AMP's Anjney Midha Plans Software Grid to Slash AI Compute Costs

AMP 的 Anjney Midha 拟构建软件网格大幅削减 AI 算力成本 由于突发的研究需求和碎片化的硬件，GPU 租赁费率飙升至其基础成本的近十倍；AMP PBC 旨在通过软件编排层汇聚闲置算力来解决这一问题。 Anjney Midha 探讨了 AI 算力基础设施面临的经济与技术瓶颈，详细说明了长期租赁与硬件异构性如何推高实际 GPU 成本。他解释了 AMP PBC 如何借鉴 Google 的 Borg，利用软件转换层将未利用的算力汇聚为标准化的公共设施服务。此外，Midha 还剖析了四阶段模型训练方案，并反驳了 AGI 赢家通吃的论调，转而描绘了专业化模型共同构成的“参差前沿”。 超大规模 AI 发展面临严峻的物理、能源和资金制约，引发了关于 AI 将商品化为基础公共设施还是会诞生单一主导赢家的争论。 (0:00) Anjney Midha 将 AMP PBC 创立为一家公共利益公司，旨在通过将碎片化的 GPU 基础设施转变为标准化公用事业来降低算力成本。 (5:07) Anthropic 早期融资情况表明，算力基础设施和硬件可用性是制约 AI 实验室的首要瓶颈。 (9:47) AI 的进步是沿着专业领域的“参差前沿”推进，而非单一的线性竞争，这使精简的实验室能够直接与科技巨头竞争。 (11:47) 模型性能是通过包含预训练、中程训练（mid-training）、后训练和持续反馈循环的四步方案训练而成的。 (11:47)

Anjney Midha argues that artificial intelligence progress is constrained primarily by compute infrastructure inefficiencies rather than capital or algorithmic limits. To address low utilization and fragmented hardware, his firm AMP PBC proposes orchestrating compute like an electric grid to radically lower operational costs. The Jagged Frontier and Compute Bottlenecks (11:47) Rather than viewing artificial intelligence development as a single monolithic race toward artificial general intelligence, Midha characterizes it as a "jagged frontier" split across domain-specific capabilities. In his view, specialized labs hold distinct leads—such as Anthropic in software engineering, OpenAI in consumer chat, and ByteDance's Seedance in video generation—often rivaling incumbent tech giants despite maintaining significantly smaller workforces. Midha points to Anthropic's trajectory to demonstrate that hardware access, not headcount, dictates frontier research velocity. After former OpenAI researchers were rejected by roughly 25 venture capital firms when pitching a $500 million raise, early backers scraped together an initial $100 million angel round before eventually orchestrating a pivotal $4 billion compute and capital partnership with Amazon. Verifiable Feedback and Grounding AI Capabilities (15:00) Model progression depends fundamentally on the nature of evaluation feedback. Midha explains that software engineering has advanced rapidly because code changes can be objectively validated using automated unit tests and peer-reviewed pull requests. In contrast, subjective domains like creative writing or therapy lack deterministic validation loops, leaving models prone to hallucinations and unverified outputs. To apply verifiable feedback to physical science, Midha incubated Periodic Labs—co-founded alongside Liam Fedus and Doyoung Chubuk—in a 40-square-foot Menlo Park facility to discover room-temperature superconductors. The lab uses AI to generate material candidates, physical robotics to synthesize samples, and X-ray diffraction hardware to verify their properties, feeding empirical physical truth directly back into model training. Grid Orchestration to Cure Compute Inefficiencies (30:00) Midha claims the prevailing compute leasing market suffers from severe structural waste. Fixed, 24/7 contracts marketed at $2.50 to $3.00 per GPU hour effectively balloon to $25 to $28 per hour because research labs must over-provision for spiky peak workloads. While Google achieves roughly 99% internal infrastructure utilization, broader independent data centers operate below 70%. As an extreme example, Midha notes that xAI's Colossus 2 deployment in Memphis—housing 500,000 GB300 GPUs—operated at less than 60% node utilization and under 11% Model FLOPs Utilization (MFU). To address this fragmentation across NVIDIA, AMD, and proprietary silicon like Microsoft's MAI chips, AMP PBC operates as an Independent System Operator (ISO). Co-founded with former Google Borg technical lead Sebastian Lobo, AMP deploys an abstraction layer that pools unutilized hardware into fungible "grid credits," aiming to lift operational GPU utilization up to 95–96% without owning physical real estate. Co-Design, Technical Literacy, and Speculation Risks (40:00) Midha cautions that business leaders cannot outsource technical literacy; executives who fail to understand underlying model architectures risk miscalculating return on investment and deploying brittle workflows. In frontier tooling like Anthropic's Claude Code, Midha rejects the idea that progress is driven purely by interface wrappers. Instead, he points to "co-design"—a dialectical engineering process where the application harness and the base model are built simultaneously to anticipate new reasoning thresholds and collapse latency. Addressing market dynamics, Midha notes that long-term GPU rental prices doubled between January and mid-2026. However, he warns against treating compute hours as speculative financial assets, arguing that secondary market financialization induces artificial scarcity for researchers. To manage supply risk without speculation, AMP instead uses an internal commodity-style forecasting desk that secures compute via call options. Post-interview reflections underline an unresolved market question: while frontier foundation models may face commoditization as enterprises route tasks to the cheapest, most efficient model rather than paying for brand prestige, building the underlying software orchestration to standardize fragmented physical infrastructure remains technically formidable.

AMP PBC创始人Anjney Midha认为，当前人工智能竞赛的核心瓶颈已从算法架构转向物理算力与基础设施效率；通过构建类似电网的软件调度层整合异构硬件，不仅能打破高昂的闲置成本壁垒，还将推动大模型从单一寡头垄断走向细分领域的“参差前沿”。 算力闲置陷阱与高昂的实际使用成本 (30:00) Midha回顾早期投资Anthropic的经历指出，硬件与算力短缺一直是制约前沿实验室的核心瓶颈，约80%的模型研发预算最终直接流向NVIDIA等芯片供应商，而台积电（TSMC）的代工产能则构成了底层增长的物理上限。由于长期租赁合同通常要求全天候支付每GPU小时2.50至3.00美元的基准费用，而科研计算需求具有高度波动的尖峰特性，大量未被充分调用的闲置算力导致实际有效成本膨胀至每小时25至28美元。 Midha批评盲目堆叠芯片规模的做法在很大程度上属于行业虚张声势。根据其引用的基准数据，xAI在孟菲斯部署的Colossus 2集群虽配备50万颗GB300 GPU，但节点利用率低于60%，模型FLOPs利用率（MFU）甚至不足11%；当前独立数据中心平均利用率普遍低于70%，远落后于Google内部系统约99%的利用效率。这种严重的产能闲置造成了极大的无谓损失。 借鉴Borg架构构建标准化算力电网 (25:00) 为解决异构硬件碎片化问题，Midha与曾担任Google Borg技术负责人的Sebastian Lobo联合创办了共益企业AMP PBC。该模式借鉴了历史经验——如同交流/直流电标准的统一解锁了全美孤立电力并催生出公共电网，AMP开发了名为“Grid”的软件翻译层，旨在将分散在NVIDIA、AMD及微软MAI系列自研芯片上的闲置计算能力抽象为标准化的“电网积分”（grid credits）。 AMP将自身的运营角色定位于电力系统中的独立系统运营商（ISO），在不自建物理数据中心的前提下，动态调度独立集群的闲置算力来承担推理任务，并保留专属容量用于高负荷训练，从而将GPU利用率提升至95%至96%。此外，针对将算力金融化为可交易投机资产的市场呼声，Midha明确表示反对，强调金融投机只会制造人为稀缺并损害科研机构，AMP选择借鉴大宗商品交易柜台模式，通过内部预测团队提前锁定集群看涨期权以对冲波动。 参差前沿与客观可验证的反馈循环 (11:47) 针对2026年顶尖大模型性能趋于同质化的观点，Midha反驳了单一线性AGI竞赛的假设，提出AI发展呈现为“参差前沿”（jagged frontier）——各实验室在细分领域各具优势，例如Anthropic领跑软件工程，OpenAI擅长大众聊天界面，字节跳动的Seedance则在视频生成上表现突出。Anthropic以不足5000人的团队规模匹敌拥有数万员工的科技巨头，证明组织专注度与工程执行力比盲目扩张人力更为关键。 在模型训练环节，Midha强调唯有“可验证反馈”才能驱动实质性能力跃升，单纯的主观指正极易引发幻觉或无效道歉。在代码开发中，代码审查（PR）与单元测试构成了客观验证闭环；其孵化的Periodic Labs在研发室温超导体时，通过机器人合成与X射线衍射仪检验真实物理特性，将硬性物理事实反馈给模型。缺乏明确客观闭环的领域（如创意写作或心理咨询）则更容易产生荒诞或具有安全隐患的输出。 协同设计演进与企业智能分层路由 (45:00) 关于Anthropic推出Claude Code等工具的产业意义，Midha指出前沿突破并非单纯依赖UI外壳（harness）包装，而是源于模型能力与外壳架构的双向“协同设计”（co-design）。在工程迭代中，研发团队能够在基础模型尚在训练阶段时预判其局限，构建外部工具接管特定任务（如复杂电子表格理解）；当下一代模型原生习得该推理能力后，外壳即可剥离第三方插件，使端到端执行延迟缩短一至两分钟。 在企业应用端，播客主持人与Midha探讨了认知断层与大模型商品化趋势。亲自动手实验的技术型高管能够清晰理解模型边界，而外包认知的管理层往往难以评估投资回报率。主持人Alloway进一步提出，随着企业客户不再盲目迷信顶尖大厂品牌，市场或将迎来大模型公用事业化；主持人Weisenthal指出，在此背景下，类似AMP的软件路由层将自动根据性价比将日常任务分派给轻量级模型，仅在关键难题上调用高规格旗舰模型，从而重塑整个AI软件经济学。 Midha所倡导的软件电网模式为打破高能耗与低利用率瓶颈提供了清晰的工程解法，但在物理芯片短缺与代工厂产能受限的大背景下，纯软件抽象层能否彻底对冲重资产云厂商的垄断优势，以及企业市场是否会完全倒向去品牌化的动态路由机制，仍取决于未来异构生态的实际标准化阻力。

## Background 本期内容由彭博社（Bloomberg）旗下的知名播客《Odd Lots》于 2026 年 6 月 13 日在 RSS 及网络平台发布。节目简介指出，曾开出 Anthropic 首笔天使支票、前 a16z 合伙人兼斯坦福爆款 AI 课主讲人 Anjney Midha，正通过其创立的公共利益公司 AMP PBC 构建标准化算力电网，旨在通过纯软件解决方案破解当前算力碎片化、利用率低下及长期合约错配的死局，从而从根本上大幅拉低算力成本，并分享了他对多元化模型前沿演进的深度预判。 ## TL;DR 访谈深入拆解了当前 AI 研发中因长期固定租赁与波峰波谷错配导致的算力严重浪费，指出研发团队扣除闲置期后的实际有效算力成本往往高达标价的十倍。针对此痛点，Midha 详述了 AMP PBC 借鉴 Google Borg 架构打造的通用软件调度层，通过将异构、分散的 GPU 资源抽象转化为标准化的“网格信用额度”（Grid Credits），将集群利用率提升至接近 95% 的极致水平。此外，他坚决否定了 AGI 赢家通吃的垄断假设，提出了多领域并存的“参差前沿”（Jagged Frontier）格局，并深入剖析了科技巨头自研芯片背后的毛利焦虑以及模型本体与 Harness 框架协同设计的工程演进趋势。 ## Atmosphere & style 整场对话在宏观产业思辨与硬核技术细节之间自如切换，既充满对“纸回形针悖论”、电力变压器瓶颈等现实物理约束的犀利审视，又兼备硅谷工程一线的生动写照。嘉宾 Midha 展现出兼具顶级投资人视野与资深系统架构师的笃定与从容，讲述深入浅出、逻辑严密；主持双人组 Joe 与 Tracy 则以标志性的默契幽默穿针引线，从“大眼夹”创伤到六岁孩子嘲笑用 AI 的家庭趣谈，在轻松愉快的笑声中举重若轻地解构了价值数千亿美元的底层算力变局。 ## Bulletin - `27:30` 前 a16z 合伙人、Anthropic 首位天使投资人 Anjney Midha 创立 AMP PBC，旨在通过类似 Google Borg 的软件网格调度层将异构闲置算力转化为标准化公用事业，将实验室利用率从 50%-60% 提升至 95% 以上。 - `32:38` AI 模型研发因需求极具突发性导致传统长期租赁合约产生巨大无谓损失：实验室名义上签下每小时 2.50 美元的合约，在扣除大量闲置期后实际有效算力成本高达每小时 25 至 28 美元。 - `12:08` Midha 驳斥 AGI 单极垄断论，提出 AI 呈现“参差前沿”（Jagged Frontier）格局，可客观验证的工作流（如编程、材料实验）进展极快，而企业真正诉求是大宗商品化的高性价比任务交付，而非昂贵的全能顶级模型。 - `36:34` 算力需求斜率已呈“垂直”爆发态势，2026 年初至年中长期算力租赁市价翻倍；科技巨头自研芯片的核心动力在于夺回被 NVIDIA 挤压的毛利率（研发支出中 80% 流向芯片商）并保障供应链安全。 ## Key takeaways - `00:00` 人类为追求更先进的 AI 正在现实中复刻博斯特罗姆的“纸回形针”悖论，电网负荷、变压器、重型硬件乃至住宅用地等物理硬约束正成为 AI 扩张的严峻瓶颈。 - `10:00` Anthropic 早期融资异常艰难，目标 5 亿美元却仅拼凑出约 1 亿美元天使轮资金，历时 18 个月敲定与亚马逊的 40 亿美元战略合作才确立算力生命线。 - `12:08` AI 前沿并非单一终点线，而是高度分化的“参差前沿”：Anthropic 领跑软件工程，OpenAI 主导大众对话，字节跳动等发力多模态视频。 - `14:15` 前沿模型研发遵循四步法：预训练通识、中训练注入专业能力、后训练对齐校准，以及基于客观场景闭环的持续反馈。 - `16:45` 客观可验证反馈是 AI 快速进化的根本动力：编程的单元测试与材料科学的物理衍射测试均提供客观硬证据，因而模型在这类领域的进展最快、幻觉最少。 - `20:00` 结构化与可验证工作将被 AI 迅速自动化接管，而无法被客观验证的真实创造力、审美与手艺领域，人类劳动者的薪酬溢价与影响力将迎来非凡增长。 - `22:50` 企业高管绝不能将“理解力”外包给黑箱模型，简单的“沙盒运行”无法防范幻觉与注入风险，领导者必须具备从土地、电力到模型全链路的技术素养。 - `27:30` AMP PBC 借鉴电力工业史标准化的过程，通过软件抽象层在消费端消除芯片与云厂商的异构性，打造全行业通用的类 Borg 算力电网。 - `30:00` 行业充斥着盲目囤卡的虚张声势，然而独立集群平均节点利用率不足 70%，马斯克 Colossus 2 集群模型浮点利用率（MFU）甚至低于 11%，利用率优化存在巨大软件改进空间。 - `32:38` 现行 24/7 长期租赁合约将利用率风险全盘转嫁给研发团队，AMP 提出“网格信用额度”（Grid Credits），通过波峰波谷动态再分配将实际成本拉回合理水平。 - `36:34` 软件调度电网能够将大块预订算力专用于模型训练，并利用训练间隙产生的碎片化空闲算力实时吞吐推理任务，实现动静混合填充。 - `36:34` 巨头竞相自研专用芯片的双重驱动力：一是夺回被 NVIDIA 挤压的毛利率自主权，二是在台积电产能瓶颈下争夺晶圆配额以确保供应链安全。 - `42:15` 企业级客户并不在乎底层是哪家模型，只关注低成本与稳定性；未来应用将把模型完全抽象化，单一助手引擎盖下可能协同调用 200 多个专业小模型。 - `46:04` 反对算力信用额度的纯投机金融化，但算力供需缺口已极其严峻，2026 年上半年长期租赁价格直接翻倍，AMP 需通过期权手段对冲供应链风险。 - `48:38` Claude Code 等突破并非单纯外部脚手架优化，而是模型能力与 Harness 的对偶协同设计（Co-design），通过预先将工具调用原生内置化实现端到端任务大幅提速。 - `51:15` 终端用户对性价比大宗商品化服务的偏好，动摇了少数全能顶级“凯迪拉克模型”赢家通吃的天量估值逻辑，软件定义调度将正面颠覆固化的物理硬件壁垒。 ## Detailed walkthrough ### `00:00` 纸回形针悖论成真：AI 算力与物理现实的剧烈碰撞 播客在开场广告后切入正题。主持人 Tracy Allaway 与 Joe Weisenthal 指出，《Odd Lots》一向热衷于探讨现实世界中的“物理硬约束”，而 AI 行业正处于这种张力最核心的位置。表面上看，AI 是一种飘渺无形、甚至有些脱离肉身的数字化技术——用户只需在 ChatGPT 或 Claude 中敲入一行提示词，屏幕那头就会给出回答。然而，在这轻巧交互的背后，是庞大无比的实体工业网络在支撑：高昂的电网负载、巨额的真实能源消耗、输配电变压器、重型硬件以及源源不断涌入数据中心的各类计算资源。 Joe Weisenthal 提到了一个经典的思想实验：哲学家尼克·博斯特罗姆（Nick Bostrom）提出的“纸回形针最大化（Paperclip Maximizer）”。在设想中，如果给 AI 设定一个“尽可能制造更多纸回形针”的目标，它就会在调动全球所有资源的过程中，将地球乃至宇宙的一切物质与能量都转化为纸回形针，甚至发射星际探测器吞噬恒星能量，最终毁灭世界。Tracy 调侃问这是否源自他对微软“大眼夹”（Clippy）的童年创伤阴影。Joe 笑称虽然一切看似闭环，但现实中的荒诞之处在于：人们为了追求制造更先进的 AI，正真实地上演着这一幕——电网接入指标、GPU、能源涡轮机、核心顶尖人才，甚至连居民住宅用地都被重新挪作它用，以换取更多算力。甚至业界已经开始严肃讨论将数据中心送入太空运行的可能性。“纸回形针”的预言正被人类以自己双手推动 AI 的形式完整复刻。 与这股算力狂热紧密交织的，是当下前沿模型市场的估值逻辑与残酷竞争。为了争夺所谓通用人工智能（AGI）第一名的头衔，巨头们陷入了“必须赢”的存在主义焦虑，并以此为由进行几乎不设上限的资本开支，导致强者恒强、大企业通吃资源。然而诡异的是，截至 2026 年 6 月，各家的前沿模型（例如 GPT-5.5 与 Claude 4.8）水平彼此极为接近，性能并未拉开鸿沟。Joe 提出了一个根本性疑问：这种胶着状态究竟是因为员工之间的信息互通、跨模型数据蒸馏与“相互借用”等市场动态导致技术不可避免地走向商品化（Commoditization），还是说在未来某一天，真的会有一家实验室突然单点突破，确立不可动摇的永久垄断？ ```mermaid flowchart TD accTitle: AI发展的物理与经济闭环 accDescr: 展现从模型竞备、资源狂热消耗到前沿模型同质化与物理约束的循环机制 subgraph Competition ["前沿军备竞赛"] A["AGI 先发存在主义焦虑"] --> B["不计代价的资本支出"] B --> C["巨头垄断关键供应链与资源"] end subgraph PhysicalConstraints ["物理约束的显化"] C --> D["电网负荷 / 变压器荒"] C --> E["GPU 基础设施短缺"] C --> F["数据中心甚至向太空扩张"] end subgraph MarketParadox ["市场现状与悖论"] D & E & F --> G["各家模型性能高度接近<br/>(GPT-5.5 vs Claude 4.8)"] G --> H["走向商品化 vs 突围垄断?"] H -.-> A end ``` ### `05:00` 从 Stanford 到 Ubiquity6：Anjney Midha 的技术与创业之路 为解答上述算力与市场机制的深层问题，节目邀请了堪称 AI 产业“王牌”的重磅嘉宾 Anjney Midha。他的履历极为亮眼：曾任知名风投机构 a16z（Andreessen Horowitz）普通合伙人、斯坦福大学客座科学家并讲授轰动全网的 AI 课程《前沿系统》（Frontier Systems）、Anthropic 的首位天使投资人，如今则是初创公司 AMP PBC 的创始人。 在开场互动中，Anjney 纠正了主持人的读音，说明公司名读作 "AMP PBC"，而后缀 PBC 代表“公共利益公司”（Public Benefit Corporation）。这与 Ben & Jerry's、REI 以及 Anthropic 类似，意味着公司受公共利益章程约束：在追求健康、适度利润以维持商业运转的同时，具备灵活调控自身利润率的自由，必须将核心使命置于单纯追求利润最大化之上。 谈及 Anthropic 的起源传奇，外界普遍流传着“他们被 25 家风投接连拒绝，直到 Anjney 伸出援手”的版本。Anjney 笑着澄清事实恰恰相反：“是我先开了第一张支票，随后我们试图说服另外 25 家风投跟投，结果我碰壁失败了。”那是一段极为痛苦且令人清醒的经历。回溯其个人成长轨迹： - Anjney 出生于印度，高中就读于新加坡，2011 年赴美进入斯坦福大学攻读本科学位。 - 彼时恰逢深度学习浪潮在硅谷席卷开来。当时 Andrej Karpathy 正担任吴恩达（Andrew Ng）的助教，校园里洋溢着一种颠覆性的思想——通过数据训练让机器自主思考，而无需人工编写刻板的规则。 - 受这股热潮冲击，Anjney 的学习重心转向机器学习，主修生物信息学（即机器学习在医疗健康领域的交叉应用）。 - 毕业后他加入顶级风投 Kleiner Perkins 工作了四年半，师从 John Doerr 与 Mary Meeker 等传奇投资家；随后在 25 岁时毅然辞职创业，创办了 3D 空间测绘 AI 公司 Ubiquity6。 Ubiquity6 曾获得来自 Benchmark、Index 等硅谷顶尖风投约 4,700 万美元的巨额融资。当时自以为“全城最酷少年”的 Anjney 很快遭遇了现实的迎头痛击：他们研发出能够精准生成三维空间地图的尖端 AI 系统，但 2020 年初新冠疫情暴发，基于线下物理空间的场景业务瞬间停摆。在经历了艰难的业务转型后，团队最终“平稳落地”，凭借其在底层构建的高性能分布式系统资产，将公司成功卖给了游戏语音社交巨头 Discord。 公司完成出售后仅一个月（2020 年底），Anjney 便接到了 OpenAI 研究主管老朋友们的电话。在旧金山湾区的机器学习圈子中，他们相交已久。对方在电话那头兴奋地告诉他：“Anj，我们刚刚训练了一个叫 GPT-3 的小模型，没人当回事，但我们觉得它是自切片面包以来最伟大的发明！我们想离开 OpenAI 独立创业，急需你来指导我们如何把这做成一家独立运转的企业。”由于当时 Anjney 必须履行将 Ubiquity6 整合进 Discord 的并购留任职责，无法全职加盟，于是他以第一位天使投资人的身份介入，利用夜间与周末时间与 Dario Amodei、Tom Brown 紧密配合，梳理商业计划与融资策略。这就是日后的 Anthropic。他们满怀信心地拿着方案走向沙丘路（Sand Hill Road）上的各大知名风投资本，原本以为会是一场轻松的胜利，却没想到迎接他们的是一场长达数月的融资冷遇。 ### `10:00` Anthropic 艰难的早期融资与亚马逊 40 亿美元算力合作 回溯 Anthropic 的初创期，Anjney Midha 讲述了一段鲜为人知的艰难历程。当时 Dario Amodei 和 Tom Brown 等核心创始人刚刚打造出震撼业界的 GPT-3，满以为凭借上一代最具轰动性的突破能够轻易获得资本青睐。然而，现实却泼了他们一盆冷水。几乎所有见过的顶级投资人都直言无法理解，甚至认为“通用人工智能（AGI）纯粹是痴人说梦”，觉得这不过是 AI 领域老调重弹的故事。 团队最初计划筹集 5 亿美元，但处处碰壁，最终只能东拼西凑募得约 1 亿美元。在 Midha 看来，这笔资金在当时听起来虽多，但与谷歌等科技巨头投入同类系统的开销相比，根本只是“四舍五入的零头”。首轮投资人全是凭借个人信念入局的天使投资人，Midha 戏称他们是一帮“形形色色的杂牌军”（cats and dogs）。在随后的 18 个月里，Dario、Tom 与团队反复推演并敲定了一项战略方案——将亚马逊引入为战略合作伙伴，最终达成了高达 40 亿美元的算力与资本合作协议。这段经历让 Midha 深刻领悟到：**算力等底层基础设施，是催生任何前沿现代 AI 实验室不可或缺的生命线**。自那之后的五六年里，他便一直专注于解决前沿研究团队所面临的算力瓶颈。 - **初期募资困境**：目标 5 亿美元，实际仅勉强拼凑出约 1 亿美元的天使轮融资。 - **里程碑拐点**：耗时 18 个月敲定与亚马逊的 40 亿美元战略合作，直接绑定海量云算力。 - **核心认知**：算力基础设施是现代 AI 实验室能否存活与突破的决定性门槛。 --- ### `12:08` 并非唯一的终点线：AI 的“参差前沿”与多元模型竞争 针对播客主持人提出的是否所有前沿实验室（如 Google DeepMind、OpenAI、Anthropic 及紧随其后的中国团队）都将走向均质化竞争的疑问，Midha 提出了强烈的反对意见。他认为业界对于“模型前沿”（The Frontier）存在根本性的认知误区：前沿从来不是单数的，而是多元且不均衡的“参差前沿”（Jagged Frontier）。 Midha 用美国西进运动打比方：在狂野西部，并不只有淘金这一个前沿；给淘金者做牛仔裤的李维斯（Levi's）同样开辟了自己的前沿，并在工业革命浪潮中成长为商业巨擘。今天的 AI 格局如出一辙： - **软件工程前沿**：Anthropic 凭借极高的执行效率占据绝对领先；其团队仅有数千人，却能打造出让拥有数万员工的谷歌等巨头都难以轻易超越的顶级代码模型。 - **大众消费级对话前沿**：OpenAI 凭借先发优势和产品打磨主导着大众聊天入口。 - **多模态与视频前沿**：字节跳动（ByteDance）等团队则在视频生成领域开疆拓拓土。 Midha 强调，日常每天“双手放在键盘上”实操写代码的工程师会非常清楚，各家模型在深层体验上绝非平手（parity）。模型的行为差异，精准映射出背后研发团队长期倾注心血的差异化使命。 ```mermaid flowchart TD accTitle: AI 前沿领域的多元分化 accDescr: 图表展示了由通用算力支撑起的不同 AI 前沿方向，包括代码工程、消费级对话与多模态视频。 subgraph Infrastructure["底层算力与基础设施"] Grid["标准化算力底座 (AMP 等)"] end subgraph FrontierBranches["参差多态的技术前沿 (Jagged Frontier)"] SE["软件工程前沿<br/>(代表：Anthropic Claude)"] Chat["消费级交互前沿<br/>(代表：OpenAI ChatGPT)"] Video["视频与多模态前沿<br/>(代表：字节跳动等)"] end Grid --> SE Grid --> Chat Grid --> Video ``` --- ### `14:15` 斯坦福前沿模型研发四步法 在 Midha 讲授的斯坦福大学 AI 课程中，他将前沿大模型的生产配方高度凝练为四个核心阶段。他直言，虽然工程落地极其繁杂，但核心逻辑非常纯粹： 1. **预训练（Pre-training）**：抓取互联网上海量的海量多模态数据，将模型训练为一个具备基础通识能力的“通用模式识别机器”。 2. **中期训练（Mid-training）**：根据特定目标领域注入高阶能力。如果希望模型具备深度的科学、数学或物理推理能力，就在此阶段集中喂入高密度的专业领域数据。 3. **后训练（Post-training）**：进行对齐、指令微调与人类偏好校准，让模型更好地理解人类指令意图。 4. **持续反馈闭环（Continuous Feedback Loop / Context Feedback）**：将模型部署到真实应用场景中，根据任务执行的成败与客观验证结果，把反馈数据回流至训练管线中，使模型在该任务分布上不断进化。 --- ### `16:45` 客观可验证反馈：代码与材料科学为何进步最快 在聊到模型交互时，主持人打趣提到，因为害怕“罗科的蛇怪”（Roko's Basilisk）思想实验中未来超级 AI 对人类的报复，自己平时总对 AI 非常客气；但每当指出模型的回答有误时，模型虽然会低三下四地频繁道歉，可重新生成时往往依然给出雷同的错误答案，缺乏实时的自我纠错机制。 Midha 对此解释道，日常对话中用户随口说的“你错了”只是一种“主观意见”（opinion），并不是真正的**可验证反馈（Verifiable Feedback）**。AI 在特定领域突飞猛进的根本原因，在于该领域能否建立起接近客观事实的严格验证闭环： - **软件开发场景**：程序员写出代码后，首先提交拉取请求（Pull Request, PR），由可信任的同事进行代码审查与同行评审；随后代码进入自动化单元测试（Unit Tests）。只有同时通过人工审查与客观测试，这段代码才被验证为有效完成任务。正是这种严密的验证闭环，使得代码生成大模型在过去两年实现了惊人的跃升。 - **物理与材料科学场景**：Midha 联合孵化的 Periodic Labs 在门洛帕克（Menlo Park）开展前沿探索，目标是预测并寻找室温超导材料（这里也引起了主持人对当年全网热议“室温超导之夏”的回忆与欢笑）。AI 系统负责预测新型候选材料，随后自动化实验室中的物理机器人现场合成该材料，并用 X 射线衍射仪（XRD）测试其晶体结构与物理特性。这种直接来自“客观物理世界”的铁证被实时反哺回模型的训练迭代中。 Midha 总结道，凡是能够获得客观、事实性验证的任务，模型的进化速度最快且幻觉最少；相反，对于没有客观事实标准的纯主观任务（例如文学创作、心理咨询），模型往往容易陷入幻觉甚至输出有害内容。主持人随即笑着附和，透露自己曾向 AI 请教如何在后院做手工搭建，结果 AI 给出的搭建步骤居然要求“违抗地心引力”，让人哭笑不得。 ### `20:00` 结构化反馈与“参差前沿”：AI 自动化的边界 对话深入探讨了为什么编程（Coding）能成为 AI 最先取得突破的白领工作。主持人指出，在 AI 普及之前，编程领域就已经拥有高度系统化的反馈闭环（Feedback loops），比如写代码、运行、报错、修改这一循环；无论是写代码还是看模型的思维链（Chain of Thought），都能直观看到它不断尝试“这个行不通，换一种试试”的试错过程。相比之下，新闻业或销售等其他脑力工作并无这种结构化的对错标准。这引出一个万亿美元级别的核心问题：编程是否属于一种特殊的白领工作？ Anjney Midha 认为，AI 进展最可预测的领域，必然是那些工作流高度结构化且其产出“可被客观验证”（verifiable）的环节。比如在 Excel 中手动填入单元格数据这种繁琐、枯燥的流水线任务，就会被极快地自动化，从而让人能够专注于建模设计与业务洞察的“艺术”本身。同样，在新闻领域，像与《纽约时报》的 Cade Metz 或《华尔街日报》的 Brad Olsen 等顶尖记者交流时，就会发现他们大量时间花在核查每句话的信源与事实（Fact checking）上，这一环节完全可以交付给 AI 工具。 ```mermaid flowchart TD accTitle: 可验证与不可验证工作流在 AI 时代的分流 accDescr: 展示结构化与不可验证工作在自动化与人类价值上的不同发展路径 subgraph TaskTypes["任务性质分类"] A["结构化 / 可客观验证的工作流<br/>（如：表格填报、事实核查、代码调试）"] B["非结构化 / 依赖技艺与创意的领域<br/>（如：深度调查、宏观判断、创意表达）"] end subgraph Outcomes["未来演进与价值分配"] C["AI 快速接管与高度自动化<br/>（参差前沿的高推进区）"] D["人类专长与薪资溢价激增<br/>（无法被客观自动验证的独特价值）"] end A --> C B --> D ``` 这就引出了所谓的“参差前沿”（Jagged Frontier）格局：凡是结果能被客观验证的工作流，未来几年内都会以极高的确定性被 AI 推进并自动化；而在那些无法用传统客观手段验证的环节（即蕴含人类真正创造力、审美与手艺之处），人类劳动者的价值反而会大放异彩，其薪酬与影响力甚至将迎来非凡的增长。 ### `22:50` 为什么领导者必须具备技术素养：拒绝外包“理解力” 针对大行 CEO 们常用的监管防御逻辑——“只要把模型关进沙盒（Sandbox）运行，监管和高管就不需要关心黑箱内部是怎么工作的”，Anjney Midha 表达了极力反对的明确立场。他指出，技术素养（Technical literacy）对于企业领导者而言绝非可选项，而是一项不可妥协的基本要求。 这正是他在斯坦福大学开设广受欢迎的《前沿系统》（Frontier Systems）课程并将教学内容全部公开的原因。该课程从第一性原理拆解 AI 系统的完整链路：从土地（Land）、电力（Power）、机房外壳（Shell）、数据中心基础设施，再到模型训练全过程；学生们的期末大作业则是打造一个属于自己的“一人前沿实验室”（One-person frontier lab）。掌握底层逻辑的意义在于： - **避免将系统拟人化（Anthropomorphize）**：在缺乏底层理解时，人们容易脱离现实地去推测 AI 的能力、商业模式甚至对社会的冲击。 - **看清多维竞争格局**：外界常以为“前沿只有三家大模型”，但在业内视角下，实际上同时存在多达 17 个细分前沿阵地，每个领域都有 3 到 4 家实力惊人的玩家。 - **识破沙盒的虚假安全感**：仅仅声称“放在沙盒里就安全”极其危险。如果员工在沙盒内滥用工具，依然会面临模型幻觉、提示词注入（Prompt injection）等风险。若高管根本不懂业务物理逻辑和模型推演的真实输入，只盲目采纳模型生成的报表，本质上就是把企业的“理解力”外包了出去。 > “你可以外包你的计算，可以外包枯燥的工作流，但你绝不能把自己的‘理解力’外包给模型。” ### `27:30` 破解算力瓶颈：AMP 要做算力领域的通用“电网” 当话题转向无论任何模型前沿都绕不开的底层瓶颈——算力（Compute）时，Anjney Midha 阐述了其新创立的公共利益公司（Public Benefit Corporation, PBC）AMP 的核心使命。当前获取算力的方式极度碎片化：有的巨头自建数据中心，小型实验室则依赖第三方机房或新兴云（Neocloud），市场处于异构且割裂的混乱状态。 AMP 借鉴了人类基础设施的发展史：无论是工业革命、互联网还是流媒体，最初的输入形态都极具异构性。比如早期的电力系统，在交直流电（AC/DC）标准化之前，全美各地都有大量孤立、闲置的“搁浅电力”（Stranded megawatts）无法流动；直到格式统一并建立起覆盖全国的“电网”（Grid），电力才真正成为即插即用的基础设施。 AMP 的目标正是为算力打造这样一个基础设施： - **算力标准化**：抹平底层不同芯片类型、不同硬件制造商以及不同云厂商之间的异质性。 - **搭建算力电网（The Grid）**：构建一个统一的算力调度与分发层，把当前闲置、分散的算力资源接入通用电网中。 针对业界在尝试创建“算力指数”或“算力期货”时屡屡碰壁的可替代性（Fungibility）难题，AMP 采取了软件层面的解法： - AMP 开发了名为“The Grid”的系统，在**消费层（Consumption layer）**实现算力的可替代化。 - 这一技术路线源自 Google 内部著名的资源调度与集群管理系统 Borg。AMP 的工程联合创始人 Sebastian Lobo（14 年前与 Anjney 在斯坦福大学同住一间宿舍的室友）曾是 Google Borg 项目的核心技术负责人之一。 - AMP 本质上是在为全行业打造一套“通用的 Borg 系统”：它作为底层的翻译转换层，让机器学习研究员无需操心底层用的是哪家芯片或架构，只需专心提交自己的训练或推理负载，所有底层硬件异构性与调度问题均由该软件系统全权接管。 ### `30:00` 算力利用率的真相：节点利用率与模型浮点运算利用率（MFU） 在探讨如何降低算力成本时，Anjney Midha 指出这完全是一个软件问题。AMP 正在构建的软件转换层，能够将许多在网格上孵化的实验室的算力利用率从原先的 50% 到 60% 大幅提升至接近 95% 或 96%。为了提供对比基准，Midha 引用了 Google 的内部数据：在 Google，算力利用率大约稳定在 99% 的惊人水平；当联合创始人 Sebastian 刚加入 Google 时，该数字仅为 62% 左右，而在他离职时已提升至 99%。在 Google 内部，如果集群利用率跌落到 96%，就已经被视作严重的系统故障（major outage）。 反观当前更广泛的独立数据中心生态，平均利用率甚至不足 70%。Midha 举例透露，埃隆·马斯克（Elon Musk）在孟菲斯部署的 Colossus 2 集群——拥有数十万颗顶级加速芯片——其实际节点利用率（Node Utilization）低于 60%，而模型浮点运算利用率（Model Flops Utilization, MFU）更是低于 11%。数据中心关注两种截然不同的利用率指标： 1. **节点利用率**：最粗浅的度量标准，单纯反映有多少芯片处于被调用状态。如果这低于 90%，没有任何借口，因为手握芯片至少应该让它们运转起来。 2. **MFU（模型浮点运算利用率）**：衡量具体工作负载中，单颗芯片理论算力究竟被压榨出了多少。在绝大多数团队中，该数值往往极低。 主持人和 Midha 展开讨论，指出业界充斥着盲目的“芯片吹嘘”（chip bragging）——很多团队喜欢高调宣扬自己囤积了数百乃至数万张 NVIDIA GPU，以此暗示自身技术实力。然而，Midha 强调“衡量应该以产出为导向，而非投入”。在创办 Periodic Labs 时（团队包括 ChatGPT 共同创作者 Liam Fedus 以及 DeepMind 前物理团队负责人），他们制定路线图时最核心的考量指标绝不是囤了多少张卡，而是真实的评测（eval）与最终产出。单纯堆叠芯片数量很大程度上只是一种虚张声势（bravado）。 ```mermaid flowchart TD accTitle: 算力利用率的两种度量层级 accDescr: 展示从集群物理芯片分配的节点利用率到芯片内部实际运算效率 MFU 的两层度量差异。 subgraph ClusterLevel["集群宏观层：节点利用率 (Node Utilization)"] A["物理芯片总数"] --> B{"芯片是否被分配作业？"} B -->|"是 (>90% 及格)"| C["被占用芯片 (Active Nodes)"] B -->|"否 (闲置浪费)"| D["空闲芯片 (Idle)"] end subgraph ChipLevel["芯片微观层：模型浮点利用率 (MFU)"] C --> E{"芯片内部算力利用深度"} E -->|"理论峰值算力"| F["理论 FLOPS"] E -->|"实际吞吐 (常低于 20%-40%)"| G["真实 MFU 产出"] end ``` --- ### `32:38` 长期租赁模式的经济死重与“网格信用额度”（Grid Credits） Midha 详细拆解了算力市场当前存在的荒谬商业形态。从经济角度来看，现行算力业务的基本原子单位是“长期租赁合约”。一位 AI 研究人员向算力供应商寻求资源，供应商通常会提供数百颗 AMD 或 NVIDIA 芯片的长期租期，要求研究人员像租公寓一样每周 7 天、每天 24 小时持续付费。哪怕标称价格看似只有每小时 2.50 美元或 3 美元，不管使用者跑不跑任务，供应商都已经锁定了未来两年的账面收入，将所有利用率风险转嫁给研究人员。 这种机制必然导致巨大的资源浪费与无谓损失（deadweight loss），核心原因在于 AI 研发具有极高的“突发性”（spiky）且极难预测： - 团队在摸索算法时往往使用较少算力； - 一旦某条技术路径跑通，团队会立即暴增需求（“大家快看，这个有效，赶紧扩容！把所有芯片都压上去，准备发布！”）； - 当模型训练完毕上线后，算力需求瞬间断崖式下跌。 因为无法根据基础负载规划，团队不得不按照“峰值负载”（peak load）进行过度配置（over-provisioning）。在这两次需求脉冲之间，存在着海量完全闲置的算力。结果就是，研发团队名义上签署了每小时 2.50 美元的低廉合同，但扣除闲置期后，实际承担的有效算力成本高达每小时 25 到 28 美元——溢价超过 10 倍。 ```mermaid flowchart LR accTitle: 长期租赁与统一网格模式对比 accDescr: 对比传统长期固定租赁导致的算力波峰浪费与 AMP 网格信贷池模式的高效再分配机制。 subgraph LegacyModel["传统长期租赁模式 (High Waste)"] L1["固定租用 24/7 芯片池"] --> L2{"突发研发需求"} L2 -->|"波峰时期"| L3["满载运行 ($2.50/hr)"] L2 -->|"闲置间歇期"| L4["完全空转 (有效成本飙升至 $25-$28/hr)"] end subgraph AMPGridModel["AMP 软件网格模式 (Grid Credits)"] G1["异构闲置算力 (NVIDIA / AMD)"] --> G2["软件抽象层转换为同质资源"] G2 --> G3["按需计费的 Grid Credits"] G3 --> G4["研发波谷自动释放，供给其它实验室"] end ``` 针对这一巨大缺口，AMP 的软件解决方案是将所有闲置算力无论底层硬件形态（NVIDIA、深受喜爱的 AMD 或其他定制芯片）汇集起来，统一转化为完全可替代的“同质化资源”（fungible resource），并以标准化格式“网格信用额度”（Grid Credits）提供给开发者。研究人员无需再关心底层具体是什么芯片架构，只需按实际使用量付费。身兼 7 家公司董事的 Midha 坦言，作为投资人，他最兴奋的就是看到投后团队从每小时承担 25 至 26 美元的沉重长租合同中解脱出来，恢复到真正符合标价的 2.50 美元实际支出，而未使用的算力则被动态调度回网格供其他实验室运转。 --- ### `36:34` 动态融合训练与推理负载，为何巨头竞相自研定制芯片 针对主持人关于该方案主要适用于模型训练还是推理的问题，Midha 明确表示网格能够同时兼顾两者。由于网格将异构计算资源抽象为同质化资源，系统能够实现极高弹性的动态填充：所有长期预订的整块算力专注用于基础模型训练，而训练间隙产生的所有碎片化空闲算力，则实时用来吞吐推理（inference）任务。 随后主持人提及当前行业的新趋势：各大实验室和科技巨头纷纷推出定制芯片，例如微软最新公布的自主 AI 芯片与大模型（如 MAI 系列）。当被问及为什么几乎每家主流公司都感到必须自研专用芯片、走“模型与芯片深度协同”（model-chip alignment）路线时，Midha 从经济与技术两个维度进行了深度剖析： 1. **经济原因（利润率控制）**： - 目前 AI 实验室在研发上每花出 1 美元，就有大约 80 美分直接流入了 NVIDIA 这类芯片供应商的口袋。这导致实验室自身的单位经济效益（unit economics）极其微薄和严苛。 - 实验室自研芯片的内在冲动是为了摆脱被单点榨取高额利润的被动局面，夺回对自身毛利率的掌控权。 2. **技术原因（供应链安全与代工厂配额）**： - 过去四五年来，全球科技界一直处于严重的算力受限（compute-constrained）状态。如果实验室无法获取所需规格和数量的芯片，就意味着完全丧失了供应链自主权，只能受制于计算供应商认为最符合其利益的配额分配。 - 即使放眼整个上游供应链，台积电（TSMC）这样的晶圆代工巨头也因产能上限，事实上升格为决定哪个算力供应商业务能够扩张的“终极裁判”。巨头唯有亲自下场参与底层设计与晶圆争夺，才能勉强维持供应链的连续性。 ### `40:00` 芯片自主与算力电网：从模型生产到 Token 消费 Anjney Midha 深入分析了巨头自研芯片背后的双重驱动力：技术上的供应链独立与商业上的单位经济（unit economics）自主权。以微软为例，Satya Nadella 并非要去建晶圆厂，而是必须自研推理芯片，因为如果完全依赖第三方芯片提供商，企业既无法掌控单次推理成本，又容易受制于供给瓶颈，进而限制产品的规模化销售。 随后话题转向算力使用的错配现象——诸如 Uber 被曝在短短四个月内就烧光了整整一年的 Token 预算。针对“更智能的模型调度是否会消除用量峰值”的疑问，Midha 明确指出必须区分“模型的生产”与“Token 的消费”： ```mermaid flowchart TD accTitle: 算力电网生产与消费解耦流程 accDescr: 展示 AMP 算力电网从底层集群调度模型训练，到端侧 Token 消费平抑波动的全流程 subgraph GridLayer["算力生产层 (AMP 作为独立系统运营商)"] A["分散异构数据中心 / 算力集群"] --> B["AMP 算力电网协同调度"] B --> C["AI 模型生产与训练峰值"] end subgraph InferenceLayer["消费交付层"] C --> D["模型托管与推理生成 Token"] D --> E["企业与终端用户调用"] end E -. 平抑周期性波动.-> B ``` - **AMP 的电网调度模式**：AMP 的本质定位是算力电网的“独立系统运营商”（ISO, Independent System Operator）。AMP 自身不拥有数据中心，也不拥有模型实验室，而是像统筹区域电网一样跨主体协调算力需求。只要终端用户生态足够多元化，算力消费的周期性波动在大规模层面上就会自然被抚平，类似于美国日常电网负荷自我平衡机制。 --- ### `42:15` 中世纪时期的技术路由：企业不在乎底层模型 在谈及模型路由器（如 Midha 投资的 OpenRouter）以及企业界对模型选择的认知时，Midha 直言：“我们目前绝对还处在这一技术的‘中世纪’阶段。”全球企业的决策层根本不在乎底层究竟是哪个大模型，他们被底层的复杂性所困扰，核心诉求极其现实：能否以最便宜、最高效、最安全可信的方式把活干完。 未来的应用架构将把模型选择完全抽象化。用户只会与一个值得信赖的品牌所提供的 AI 伴侣/助手对话，而在引擎盖之下，系统可能同时编排调用了 200 多个不同的小模型与专业模型。这也印证了 Midha 的预判：未来绝不可能只剩下两三家前沿大模型独大，而是一个百花齐放的庞大生态圈。主播 Tracy 对此戏称为她的“咖啡胶囊理论”（Coffee Pod Theory of AI）。 --- ### `43:56` CEO 认知两极化：“杠铃分布”与现场育儿趣谈 近期媒体频频报道企业对 AI 支出的反思以及投资回报率（ROI）不达预期的质疑（包括 Uber 首席运营官对 AI 支出的公开发难）。Midha 观察到，企业高管对 AI ROI 的认知呈现出极其鲜明的“杠铃分布”（barbell distribution）： - **亲自动手的领跑者**：一端是亲自上手使用工具的 CEO。他们深切体会到了 AI 能力的“参差前沿”（jagged frontier），明白模型在某些任务上堪称神技、在另一些任务上漏洞百出，因此制定的战略与提出的问题非常切中要害。 - **外包认知的旁观者**：另一端是不碰工具、把理解外包给下属或儿女的高管，往往只会随口问孩子“那个 ChatGPT 挺厉害的对吧”。 这番话引来了 Joe Weisenthal 的插科打诨。Joe 笑称自己的孩子不仅没觉得 AI 厉害，反而觉得自己很蠢：“他们才六岁，什么都不懂，只是单纯觉得老爸用 AI 很蠢！”Midha 随即大笑，热情邀请 Joe 把孩子送到他们家在伦敦开办的夏令营“Camp Midha-Shen”（由他和妻子 Shen 的姓氏命名），自告奋勇当“好玩的叔叔”，现场洋溢着轻松愉快的笑声与调侃。 --- ### `46:04` 算力金融化的隐忧与垂直飙升的需求曲线 针对闲置算力信用额度是否会演变为二级衍生品交易市场的讨论，Midha 表达了高度谨慎甚至反对的态度。他强调，一旦把纯投机机制引入生产资料，将制造出全新维度的资源稀缺，金融投机客追逐溢价的狂欢最终会严重伤害一线科技研发团队。 - **对冲需求而非投机炒作**：AMP 内部确实组建了一支类似对冲基金的供需预测团队，通过算力集群的期权（call options）提前锁定算力产能。但这种操作的逻辑等同于大型钢铁企业设立内部交易台来锁定铁矿石现货，目标是为了保障生产连续性与成本可控，而非资本套利。 - **算力需求曲线“垂直冲天”**：当被问及当前算力需求斜率相比一年前的变化时，Midha 用“垂直”（Perpendicular）来形容其爆发性。从 2026 年初到年中的短短六个月内，长期算力租赁的市场价格直接翻了一番；AMP 在 1 月份锁定的 2026 年算力配额，如果现在转手卖掉，转眼即可获得 100% 的净溢价。 --- ### `48:38` Claude Code 的启示：模型与 Harness 的协同设计 节目尾声探讨了近期 Claude Code 引发的技术狂潮。针对主持人提出的“实验室投资重点是否正在从纯模型缩放（scaling）转向外部工具链与 Harness 开发”这一猜想，Midha 明确予以纠正。他指出，这绝非单纯的 Harness（外部脚手架/运行环境）层面的突破，而是模型能力与 Harness 之间深度协同、对偶演进的产物： - **协同设计（Co-design）**：在 AI 工业界，下一代 Harness 是由系统工程团队与研发前沿模型的研究员并肩协同设计的。 - **可预测的研发流水线**：由于前沿模型的训练演进曲线具有可预测性，团队在模型尚未出炉前，就已经精准预判了它将在哪些逻辑分支与编程环节展现出突飞猛进的能力，从而量身定制出让模型能像“电脑里的微型电脑”一样自主运行的 Harness。模型核心能力与外围调度体系如同交响乐般协同演进，两者缺一不可。 ### `50:00` 模型与 Harness 的协同设计：将工具能力内置化的工程闭环 在探讨模型迭代时，Anjney Midha 深入解析了预训练（Pre-training）、中训练（Mid-training）与持续反馈闭环之间的紧密联动。他指出，一旦研发团队在整个训练流程中拥有了全链路的能见度，就能做出极具前瞻性的工程规划：“我们明确要在未来三个月内攻克模型在某一类特定任务上的能力缺陷，那么现在就可以同步着手为该能力的落地设计外挂框架（Harness）。”在这种协同研发模式下，外挂脚手架在架构设计之初就清晰预判：哪些能力在几个月后能被模型本体原生消化，哪些依旧需要依赖外部第三方工具。 ```mermaid flowchart TD accTitle: 模型与外挂脚手架的协同设计演进 accDescr: 展示模型训练将外部工具调用能力内置化并联动更新Harness以缩减延迟的过程 subgraph Before ["初期阶段：模型能力受限"] direction TB A1["用户输入请求"] --> B1["Harness 调度判断"] B1 --> C1["调用第三方表格工具"] C1 --> D1["返回结果（高延迟，多耗时1-2分钟）"] end subgraph Transition ["三个月协同研发闭环"] direction TB T1["识别能力短板"] --> T2["中训练注入表格推理能力"] T2 --> T3["同步重构 Harness 路由逻辑"] end subgraph After ["协同进化：模型原生推理"] direction TB A2["用户输入请求"] --> B2["Harness 识别原生支持"] B2 --> C2["模型内部直接完成推理"] C2 --> D2["极速返回响应（体验大幅跃升）"] end Before ==> Transition ==> After ``` Midha 举了一个生动的具体案例：假设三个月前，底座模型在理解和处理电子表格（Spreadsheet）方面表现糟糕，整个系统必须依赖外部 Harness 调用第三方工具来代为处理；而在过去三个月里，团队通过专项训练，直接将针对电子表格的结构化推理能力注入到了模型本体（"in the model"）之中。此时，外部 Harness 随之敏捷更新，裁撤对第三方工具的冗余调用。这种架构演进直接省去了跨服务交互开销，将端到端任务耗时“瞬间缩短了一到两分钟”。Midha 感慨道，只有当模型权重与外围 Harness 像这样紧密协同设计（co-design）时，软硬件体系才算真正奏响了“交响乐（create a symphony）”，产品的用户体验也会产生质的飞跃。 访谈在极其融洽的气氛中收尾。主持人 Joe Weisenthal 与 Tracy Alloway 诚挚感谢了 Midha 的深度分享，并热烈向听众推荐他在斯坦福大学主讲的爆款 AI 系列公开课（YouTube 上的 CS153 / 相关课程代码）。Joe 开玩笑称自己正准备搭乘一趟长途航班，打算在飞机上把这些视频全部下载看完。 - **端到端提速效应**：将外部工具调用逻辑转化为模型内生推理后，复杂任务的处理时间通常能直接削减 1 至 2 分钟。 - **协同设计核心**：模型中训练（Mid-training）目标与工程化 Harness 需并行演进，而非在模型训练完成后再被动适配。 --- ### `51:15` 主持人复盘：参差不齐的前沿、模型大宗商品化与软件定义算力 节目结束后的复盘环节中，两位主持人对 Midha 提出的诸多行业洞见进行了深入拆解。Joe 首先强调了 Midha 在斯坦福开设的课程在硅谷学术圈的狂热影响力，形容学生们几乎是“踏破门槛（beating down the door）”去争抢选课名额。随后，Tracy 将话题引向访谈中最令她震撼的两个概念：其一是“参差不齐的前沿”（Jagged Frontier），这意味着基础模型领域绝非简单的“单极独霸”或赢家通吃（Winner-takes-all）格局，不同模型在各类垂类能力上各有所长；其二是下游企业的实际使用诉求——绝大多数公司根本不在乎模型背后的品牌光环，他们只想要成本最低、刚好能稳定交付任务的实用工具。 这种市场心理直接导向了大模型服务的“大宗商品化（Commodification）”。Tracy 直言，如果终端用户和企业追求的只是性价比极高的平价实用方案（就像随冲随饮的 Keurig 胶囊咖啡机），那么当前资本市场上针对少数昂贵头部模型——即所谓的“凯迪拉克（Cadillac）”模型——给出的天量估值逻辑就显得十分脆弱。如果所有估值模型都预设所有人最终都会为最顶级的全能模型买单，甚至声称其潜在市场空间（TAM）“不仅能覆盖整个地球，还能席卷全宇宙”，这种狂热预期显然存在巨大泡沫。 ```mermaid flowchart LR accTitle: 软件层算力网格调度示意图 accDescr: 展现异构物理硬件通过软件调度层虚拟化为标准化统一服务的架构 subgraph Physical ["物理基础设施层（高度异构碎片化）"] H1["单卡 / 8卡机箱"] H2["CoreWeave 等算力云"] H3["各类规格 GPU 集群"] end subgraph Software ["AMP PBC 软件网格层"] S1["动态虚拟化与资源抽象"] S2["智能化按需路由调度"] end subgraph Market ["企业与终端应用层"] C1["高性价比大宗商品化服务"] C2["定制化垂类业务需求"] end Physical --> Software --> Market ``` 两位主持人进一步探讨了 AMP PBC 商业模式的底层逻辑。Joe 指出，用户表面上采购的是高度标准化的廉价大宗商品服务，但在技术底座之下，系统必须依赖一套极其复杂的调度引擎，在形态各异的各类模型间敏捷路由。两人联想到此前与 CoreWeave 联合创始人 Brannin McBee 等算力基础设施先驱的对话，通常业内都将算力瓶颈与碎片化视为难以逾越的“物理世界难题”（涉及芯片排布、网络拓扑、机箱异构等极其固化的硬件限制）；然而，Midha 及其团队却坚信这一死局完全可以通过“纯软件定义调度”的方式予以攻克。虽然像 Google 这样具备顶尖工程实力的巨头内部早已实现了类似调度，但若能将这种软件层面的算力网格推广为通用基础设施，必将彻底重塑整个 AI 算力市场的经济模型。 - **大宗商品化趋势**：企业级客户更倾向于采购具备成本确定性的低门槛服务，而非盲目溢价采购全能顶级模型。 - **TAM 估值反思**：基于“通用凯迪拉克模型通吃全球”所推导的超高估值，严重忽视了模型分化与垂类替代品的存在。 - **软件破局物理异构**：传统云基础设施厂商强调的芯片配置物理壁垒，正面临通用软件虚拟化与智能分配调度的正面颠覆。 ## Notable quotes > `05:00` 是我先开了第一张支票，随后我们试图说服另外 25 家风投跟投，结果我碰壁失败了。 — Anjney Midha > `10:00` 算力等底层基础设施，是催生任何前沿现代 AI 实验室不可或缺的生命线。 — Anjney Midha > `22:50` 你可以外包你的计算，可以外包枯燥的工作流，但你绝不能把自己的‘理解力’外包给模型。 — Anjney Midha > `30:00` 衡量应该以产出为导向，而非投入。 — Anjney Midha > `32:38` 大家快看，这个有效，赶紧扩容！把所有芯片都压上去，准备发布！ — Anjney Midha > `42:15` 我们目前绝对还处在这一技术的‘中世纪’阶段。 — Anjney Midha > `43:56` 他们才六岁，什么都不懂，只是单纯觉得老爸用 AI 很蠢！ — Joe Weisenthal > `50:00` 我们明确要在未来三个月内攻克模型在某一类特定任务上的能力缺陷，那么现在就可以同步着手为该能力的落地设计外挂框架（Harness）。 — Anjney Midha ## Mentioned Anjney Midha, Joe Weisenthal, Tracy Alloway, Nick Bostrom, Andrej Karpathy, Andrew Ng, John Doerr, Mary Meeker, Dario Amodei, Tom Brown, Sebastian Lobo, Liam Fedus, Cade Metz, Brad Olsen, Elon Musk, Satya Nadella, Brannin McBee, Anthropic, OpenAI, AMP PBC, Google, Google DeepMind, Amazon, Microsoft, NVIDIA, AMD, TSMC, ByteDance, Discord, Ubiquity6, Periodic Labs, CoreWeave, Kleiner Perkins, a16z, Benchmark, Index, Uber, OpenRouter, Claude, Claude Code, Claude 4.8, ChatGPT, GPT-3, GPT-5.5, Borg, MAI ## Video description <p data-prosemirror-content-type="node" data-prosemirror-node-name="paragraph" data-prosemirror-node-block="true" data-pm-slice="1 1 []">Anjney Midha wrote the first check to Anthropic. He teaches a viral course at Stanford on how AI works. And he was, until recently, a partner at a16z. In other words, he is AI-industry royalty. Midha's new project is AMP PBC, a company that believes it can radically lower the price of compute. To accomplish that, he is working on building a compute grid that turns GPUs into a standardized utility. But right now, compute is too fragmented. It's too heterogeneous. And given the way contracts are structured, he says that labs are being forced to spend money on capacity that often goes unused. In other words, small labs are forced to pay up for big, long-term contracts, even though their own demand (particularly during model training) may be very spiky. On this episode, Midha explains how the market for compute currently works and why he believes there's a software solution that could significantly improve compute utilization. He also tells us why he does not anticipate one company will emerge as the dominate player and that instead we'll have a wide range of models, each optimally used in specific applications.</p> <p data-prosemirror-content-type="node" data-prosemirror-node-name="paragraph" data-prosemirror-node-block="true">Read more:<br data-prosemirror-content-type="node" data-prosemirror-node-name="hardBreak" data-prosemirror-node-inline="true"><a href="https://www.bloomberg.com/news/articles/2026-06-11/amazon-says-its-data-centers-use-2-5-billion-gallons-of-water?utm_medium=referral&utm_source=podcast&utm_campaign=odd_lots&utm_content=article" data-prosemirror-content-type="mark" data-prosemirror-mark-name="link">Amazon Says Its Data Centers Use 2.5 Billion Gallons of Water</a><br data-prosemirror-content-type="node" data-prosemirror-node-name="hardBreak" data-prosemirror-node-inline="true"><a href="https://www.bloomberg.com/news/articles/2026-06-10/oracle-reports-higher-than-expected-data-center-spending?utm_medium=referral&utm_source=podcast&utm_campaign=odd_lots&utm_content=article" data-prosemirror-content-type="mark" data-prosemirror-mark-name="link">Oracle Falls Most in Six Months on Mounting Data Center Costs</a></p> <p data-prosemirror-content-type="node" data-prosemirror-node-name="paragraph" data-prosemirror-node-block="true">Only <a href="http://Bloomberg.com" data-inline-card="" data-card-data="">http://Bloomberg.com</a> subscribers can get the Odd Lots newsletter in their inbox each week, plus unlimited access to the site and app. Subscribe at&nbsp; <a href="https://www.bloomberg.com/subscriptions/oddlots?in_source=oddlotspodcast" data-prosemirror-content-type="mark" data-prosemirror-mark-name="link">bloomberg.com/subscriptions/oddlots</a></p> <p data-prosemirror-content-type="node" data-prosemirror-node-name="paragraph" data-prosemirror-node-block="true"><a href="http://bloomberg.com/subscriptions/oddlots" data-prosemirror-content-type="mark" data-prosemirror-mark-name="link">Subscribe to the Odd Lots Newsletter</a><br data-prosemirror-content-type="node" data-prosemirror-node-name="hardBreak" data-prosemirror-node-inline="true"><strong data-prosemirror-content-type="mark" data-prosemirror-mark-name="strong">Join the conversation:</strong> <a href="https://discord.gg/oddlots" data-prosemirror-content-type="mark" data-prosemirror-mark-name="link">discord.gg/oddlots</a></p><p>See <a href="https://omnystudio.com/listener">omnystudio.com/listener</a> for privacy information.</p> [Source](https://podtrac.com/pts/redirect.mp3/tracking.swap.fm/track/UVBrz8bN8aM2Xe47PEPu/traffic.omny.fm/d/clips/e73c998e-6e60-432f-8610-ae210140c5b1/8a94442e-5a74-4fa2-8b8d-ae27003a8d6b/90cd8aa8-11c9-4cd3-813a-b467011c7f57/audio.mp3?in_playlist=982f5071-765c-403d-969d-ae27003a8d83)

[5:00] Anjney Midha: 是我先开了第一张支票，随后我们试图说服另外 25 家风投跟投，结果我碰壁失败了。 [10:00] Anjney Midha: 算力等底层基础设施，是催生任何前沿现代 AI 实验室不可或缺的生命线。 [22:50] Anjney Midha: 你可以外包你的计算，可以外包枯燥的工作流，但你绝不能把自己的‘理解力’外包给模型。 [30:00] Anjney Midha: 衡量应该以产出为导向，而非投入。 [32:38] Anjney Midha: 大家快看，这个有效，赶紧扩容！把所有芯片都压上去，准备发布！ [42:15] Anjney Midha: 我们目前绝对还处在这一技术的‘中世纪’阶段。 [43:56] Joe Weisenthal: 他们才六岁，什么都不懂，只是单纯觉得老爸用 AI 很蠢！ [50:00] Anjney Midha: 我们明确要在未来三个月内攻克模型在某一类特定任务上的能力缺陷，那么现在就可以同步着手为该能力的落地设计外挂框架（Harness）。

[9:47] Anjney Midha: They were like, 'nobody cares, but we think it's the best thing since sliced bread.'... We tried to raise $500 million. We couldn't. We instead scraped together about 100 million... It was all angels in that first round. [20:00] Anjney Midha: You can outsource your thinking. You can outsource part of the tedious workflows, but you can't outsource your understanding. [40:00] Anjney Midha: When you add speculation to production goods, it creates scarcity of a different kind... I'm not that thrilled about the financialization of these products if it ultimately results in more speculation. [45:00] Anjney Midha: It was not just a harness innovation... It's a symphony of improvement... a dialectic between the model capability and the harness. In the industry, we call this co-design.

[0:00] Whatever your goal, trade show giveaways, client gifts, or team gear, 4imprint has the promo products to match. With thousands of options, from apparel and drinkware to tech and totes, it's easy to find the right fit for your brand and budget, with standout choices at every price point. And with their 360-degree guarantee, you can be 4imprint certain your order will show up just right, right on time. Explore more at 4imprint.com. 4imprint. 4certain. So there's a lot of noise about AI. But time's too tight for more promises. So let's talk about results. At IBM, we work with our employees to integrate technology right into the systems they need. Now, a global workforce of 300,000 can use AI to fill their HR questions, resolving 94% of common questions. Not noise. Proof of how we can help companies get smarter by putting AI where it actually pays off. Deep in the work that moves the business. Let's create smarter business. IBM. When you're running a business, the best days are the ones where priorities stay on track. For midsize and large companies, risk can affect multiple parts of the organization at once, from property and liability to cyber and regulatory challenges. At that level managing risk becomes an ongoing discipline At the Hartford the focus is on helping businesses manage risk before it turns into something more disruptive And when losses do happen, that work is paired with insurance coverage shaped by years of underwriting, risk engineering, and claims experience. Learn more at thehartford.com slash risk mitigation. Policies provided by Hartford Fire Insurance Company and its property and casualty affiliates, Hartford, Connecticut. Bloomberg Audio Studios, podcasts, radio, news. Hello and welcome to another episode of the Odd Thoughts podcast. I'm Tracy Allaway. And I'm Joe Weisenthal. Joe, we like to talk a lot about physical constraints on this show, right? And this is one reason why AI is a really fascinating area for us right now, because there are a lot of physical constraints on what is ultimately the sort of ephemeral technology. And I think that the tension between those two things is really interesting, right? Like you type a prompt into chat GPT or Claude or whatever and it the sort of like disembodied digital platform You don necessarily think about the power usage the real resources the transformers that have to go into data centers to get compute The thing that I've been on my mind lately, and I've written about it, and I plan to write more, is this idea that the canonical AI thought experiment is what happens if you tell an AI to make a lot of paper clips, and then it destroys the world. Because in the pursuit of marshalling all of the world's resources. It just turns everything into paperclips because it doesn't know. I have to ask, is this canonical example? Is this based on your traumatic fear of Clippy? No, but that is, you know, it all comes back full circle. But what we are seeing in real life is that everything from access to electrical grids, GPUs being the big example, energy turbines, talent, and now even including residential real estate are being reproposed to make more and more advanced AI. And in the original paperclip thought experiment, they envision, or at least in one version, the philosopher Nicholas Bostrom envisions the AI having exhausted all of the world's resources, then sending a probe into outer space to consume star energy to build more paperclips. Just eat the universe. And to this point we even talking about going into outer space for data centers to build more AI So every version of the thought experiment is being replicated except it just more and more resources to build the AI by humans rather than paperclips by AI There's this other connected theme here. So we've talked before about how one of the reasons valuations seem to be getting insane in the market is because all of this activity is being driven by like this existential need to become number one in frontier models and this new technology. And so if you say you absolutely have to be the first to invent AGI, then you can justify any amount of spending on Earth. Right. And so what we tend to see is like the biggest companies just keep getting bigger. They're the ones that can get resources for all this stuff. And I think one of the most fascinating things right now is that at least as of right now, June 4th, 2026, the frontier models are really close to each other. Right. So and the 4.8 GPT 5.5, like they're not that different. And one of the things I'm curious about is, is there something inherent in market dynamics in this space that will always keep, you know, whether it's being able to distill results from another model and quasi-steal them, whether it's information sharing. [5:00] among employees. Is there some inherent reason why we've seen the stability? Or could it be that at some point one lab just like breaks out and establishes permanent possibility? But like I am personally on the side of commodification and everything just becomes kind of basic. Well, everyone's basically available. I know. OK, well, I'm just kidding. All right. Thank you, Joe. All right. That's a polite prompt to get to the guest. We do, in fact, have the perfect guest. we're going to be speaking with Anjane Mida. He is, of course, a former general partner at Anderson Horowitz, a Stanford University visiting scientist who teaches the viral AI lecture called Frontier Systems. Also, one of the first guys to write a check for Anthropic. He is now the founder of a new company called AMP PBC. So thank you so much for coming on AdLots, Anj. Thanks for having me. One correction, it's pronounced AMP PBC, but that's everything else you got perfect on the intro. Amp would make sense, wouldn't it? As in energy. Yeah. And just remind us, the PBC is Public Benefit Corporation. That's right. So you're doing this for the public benefit. We're governed by a public benefit charter, which means everything we do has to follow our mission We have a public charter mission We are for in the same way Ben Jerry or REI and Anthropic are public benefits So we aim to make a healthy modest amount of profits that can sustain our mission. But we have the flexibility to choose what that margin is. Can I just start? I want to establish your credentials, although I feel like that very long list did a pretty good job. But writing the first check for Anthropic, tell us that kind of origin story? Because the anecdote that you hear is like 25 VCs turned them away initially, and you said yes. It was a little bit of the other way around. I said yes, then we tried to get another 25 VCs to say yes, and I failed. It was a harrowing experience. It was a bit of a wake-up call. It was late 2020. I had just sold my last business. It was called Ubiquity 6. It was a 3D mapping business. It was an AI business that we had founded in 2017. And I felt like a failure at the time because I was in San Francisco. Just as big picture, my life stories. I was born in India. I went to high school in Singapore. And I came out of college to the United States at Stanford for my undergraduate degree. And then when I arrived at campus in 2011, deep learning had just started taking over the world in Silicon Valley Andre Carpathy was a computer science TA to Andrew Ng who was one of the I would say modern sort of founding fathers of deep learning this idea that you can teach machines to think without having to give them prescriptive rules. And so I went into sort of machine, I got swept up in that moment and started studying. A lot of my coursework was in machine learning. My primary department at Stanford was in bioinformatics, which was machine learning applied to healthcare. I got sidetracked to a venture firm called Kleiner Perkins for about four and a half years where I got the chance to work for some of the great investors like John Doerr and Mary Meeker. And then I left and started my own company. And as is the case in Silicon Valley, when you start, I mean, I was 25, I went and raised about 47 or so million dollars from some of the usual suspects like benchmark and index and so on. And I thought I was the coolest kid in town and I got the beat out of me because we we built this incredible technology which is this AI system that could map any location in 3D and then the pandemic hit and so location-based mapping 3D mapping you can the only thing you can control is how you react to what happens and so I did feel for a moment like it was bad luck and then you just have to pick up the pieces and make the best of it so i i did with my co we we figured it out it was it was a tough um few years where we had to pivot the business but we we landed the plane we we essentially a lot of the distributed systems we'd built on the back end side ended up being quite valuable we sold that to a company called discord which is a chat up for gamers yeah we have an all-life discord time to plug that with our fans on there yeah awesome our listeners so you know about a month after i sold the business I got a call from some friends who were running research at OpenAI. And we'd all been friends in the machine learning community in the Bay Area. And they said, Ange, we've trained a little model called GPT-3. And we think it's the best. Just a little model. Yeah. Nobody really paid attention. They were like, nobody cares. But we think it's the best thing since sliced bread. And we want to leave and turn this into a standalone business. But it would be helpful to get some of your advice on how to do that. And I couldn't really come on board full time at the time with them because I had to integrate my company to the acquirer. Sure. But I came on as their angel and nights and weekends, I worked with them on the business plan and who we should raise from. That company is Anthropic. Dario and Tom and I started doing these weekly working sessions in early 2021. And yeah, I assumed that, you know, if we went and talked to a bunch of venture capitalists on Sandhill Road, especially some of the ones who were [10:00] involved in the biggest hits of the last decade before that, they would get it. These are the creators of GPT-3. And they were like, we just don't get this. We've heard the whole AI story before. This whole general intelligence thing is a pipe dream. And it was painful. We tried to raise $500 million. We couldn't. We instead scraped together about 100 million, which I know sounds like a lot, but at the time was a rounding error compared to how much Google has spent on the same kind of systems. And it was all angels in that first round, a bunch of cats and dogs, all of us who believed in the mission. And then over the next 18 months, Dario, Tom and team put together a plan that we kind of workshopped on getting Amazon involved as a strategic. And that resulted in a $4 billion compute and capital partnership that made me realize infrastructure, especially compute infrastructure, was just a key requirement to create any kind of modern AI lab. And so since then I've spent the past five, six years figuring out how to unblock that compute bottleneck for research teams. Amazing. Well, obviously an incredibly well-timed... It just like emphasizes how much things have changed, right? Where like people are literally throwing money at like almost any model now versus like a few years ago going like ah AGI I don really know Well let me ask you this question because this is a very top of mind question for me And we getting we can skip around on the timeline here But there are three labs that are seen as like genuinely at the frontier right now. And that is obviously DeepMind within Google, OpenAI and Anthropic. And then, of course, you know, a lot of people say that the Chinese labs are very close, if not quite there. Maybe they're a few months behind. Is this, is there, you know, when we think about like part of your mission is like you say, okay, a new lab should be able to get access to compute. If you're really bright, like that shouldn't be the bottleneck. Does that imply, therefore, that you expect more labs to be able to, were they to have access to the compute, also reach the frontier, and that there is something inherent about like this sort of seeming stability or parity that we see among frontier models? So the answer to your first question is yes, there are many frontiers to be conquered and pioneered. And it's not just one frontier. I think that's a fundamental misunderstanding people have about the frontier. The jagged frontier. Exactly jagged intelligence In a poetic sense in a historical sense if you think about the Wild West or the Western frontier it wasn just one frontier There was a frontier of gold and there was a frontier of jeans It turns out Levi turned out to be a new modern behemoth of a company I mean, there were so many new businesses founded in the Industrial Revolution. And I think that's the reality is the software engineering frontier, which is where Anthropic is clearly a leader, is one frontier. I think the chat frontier, the sort of consumer chat frontier is another frontier where OpenAI has been a leader. arguably bite dances at the video frontier with seed dance right absolutely yeah you know and so i think there's just many many frontiers to be conquered or pioneered rather i think anthropic is clearly a role model for the rest of the community on how to do it in an efficient way they're you know i think fewer than 5 000 people and they've been able to put out state-of-the-art models that you know teams like google which have 60 000 people are close to but not yet quite there Actually, I don't really agree with your assessment that they're all at parity. If you use the models day in and day out, they're quite remarkably different in meaningful ways to the person with hands on the keyboard doing the engineering work. And I think those differences reflect the focus of the teams, right? What is the actual mission that the team working on that domain cares about day after day after day So in the Stanford class I teach the first lecture was a breakdown of how frontier models are even created And it's actually quite simple. The recipe is super simple. There's basically four steps. There's pre-training, mid-training, post-training, and then what we call the continuous feedback loop. So pre-training just says, hey, you collect a bunch of data from the internet and train a model to be a generally good pattern recognition machine. You then do mid-training, which is to say, in a particular domain that you really care about, you inject more capabilities. So if you want this model to reason about science or math or physics, then you give it science or math or physics data. And then you get a pretty good model that's specialized in that domain. And then you deploy it to the real world where you have people using it. And the context feedback, which is when the model is able to do a task well or not, and you can verify whether that task was done correctly, gives the model the data it needs to keep improving on that task, on that distribution. [15:00] Whether you're planning a big tech event, launching a new campaign, or just stocking up on team gear, finding the right promotional products makes all the difference. 4imprint offers thousands of options, from on-trend apparel and premium drinkware, to tech, totes, and giveaways, so you can find the right fit for any audience, purpose, or budget. You can customize it all, your logo, your message, your look, and many items come with no setup charge to help you save. And if you're really watching the bottom line, you'll find standout choices at every price point, so you can make a real impact while staying on budget. Plus, you'll get expert help, fast turnaround times, and their 360-degree guarantee. So you can be 4imprint certain your order will arrive on time and look exactly right. Whatever your goal, 4imprint makes it easy to find your perfect promo match. Explore the possibilities today at 4imprint.com. 4imprint. 4certain. The impact of myasthenia gravis symptoms can pile up, plans getting rescheduled, Hobbies being left behind Everyday tasks going unfinished If MG keeps disrupting your life the burden may be bigger than you realize Learn how to assess and track the impact of your symptoms so you can talk about it with your healthcare team Take the next step to better understand how MG symptoms may be impacting your life Visit mgsymptomimpact Ask yourself, what are your best people spending their time on right now? Expense reports, receipt chasing, month-end close that takes weeks. You become what you spend on, and that's not what you're building toward. Brex is the intelligent finance platform that eliminates that work before it starts. AI agents that handle the manual stuff automatically, so your team can spend their time on what actually compounds. It's time to get Brex AF. Learn more at brex.com slash AF. This is slightly tangential, but I give a lot of feedback to the models because Joe made me paranoid about the basilisk theory. So I want the models to appreciate me once they take over the world. But when you give them feedback, like if they spit out a wrong answer and you say that's wrong, they immediately apologize and fall over themselves to say that they're sorry. But then you ask them, like, give me another output or like, would you do it again the same way? And they like they often say yes or they give like a very similar answer. They don seem to be responding in real time Correct So when I say feedback I in a very specific kind of feedback which I which I call verifiable feedback So when you say that wasn right or that was wrong that an opinion Verifiable feedback is when you can have as close to factual verification as possible. The reason... So what does that actually look like? That's a great question. So let's take reason by example in two or three cases. In the case of software engineering, the way software engineers actually code is you write a piece of code and then you submit it to the main code base And then you usually have a peer on your team review the code and approve it or reject it. And if it gets approved, that's the first step. That's called a PR, a pull request. And if another human on your team that you trust approved it, that's one kind of verification of quality. And then two, before that piece of code usually gets deployed to a production system, you have unit tests. And those are quite objective tests of is this code performing the function we need it to? And if it passes both those tests, it's a verifiable piece of code that accomplished the goal. So in software engineering, the reason we've seen such a dramatic improvement in capabilities is that a lot of these labs are using feedback from that verification loop. In the case of another lab I incubated called Periodic Labs which we started a year ago and you should come by sometime We got 40 square feet in Menlo Park where we got AI models that are predicting new The goal is to try to find a room temperature superconductor And so these models predict Oh, I forgot about that. The summer of room temperature superconductors. I forgot about that. That was a fun summer. Yes. This time we will verify that... If we ever put something on, you will know it's not... It'll be real. That's not going to be us. But the AI system predicts new materials candidates. then we have robots that synthesize the new material in the lab and then use x-ray diffraction machines to test whether the material has the properties the AI said it would. And that's verifiable feedback from reality, from physics. And then we pipe that data back into the training loop over and over again. That context feedback is very factually verifiable. And that's where progress is the fastest today because that feedback doesn't result in the kind of hallucinations that you often experience with these models on more subjective tasks. It's also, by the way, why the models are terrible at subjective tasks like creative writing. And sometimes it can get quite toxic, to be honest, if you get them down the wrong loop. I don't know if you've been using it as a therapy bot and so on. I have not, just for the record. That's great. It did ask me to defy the laws of gravity at one point because I was trying to create something in my backyard and I was asking how to do it. And it was like, then just set this up like the following way. And I was [20:00] That's not within the laws of physics. Yeah. Yeah. Whatever. Well, what's interesting, and this is actually a trillion dollar question from just a very broad standpoint, is, as you point out, even prior to AI, the field of coding had a very systematized approach to the feedback loops already. Yes. And so then it's like, hey, I can sort of replicate that. Anyone who's done your vibe coding can see in the chain of thought sometimes, oh, that didn't work. Let me try this. That didn't work. Let me try this. Yes. Most fields don't really have that by and large. Journalism doesn't have that. I mean, there are outputs that are better and worse. We don't really have that sort of like formalized approach to the yes, no. Does that just zooming out to my mind, that would imply that maybe at least to some extent coding is a little bit special from a sort of white collar knowledge work that in terms of like, is it going to be as good as say sales or something like that? Because it has a coding as a long history. It's formulated and structured. That structured pipeline. Yeah, that's a great point. So where progress will be made most predictably is in parts of knowledge work where the task is essentially a workflow that fairly structured And so somebody who spends most of their day inputting cells into an Excel spreadsheet, well, that part of the job will get automated pretty fast because that's actually verifiable. And you know what? That's frankly often the most tedious part of the job anyway. And so I'm quite excited to see that progress because I'm terrible at spreadsheets. and I think if we could free up more of my time and hopefully other people's time to focus on the art of the spreadsheet, not the tedious part of it. The entry and retrieval, yeah. Yeah, exactly. And in journalism, I think it's the same thing. There's so much craft that goes into the verification of a story before it goes out that's not legible to the world. I've had a chance to spend some time with some of the journalistic institutions of the barrier like Cade Metz or Brad Olsen at the journal. And as you spend time with them, you realize, I mean, they're verifying every sentence that goes into each. Fact checking. So fact checking. That's an example where I think we should be leaning on these tools and you should expect more progress. And the parts then that will be more to borrow your jagged frontier framing there that that we will be in a regime of jagged frontier progress where wherever parts of workflows that are verifiable factually will essentially you see progress there very predictably over the next few years And consequently, wherever that progress, the workflows are not verifiable is actually where humans are going to shine. And I think that's where parts of the economy are. You're going to see extraordinary gains in the wages of humans who have creativity and craft that are not typically verifiable through traditional objective means. Does that make sense? Yeah, it does. And it dovetails with a lot of what we've been talking about on the show recently. Just going back to verifiable feedback. So, OK, the model spits out something and you can check whether it's right or wrong. Right. Is it important to understand how the model actually got to that answer? Because we have discussions with like big bank CEOs who are using more AI and their response to this question is always like, well, if we can put restrictions around the AI, if we make sure that it's like released into a sandbox before it's released into the wider world, we're all set from a regulatory perspective and regulators don't actually need to know what's in the black box model and how it working But like this seems a bit concerning to me Yeah No I am quite strongly opinionated about this one which is that technical literacy should be non It the reason I spend so much time teaching this class at Stanford putting it up online And the idea of the frontier systems class is that end to end, it's a full, simple, but first principles breakdown of how these AI systems are built from scratch, from land, power, shell, like the energy, where do we get them, the data centers, then how do we train the models? And the final project, the class with the kids was actually the one person frontier lab, which is at the end, they're creating their own models and so on. Because the idea is that a person with the right tools today can scale themselves infinitely, but they need to know how to use the tools, what the limitations are, when to lean on them versus not. And I think this is a generalizable piece of technical literacy that all leaders should have. It's like saying, you know, in the 90s, I imagine if you knew, you could use the internet without really knowing how it worked. But on the margins when the page doesn't refresh or you're like, this cookie thing is annoying me. Over time, people who are more technically literate just realized sometimes you've got to debug the browser. And those of us who've learned over time to do knowledge work are more adept at leaning on them versus not. Just now when I was trying to get onto the internet, I realized, okay, there's this... [25:00] you know, Wi-Fi, password, whatever. And then you don't end up relying on them in ways that they can't fulfill your need anyway. And what's a little bit more dangerous with these systems is because we tend to anthropomorphize them without the technical literacy that I wish all leaders had about reasoning about how these systems were built. What you end up doing is projecting out in your mind what the capabilities are in ways that are inaccurate. You project out their impact on society that are not accurate. You project out their business models in a way that are not accurate. I mean, the very fact that when you started this conversation, I don't blame you for it. You're like, Anj, there's three models at the frontier. I'm like, well, which frontier and which three models? Because from where I'm sitting, there's like 17 different frontiers right now and there's four different players in each one and the businesses of all of them are kind of breathtaking. So I think that technical literacy should always for leaders be a basic requirement. And then if you're deploying these systems at Goldman Sachs, you won't oversimplify and get tripped up later when two years later, you realize half your employee base has been leaning on this sandbox framing When in reality inside the sandbox they were using the tools in ways that were prone to hallucination prone to risks prompt injection They were leaning on it in ways that were not informed in the appropriate ways Is this making sense? Yeah, like at a minimum, they would not be using it in the optimal way. Correct. Or relying too much on it. You can't outsource your understanding to a model. You can outsource your thinking. You can outsource part of the tedious workflows. but you can't outsource your understanding. And if you keep thinking, if you say, if you create these simplistic frameworks of, oh, here's a sandbox and this is safe, you have to use that sandbox in the right way because if you say, well, now everything that happens in the sandbox is totally fine. If the model says, use the spreadsheet, the spreadsheet is good, it's deployed on our servers, but you didn't actually check the spreadsheet and what went into the spreadsheet and did the model actually understand the particular structure of the business, the physics of the business that you're trying to model? out, then you've outsourced your understanding to it. Does that make sense? Absolutely. Let's talk about AMP. Yes. And because you're never going to get the frontier in anything unless you have access to compute. It seems pretty obvious. And there are various arrangements for acquiring compute You have companies building their own data centers You have smaller labs and maybe they use someone else data centers or a neocloud etc What are you building at AMP such that at least as part of this story is trying to solve the compute bottleneck specifically Yeah, we are. It's very simple what we're doing at AMP. We're doing two things. We are trying to standardize the format for compute, which today is super fragmented. So in the history of infrastructure, if you look at whether it was the Industrial Revolution, the internet, streaming, there were usually formats of inputs that were quite heterogeneous. They were fragmented. And then to unlock productivity, you had to standardize a format. So in the case of electricity, until ACDC was standardized, megawatts would just sit in stranded pockets around the United States being unused. And then once we standardized the format to ACDC, then the question was, okay, great. Now we turned all these stranded pockets of electricity into one sort of interoperable universal format. Now how do we distribute it to everybody who needs it? And we came up with this distribution layer in the United States called the grid That all we doing You building a grid for compute Correct We trying to standardize the compute layer today Different chip types different manufacturers different clouds I mean, it's a complete mess. Say more about how you plan to do this, because we've talked before about, you know, there are various people out there that want to create indices of compute, futures potentially on compute. And the issue that always comes up is fungibility. Right, exactly. So we've got a couple of ways we solve the fungibility problem. This is a pretty thorny challenge. We solve it in two or three ways. The first is we have a system called the grid, which actually makes the compute fungible at a consumption layer. So under the hood, we have a bunch of different chip types. We support various different manufacturers. And there's a system that was built to do this already inside a little company called Google. and one of the technical leads on that on that project was called Borg internally at Google is my co-founder Sebastian Lobo he was my roommate at Stanford 14 years ago he's my engineering co-founder and we're building Borg for everybody else which is essentially a translation layer that says no matter what the underlying chip type is the machine learning researcher who's using the chip just has to worry about the workload and we handle everything else underneath the hood when you when you say system yeah is this hardware or software that's [30:00] It's all software. Okay. Yeah. So we handle that translation layer in software. Huh. And it's a pretty gnarly challenge. But today, we're able to do that in ways that improve utilization sometimes from 50%, 60% at labs that we have incubated on the grid to close to 95%, 96%. At Google, the utilization is roughly 99%. When Sebastian arrived at Google, it was about 62%. By the time he left, it was roughly at 99%. At Google, if utilization is at 96%, that's considered a major outage. Today, the average data center in the industry, in the ecosystem, in the independent ecosystem, is running at less than 70% utilization. The Colossus 2, which is running in Memphis, Elon's 500,000 GB300s, was running at less than 60% node utilization and less than 11% MFU. Model flop utilization is how much of the chip is actually being used. So there's two kinds of utilization people care about in a data center. First is how many chips are being used. That's the highest. That's just the most naive measure. If that number is not at 90 plus percent, no excuses. So you have the chips. They should at least be doing something. And then within the chip how much of the chip is being used Within a workload That number is usually much lower I very intrigued by this latter point about that even if that even like the chip itself may not be even used at full capacity Yes Because I see these numbers and you say like a lab has like a we have 200 chips we've acquired 800 GPUs etc. And when I see these headlines I assumed that oh utilize optimal utilization techniques must be so good that you can infer someone's capabilities simply by how many NVIDIA GPUs they've acquired. What you're saying is that there is actually quite a bit of heterogeneity about the techniques and approaches to getting the most juice out of AdChip. Yes, you have to measure what matters. And what matters is output. Okay. Anytime I start a new lab, in the case of periodic labs, we started with Liam Fettis, who was the co-creator of ChatGPT and Doish Chubbuk who led the physics teams at DeepMind. And when we sat down and we planned out the company's roadmap, the most important thing to us to measure was not the number of chips we had. Yeah. It's the eval, what we call... So all this chip bragging, they're like, oh, we acquired, it's just a sort of... It's a lot of bravado. Yeah, all right. This is helpful You don measure the inputs You should be using the outputs No I agree I actually fascinated that there is a software solution to what I perceived in my head as a very physical constraint How does this actually work? Feel free to get technical here. I want to understand the system. Yes. So let me give you the technological answer and the economic answer. The economic answer actually is a simple reason about. But the way the compute business works today is primarily on the construct of the atomic unit of long-term leases. So I'm a researcher. I need some compute. I show up to a compute provider and say, hello, I would like some compute, please. And the compute provider says, no problem. Here's, you know, 500 AMD chips or NVIDIA chips that you can lease from me on various timescales. And you've got to pay for it 24-7. It's like leasing an apartment. And whether you use it or not, that's your problem, but it's $2.50 per hour, $3 an hour. So instead, you take a long-term lease, and now the cloud provider, the compute provider said, great, I just booked revenue for the next two years that this guy rented from me. Now, what happens with that compute, whether it's used or not, is the researcher's problem. They've outsourced that problem. As a result of this wastage that we talking about and I happy to go into why it hard for individual teams to utilize most of the capacity The primary reason is because research is spiky It hard to forecast So you over-provision for your peak, not your base load. Because what happens, you're researching on these algorithms. And the minute one is working, you go, guys, let's scale. We want to ship this thing. So let's throw as many chips at it. And then once we ship it, the needs go down. And so between these spikes, there's just huge pockets of unused compute. As a result, the effective price per hour that you're paying is closer to $25 to $28, whereas the marketed rate that you think you're paying is $2.50. So that spread due to wastage is just insane. So from an economic perspective, that's the wastage. That's the deadweight loss. Yeah. Right. Okay, so now how do we, from a technological perspective, how do we utilize that opportunity? Literally, all we do is, from a software perspective, we take all of that unutilized compute, no matter what format it is, it might be NVIDIA, it might be AMD, we love AMD, it might be some other chip, and we turn it into one fungible resource. And we standardize the format on something we call grid credits. So researchers don't even need to think about what chip type is under the hood specifically. [35:00] They're just paying what they need or what they use. And so from a fiduciary perspective, I'm on seven boards. As an investor, I get very excited when teams switch from this sort of long-term lease model where they're paying $25, $26 per GPU hour, to now they're actually only paying the $2.50 that was marketed because everything they're not using gets reallocated to the grid and other research labs can use that resource. Thank you. Thank you Help fast turnaround times and their 360 guarantee So you can be 4imprint certain your order will arrive on time and look exactly right Whatever your goal 4imprint makes it easy to find your perfect promo match Explore the possibilities today at 4imprint.com. 4imprint, 4certain. The impact of myasthenia gravis symptoms can pile up. Plans getting rescheduled. Hobbies being left behind. Everyday tasks going unfinished. If MG keeps disrupting your life, the burden may be bigger than you realize. Learn how to assess and track the impact of your symptoms so you can talk about it with your health care team. Take the next step to better understand how MG symptoms may be impacting your life. Visit mgsymptomimpact.com. Ask yourself, what are your best people spending their time on right now? Expense reports, receipt chasing, month-end close that takes weeks. You become what you spend on, and that's not what you're building toward. Brex is the intelligent finance platform that eliminates that work before it starts. AI agents that handle the manual stuff automatically, so your team can spend their time on what actually compounds. It time to get Brex AF Learn more at brex slash AF This is the problem you mostly solving for is the training part because they training right Training and interest. No, no, no. Or is it both? Both. Yeah. So the beauty about having diverse types of compute on our grid is that once you make the resource fungible, you can do any workload. You just fill all the unutilized pockets with inference and then all the reservations with training. Got it. So can you explain, like, why is it that every lab also seems interested right now in customized silicon, including Microsoft announcing a chip that says, like, oh, our new MAI, I don't know, MAI, I don't know how it's pronounced. Oh, MAI, I believe. We had Sachin in the class yesterday at Stanford, and he pronounced it as MAI. Okay, their new MAI model. And he's like, oh, and we also have a new MAI 200 chip or something that's optimized with it. Why is it that so many labs or companies that are in a lab, I guess, feel impelled to also design a chip that goes along with the model? And long term is what you're doing saying like this really is not necessary to have that sort of model chip alignment. Yeah there two technological reasons and two economic reasons The first is from an economic perspective about 80 cents of every dollar a lab spends today on their R close to a chip provider like NVIDIA Okay. Right? And so as a result, your margins are just super, super rough. So from a unit economic perspective, you want more control over your margins. And therefore, when you look at your unit economics, you're going, wait, wait a minute. For every dollar we make, there's this massive chunk that's going to somebody else. So instead of spending 80 cents to NVIDIA, you spend 78 cents to TSMC and keep that two cents for yourself? Well, I think that the better our software gets, the more that margin should flow actually to the researcher. Because that's where the value will be captured. But wait, sorry, you were going to say, what's the technical reason why they're trying to do optimal model chip alignment? On the technical side, the primary reason is you want control of your supply chain. because today in a compute, well, we've been a computer constrained world now for at least four or five years. But if you can't get the chips you need, you're not in control of your own supply chain. So you're dependent on compute allocations that the compute manufacturer thinks is optimal, right? By the way, that's how it works at the foundry level. Today, TSMC gets to decide which compute provider, compute providers business grows or not because they only have so much [40:00] capacity. And so the technological reason is you want supply chain independence. And so when you want economic independence, unit economic independence, and you want supply chain independence, you want as much control over your own chip. But Microsoft doesn't have a fab. That's not what I'm saying. What I'm saying is in inference, for example, Satya would like more control over his unit economics. So he's making an inference chip. Right. Because if you are dependent on a third party to give you the inference chip. Okay. And if you don't have an inference chip, you can't sell more product. You want more Is it about having a predictable supply of just for you rather than a predictable supply of. OK, yes. So there's a lot of discussion right now about more efficient model allocation. So this idea that like you do not have to be using the latest model to ask, like what the weather is going to be tomorrow or something like that. And you also don't want to blow through your entire one year token budget in the space of four months, as Uber apparently did. So the spikes in usage that you're seeing that allow you to do the system and you know have grid credits Does some of that go away if people become smarter about which models they're actually using? Okay, so There an embedded assumption I think I should tease apart in your question Usage is different from the production of the model So what happening right in terms of the pipeline is you use the grid to produce the system, the model, and then the model produces tokens. If the end user is only using tokens, then as long as everybody, we have enough diversity in the end user base using models hosted on the grid, things actually even out. Okay. That cyclicality in the same way electricity in America evens out if you have enough scale at scale, basically, except when like there's a heat wave, exactly. So some of that infrastructure, we are having to reboot, but you can think about AMP in the broadest sense as a utility company where what's called an independent system operator of the grid. So we don't own our own data centers. We don't own our own labs, but we coordinate the capacity needs across different parties. and at sufficient scale, those usage patterns actually just get evened out. Does that make sense? Yeah. Well, you are an investor in OpenRouter, I believe, which I think is an interesting company. Do you see, setting aside AMP for a second, do you think that there is at this point still within say corporate America a certain lack of savviness about knowing which model to route to for the query and that there will be an improvement and learning within companies within users so that you don have these incidents for like massive token consumption because perhaps everyone was using the wrong, the Cadillac model and the Ford model would have been just as fine for that purpose. Oh yeah, we're absolutely in the medieval ages of this technology. I think what will happen is increasingly, based on my conversations with corporate American leaders and corporate leaders across the world, they don't really care about the models. They don't care about the underlying model, the technology. They just don't care. It's like too much complexity. We just want the work done. Can you guys please figure out how to get the work done in the cheapest way, in the most efficient way, in the most secure and trusted way? And increasingly what you'll find is that which particular model is helping you out in a particular task will just be abstracted. You won't even think about that it'll just be a companion you're just going to talk to it it'll be a companion provided by a brand you trust and under the hood they might be using 200 different models to orchestrate your task and over time that efficiency will get better and better and better and better and that's why i just don think there only three frontier models that are going to win well it going to be an ecosystem this is i know you don want my take Joe but this is my coffee pod theory of AI I want your take Tracy I want your take I love your take I'll save it for the outro. Actually, on this note, we have seen some headlines recently. Obviously, there's the Uber one about token spending, and I think it was the COO said he wasn't sure if the ROI was there on Uber's AI usage. And we've seen, there was a good Vox article recently about a corporate reckoning with AI spend. Since you're going out and talking to CEOs, do you see any like, has anything shifted in the past couple months or so in the way people are thinking about the return on this initial investment or the return on spending on tokens? Yes, I think it's a barbell distribution. So there's two types of CEOs, broadly speaking. The first is the CEOs who are using the tools themselves. And those folks are going, aha, I understand the jagged frontier. When they understand the jagged frontier we talked about, their strategies, their questions they ask me are completely different from the CEOs who are outsourcing their understanding. They're not trying the tools. They're mostly asking their kids like, hey, kiddo, this chat GPT thing, like, it's good, right? And your kid is like... [45:00] That's pretty good, Dad. My kids think it's really dumb, by the way. Yeah, so that's the other thing, right? So the kids are super smart, and they're using the tools, and they're like, it's good at this thing, but not at that. So they understand the jagged frontier part as well. Actually, you know what? They think I'm dumb for using it. They're like, Dad, you're not doing anything smart. They don't think the models are dumb. They think it's dumb of me to use them. Exactly. They might be going, the way you're using it is not optimal. so so the see no what i'm saying is my kids are six and they have no intent and they have no idea about anything and they just think i'm dumb that's the only that just might be a generalizable that's it that's really the only point i'm trying to make i see okay well you can send them over to me anytime i'm happy to be the fun uncle yeah that would be great you can show them that actually this is fun to play my wife and i are happy to host your kids okay that's great that's really what i'm trying to get at uh it's it's the summer right we have two nieces in london and we call it Camp Mida Shen. My last name is Mida. My wife's name is Shen. And so you're welcome to send them to Camp Mida Shen anytime. That's amazing. But that's the bifurcation is leaders who are actually trying the tools out. They realize they're extraordinary at some things and not at others. And so depending on whether you get it or not or you actually getting your hands dirty or not I find the questions are completely different So this has been an incredibly helpful conversation in terms of understanding basically the problem of essentially tons of money is being spent And your thesis is that it's massively suboptimally used up and down the stack. You mentioned this, okay, you get a credit, et cetera. Do you actually see that being financialized in a way? I mean, okay, I bought this capacity, have a lot of unused time. I don't always have a research idea that is going to require a big model run test. I can resell that. Is that something that you see, something that genuinely resembles a financial market? I hope not, because when you add speculation to production goods, it creates scarcity of a different kind, right? Because then you have financial traders and markets trying to trade the speculative value of the asset. And that's going to hurt a lot of our research teams in technology. On the other hand, I think that creates a need for innovation inside of the research teams. And so one of the core operating functions we have inside of our business is a forecasting capability where we have a team that very similar to actually the kind of forecasting team you have inside of a hedge fund We constantly predicting demand and supply And then we actually procuring capacity in advance to call options on compute clusters But our needs are similar to the kind of internal trading desk you'd have inside of a large steel company, right, where they need to lock up iron ore and so on for their production needs. So I'm a big fan of efficient markets. And I'm trying to actively invest in and help entrepreneurs out and teams out who are trying to drive more efficiency in the service of more productivity in science and engineering. I'm not that thrilled about the financialization of these products if it ultimately results in more speculation. Does that make sense? Yeah. I'm just curious, since you're tracking demand in that way, like if you were going to describe the slope of demand right now versus, say, like a year ago, is it steeper? Is it starting to plateau? Perpendicular. Oh, wow. Okay. If you look at the compute prices of long-term rentals over the last six months, between January and now, they're trading up 2x. So we started, for example, for 2026, we started securing our capacity in January at these long-term rates. We could resell that at a 2x markup if we wanted to Part of the reason that 2026 has become just totally AI has consumed everyone mind I think is because people got very excited about Claude Code specifically Yeah. But that was a breakthrough at the harness level, not the model level, right? Suddenly, like the really excited is like, wow, this is just so fun. It's just so easy. You have a computer inside your computer. That was a harness breakthrough. Do you see like when you think about investment among AI labs, do you see any shift in allocation away from pure scaling and improving the model towards sort of like tooling and harnesses as a way to get more juice out of the models? No, I'm sorry. I have to correct you there. It was not just a harness innovation. Those two things go hand in hand. It's a symphony of improvement between, it's a dialectic between the model capability and the harness. That harness was designed specifically for the capabilities that the new model was going to have. And so when you design these things, in the industry, we call this co-design. So you have the harness designed side by side with the researcher who is designing the next generation capabilities in the model. And you get a little bit of visibility in where the model is going to be good because, as I described earlier, the pipeline is actually quite predictable. Thank you. [50:00] Pre-training, mid-training, continuous feedback loop. Once you have that visibility, you go, aha, we specifically want to improve the capabilities on this type of task. It's going to take us about three months to get there. Start designing the harness for that improvement. By the time they show up, then you can have the harness assume that the model will be able to do X, Y, Z on its own, whereas A, B, C, it's going to need third-party tools. so then the harness says remember that three months ago you were terrible at understanding a spreadsheet yeah so then we had to go right like go use a third-party tool to use a spreadsheet in the last three months what we've done is added the ability to actually reason about a spreadsheet in the model in the model not so now you don't need to use a third-party spreadsheet okay and so then the harness gets updated to say don't go out and use a third-party spreadsheet which by the it collapses the time required to do that task by like sometimes a minute to two minutes. Now suddenly I've improved the user experience. And that's when things really sing. It's when both of those parts, the model and the harness are co-designed to create a symphony. Does that make sense? Yeah, absolutely. All right Anjaneh Mida of AMP PBC Thank you so much for coming on OddLots Really appreciate it Thanks for having me And everyone go out and check out the Stanford lecture series It on YouTube right It is CS153 Perfect I have a big flight coming up, so I'll watch it then. You should download all the lectures. There's quite a few. Thank you so much, Anjan. That was fantastic. That was great. All right, Joe, that was a great discussion. Yeah. I should emphasize just how big a deal that lecture series actually is at Stanford. Like students are beating down the door basically to get into that. And if it's free on YouTube, you should definitely check it out. I just want to establish that if I had given you the A, the insinuation that I didn't want to hear your take, or B, the idea that I would have wanted to hear Anjaneh's take instead of yours, I want to hear your take. No, it's fine, Joe. I realize that most listeners are here for the guest takes. I get it. But I thought his point about the jagged frontier was an important one. And this idea that like maybe the future, it's not going to be a winner takes all thing in terms of models. You're going to have a bunch of different models doing different things that might suit different companies And also the idea that like a lot of companies aren going to care about which specific model they using they just want the cheapest one that basically gets the job done in my mind that sounds like more of a it's commodified the modification yeah a commodified market right rather than like oh people are going to pay up for as you said the cadillac well so what i would say is by listening to Anjanae and Amp is that people will want a commodified service, but that under the hood, I mean, this just sounds like what she's really trying to solve. And it's very interesting. I, as a user or a company, buy a commodified service, but under the hood, the commodity has an incredible amount of variety of models through which it can route. Some of which will be the Cadillac. Some of it will be the Keurig coffee cup. Yeah, absolutely. But like my point is maybe in terms of valuation. Sure Right Like if everyone is assuming that the Cadillac is going to be like the one that everyone is going to get and the total available market the TAM infamously is like not just the world but potentially the universe like that seems a stretch to me Totally And just generally I thought that was super interesting And the idea we done a couple episodes recently specifically learning more about both chip level and box level optimizations both how many chips you're using and how well you're using a chip. Definitely we have more to do on that. It still blows my mind that this is a problem that can be solved with software rather than like something physical. You just come up with a way to efficiently allocate the compute. Because in my mind, it's such a physical problem. And we've talked to previous AI market participants like Brandon McBee at CoreWeave. And they talk about, oh, it's difficult to standardize because of the configurations of chips and things like that. But if you could solve it just through a software system, that's pretty crazy. I guess Google's already done it. Yeah. All right. Shall we leave it there? Let's leave it there. This has been another episode of the All Thoughts Podcast. I'm Tracy Alloway. You can follow me at Tracy Alloway. And I'm Jill Weisenthal. You can follow me at The Stalwart. Follow our guest Anjane Mida at Anjane Mida. Follow our producers Carmen Rodriguez at Carmen Armid. Dashiell Bennett at Dashpot. Cale Brooks at Cale Brooks. And Kevin Lozano at Kevin Lloyd Lozano. And for more All Thoughts content, you should check out our daily newsletter. You can find that at Bloomberg.com forward slash All Thoughts. [55:00] You can chat about all these topics 24-7 in our Discord, discord.gg. And if you enjoyed this conversation, then please leave a comment or like the video, or better yet, subscribe. Thanks for listening. Thank you. And with their 360 guarantee you can be 4imprint certain your order will show up just right right on time Explore more at 4imprint 4imprint 4certain Ask yourself, what are your best people spending their time on right now? Expense reports? Receipt chasing? Month-end close that takes weeks? You become what you spend on, and that's not what you're building toward. Brex is the intelligent finance platform that eliminates that work before it starts. AI agents that handle the manual stuff automatically so your team can spend their time on what actually compounds. It's time to get Brex AF. Learn more at brex.com slash AF. Okay, picture your market as a living map. Just think of it. Customers moving, competitors shifting, conditions constantly changing. Where do you see opportunity? Where do you see risk? Esri turns those questions into business strategy. You can see the bigger picture and act with confidence because until you see where everything connects, you never truly know how to move forward. Discover the power of where at ESRI.com. That's ESRI.com. Esri, helping to turn questions into business strategy.