一张现场球票；风投机构 Thrive Capital 也正在收购旧金山巨人队（San Francisco Giants）的股权。但与此同时，另一组数据展示了数字化娱乐对实体经济的剧烈蚕食：上周公布的统计数据显示，体育博彩（Sports Betting）的交易量已经超过了电影票、实体剧场、主题乐园等数个主要线下体验类别的销售总和。 ```mermaid flowchart TD accTitle: 现代娱乐消费的杠铃分化 accDescr: 展示数字化无尽内容、高溢价线下体验与数字博彩之间的竞争动态 subgraph Digital["线上与数字经济"] A["无尽数字娱乐<br/>(TikTok/流媒体/个性化内容)"] B["线上体育博彩<br/>(下注交易量碾压线下票务总和)"] end subgraph IRL["实体沉浸体验 (IRL)"] C["传统主题乐园<br/>(资本极重/折旧长达20年)"] D["顶级现场赛事与高溢价体验<br/>(如 5,000 美元现场球票/球团股权投资)"] end A -.->|"分流大众休闲时间"| C B -.->|"资金吸纳量压制"| C C <-->|"杠铃策略分化"| D ``` --- ### [31:29](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=1889s) “全明星好友团”集结与节目尾声 节目的最后，两位主持人关注了一笔硅谷风险投资圈的重磅交易：查马斯·帕里哈皮蒂亚（Chamath Palihapitiya）为其新项目（8090）完成了 1.35 亿美元的 A 轮融资。值得玩味的是本次融资的投资方阵容，不仅吸引了 Salesforce Ventures 和 WonderCo，更集齐了播客节目《All-In Podcast》的“铁哥们”（The Besties）：戴维·萨克斯（David Sacks）掌舵的 Craft Ventures、杰森·卡拉卡尼斯（Jason Calacanis）的 Launch，以及戴维·弗里德伯格（David Friedberg）旗下的 The Production Board。 主持人幽默地清点道，查马斯成功把《All-In》播客的所有好兄弟全都绑上了自己的战车，实现了全员集结。在轻松的调侃氛围中，主持人向观众道别，宣布今天的 Diet TBPN 到此结束，并提醒观众明天同一时间继续收看直播。 ## Notable quotes > [00:02](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=2s) 随着安全模型取得突破，中国重置中美 AI 竞赛格局 — 《华尔街日报》 > [03:13](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=193s) GLM-5.2 这名字听起来实在太像某种地下黑市兜售的多肽补剂 — Ryan > [06:32](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=392s) 基础模型的未来终属于闭源 — John Luttig > [09:32](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=572s) 泰坦尼克号上重新摆放甲板躺椅 — 乔治·霍茨 (George Hotz) > [12:26](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=746s) 难道在按下开始训练键之前，大家就不知道先全选‘查找替换’一下原模型实验室的名字吗？ — 主持人 > [15:45](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=945s) 难道我需要一个通用超级智能（AGI）来告诉我刚刚在咖啡店花了 10 美元吗？不需要，它只需要做好基本的 OCR 信息抽取就行了 — 主持人 > [21:35](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=1295s) Meta可能正在直接监听你的思想 — 主持人 > [23:50](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=1430s) 狂妄拒绝10亿美元的年轻人 — 《连线》(Wired) > [27:57](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=1677s) 现在的价格甚至比 CPU 还要贵 — 主持人 > [27:57](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=1677s) 我猜是零。 — 泰勒 (Tyler) > [29:25](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=1765s) 难道就没有机会以现代技术栈打造一个全新的主题乐园，做‘主题乐园界的 Anduril’吗？ — 主持人 > [29:25](https://www.youtube.com/watch?v=jgJYskq-1Yo&t=1765s) 这绝对是一门极其残酷的生意 — 主持人 ## Mentioned 智谱 AI (Zhipu AI), GLM-5.2, OpenRouter, Alex Atallah, Semgrep, Claude Opus 4.8, CAISI, Epoch AI, John Luttig, Mark Zuckerberg, Meta, Llama, DeepSeek, George Hotz, Dario Amodei, Anthropic, CrowdStrike, Palo Alto Networks, Mythos, GPT-5.5 Cyber, Ramp, DeepSeek Flash, Google, Gemini, Brain-to-QWERTY V2, Rob Toews, Terry Semel, Yahoo, Facebook, Micron Technology, Samsung Electronics, SK Hynix, Apple, Comcast, NBCUniversal, Sky, Universal Studios, Disney, Thrive Capital, Chamath Palihapitiya, All-In Podcast, David Sacks, Jason Calacanis, David Friedberg, Craft Ventures, Salesforce Ventures ## 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=jgJYskq-1Yo)

[0:02] 《华尔街日报》: 随着安全模型取得突破，中国重置中美 AI 竞赛格局 [3:13] Ryan: GLM-5.2 这名字听起来实在太像某种地下黑市兜售的多肽补剂 [6:32] John Luttig: 基础模型的未来终属于闭源 [9:32] 乔治·霍茨 (George Hotz): 泰坦尼克号上重新摆放甲板躺椅 [12:26] 主持人: 难道在按下开始训练键之前，大家就不知道先全选‘查找替换’一下原模型实验室的名字吗？ [15:45] 主持人: 难道我需要一个通用超级智能（AGI）来告诉我刚刚在咖啡店花了 10 美元吗？不需要，它只需要做好基本的 OCR 信息抽取就行了 [21:35] 主持人: Meta可能正在直接监听你的思想 [23:50] 《连线》(Wired): 狂妄拒绝10亿美元的年轻人 [27:57] 主持人: 现在的价格甚至比 CPU 还要贵 [27:57] 泰勒 (Tyler): 我猜是零。 [29:25] 主持人: 难道就没有机会以现代技术栈打造一个全新的主题乐园，做‘主题乐园界的 Anduril’吗？ [29:25] 主持人: 这绝对是一门极其残酷的生意

[0:02] Unfettered intelligence. [4:20] GLM-5.2 sounds too much like a gray market peptide you're taking. [19:04] If you thought Instagram was listening to your conversations, now you can have a new conspiracy at home, which is that they might be just listening to your thoughts. [26:31] We are witnessing an extraordinary transfer of cash from the providers of AI, and perhaps one day AI users, to memory chip makers.

[0:02] Well, on the front of the Wall Street [0:04] Journal today, this is how you know this [0:07] is the whole AI 2027 Washington waking [0:10] up. The AI stories are making it to the [0:13] front page, the the the world news [0:15] section, not just the business and [0:17] finance section, more and more. So, the [0:20] picture is about the heat wave, but the [0:22] lead, the the story with the largest [0:25] text, is about artificial intelligence. [0:27] China resets the AI race with the United [0:30] States as security models mark gains. [0:34] We're going to get into it. This is a [0:35] fascinating debate because I thought [0:37] that we'd have a conclusion to the [0:39] open-source AI debate by now. Either [0:43] they would the the frontier would have [0:44] collapsed and there would be, you know, [0:46] perfect commoditization, or they would [0:49] have fallen so far behind. [0:51] >> It'll It'll just go, "It's over. We're [0:53] so back. It's over." [0:55] >> If you're in open-source AI, that's [0:56] exactly how it feels. The big The big [0:58] story is centered around GLM 5.2 from [1:01] Z.ai, [1:03] uh is officially released June 13th. So, [1:05] it's taken a couple weeks for it to [1:07] really break through to the front page [1:09] of the Wall Street Journal, but uh [1:11] they're seeing some strong performance [1:12] on benchmarks, some positive reviews [1:14] from developers. I have a whole review [1:16] from Tyler we can go through in a little [1:17] bit, but we're now entering another [1:19] round of debates around open-source AI. [1:21] What can the model actually do? Is this [1:23] a threat to national security? What are [1:26] the geopolitical ramifications here? And [1:28] so, I'm sure this will be an ongoing [1:29] conversation throughout this week, [1:31] probably next week. We have some guests [1:33] lined up to help contextualize it. But, [1:35] uh laying down the facts from the [1:37] journal, "Security researchers said that [1:39] a new AI model released this month by [1:41] China's Zhipu AI, also known as Z.ai, [1:45] can match the latest US models when it [1:47] comes to finding security bugs, a [1:48] development poised to reset the global [1:51] tech race and pressure the White House [1:53] in its overhaul of US AI policy. So, [1:57] unlike models from Anthropic or OpenAI, [2:00] Zippo's GLM 5.2 is open weight. You can [2:03] just download it, run it anywhere. You [2:05] don't need to go to an API. You don't [2:07] need to go to a private company and pay [2:09] them. You can run it on your own server, [2:11] provided you have the electricity and [2:13] GPUs to do so. It is expensive to run, [2:16] as we'll go into, but it is open weight. [2:18] That means it can be downloaded right on [2:20] hardware operated by anybody and can be [2:22] modified and used without supervision. [2:24] Scary stuff. Open weight models are [2:26] ideal for users who want unfettered [2:28] access to systems they control, uh but [2:30] they're also ideal for hackers who want [2:32] to run them in the shadow. [2:34] >> Unfettered intelligence. [2:36] >> Unfettered Oh, that's a good [2:37] >> completely out of We are completely out [2:39] of names for new Neo Labs. [2:41] >> That's a good Neo Labs name, yeah. [2:43] >> Unfettered intelligence. [2:44] >> Unfettered intelligence is good. [2:46] Uh GLM 5.2 has ranked as one of the top [2:50] 10 most used AI models, according to [2:52] data from OpenRouter, a company that [2:54] provides access to more than 400 AI [2:56] models. And what a fantastic business, [2:59] uh Alex Atallah over there. Absolutely [3:01] cooking it OpenRouter. It's such an [3:02] exciting way to plug into the AI [3:06] the AI race without actually needing to [3:08] play the play the benchmark game so [3:10] much. Be the uh the front door. Anyway, [3:13] uh in some benchmarking tests, according [3:14] to cybersecurity company Semagrap, GLM [3:17] 5.2 bested Anthropic's Claude Opus 4.8 [3:21] model, which was released in May. When [3:23] given further instructions, Opus 4.8 and [3:26] GLM 5.2 can match Mythos in bug binding [3:30] ability, according to researchers. So, [3:31] prior to this launch, uh and there's a [3:34] chart that we should pull up here about [3:35] overall AI capability, and we can talk [3:37] to Tyler about what this chart actually [3:38] means, but there was this narrative [3:40] brewing that open-source AI was slowing [3:43] down relative to the closed-source [3:45] frontier. And I saw a lot of uh American [3:48] AI fans sort of cheer for this. Hey, the [3:53] we have the capital markets, we have the [3:55] data centers, we have the researchers, [3:57] and so we are able to push the frontier [4:00] at a different rate. And if we're [4:01] actually growing at a faster rate in [4:03] America within the closed-source labs, [4:06] that will compound and there will be a [4:08] stronger takeoff in the American [4:10] closed-source AI industry. Now, this [4:13] chart sort of goes back and forth and [4:15] there's some debate over it. It's in the [4:17] newsletter. You can go sign up at [4:18] tbpm.com. [4:20] >> While we're pulling that up, let me tell [4:21] you about Codex. Codex is a powerful [4:23] workspace for getting work done with AI [4:25] agents. Whether you're writing code, [4:27] analyzing data, creating content, or [4:29] automating business workflows, Codex [4:31] helps you move projects forward from [4:32] start to finish. This chart, which we [4:34] can pull up, shows progress from GPT-40 [4:37] to 01, 03 mini, 03 Opus, 4 GPT-5, 5.2, [4:43] Opus 4.6, GPT-5.4, GPT-5.5, [4:47] showing [4:48] you know linear trend in this Elo. [4:50] >> Ryan says GLM-5.2 sounds too much like a [4:54] gray market peptide [laughter] you're [4:55] taking. [4:56] >> It actually does. [4:58] It does sound a lot like that. And and [4:59] then you can see the red line are the [5:02] are the Chinese models, which are also [5:05] improving over time but at a slightly [5:08] lower rate. And so the question was are [5:10] they going to plateau while America's [5:12] progress continues to advance? This [5:15] latest model GLM-5.2 seems it's very [5:18] hard to apply it to this particular [5:20] benchmark because this Elo was Can you [5:22] give us some background, Tyler, on where [5:24] this chart came from, what this is [5:26] demonstrating? [5:27] >> Yeah, so this is by Casey. I think [5:29] that's how you pronounce it, the Center [5:30] for AI Standards and Innovation. They [5:31] they have this way to calculate like the [5:33] Elo of models. It's basically [5:34] approximation of a bunch of different [5:35] benchmarks. Um some of those like are [5:37] proprietary like they're they're they're [5:38] not open. So it's actually hard to run [5:40] these. Also cuz I was basically trying [5:42] to bench like all the recent models [5:44] since this was published. It it I I want [5:46] to say May 1st. [5:47] >> Yeah, it'd be great to throw 5.6 Soul, [5:49] Mythos, and Fable. All all I it would be [5:52] great to just continue this chart cuz [5:54] it's an interesting trend. [5:55] >> a lot of those benchmarks aren't [5:56] actually public, so it's very hard to [5:57] estimate. Um but I I tried I got you can [6:01] look at like some of the benchmarks that [6:02] that are public that you can reference. [6:04] You can kind of match them up to [6:05] previous models. 5.2 looks like it it it [6:07] it is like a big step up from the like [6:11] Chinese trend line, right? I think that [6:13] the the group of benchmarks that were [6:14] chosen for this ELO definitely [6:16] accentuate that the gap between US and [6:19] Chinese labs. Um I I think there's a [6:21] bunch of other like groups like Epoch AI [6:24] has done a chart. They basically a [6:26] relatively stable gap between close [6:28] source and open source models since like [6:30] 2023, like a long time. [6:32] >> Yeah. And perhaps at this point the the [6:35] discussion should be more centered [6:37] around cost per task more than cost per [6:40] token. [6:41] >> Yes. Yeah, definitely. Because even [6:42] like, you know, new models a lot of [6:44] times when when they come out like, [6:45] "Okay, maybe the token price is actually [6:47] the exact same, but the token efficiency [6:49] is much better." So, then when you do a [6:51] lot of these tasks, it's like it it's [6:52] not the the the price per token, it's [6:54] price per per, you know, something [6:56] completed. And then you actually see it [6:58] go down. [6:58] >> And there's a lot of test time scaling [7:00] laws where you can just throw a million [7:02] dollars of compute at a particular [7:03] problem and [7:05] all the models do really well at it, but [7:07] it's completely non-viable for any real [7:10] enterprise use case and probably not [7:12] even viable if you're trying to [7:13] be a nefarious hacker. [7:15] >> Yes. Most people are saying like 5.2 is [7:17] very token hungry, right? So, it uses a [7:18] lot of tokens. So, maybe it like it [7:20] definitely is much cheaper than the [7:21] frontier models. [7:22] >> On a per token basis. [7:23] >> On a per token basis. [7:24] >> per task basis, it might be more [7:27] expensive. [7:28] >> Yeah. I mean, on that that's still it's [7:29] generally not, but on on specific tasks, [7:31] you can get you know, if you have low [7:33] thinking models, low thinking mode on [7:35] the close source ones, you can see. [7:37] >> Well, let's revisit John Ludwick's post [7:40] from 2024, May 2024. This is pre-deep [7:43] seek talking about his [7:47] prediction about why the future of [7:48] foundation models is closed source. He [7:50] got a lot of pushback from this because [7:52] a lot of people like open-source models, [7:54] but he laid out a [7:56] thesis around closed-source data [7:59] flywheels, exponential capex intensity [8:02] of training. And he said open-source [8:04] will have a home wherever smaller, less [8:07] capable, and configurable models are [8:09] needed. Enterprise workloads, for [8:11] example. But the bulk of the value [8:13] creation and capture in AI will happen [8:15] using frontier capabilities. The impulse [8:18] to release open-source models makes [8:20] sense as a free marketing strategy and [8:23] as a path to commoditize your [8:24] complements, but open-source model [8:26] providers will lose the capital [8:28] expenditure war as open-source ROI [8:30] continues to decline. And that was the [8:32] thesis around the time that [8:36] the open-source AI discussion was [8:37] primarily driven by Mark Zuckerberg's [8:39] work at Meta on the llama family of [8:41] models. [8:42] The idea was that Meta would benefit [8:44] from attracting talent. It was good [8:46] marketing. It told the story that Meta [8:49] has an AI story and has AI talent [8:51] in-house, even if they weren't [8:53] monetizing it and sharing, you know, a [8:56] really fast takeoff in ARR around those [8:58] models. It showed that, "Hey, they're [9:00] able to develop these models." And and [9:02] that might help them cut their costs in [9:04] the long term. Very interesting that [9:06] that wound up being very different in [9:08] 2026 looking at the news today, which [9:11] we'll go into about them spending a lot [9:13] on Gemini. There's been reports about [9:15] them spending a lot with other [9:16] closed-source frontier labs that they [9:19] should have commoditized with their [9:20] open-source plan. But nonetheless, that [9:23] was the idea with Meta. But then China [9:25] sort of woke up and the deep seeks and [9:27] deep seeks launch at the start of 2025. [9:30] And the game theory became way more [9:32] complicated. So George Hotz sort of [9:34] summed this up nicely. He has a take in [9:36] AI will be massively deflationary a post [9:39] from just a few weeks ago as to why [9:41] China benefits from investing in open [9:43] source more than American firms. He says [9:46] this explains why Chinese the Chinese [9:48] are giving the much more moderate [9:50] resources to train models away for free. [9:52] They love to see deflationary economics [9:55] in the US. It is not as much less of a [9:58] service-based economy and so if they can [10:00] go and give away free tools that deflate [10:04] the value of the service sector that is [10:05] an advantage to the Chinese economy in [10:08] his formulation. [10:10] He says even if you don't regulatory [10:12] capture the US government, nobody is [10:13] getting a monopoly on AI. We don't live [10:15] in a unipolar world anymore and so he [10:18] likens what's happening in DC to sort of [10:20] rearranging deck chairs on the Titanic. [10:22] It's a very [10:24] fun fun means. [10:26] So we're back to this discussion [10:28] what are the consequences and the [10:29] impacts of open source models [10:31] particularly in the United States. And [10:33] there's been this clip that's [10:34] resurfacing from Dario Amodei when he [10:36] was testifying in front of Congress in [10:38] 2023 [10:40] and is now recirculating and it was [10:41] reposted like he just said it and he did [10:44] not. So be clear about that. This is [10:46] from 3 years ago, but some of his [10:48] predictions were very prescient as of [10:51] where the frontier is today. So he said [10:53] I'm very concerned about where things [10:55] are going. If we talk about 2 to 3 years [10:58] for the frontier models for the bio [11:00] risks sort of [11:02] a bad transcription of what he was [11:03] saying. He's talking about 2025 2026. [11:06] Remember he was saying this in 2023. [11:09] We're there now. I think the path that [11:11] things are going in terms of the scaling [11:13] of the open source models. I think it's [11:15] going down a very dangerous path. And [11:16] again, if the path continues, I think we [11:18] could get to a very dangerous place. So [11:20] he was worried about cybersecurity and [11:22] bio risks being open sourced and then [11:26] not having a counterweight to that. Now [11:29] the good news is that uh talked to the [11:31] CEOs of cybersecurity firms like [11:33] CrowdStrike and Palo Alto Networks and [11:36] uh they've been working with Mythos and [11:37] GPT 5.5 Cyber for months now to harden [11:40] systems from LLM-driven attacks. And so, [11:43] there's still this gap between [11:44] closed-source and open-source models, [11:46] and that gap allows white-hat hackers to [11:48] implement fixes before black-hat hackers [11:51] have a chance to exploit easy bugs. [11:53] There still will be a bigger discussion [11:55] here though in DC over the next few [11:57] months as the frontier models roll out, [11:59] and the gap doesn't appear to be [12:00] widening at the moment, so security [12:02] stances must adjust. It's not a a [12:06] closed source is falling behind, so it's [12:07] never going to be an issue. Uh there [12:09] will be this gap and how uh the American [12:11] cybersecurity industry and eventually [12:13] the biosecurity industry implements [12:15] changes and fixes before open source [12:18] catches up or or commoditizes and makes [12:21] that particular capability widely [12:22] available is going to continue to be [12:25] important. Uh so, let's go over to [12:27] Tyler's quick review of GLM 5.2. Why [12:30] don't you take me through your bullet [12:31] points, uh and you can tell us like what [12:33] is the shape of this model? How are the [12:35] reviews? [12:36] >> Yeah, so I I think so far um one of the [12:39] main things is like people are saying [12:40] it's oh it's distilled, right? This is [12:41] this has been a big thing with a lot of [12:42] these open-source models, especially the [12:44] Chinese ones. Oh, the only reason that [12:45] they're good is cuz they're distilled. [12:47] It It's very hard to actually figure out [12:48] how true this is. It certainly seems [12:51] like there there there's some uh you [12:53] know [12:53] >> I think Anthropic's of of Anthropic [12:55] models. [12:56] >> wait oh Anthropic openly accuse Alibaba [12:59] of distilling models? [13:01] >> of these these labs [13:02] >> Yeah, and there's also been a big like [13:04] professionalization of the gray market [13:06] where a whole bunch of different uh sort [13:09] of individual groups will connect a [13:12] whole bunch of different entities and [13:14] users accounts and subscriptions and [13:16] APIs to then create a front end to like [13:20] the model that can be served at a very [13:22] high rate through a VPN most likely. Uh [13:25] what's interesting is that you'd think [13:27] that if you were going to do a training [13:29] run, you would just find and replace on [13:33] the other lab's name before you hit run. [13:36] Is that not something people can do? I [13:38] don't understand. [13:40] >> Yeah, I mean, it also depends on what [13:41] you're actually like maybe you're not [13:42] directly distilling on the API, but you [13:44] know, you're training on [13:46] you know, public GitHub, you know, repos [13:49] and those were all used those were all, [13:50] you know, made with with with open [13:52] source models. Then you're kind of like [13:53] distilling, but it's not really [13:55] like is this really kind of distilling? [13:57] I don't know. [13:57] >> Yeah. [13:58] >> But so so if you are like if you're [13:59] convinced that these are like super [14:01] distilled, the only reason that they're [14:02] good is is because they're just, you [14:03] know, basically taking the closed [14:05] sourced [14:05] >> Yeah. [14:06] >> like labs. There's also this weird thing [14:07] with distilling where as more and more [14:09] of the public internet and GitHub [14:11] broadly and open source repos become [14:15] LLM outputs, you if you train on that, [14:19] you are in some ways distilling because [14:22] an LLM has a quirk like it's not this, [14:24] it's that in text and you'd wind up [14:26] training on a whole bunch of Amazon [14:28] Kindle books, you're going to wind up [14:30] learning it's not this, it's that. And [14:32] the same thing applies for different [14:34] code conventions in open source repos [14:36] that have effectively been completely [14:38] been rewritten by closed source models. [14:40] >> Yeah, and so so I think it's safe to say [14:41] that like we've generally seen that [14:43] distilled models [14:45] generally will generalize worse, right? [14:48] So you'll see really good benchmark [14:49] scores. And maybe they're benchmarks, [14:51] maybe they're not, but even if they're [14:52] not like directly benchmarks, you still [14:54] find that they generally [14:55] >> Yeah, they're kind of accidentally [14:56] benchmark sing. [14:57] >> Yeah. [14:57] >> Yeah. [14:57] >> So so you should always so I think [14:59] initially you should just be a little [15:00] bit suspicious of these super high [15:01] benchmark scores. [15:02] >> Yeah. [15:02] >> Um [15:03] >> But they lack that big model je ne sais [15:05] quoi. [15:06] >> Yeah, and this is like anecdotally [15:07] reinforced. A bunch of people have been [15:08] saying, you know, for coding these [15:11] models are really great. GLM is it's a [15:12] very good model, you know, for creative [15:14] writing or something like this where [15:15] you'd imagine it's a bit harder to to [15:17] kind of benchmark this. [15:18] >> I wonder hey have have people have [15:20] testing it with the like Tiananmen [15:22] Square bench. Like it does it reject [15:24] that stuff or because it felt like that [15:26] was something that was [15:28] like widely misunderstood by American [15:30] audiences that in fact that might not be [15:33] the biggest deal for the CCP and [15:36] >> Yeah, also I I think you know even if [15:37] that's true like the model is open [15:39] source. You can kind of just fine-tune [15:40] it to like not that. And maybe maybe [15:42] it's a bit harder than that but I think [15:44] you can kind of get around like that [15:45] kind of stuff. [15:45] >> Okay. Yeah. So we talked about the token [15:47] hunger and the API price [15:51] and in general I mean you said I'm not [15:53] convinced that there's a big market for [15:54] this class of model especially as [15:56] frontier models get more efficient. If [15:58] you look at open router the most used [16:00] models are the smallest open source [16:02] models presumably being used for [16:05] specific tasks that need to be repeated [16:07] over and over again. [16:09] >> Yeah, so I think like what we've seen is [16:11] you know a marginal IQ point of the [16:13] models is like extremely expensive. [16:15] Frontier models are are getting very [16:16] expensive. People have to cut back. You [16:18] know they're they're token max and this [16:19] is like massive bill in their balance [16:21] sheet or whatever. It seems like there's [16:22] there's now basically like two classes [16:24] of models that that people really use. [16:26] There's like the frontier ones and [16:28] they're they're using coding agents. [16:30] They need the best thing. If you're [16:31] doing cyber like you you just need the [16:32] the best model because you know the risk [16:35] of of someone hacking you it's so great. [16:36] You just need the best thing. You pay [16:37] whatever it is. [16:38] >> Yeah. [16:38] >> And then there's the second class which [16:40] is like these very small [16:42] very fast very cheap models that you can [16:44] use for these kind of point solution [16:46] things. Maybe you have some [16:47] orchestration where you're using a [16:48] really big model to to have these like [16:50] little agents using these very cheap [16:51] models. I I think in the middle it's [16:53] hard to actually figure out what is the [16:55] the real use case. Maybe it's like [16:56] hobbyists using these coding agents and [16:59] and they don't want to pay the super [17:00] expensive tokens of the closed source [17:02] labs. You see this on open router where [17:04] like what are the top models by by token [17:07] like usage. It's these very small [17:09] models. It's like [17:10] you know [17:11] deep seek flash [17:12] >> Yeah, cuz you're spamming them for like [17:14] you know every receipt that goes into [17:16] ramp gets processed by an LLM at this [17:18] point. Does it need to be a frontier [17:21] model telling me that I spent $10 on a [17:24] coffee? No. Yeah. It can just do [17:26] standard OCR. [17:27] >> That'd be my preference. [17:29] >> [laughter] [17:29] >> Yeah, you want you want super [17:31] intelligence overseeing your expenses [17:33] most likely. Um but no, you use the [17:36] right tool for the job and that's [17:37] clearly what's happening right now. [17:38] >> Yeah. But also I I think like it is a [17:40] very good model, right? Like we should [17:41] not fully dismiss I I think that the [17:43] idea that oh, the gap is widening. We we [17:45] really don't have to worry about these [17:46] these open source models. I think they [17:48] are like very good. [17:49] And maybe if you're super worried about [17:51] distillation, maybe something changes if [17:53] if the models are are, you know, kept to [17:55] these big partners, right? Like what [17:57] we've seen recently with with government [17:58] coming in, but [17:59] >> Yeah. [18:00] >> I think we we can't really fully dismiss [18:02] these these labs. [18:02] >> Yeah, it it throws it throws a little [18:04] bit of a wrench in the [18:07] like the monetization potential, like [18:09] how long can you monetize a new frontier [18:11] model? That's more tricky. And then the [18:15] other one is just like if you're going [18:16] to [18:18] keep a model behind KYC or behind a an [18:22] approval for specific companies like the [18:24] government has been sort of edging [18:26] towards and moving towards, it gets a [18:28] little bit tricky if all of a sudden [18:32] you just wait 3 months and oh, I was [18:34] waiting to get approved for this one for [18:37] like [18:38] GPT-7 or whatever, but by the time I the [18:41] government got back to me, my company [18:43] got access to GLM-6 and it's [18:47] close enough. And so [18:49] that in that just throws another wrench [18:51] that I think the government will have to [18:54] figure out how it puzzles together with [18:55] the rest of the strategy, which has [18:57] been, yeah, back and forth as always. [19:00] >> Google caps [19:04] Meta's Gemini use as AI demand strains [19:08] capacity in the Financial Times. Surging [19:10] appetite for advanced models is turning [19:12] computing power into [19:14] And they have a picture here of a Google [19:16] Gemini bicycle, which looks fantastic. [19:19] >> What what does that have to do with with [19:22] uh meta though? [19:23] >> I think that was just the best Google [19:26] Gemini picture. [19:27] Google has put limits on Meta's use of [19:29] its Gemini AI models after the social [19:31] media giant sought more computing [19:33] capacity than the rival tech group could [19:36] provide in the latest evidence of the [19:38] infrastructure constraints facing even [19:39] the world's largest AI providers. Google [19:41] told Meta around March that it could not [19:43] provide all of the Gemini capacity the [19:44] company wanted to purchase, according to [19:46] three people familiar with the matter. [19:48] In a move that has disrupted and delayed [19:51] some of Meta's internal AI projects. So [19:54] >> I don't understand how this is possible. [19:56] >> yeah. So so one [19:58] >> Google spent $200 on CapEx. [20:01] >> Okay. So so So of course like around [20:03] this time token maxing was becoming a [20:06] thing. A lot of every company in the [20:07] world, at least every tech company in [20:09] the world kind of going a little bit of [20:11] crazy from a spending standpoint. You [20:13] know, I could see Meta going and like [20:15] wanting to basically buy a bunch of [20:16] capacity and then being told like, "Hey, [20:19] we we can't fulfill that." But I'm I'm [20:22] wondering is it worth reading [20:24] >> I mean it it sounds extremely bullish [20:25] for Google. Like if they're [20:26] >> Yeah, no. [20:27] And this this tracks with what they talk [20:29] about on on earnings calls. [20:31] Um [20:32] >> Yeah, Google Cloud. [20:33] >> But [20:34] you do have to wonder like [20:37] could distillation be be part of this [20:40] this story? Is is that Could that be a [20:42] factor here? I have I have no idea. [20:44] >> know. Zero Hedge said Meta puts limits [20:46] on Claude and Codex fearing [20:47] distillation, the information. [20:49] >> But but but so this story is different. [20:51] This is Meta telling its own employees, [20:53] "Don't use Claude and Codex in certain [20:57] parts and certain parts of our business [21:00] because we don't want we don't want to [21:02] accidentally do distillation." [21:04] >> Oh. [21:04] >> Meta is saying. So So, that's different. [21:07] I was wondering like is Google thinking [21:09] like well, that's a lot of [21:10] you know. [21:11] >> Yeah, yeah. [21:11] >> Cool it, you know. Owing to the [21:13] restrictions which remain in place as [21:15] well as a broader push to streamline AI [21:16] costs, Meta has encouraged staff to be [21:18] more efficient with AI tokens. [21:20] Uh several other Google clients have [21:22] been affected by the restrictions [21:23] although to a lesser extent. Meta has [21:25] been particularly impacted because of [21:27] its exceptionally high demand for [21:29] Google's models. [21:30] >> Interesting. [21:31] >> Um [21:32] very interesting. On the topic of Meta [21:35] >> Yeah. [21:35] >> Meta shared this morning [21:37] >> What they do? [21:38] >> a new milestone. It is a mind reader. [21:40] >> Mind reader. [21:41] >> Non-invasive brain detects decoder [21:43] research brain to QWERTY V2 building on [21:46] V1 which was published today in nature [21:48] brain to QWERTY V2 is the highest [21:50] performing end-to-end pipeline capable [21:52] of real-time sentence decoding from raw [21:55] brain signals. It advances beyond [21:58] character level performance to decoding [22:00] words and semantics enabling accuracy [22:03] for overall communication. So, if you [22:06] thought Instagram was listening to you [22:09] >> [laughter] [22:10] >> It's going to [22:10] >> Uh if you thought it was listening to [22:12] your, you know, conversations, now you [22:14] can have a, you know, new conspiracy at [22:17] home which is that they might be just [22:19] listening to you to your thoughts. [22:21] >> you know do you know the device they [22:22] they say this is a non-invasive device? [22:24] I just shared an image of this device [22:26] and I want you to tell me, do you [22:27] consider this non-invasive or invasive? [22:31] Look [laughter] at this image of the [22:33] magnetoencephalography [22:36] uh graphy device. [22:38] >> No, you got to go [22:39] you need to [22:40] scroll up a little bit cuz you can't [22:43] even see the whole thing here. [22:45] >> [laughter] [22:46] >> It's not invasive. [22:47] >> Cuz it looks like the device could [22:48] actually potentially carry on for like a [22:50] whole half of a football [laughter] [22:51] >> It really does seem like it's a just put [22:53] yourself in this in this room-sized [22:55] device. No, of course this will shrink [22:57] down. [22:57] >> giving him I'm giving him credit here. [22:59] Non-invasive. [23:01] >> Okay. As long as he [23:03] >> You're You're putting this thing on? [23:04] You're daily driving this thing. [23:05] >> I don't know if I'm ready to daily it. [23:07] >> [laughter] [23:09] >> I don't know if I'm ready to daily it. [23:11] >> Yeah. [23:12] >> This will be a cool demo. Like this this [23:14] will actually when when [23:16] when you can just walk in, sit down in a [23:19] chair, and see your thoughts on a [23:21] screen. [23:21] >> No, we were debating earlier. My buddy [23:22] Rob Tev has been on the show twice, [23:24] uh dropped five predictions in Forbes [23:26] recently. We can go through them at some [23:27] point. He's going to come on the show. [23:28] But, uh [23:30] four of the five were very very like [23:33] reasonable, you know, Anthropic's going [23:35] to be bigger, and uh you know, TSMC is [23:38] going to face more comp- more [23:39] competition. And then he predicts that [23:41] in 2030 telepathy will be commonplace, [23:45] which is a very aggressive prediction in [23:47] my it it in my estimation. How Terry [23:50] Semel fumbled Yahoo's Facebook deal. Uh [23:53] how much is Facebook worth? 5 billion, [23:55] 10 billion, 15 billion, whatever the [23:57] number, it's probably a lot more than [23:59] the 1 billion that Yahoo could have [24:02] bought it for a year ago. As Yahoo [24:04] continues its soul-searching, here's an [24:06] unpleasant rendition of Semel's [24:08] catastrophic decision, courtesy of [24:11] Wired. When Yahoo came calling with a [24:13] bid of $1 billion in cash, the pressure [24:16] became too much. Zuck relented in July [24:19] of 2006. He was just like 18 months into [24:22] building the company, something like [24:23] that. Verbally agreeing to sell Facebook [24:26] to Yahoo. He said yes. He said he was [24:28] going to sell Facebook [24:29] >> allegedly. [24:30] >> to Yahoo, allegedly. Uh strategically, [24:32] it [snorts] seemed like a good match. [24:34] Yahoo had hundreds of millions of users, [24:36] but its foray into social networking was [24:38] struggling. Facebook had cool tools and [24:40] was looking for a mass audience. The [24:42] timing, however, could not have been [24:44] worse. In the days after Zuckerberg [24:46] agreed to sell, Yahoo announced it was [24:48] projecting slower sales and earnings [24:50] growth, and that it's that the launch of [24:52] its new advertising platform would be [24:54] delayed. Its stock price tumbled 22% [24:58] overnight. Terry Semel, Yahoo's CEO at [25:00] the time, reacted by cutting his offer [25:02] from 1 billion to 800 million. He just [25:05] took 20% off, but Zuckerberg, who had [25:07] been warned about Semel's reputation for [25:09] last-minute renegotiations, walked away. [25:12] And that's probably reasonable. I mean, [25:13] if they're cutting the price there, you [25:15] have to imagine that as it gets papered, [25:17] you get cut down again, then the earn [25:19] out, you get cut down again, and all of [25:20] a sudden you're walking away with barely [25:22] anything. Uh but, 2 months later, Semel [25:25] reissued the original 1 billion-dollar [25:27] bid, but by then, Zuckerberg had [25:29] convinced his board and executive team [25:30] that Yahoo wasn't a serious partner, and [25:32] that Facebook would be worth more on its [25:34] own. He rejected the offer and became [25:36] famous as the cocky youngster who turned [25:39] down 1 billion dollars from Wired. [25:41] >> Legendary. [25:42] >> Legendary. [25:43] Um [25:44] it's so interesting to imagine the road [25:46] not traveled there, because the the the [25:49] the dynamic, the way Facebook is built [25:52] as a social network, like could it have [25:55] been successful under Yahoo's [25:57] stewardship, or would it have been uh [26:00] less exciting, [26:02] attract less talent, ultimately been [26:04] disrupted, and would they have had the [26:06] capital and the guts to go and buy [26:09] WhatsApp, and then also buy Instagram? [26:11] Um you know, to actually maintain the [26:14] the dominant position in social [26:15] networking. [26:16] >> What do you [26:16] >> I think Yahoo should make another offer. [26:19] I would like to see Yahoo make another [26:21] bid. [26:21] >> Hey, Meta is trading down, just keeps [26:23] going. [26:24] >> If it continues at this [26:25] trend [26:28] >> might be able to pick it up [26:29] >> at at this trend [26:30] >> Anyway. [26:31] >> Uh chip makers are profiting off AI at [26:33] the expense of just about everyone else. [26:36] >> on the cover of the business and finance [26:38] section today. [26:39] >> Uh we are witnessing an extraordinary [26:41] transfer of cash from the providers of [26:43] AI, and perhaps one day AI users, to [26:45] memory chip makers. Take us away, John. [26:48] >> Yeah, the explosive growth in Micron [26:50] Technology's profit in the latest [26:52] quarter is extraordinarily good news for [26:54] its shareholders, but it comes at the [26:56] expense of the artificial intelligence [26:58] companies to which it sells fast memory [27:00] chips. Micron, along with the Korea with [27:03] Korea's Samsung Electronics and Sam and [27:06] SK Hynix are to AI what oil producers [27:09] are to the airlines, makers of an [27:12] essential input that this year suddenly [27:14] became much more pricey. Because there [27:16] is extremely limited capacity to make [27:18] the high bandwidth memory that AI needs, [27:21] and it takes years to build production [27:23] facilities. Soaring data center demand [27:25] simply jacked up prices. Micron's [27:27] soaring profits are for its customers [27:30] soaring costs. "We are witnessing an [27:32] enormous transfer of cash," they said. [27:34] "Profit shift of this scale are rare [27:37] events, and investors should be paying [27:39] attention to where the money's coming [27:41] from, where it's being spent, and how [27:43] long it will keep flowing." In the [27:44] quarter ended May 28th, Micron increased [27:47] prices for DRAM chips more than 60% on [27:51] the previous 3 months while increasing [27:52] shipments by a low single-digit [27:55] percentage. It said last week prices for [27:57] NAND flash memory also used in data [28:00] centers jumped more than 80%. Usually, [28:02] memory doesn't matter that much, but for [28:04] Micron customers paid $18 billion more, [28:07] and that was just in the quarter. Prices [28:09] quadrupled in a year, and it's hurting [28:11] outside AI, too. Apple last week raised [28:14] prices for MacBooks more than 15% Closer [28:17] to home for me, the memory I bought on [28:20] amazon.com a year ago to build a super [28:22] quiet computer, I hate fan noise, good [28:25] good color commentary here, uh has [28:27] tripled in price [28:28] and now costs more than the CPU. For an [28:31] industry [28:32] in which prices usually drop every year, [28:34] it's a huge turnaround. And consumer [28:36] electronics passing on higher prices [28:38] helps limit demand for chips just as [28:40] higher oil prices reduce consumption. [28:42] But the AI companies aren't passing on [28:44] higher prices because they're able to [28:46] throw money at supply problems. The [28:48] problem in AI is that the end users [28:50] aren't covering the cost of the service [28:51] with big losses being recorded by AI [28:54] model producers. Everything is still [28:56] priced to bring in new customers, yet [28:58] not yet to make money. So, higher input [29:01] costs create a nasty problem. Either [29:03] losses will either be bigger or higher [29:05] prices will be needed, putting off [29:07] potential customers. And you can see the [29:09] price of Micron's stock price has been [29:12] through the roof as the company joins [29:15] >> the $1 trillion club. [29:16] >> Tyler, how many trillion-dollar [29:18] companies are there in Europe out of [29:19] curiosity? [29:22] >> I'm going to go with zero. [29:24] >> That's true. [29:25] >> Uh NBCUniversal and Sky will separate [29:28] the company's connect connectivity [29:29] business from its film, theme park, and [29:31] streaming operations. Oh, yeah. [29:33] Universal Studios. Uh Comcast plans to [29:36] separate its media and connectivity [29:38] businesses. [29:39] >> the Anduril of theme parks? [29:42] >> It does seem like a [29:43] >> Could there not be an opportunity to [29:44] create a a net new theme park business [29:49] with with modern a modern technology [29:52] stack? [29:52] >> It's very expensive. Everything needs to [29:54] be Like, the modern technology stack in [29:57] theme parks is expensive. [29:59] >> in the theme park capital markets? [30:01] >> know. I I I I I I know I've known people [30:03] that have worked on theme parks at [30:04] Disney, and it's tricky because you you [30:08] have to amortize a ride over like 20 [30:10] years. And so, you'll go and [30:12] >> like an absolutely brutal business. [30:14] >> Yeah. [30:15] >> That is probably harder today because at [30:17] the time that a lot of these parks were [30:19] built, like, you didn't have like [30:21] infinite online entertainment for every [30:24] single sub [30:26] niche. [30:27] >> Yes, but instantly available. [30:28] >> I mean, there's a whole bunch of trend [30:30] pieces right now about how IRL [30:32] experiences are seeing higher than ever [30:35] pricing in the face of you could just [30:38] watch the Knicks game on TikTok [30:41] highlights, but people still forked over [30:44] $5,000 to go see the game. And so, you [30:47] know, you have that like barbell [30:49] strategy where Thrive is buying a stake [30:51] in the San Francisco Giants, a baseball [30:53] team that should face competition. [30:56] >> uh the the NBA team to Vegas. [30:58] >> Yeah. [30:58] >> Uh [30:59] but at the same time at the same time, [31:01] John, there was a stat that came out [31:02] last week that came out. There's more [31:04] sports betting volume than all sales of [31:08] uh movie tickets, theaters, theme parks, [31:11] and like a couple other these IRL [31:12] categories. [31:13] >> Was it up or down? [31:14] >> Uh lat uh less [31:16] >> Less. Like in in in in the the stat was [31:19] like volume. [31:20] >> Yeah, yeah. [31:21] >> Uh and so it's not exactly like [31:24] uh a proxy for like revenue, but still [31:29] meaningful. Chamath raised $135 million [31:32] series A for 890. They got sale They got [31:35] Salesforce Ventures. They got Wonder [31:38] Cow. They got Craft, and they got [31:40] Launch. [31:41] >> The besties. [31:41] >> They got the besties. [31:42] >> the besties together, too. [31:43] >> You think Fried [31:45] Friedberg Friedberg's got to be in. [31:47] >> Yeah, that's the production board. [31:48] >> Oh, the production board. [31:49] >> Yeah. No, Friedberg's fine. [31:50] >> Oh, great. [31:51] >> So, yeah, you actually have all three of [31:53] the other besties. And we will see you [31:55] tomorrow. [31:56] Goodbye.