Hudson River Trading Hunts Megawatts as Power Limits Constrain AI

电力限制约束AI发展，Hudson River Trading寻求兆瓦级电力 AI负责人Iain Dunning表示，数据中心电力短缺已取代GPU紧缺，成为高频交易机构面临的主要瓶颈。 在本期现场节目中，Iain Dunning探讨了Hudson River Trading如何利用AI模型，并应对严峻的数据中心电力限制、不断上升的硬件成本以及不断演变的计算基础设施。他详细解析了内部基准测试实践、诸如Meme股聚类等模型涌现行为，以及员工每日100至1,000美元Token预算的财务现状。此外，Dunning还阐述了AI如何重塑量化招聘实践，使机构能够在调整面试流程以适应AI辅助工作流的同时，赋能理论思考者。 在City Winery的现场录制中，Hudson River Trading的AI负责人Iain Dunning在他首次做客节目七个月后，探讨了AI正在如何重塑量化金融。 (0:00) 尽管像HRT这样的量化机构拥有巨大的算力和迭代研究能力，在理论上有可能匹敌DeepSeek等开源模型，但前沿级别的训练依然极其消耗资本。 (0:00) 算力的指数级增长使HRT能够同时在全球股票、期货、期权和加密货币市场上运行统一的机器学习模型。 (5:00) HRT通过与人类研究员进行内部正面对比来对新的前沿模型进行基准测试；目前，顶尖模型的表现大致相当于一名优秀的量化研究实习生。 (8:30) 神经网络擅长在不需要人类可解释叙述的情况下提取原始信号，模型能够自主发现诸如Meme股聚类等涌现模式。 (10:00)

Hudson River Trading Head of AI Iain Dunning argues that the rapid acceleration of artificial intelligence is fundamentally rewiring quantitative finance into a post-narrative, compute-intensive discipline. However, realizing these compounding efficiency gains increasingly depends on navigating severe physical power shortages and adapting institutional hiring to favor high-level conceptual articulation over routine coding. Power Constraints and Infrastructure Bottlenecks (17:35) According to Dunning, procuring individual GPUs like Nvidia's Blackwell architecture is no longer the primary hurdle in scaling AI capabilities; the defining bottleneck is securing data center sites with sufficient electrical power. Deploying clusters on short notice has become virtually impossible without inheriting forfeited leases, and forthcoming hardware such as Nvidia's 2027 Rubin chips is expected to be heavily backordered. To maintain its exponential internal compute growth, Hudson River Trading (HRT) has had to aggressively lease capacity across hyperscalers and specialized NeoClouds, committing to three- to five-year contracts for footprints as large as 8,000 GPUs while evaluating counterparty credit risk and even contemplating credit default swaps. While quantitative trading firms are partnering with designers like Broadcom to explore custom in-house inference silicon, Dunning emphasizes that Nvidia retains an overwhelming moat in frontier model training. Concurrently, efforts to financialize computing capacity through standardized compute futures face steep structural obstacles. Bespoke data center requirements—including networking latency, interconnectivity, and minimum trading blocks of thousands of GPUs—complicate delivery, even as volatile memory prices demonstrate the financial risk of procurement delays. Token Economics and Compounding Productivity Gaps (27:30) Daily API token expenditures on Dunning's team average between $100 and $200 per employee, occasionally surging to $1,000 per day during intensive research periods. HRT views these spikes as productive exploratory signals rather than budgetary infractions, operating on the premise that even a 50% increase in researcher output easily justifies the cost. Dunning warns that this dynamic creates a self-reinforcing divide between 'token rich' and 'token poor' organizations, where firms with the capital to subsidize heavy model usage compound their operational velocity over competitors. Nevertheless, heavy reliance on frontier models introduces distinct behavioral side effects. Dunning notes that LLM sycophancy can reinforce ungrounded ideas, citing a winter internship program where incoming interns rapidly adopted obscure, hyper-technical academic terminology generated by models that seasoned traders never used in practice. HRT systematically benchmarks new models against internal research environments rather than public metrics, finding that while tools like Claude Opus 4.5 currently match the output of a capable research intern, capability trajectories are advancing rapidly. Post-Narrative Market Modeling and Risk Controls (10:00) Quantitative finance is increasingly abandoning human-interpretable economic factors, such as traditional value or momentum narratives, in favor of multi-asset deep learning models. Dunning explains that neural networks excel at detecting statistical relationships directly from raw order book and market data across equities, futures, options, and digital assets without requiring intuitive economic explanations. Diagnostics of HRT's high-dimensional models revealed emergent capabilities, such as autonomously clustering speculative retail meme stocks and crypto equities together without fundamental inputs. Because these hyper-dimensional representations cannot be fully inspected or rationalized by human researchers, HRT maintains strict automated risk boundaries around execution. Dunning notes that hard, deterministic guardrails are comparatively straightforward to enforce in ultra-short-term market making, ensuring safety without overriding the model's empirical signal generation. Hiring Shifts from Implementation to Articulation (32:20) Recruiting dynamics between Wall Street trading desks and frontier AI research labs have shifted as major AI labs reach multi-trillion-dollar valuations and adopt standard big-tech corporate hierarchies, diminishing their equity upside and giving quantitative firms a competitive hiring opening. Meanwhile, HRT has moved toward open-book technical interviews that allow candidates to use AI freely, reflecting the reality of modern engineering workflows. Because code generation models increasingly automate routine software implementation, Dunning observes that the quant industry can pivot toward hiring theoretical thinkers and pure researchers. He describes this as a transition from 'shape rotators'—engineers focused on structural software execution—to 'wordcels,' whose core strength lies in precise linguistic articulation. Communicating requirements clearly and prompting models without introducing confounding instructions has emerged as a scarce, highly leveraged skill. While aggressive token spending delivers tangible short-term productivity gains in quantitative research, the industry confronts unresolved structural constraints: physical data center power ceilings restrict hardware deployment, bespoke facility architectures impede the creation of liquid compute hedging markets, and quant firms must determine whether custom silicon can remain cost-effective against consolidating hardware incumbents.

Hudson River Trading（HRT）人工智能主管Iain Dunning指出，量化交易机构正全面接入前沿AI模型，通过激进的Token消耗和指数级算力扩张重塑投研生态，但在物理基础设施约束与算法黑盒化背景下，行业竞争优势正加速向算力资本充沛且具备精准表达能力的机构倾斜。 模型演化与后叙事交易体系 (8:30) 在内部投研流程中，HRT系统化测试前沿基础模型对代码编写、实验构思和自动化监测的辅助能力。Dunning披露，团队在早期测试Claude Opus 4.0时曾经历过误以为触及通用人工智能（AGI）的“虚假开局”，直至Opus 4.5推出才真正带来显著的生产力跃升。为了对模型进行客观量化评估，HRT构建了专门的测试基准，让AI智能体与人类量化研究员直接对抗生成交易信号。Dunning评估认为，当前前沿模型的能力水平大致相当于一名合格的量化实习生，但其进步速度预示着未来一年的表现将极具颠覆性。 与传统依赖“价值”或“动量”等人类可理解叙事的因子投资不同，现代量化投资已全面转向由神经网络直接从原始市场动态中提取信号的“后叙事”范式。HRT在高维模型表征诊断中观察到，模型在未输入任何基本面先验标签的情况下，能够自发地将散户抱团股（Meme stocks）、加密货币相关股票等投机资产聚类在一起。尽管模型展现出复杂的高维推断能力且人类难以对其内部逻辑进行全面解释，但HRT在执行层施加了极其严苛的自动化硬性风险拦截与安全护栏，以确保高频交易环境下的绝对风控底线。 电力瓶颈、定制芯片与算力金融化 (17:35) Dunning指出，当前算力扩展的核心瓶颈已从单纯的GPU芯片采购，全面转移为数据中心电力及机房空间的极度短缺。例如在北美地区寻找具备充足电力以容纳6000张Blackwell GPU的数据中心机位极为困难，而英伟达计划于2027年交付的Rubin架构GPU预计将在早期被预订一空。面对长期需求的激增，HRT不得不放弃单一数据中心模式，在全球范围内争夺兆瓦级电力，并与新兴算力服务商（NeoClouds）及超大规模云厂商签订为期3至5年、规模达8000张GPU的长期租赁协议。为防范高杠杆算力供应商违约引发的运营中断风险，HRT甚至曾考虑购买相关企业的信用违约互换（CDS）。 在硬件架构层面，推理芯片的设计复杂度远低于训练芯片，促使HRT等量化机构组建内部专用硬件团队，并与Broadcom等半导体设计商合作定制自研推理芯片。相比之下，英伟达在模型训练领域依旧保持垄断壁垒，并通过收购低延迟推理公司Groq进一步巩固市场地位。与此同时，虽然市场开始探讨通过算力交易所（Compute Exchange）将计算容量期货化以对冲2026年出现的存储价格暴涨等供应链波动，但Dunning强调，由于网络延迟、存储配置各异以及高频机构动辄数千张GPU的大规模交割需求，算力合约的标准化与物理交割目前仍面临巨大阻碍。 “Token富人”分化与工作流异化 (27:30) 在HRT内部，员工日常调用大语言模型API的Token开销平均为每人每天100至200美元，而在高强度攻坚实验期间，单人单日消耗甚至可能激增至1000美元。HRT管理层将这种偶发的开销激增视为探索创新工作流的“AI谵妄（AI delirium）”，不仅不予惩戒，反而将其作为发现高价值工程技巧的前沿信号。Dunning认为，即使团队生产力仅获得50%的实质性提升，每日数百美元的Token成本也是极具性价比的投资。这种资金投入上的差异正在技术人员之间划出“Token富人”与“Token穷人”的深层鸿沟，能够无上限使用前沿模型的工程师将获得滚雪球式的研发效率优势，从而加剧赢家通吃的行业马太效应。 然而，重度依赖AI也催生了特定的人工智能心理副反应。由于量化金融缺乏标准化教科书，新入职实习生高度依赖模型解答疑问。全职员工随后发现，模型在交互中过度迎合用户并输出晦涩高深的学术黑话，导致实习生在日常交流中频繁使用正式员工从不使用的冷门术语，犹如在群体内部传播了一种由AI播种的“思想病毒（mind virus）”。 招聘转型：从空间构造者到文字工匠 (32:20) 随着顶尖前沿AI实验室估值迈向万亿美元门槛并逐渐蜕变为大型官僚化科技公司，其新员工的潜在财务回报有所收窄，这一文化与经济转变降低了HRT吸引顶尖工程人才的相对阻力。面对AI能够胜任初级工程师水平代码编写的现实，HRT在技术面试中全面推行“开卷考试”哲学，明确鼓励候选人直接调用AI工具，因为在真实业务中脱离AI工作的假设已不切实际。 代码编写门槛的降低从根本上重塑了量化人才的筛选偏好。量化机构历来偏爱具备极高数学造诣或弦理论背景的研究者，但以往许多顶尖理论家受制于编程实现能力而无法落地其创意。如今，代码生成模型填补了实现层面的鸿沟，使机构能够更大胆地吸纳专注构思纯理论创新的“梦想家”。Dunning将这一趋势形象地概括为从专注于工程编码落地的“空间构造者（shape rotators）”，向依赖严谨表达与精准描述需求的“文字工匠（wordcels）”转移——用无歧义的文字明确界定复杂任务的技术表达能力，已成为当下极度稀缺且价值攀升的核心生产力。 Dunning的观点描绘了量化金融在AI技术驱动下的结构性变革，但也暴露了行业深层的未决问题与潜在风险。尽管黑盒神经网络在资产模式提取上展现出强大效能，且模型对各类资产预测时间维度的拉长可能持续带来超额收益，但整个行业在算力军备竞赛中的过度资本开支、对数据中心电力的极端依赖以及模型“后叙事”特征在极端行情下的稳健性，仍是缺乏长期历史周期检验的未知数。

## Background 本期节目由彭博社（Bloomberg）旗下知名播客《Odd Lots》于 2026 年 6 月 5 日发布，主持人 Joe Weisenthal 与 Tracy Allaway 在纽约 City Winery 现场录制。节目邀请到全球顶级量化做市商 Hudson River Trading（HRT）的 AI 主管 Iain Dunning 返场对谈，深入探讨了该机构对前沿 AI 技术的最新落地实践、急剧膨胀的 Token 消耗开销、算力与电力基础设施瓶颈，以及身处技术狂飙中的“AI 诱发型精神谵妄”。 ## TL;DR 随着深度学习全面接管跨市场高频做市，HRT 正在利用统一的神经网络模型延伸至更长周期的市场预测，逐步摒弃传统量化依赖的人类叙事，同时基准评测显示前沿模型已稳达“量化实习生”水准。由于算力供应链极度紧缺，瓶颈已从单纯的芯片采购转向通电机房与电力配额，促使 HRT 搜刮一切零散算力并与高杠杆 NeoCloud 展开复杂的双向信用博弈，同时将自研硬件重心聚焦于推理芯片。此外，团队内部员工单日 Token 消耗高达 100 至 1,000 美元，巨额资本开销引发了生产力指数级拉开的“Token 贫富分化”；而在人才市场上，AI 辅助面试已成常态，能够精准向模型传达意图的严密语言表达能力正在取代基础编码成为新的核心优势。 ## Atmosphere & style 现场 350 名观众的热烈氛围与硬核深度的行业内幕碰撞，营造出极具临场感与张力的对谈气场。嘉宾 Iain Dunning 以务实工程师的严谨与自嘲幽默，频繁用“AI 诱发型谵妄”调侃自己从怀疑论者向技术信徒的滑落，与两位主播节奏紧凑的层层逼问相得益彰。整期交流在荒诞的科技幽默与冷酷的资本算计之间穿梭，让听众深刻体会到身处指数级狂飙风暴中心时，那种混杂着极度兴奋、认知眩晕与生存焦虑的时代战栗感。 ## Bulletin - `28:13` Hudson River Trading（HRT）团队日常单人 Token 消耗高达 100 至 200 美元，重度实验期甚至飙升至单日 1,000 美元；但管理层认为 AI 带来的约 50% 生产力跃升完全证明了这一新型开销的超高投资回报率。 - `18:30` AI 算力竞赛的核心瓶颈已从芯片供应全面转移至电力与数据中心机房选址。目前直接采购英伟达 Blackwell 芯片并非不可能，但市场上几乎无法找到能够承载 6,000 张高功率密度芯片并即刻通电的交钥匙机房。 - `05:00` 量化交易正步入全资产统一模型范式，同一套 AI 技术架构已能跨越美股、期货、期权及加密市场进行建模，甚至在没有基本面输入的情况下自发在高维特征空间中将散户模因股、量子计算与加密概念股聚类在一起。 - `21:28` 高频量化机构与新兴算力云（NeoCloud）的长期锁定合约正面临双向信用风险，由于部分算力分销商背负高杠杆，HRT 甚至在商务考量中精细计算是否每小时多花 10 美分购买信用违约互换（CDS）进行对冲防雷。 ## Key takeaways - `03:30` HRT 具备训练前沿模型的工程底色，但明确放弃争夺基础通用大模型最前沿，团队认为入局窗口期已过，且追赶极度消耗资本的闭源前沿并不符合其做市主业的商业回报逻辑。 - `05:00` 量化交易正在告别为各个孤立市场人工编写因子的传统模式，迈向以海量算力为基础、贯穿全球跨资产类别的统一预测架构，且模型的预测时间视界正向远端持续延伸。 - `08:30` HRT 通过客观基准与人机对抗评估模型能力，最前沿的 AI Agent 在信号挖掘与量化实验方案构建上的自主水平已稳稳达到“量化实习生”级别。 - `10:00` 面对复杂高维的深度学习模型，量化交易正在放弃传统经济学叙事与人类直觉解释；只要严格的样本外回测有效，机构已能心安理得地接受纯数据驱动的黑盒规律。 - `12:35` 高频做市商敢于部署“魔法黑盒”模型的底气在于其高频对冲与毫秒级全自动微观风控机制，能够避免单只标的数月以上的特质性长周期敞口爆仓风险。 - `20:00` 算力基础设施陷入极其紧绷的搜刮式采购状态，顶级量化机构彻底放弃挑剔心态，哪怕市场上仅释放出 1 兆瓦（MW）的电力容量也会立即全盘接纳。 - `21:28` 数千张 GPU 的 3 至 5 年期长约伴随着严重的交易对手信用博弈：运营商担忧未发债的金融机构中途违约，而量化机构则密切监控高杠杆算力商暴雷的可能性。 - `23:45` 推理芯片门槛较低促使量化基金布局专用硬件团队甚至探索自研，但在大模型训练生态上英伟达仍旧维持着不可撼动的绝对垄断地位。 - `25:40` 将算力作为期货衍生品进行标准化金融交易面临巨大实操壁垒，一方面由于各数据中心网络与存储的高度异构，另一方面量化研发存在以“数千张 GPU 起步”的极端颗粒度门槛。 - `30:00` 算力消费的分化正在拉大团队间的加速度差距，“Token 富豪”能凭借资本开销换取指数级复利加速，使资源受限的“Token 贫民”面临极其残酷的竞争劣势。 - `31:35` 随着万亿美元大模型实验室因规模膨胀出现大公司病且期权上行空间收窄，量化机构在争夺顶级算法与系统工程人才时迎来了更有利的窗口。 - `32:55` AI 接管底层代码实现打破了纯理论家“有构想无落地能力”的壁垒，量化面试转向 AI 辅助的开卷考核，能够精准严密使用自然语言表达意图的“语言掌握者（Wordcel）”正在迎来时代溢价。 ## Detailed walkthrough ### `00:00` City Winery 现场活动与“交易的未来” Odd Lots 主播 Joe Weisenthal 与 Tracy Allaway 在纽约 City Winery 举办了播客有史以来规模最大的一场线下录制活动，现场汇聚了约 350 名观众。虽然并非刻意预先规划，但整场活动自然而然地围绕着一个核心主题展开——“交易的未来”（The Future of Trading）。在当前市场环境下，AI 与金融市场的交织已成为不可回避的焦点，技术爆炸不仅重塑了交易本身的形态，也催生出各类全新标的。 此次现场秀圆了 Joe 长期以来想为嘉宾做“上下集”专访的心愿。作为当晚的第一位登台嘉宾，Hudson River Trading（HRT）的 AI 主管 Iain Dunning 继大半年前首次做客 Odd Lots 后再次受邀返场。上次访谈讨论了高频做市商如何看待与部署 AI 和 GPU；而这一次，随着算力竞赛与大模型迭代的急剧加速，讨论的基调变得更加硬核，也直接切入了技术发展带来的一线冲击与现实拷问。 --- ### `03:30` HRT 会成为下一个幻方量化（High-Flyer）吗？ 访谈伊始，Joe 抛出了一个积蓄已久的疑问：中国顶级量化私募幻方（High-Flyer）孵化并开源了 DeepSeek，凭借顶尖算力与模型工程震惊全球；那么，同为全球顶尖高频做市巨头的 Hudson River Trading，在拥有类似基础设施与技术栈的前提下，理论上是否也能推出自己的基础大模型并打造一个深度竞争对手？ Iain Dunning 坦言 HRT 确实具备相应的工程实力：团队极度擅长模型训练，手握庞大的算力储备，研发团队也非常熟悉追赶模型前沿所需的完整科研迭代流程。理论上，若全力以赴，HRT 能够追上类似早先开源模型（如 DeepSeek）的技术水平。然而，他明确表示 HRT 不会去争夺大模型的最前沿（Frontier），因为这是一项极其消耗资本的无底洞。团队内部确实私下讨论过这个可能性，但 Iain 认为他们“大概已经错过了入局的最佳窗口期”。随着美中两地大量极具竞争力的开源模型接连面世，整个开源生态已经从最初 DeepSeek 带来的“平地惊雷”转变为多方混战。当行业最前沿以前所未有的速度向前狂奔，而多数开源模型相比最顶尖闭源前沿仍有肉眼可见的差距时，后入局者要如何持续追赶，成了一个巨大的问号。 --- ### `05:00` 指数级狂飙与“AI 引起的精神谵妄” Tracy 将话题引向 Iain 在登台前最新发布的一条推文。Iain 在推特中写道：“每天都能感受到这一点，担心这是某种‘AI 引起的精神谵妄’（AI-induced delirium），但话说回来，各种实证指标看起来都呈指数级增长，所以似乎最好假设我们正在疾速冲向某种终局。”Tracy 现场打趣让他自证没有陷入谵妄，并追问他口中的“终局”（endgame）到底意味着什么。 面对台下笑声，Iain 调侃自己现在听起来太像旧金山的 AI 狂热信徒了。他回顾道，自己从 2016 年在 DeepMind 时就开始接触 AI，当时身边充斥着“技术真信徒”，而他自己性格偏向务实工程师，长年扮演着坚定的怀疑论者。但近年来，无论是外部世界的进展，还是 HRT 自身的算力扩张，都实打实地呈现出指数级曲线——明年将拥有的算力对比今年、今年对比去年，增长全部是指数级的。很多今天 HRT 正在执行的系统和能力，换在几年前他根本不敢想象。“我不是那种能一眼看穿未来的愿景家，我只是个务实的工程师，习惯一步一个脚印，但过去这一年的巨变让我不得不惊呼：天哪，这一切居然在短短一年内就成了现实。” ```mermaid flowchart LR accTitle: AI in Quant Trading Capability Shift accDescr: Comparison between past quantitative trading methods and modern unified multi-asset AI models. subgraph Past["传统量化模式"] A1["针对各市场手工编写指标"] A2["割裂的数据处理与策略逻辑"] end subgraph Modern["2026 统一模型范式"] B1["海量算力预训练与推理"] B2["单一架构打通跨市场跨品种\n(美股/期货/期权/加密)"] B3["更长周期的精准市场预测"] end Past --> Modern ``` 当被要求给出具体行业实例时，Iain 指出，最不可思议的转变在于**算力投入规模**与**技术通用性**的跃升： - **全资产统一架构**：在 2024 年还无法预料的是，到了 2026 年，同一套核心 AI 技术架构能够直接贯穿全球所有的股票、期货、加密货币和期权市场，实现跨资产类别的统一建模。 - **预测视界向远端延伸**：HRT 与同业顶级量化机构不仅在同一时间段进行天量资本投入，而且在预测市场走向的能力上，能够准确预测的未来时间周期正被推得越来越远。 这种狂飙也带来了冷酷的博弈反思：如果所有人都在以指数级速度把预测能力做得越来越强，市场是否会走向某种残酷的“高地人模式”（Highlander——“到最后只能剩下一个”），直到利润被超额资本支出压缩至零？面对 Anthropic 等头部实验室发布新模型的间隔时间被急剧压缩，Iain 感慨行业内的每个人其实都能感受到这种如坐针毡的“发烧感”（fever）。 --- ### `08:30` HRT 如何评测新模型：从“虚假 AGI”幻觉到量化实习生对抗 对于像 HRT 这样的高频交易机构，每逢市面上发布全新的前沿基础模型，团队的第一反应和评估工作流高度聚焦于两点：辅助编程以及全流程的量化科研加速（包括构思实验方案与实时监控实验运行）。 Iain 透露了团队此前经历的一次戏剧性心理过山车： 1. **Opus 4.0 的“假动作”**：去年团队在初测 Anthropic 的 Opus 4.0 时，粗略评估一度让他们产生了“技术临界点已破、感受到了 AGI（通用人工智能）降临”的狂热周（feverish week）。但深入使用后，团队大失所望，发现模型依然无法真正有意义地增强人类资深研究员的工作，留下极大的空虚感。 2. **Opus 4.5 的实质跨越**：半年后 Opus 4.5 问世，原先期望在半年前实现的能力终于真切兑现，模型的错误率肉眼可见地大幅收缩。 针对最新的模型迭代，HRT 不再仅依赖主观感受，而是在过去数周内开发了一套客观基准评测体系： - **模拟量化研究人机对抗**：过去量化研究严重依赖研究员手工构建 Alpha 信号与统计指标，现在 HRT 让 AI Agent 承担相同的任务，让模型与人类量化研究员进行同台竞技。 - **当前能力定级**：根据 HRT 的基准测试评估，目前最前沿的 AI 智能体在量化研究和信号构建上的自主表现，已经稳稳达到了“量化实习生”（intern level）的水准。 ### `10:00` 抛弃人类叙事：当“AI 谵妄”接管交易决策 在传统量化交易的漫长岁月中，策略背后往往依托着某种人类直觉或故事支撑——哪怕是简单的“便宜股票表现更好”，虽然学界和业界对价值因子的具体成因争论不休，但至少在逻辑上符合人类常识。然而，如今的市场正滑向一种“后资本主义”语境：今年夏季 IPO 的估值动辄令人匪夷所思，让人忍不住发问“到底什么是基本面？一切究竟还有何意义？”从某种冷酷的角度来看，整个公开市场早已变成一场巨型博弈，价格涨跌本质上只是资金买卖流向的体现，它与基本面完全脱节，而复杂的深度学习模型恰恰极度擅长从海量数据中挖掘这种流动性规律。 这引发了一个尖锐的问题：量化机构是否已经到了完全不需要人类叙事的节点？如果严密的样本外回测证明“以字母 P 开头的股票每逢周二表现优异”，在过去这会被视作荒谬的过拟合而直接丢弃；但在今天，人们会不会仅仅因为回测有效，就毫不犹豫地“听从 AI 的指令”砸入真金白银？Iain Dunning 坦言，这听起来像是一种疯狂的“AI 谵妄（AI delirium）”，但逻辑上确实正在变成现实。在极短的时间尺度上，人类早已承认了自己的无能——即使把订单簿盯穿，也没有任何交易员能预测下一分钟的微观价格，人们心安理得地将这一任务交给神经网络。那么，为什么这种逻辑在更长的时间跨度上就必然失效？如果模型能够吞吐所有数据并维持超越人类认知的全局上下文，人类凭什么认为自己有能力理解它？这种“失控感”固然让人不安，但事实或许在于，人类本身就不擅长数学和跨越数千种标的的复杂决策，过去某些人自诩“擅长市场预测”，可能仅仅是技术受限时代的一场暂态错觉。 ```mermaid flowchart TD accTitle: 量化交易范式转移与认知演进 accDescr: 展示从依赖人类直觉的传统量化到纯数据驱动AI黑盒决策的逻辑流转 subgraph Classical["传统量化范式"] A["市场数据"] --> B["人类直觉与经济学叙事<br/>(如价值因子)"] B --> C["回测验证与实盘"] end subgraph Modern["高维 AI 范式"] D["全维度流动性与高频数据"] --> E["深层神经网络<br/>(高维上下文建模)"] E --> F["非直觉性模式提取<br/>(纯数据规律驱动)"] end Classical -.->|"放弃人类叙事解释"| Modern ``` ### `12:35` “魔法模型”的涌现特征与高频风控防线 在谈及模型的可解释性与一度被戏称为“魔法黑盒”的算法架构时，Dunning 指出，随着团队近半年的深入探索，他们确实能通过某些特定维度的诊断切片窥见模型的内部逻辑。然而，这种理解往往是局部且带有虚幻感的。在一个极其高维且复杂的对象面前，人类只能观察到极少量的侧面。令人惊奇的是，模型在没有任何基本面输入的情况下，展现出了从第一性原理“自发理解”模因股（Meme Stocks）的涌现行为。在模型构建的高维特征空间中，量子计算概念股（Quantum stocks）与加密货币概念股（Crypto stocks）以及 Wall Street Bets 论坛的热门追捧标的，竟然被紧密地聚类在相近的向量空间中。从传统财务和基本面角度来看，这些公司毫无关联，但模型却敏锐捕捉到了它们背后高度重合的散户投机与资金流动性网络。当面对其他完全无法用人类逻辑解释的高维切片时，交易员也很难反驳——“你凭什么去推翻模型？” 然而，Hudson River Trading（HRT）之所以敢于将资金交给这些“魔法模型”，关键在于高频交易（HFT）天然具备严苛的防御纵深。高频交易本质上是一个良构问题（well-posed problem），系统只做超短周期的频繁多空仓位互换，并嵌入了全自动化的毫秒级风险熔断与仓位限制，绝不会在单只股票上承担长达数月的特质性风险。Dunning 直言，“正是因为拥有这套风控系统，我们晚上才能睡得着觉。”相比之下，他极难想象如何将这种纯粹交出控制权的黑盒 AI 逻辑复制到基本面主观多空策略（Fundamental Long/Short）中：如果你必须建仓并承担三个月以上的定向巨大风险，一旦模型出现幻觉，根本无法在放弃逻辑解释的同时提供清晰的风控闭环。 - **高维空间的“模因股”聚类**：模型在无监督学习中自发将量子计算、加密货币及散户投机标的聚为临近簇，证明其从微观订单流中学习到了底层行为模式。 - **局部解释与认知幻觉**：研究人员能够理解模型极小部分的切片，但无法完全还原高维全貌，依赖人类常识去强行修正模型往往适得其反。 - **高频对冲与长线敞口的风控差异**：高频交易依靠快速换手和全自动微观风控筑起防御底线，规避了长周期特质风险带来的爆仓隐患。 ### `18:30` 算力瓶颈转移：买得到 Blackwell 芯片，找不到通电的机房 在围绕“AI 交易”的狂欢中，资本市场始终在密切追踪产业链的下一个关键瓶颈。针对此前谈到的“芯片不是最大限制，电力与机房选址才是”的观点，Dunning 进一步阐述了当前基础设施面临的严峻现实。如果当下有机构试图挖角团队并自立门户采购硬件，他们会发现真正的阻碍在于交付周期极短的机房托管与电力接入能力。 以具体数字为例，如果有人现在进入市场，要求在第四季度交付一套部署在北美某处机房、包含 6,000 张英伟达 Blackwell 架构 GPU 的开箱即用集群，市场上在合理价格区间内几乎不存在现成方案。除非恰好有其他租户违约退租并接盘，否则纯粹寻找能够承载如此高功率密度的场地和充沛电力供应，难度极大。Dunning 澄清，现实情况是：只要手里拥有现成的电力指标和变电基础设施，今年拿到 Blackwell 芯片本身并非天方夜谭；但要想获得“芯片 + 机房 + 供电”的一体化交钥匙解决方案，几乎不可能。而这种电力与数据中心容量的紧张局面，甚至已经提前蔓延到了针对 2027 年下一代英伟达 Rubin 架构 GPU 的前期规划之中。 ### `20:00` 算力争夺战：绝不能挑剔的 GPU 采购困境 在算力供应链极度紧绷的背景下，GPU 产能几乎处于长期售罄状态。Iain Dunning 直言，哪怕放眼到 2027 年的交付周期，企业或许更容易找到数据中心机房空壳（shell），但要想尽早拿到芯片，现在就必须进入排队序列。Iain 坦言自己过去最大的误判与失败之一，就是低估了公司在长周期内对 GPU 的海量需求；这种认知上的怀疑态度带来了惩罚性的后果，导致团队不得不时刻处于“被动追赶”的状态。 这种紧缺在量化交易圈内是普遍现象。Iain 提到某家竞争对手（Show notes 提及为 Jane Street）近期在播客中表示，他们原本只规划了一个核心数据中心，但随着对算力极度饥渴，最终不得不演变成“在全球各地到处搜刮算力”。Hudson River Trading（HRT）的处境完全一致，在算力面前没有任何挑剔的资本：哪怕对方只能提供 1 兆瓦（MW）的电力容量，HRT 也会毫不犹豫地全盘接下，哪怕合同条款并不十分有利。即便 HRT 在挪威等地拥有极其庞大的数据中心，面对当下的模型需求依然远远不够。 - **远期交付排队**：即便面向 2027 年的交付排期，提前锁定 GPU 供应配额也是唯一可行的途径。 - **搜刮式采购策略**：从前集中式单一数据中心的架构已无法满足需求，只要市场上出现散落的电力或机柜额度（哪怕只有 1 MW），顶级高频量化机构都会立刻吃进。 --- ### `21:28` 与 NeoCloud 的博弈：长期合约与双向信用风险 为了筹集算力，HRT 必须与超大规模云厂商（Hyperscalers）以及新兴算力云提供商（NeoClouds）展开持续谈判。然而在这些 NeoCloud 背后，正上演着一场争夺土地空间与电力供应的隐秘暗战。由于底层资源高度稀缺，算力分销商的节奏极其紧迫，往往会突发通知 HRT：“刚刚空出一个租约，请在今天下班前确认是否签署长期承诺协议。” 与普通云计算不同，这类交易完全不是现货弹性算力（Spot compute），而是极度重资产的长期大单——例如一次性锁定 8,000 张 GPU，租期长达三年、四年甚至五年，涉及分期付款、首付半数等复杂的商业条款，并伴随着巨大的双向交易对手信用风险（Counterparty Risk）： ```mermaid flowchart LR accTitle: HRT 与 NeoCloud 的双向信用博弈 accDescr: 展示量化机构与新兴算力云提供商在长周期重资产合同中的相互信用审核与对冲机制。 subgraph NeoCloud["NeoCloud / 算力出租方"] NC1["担忧租客违约<br/>审核 HRT 债券评级"] NC2["防止电力配额闲置<br/>控制单一租户占比"] end subgraph HRT["HRT / 算力租户"] H1["数千张 GPU 长期锁定<br/>(3-5 年巨额合约)"] H2["购买 CDS 信用违约互换<br/>防范高杠杆算力商暴雷"] end H1 -->|"承诺数年付款 / 巨额首付"| NC2 NC1 -.->|"质疑非硅谷巨头背景"| H2 H2 -.->|"每小时多付 10 美分对冲风险"| NC1 ``` - **数据中心对机构的疑虑**：在纽约金融圈 HRT 妇孺皆知，但在西海岸的数据中心运营商眼中，HRT 并非家喻户晓的科技巨头。运营商会反复审查 HRT 是否发行过债券、信用评级如何，甚至不敢把单一园区的全部供电容量一次性租给 HRT，生怕万一租客倒闭，漫长的空档期将导致其无法转租。 - **HRT 对高杠杆供应商的防备**：许多 NeoCloud 本身背负着高额债务杠杆。HRT 会紧密监控这些运营商的信用违约互换（CDS），甚至在商务考量中精细计算：是否应该在每小时 3.50 美元的算力租金之外，额外支付相当于每小时 10 美分的成本购买 CDS 保险，以防算力供应商突然破产导致业务中断。 - **唯一的确定性压舱石**：整条混乱供应链中，唯有资金极其充裕、芯片源源不断的 Nvidia 是不可撼动的基石，HRT 与 Nvidia 保持着非常深厚的合作关系。 --- ### `23:45` 自研芯片的可能性：推理分化与训练垄断 在探讨硬件话语权时，主持人提问 HRT 是否能自由针对“延迟（Latency）”或“吞吐量（Throughput）”定制芯片，或者是否会走上自研硬件之路。Iain 指出，业界如今都在全力研发专用的**推理芯片（Inference chips）**，因为在工程设计上，推理是一个比大模型训练简单得多的技术命题。 HRT 及同行顶尖量化对冲基金大多早已建立了自己的专业硬件工程团队，且芯片设计流程中的部分环节可以外包（例如业界普遍传出与 Broadcom 展开深度合作）。这意味着未来很可能会出现 HRT 自主架构的专用芯片，但这类芯片仅供内部量化系统使用，绝不会对外出售。然而，芯片巨头并没有留给挑战者太多舒适区：Jensen Huang（黄仁勋）“从不睡觉”，Nvidia 收购了 Groq 资产并迅速推出了极具竞争力的新产品线，初创企业 Etched 也在特定架构上发力。 - **推理架构的百家争鸣**：推理芯片设计空间相对有限且技术门槛较低，随着第三方竞争者增加，未来不一定必须依赖全自研。 - **训练生态的绝对护城河**：模型训练领域依然是 Nvidia 绝对垄断的天下。Google 拥有出色的自研 TPU，但使用 TPU 意味着与 Google 云生态深度绑定（Vendor lock-in）。对于算力极度饥渴的前沿实验室（如 Anthropic），唯一的策略就是“照单全收”——TPU、AWS Trainium 以及 Nvidia GPU 统统并网作战。 --- ### `25:40` 算力期货与金融化：标准化难题与“以千卡为起跑线” 针对随后即将登场的嘉宾 Carmen Lee（Compute Exchange 首席执行官）试图推动的“算力金融市场”（将算力像石油一样期货化交易），Iain 表达了审慎的观点。从对冲远期风险的角度来看，算力期货确实具备理论吸引力：例如 2026 年内存价格出现不可思议的暴涨，让许多机构后悔“没有早一个月下单”，如果能有远期金融合约锁定未来算力交付价格，确实能有效对冲规划失误的成本。 但谈及短期算力期货或实物交割，Iain 认为存在巨大的实操障碍： 1. **算力资产的高度异构性**：算力无法简单归纳为一个纯粹的算力单位。当 HRT 接入一个 NeoCloud 节点时，背后涉及极其复杂的工程改造：如何将异地算力互联回自有系统、各家完全不同的网络架构、分布式文件系统的兼容性、海量数据究竟是就地存储还是存放于相邻存储节点等。 2. **极高的最小交易门槛（Lot Size）**：量化研究对于算力集群规模的要求极为苛刻。Iain 直言：“128 张 GPU 对我来说毫无意义。我的基本单位是数千张 GPU，要么成千上万，要么拉倒（thousands of GPUs or bust）。”这种庞大的颗粒度让算力很难拆解为标准化的小额金融衍生品。 --- ### `28:13` 疯狂的 Token 消耗：员工的“AI 谵妄”与量化实习生的思想病毒 关于日常业务中的算力开销，Iain 透露了 HRT 内部惊人的 Token 消耗水平： - **日常人均花销**：在其所在的技术团队中，普通员工平均每天的 Token 消耗费用在 **100 至 200 美元** 之间。 - **重度极端情况**：部分处于高强度实验状态的研究员，单人单日消耗甚至能够飙升至 **1,000 美元** 这一惊人区间。 Iain 幽默地将这种短时间内疯狂调取大模型算力的现象称为“突发实验或 AI 诱发型谵妄（AI delirium）”。管理层并不打算刻意压制这种花销，而是努力探究这些高消耗员工是否挖掘出了其他人尚未掌握的全新生产力范式。对于对冲基金而言，这完全是一项过去从未存在过的新型支出类别。 聊到通用大模型对人类的“阿谀奉承（Sycophancy）”——例如 AI 总是奉承用户“你太聪明了，离破解市场密码只差一步，再多生成一个 Token 吧，兄弟！”，或者像 Claude 总是喜欢赞叹“你这个论点确实立大功了（doing some real work）”——Iain 确认工程与量化领域同样无法幸免，并分享了一个生动的实习生案例： 量化金融领域高度私密，几乎没有教科书可以系统学习，新入职的实习生往往面临极高的认知门槛，因而极度依赖向大模型求助。在最近的冬季实习项目中，全职研究员们惊奇地发现，这群新实习生口中频繁蹦出一堆极其冷僻生硬的高级量化金融学术术语，而这些词汇在经验丰富的全职量化交易员日常沟通中根本无人使用。这正是大模型在解释量化知识时对特定词汇的异常调用频率所致，AI 仿佛在实习生脑海中悄悄播下了一颗特殊的“思想病毒之种（the seed of the mind virus）”。 ### `30:00` Token 富豪与 Token 贫民：AI 赋能下的复利加速与狂热 Hudson River Trading 在 Token 上的开销必将持续飙升，但 Iain Dunning 坚信这项投资物有所值。尽管当前未必能达到人人生产力翻倍（2x）的理想状态，但他交流过的团队中有人反馈获得了 50% 的生产力提升。在量化交易领域，为每位员工每天支付 100 美元来换取这种程度的效率飞跃，几乎是一笔毫不犹豫的划算买卖。这也引发了一个更深层的阶级分化问题：那些因预算或资源受限而沦为“Token 贫民”（token poor）的人或团队，究竟该如何与资源充沛的“Token 富豪”（token rich）竞争？ 这种资源差距带来了令人眩晕的“加速度感”。当两个天资和能力不相上下的工程师站在同一起跑线上，其中一人却配备了一个强大的 AI 副驾驶并获得 50% 的额外生产力增益，而获取该优势的唯一门槛仅仅是“花钱”，这种“有产者”与“无产者”之间的鸿沟便会以惊人的速度自我强化。团队越成功、赚到的钱越多，就越乐意毫不犹豫地将每人每天的 Token 预算追加到 1000 美元；更充沛的算力进而带来更快的迭代和更大的战果。Iain Dunning 坦言，这种指数级复利的加速感或许带有某种时代狂热带来的“谵妄”（delirium），但在赢家通吃、速度至上的激烈博弈中，它确实勾勒出了一个极其真实且残酷的现实逻辑。 - **生产力提升与成本收益**：团队反馈 AI 带来了约 50% 的生产力跃升，完全证明了每人每天 100 美元甚至更高 Token 开销的合理性。 - **贫富差距与复利加速**：充沛的资金转化成算力与 Token 优势，使得领先者能够在快节奏的研发竞赛中以指数级优势甩开对手。 --- ### `31:35` 人才争夺战的变局：VC 狂热、巨头大公司病与本科生的焦虑 随着顶级 AI 工程师成为全行业追逐的对象，顶尖人才的竞争格局也在悄然演变。Iain Dunning 指出，当前的创投圈依然弥漫着强烈的错失恐惧（FOMO）：只要创业者拥有光鲜的履历背景，哪怕只有一个雏形概念甚至没有任何成熟想法，也能轻松从风险投资机构拿到令人咋舌的高额资金。这种热潮甚至引发了一种另类的竞争——许多工程师都在暗自盘算“我为什么不直接去创办一家公司”。 ```mermaid flowchart TD accTitle: 顶尖人才流动动向图 accDescr: 展示了当前顶尖工程师在初创企业、万亿级大模型实验室与量化自营交易公司之间的流动考量 subgraph Startups ["风险投资与初创企业"] A["顶尖履历人才"] -->|"极高估值与 FOMO 资本"| B["低门槛立项创业"] end subgraph BigLabs ["前沿模型大厂"] C["万亿美元级估值实验室"] -->|"剩余上行空间有限<br/>组织迅速膨胀、大公司病"| D["核心员工选择离职"] end subgraph QuantTrading ["量化交易机构（如 HRT）"] E["更公平的竞争环境"] --> F["吸引顶尖算法与工程人才"] end D -->|"人才外流"| F ``` 与此同时，前沿大模型实验室对人才的吸引力正在发生微妙分化。其中两家头部实验室的估值已飙升至万亿美元级别，这就给求职者带来了一个现实疑问：未来的期权上行空间究竟还剩多少？此外，这些实验室在极短时间内急剧扩张，无可避免地染上了传统科技巨头的官僚习性与“大公司病”，企业文化的大幅转变促使不少核心人员开始寻找下家。这一变化让曾面临极其激烈抢人压力的量化交易公司迎来了相对均衡的竞争窗口。然而，下层的年轻群体却笼罩在悲观情绪中，Dunning 提到许多在校本科生对未来感到极度迷茫与焦虑，面对飞速进化的模型，他们对自身的职业前景缺乏安全感。 - **风投资本的狂热**：创业门槛被资本 FOMO 推得极低，仅凭名校或名企背景就能凭模糊想法拿到巨额融资。 - **万亿美元大厂的瓶颈**：随着组织急速膨胀和估值见顶，头部大模型实验室的股权吸引力与企业文化吸引力正在减退。 - **应届毕业生的焦虑**：面对模型能力迅速达到初级工程师水平的现实，大量计算机专业本科生对未来就业前景深感忧虑。 --- ### `32:55` 从“图形旋转者”到“语言掌握者”：AI 时代的面试哲学与理论家崛起 当 AI 模型已经能够熟练解决大量初级工程问题，企业对人才的定义正在发生根本性重塑。像 Anthropic 这类顶级机构的面试甚至开始要求应聘者现场编写 GPU 底层内核（GPU Kernel）或优化数据中心集群拓扑。而在 Hudson River Trading，面试理念正在向“开卷考试”全面转变。Iain Dunning 认为，假装员工在日常工作中不用 AI 是脱离现实的自欺欺人，让候选人在 AI 辅助下展现能力已成为必然趋势。 在传统的量化交易领域，一直流传着某种纯粹理论家的神话——某位精通弦理论的数学天才隐居在长岛某处，闭关推导就能带回超额收益（Alpha）。但现实中，这种理论家的成效往往好坏参半，因为再绝妙的数学想法，如果无法转化成实际代码去落地，最终也只是一纸空文。然而，随着 Claude 等高阶 AI 能够高效接管具体的代码实现工作，量化行业或许终于可以张开双臂接纳那些富有远见的“理论家”和“梦想家”。 Iain Dunning 借用网络上流行的“图形旋转者”（Shape Rotator，偏向数理几何与硬核空间工程能力）与“语言掌握者”（Wordcel，偏向文字掌控与清晰表意能力）迷因调侃道，当前这个时代很可能会迎来“语言掌握者的红利期”。尽管“提示词工程”（Prompt Engineering）在许多极客口中已被视作过时的老生常谈，但能够毫无歧义、剔除干扰因素、清晰精准地向系统阐明意图，本身就是一种极具价值且分布极不均匀的核心能力。在 AI 时代，这种严密的逻辑表达力身价倍增——尽管 Dunning 随后笑着自嘲，把自己划入这类人群或许同样也是一种 AI 引发的“谵妄”。 - **开卷面试成为常态**：顺应技术现实，允许并鼓励候选人在 AI 协同工具的辅助下进行技术与算法面试。 - **理论家门槛的瓦解**：过去“只有构想、缺乏工程实现力”的短板被 AI 抹平，纯数学理论家和直觉型研究员的价值有望被重新激活。 - **清晰表意的高溢价**：能够无混淆、严密精确地使用自然语言向 AI 传达复杂意图（“Wordcel”能力），已成为 AI 时代最稀缺的核心技能之一。 --- ### `34:30` 尾声与致谢：纽约现场录制落幕 随着时间的流逝，主持人 Joe Weisenthal 感叹这场关于算力、Token 与人才选拔的交流意犹未尽，完全能再深入探讨两个小时。在现场观众的热烈掌声中，Iain Dunning 对受邀登台表示感谢。本期节目由 Bloomberg 的 Joe Weisenthal 与 Tracy Allaway 在纽约 City Winery 现场联合主持录制，并特别致谢了 Carmen Rodriguez、Dashiell Bennett、Kael Brooks 以及 Kevin Lozano 等制作团队成员。节目最后向听众介绍了 Odd Lots 每日通讯、官方 Discord 讨论社区以及 Apple Podcasts 彭博无广告收听渠道，并简要播出了技术与医疗等赞助商信息。 ## Notable quotes > `03:30` 大概已经错过了入局的最佳窗口期。 — Iain Dunning > `05:00` 每天都能感受到这一点，担心这是某种‘AI 引起的精神谵妄’（AI-induced delirium），但话说回来，各种实证指标看起来都呈指数级增长，所以似乎最好假设我们正在疾速冲向某种终局。 — Iain Dunning > `05:00` 我不是那种能一眼看穿未来的愿景家，我只是个务实的工程师，习惯一步一个脚印，但过去这一年的巨变让我不得不惊呼：天哪，这一切居然在短短一年内就成了现实。 — Iain Dunning > `10:00` 到底什么是基本面？一切究竟还有何意义？ — Joe Weisenthal > `12:35` 你凭什么去推翻模型？ — Iain Dunning > `12:35` 正是因为拥有这套风控系统，我们晚上才能睡得着觉。 — Iain Dunning > `25:40` 128 张 GPU 对我来说毫无意义。我的基本单位是数千张 GPU，要么成千上万，要么拉倒（thousands of GPUs or bust）。 — Iain Dunning > `28:13` 你太聪明了，离破解市场密码只差一步，再多生成一个 Token 吧，兄弟！ — 大模型拟人化表达（由节目引述） > `28:13` 思想病毒之种（the seed of the mind virus）。 — Iain Dunning > `31:35` 我为什么不直接去创办一家公司。 — AI 工程师普遍心态（由 Iain Dunning 引述） ## Mentioned Hudson River Trading (HRT), Iain Dunning, Joe Weisenthal, Tracy Allaway, Odd Lots, Bloomberg, 幻方量化 (High-Flyer), DeepSeek, DeepMind, Anthropic, Claude, Opus 4.0, Opus 4.5, Nvidia (英伟达), Blackwell, Rubin, Jensen Huang (黄仁勋), Jane Street, NeoCloud, Broadcom, Groq, Etched, Google, TPU, AWS Trainium, Carmen Lee, Compute Exchange, Wall Street Bets, City Winery ## Video description <p data-renderer-start-pos="95">Today&rsquo;s episode, which was recorded at our recent live show at New York&rsquo;s City Winery, follows up on a conversation we had with Iain Dunning, head of AI at Hudson River Trading. Last year, we talked about how his firm uses AI. Now, some seven months later, we follow up on how one of the biggest market makers around is deploying this technology. We talk about the price of memory, bottlenecks in compute, how much HRT employees are actually spending on tokens, why the firm might develop its own chips, as well as AI-induced delirium.</p> <p data-renderer-start-pos="641">Read more:<br><a href="https://www.bloomberg.com/news/articles/2026-06-04/jane-street-plans-new-data-center-as-compute-power-runs-scarce?utm_medium=referral&utm_source=podcast&utm_campaign=odd_lots&utm_content=article" data-renderer-mark="true" data-is-router-link="false" data-testid="link-with-safety">Jane Street Plans New Data Center as Compute Power Runs Scarce</a><br><a href="https://www.bloomberg.com/news/articles/2026-06-04/nvidia-backed-robotics-startup-generalist-ai-valued-at-2-billion?utm_medium=referral&utm_source=podcast&utm_campaign=odd_lots&utm_content=article" data-renderer-mark="true" data-is-router-link="false" data-testid="link-with-safety">Nvidia-Backed Robotics Startup Generalist AI Valued at $2 Billion</a></p> <p data-renderer-start-pos="782">Only <a href="http://bloomberg.com/" data-testid="inline-card-resolved-view">Bloomberg - Business News, Stock Markets, Finance, Breaking & World News</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-renderer-mark="true" data-is-router-link="false" data-testid="link-with-safety">bloomberg.com/subscriptions/oddlots</a></p> <p data-renderer-start-pos="954"><a href="http://bloomberg.com/subscriptions/oddlots" data-renderer-mark="true" data-is-router-link="false" data-testid="link-with-safety">Subscribe to the Odd Lots Newsletter</a><br><strong data-renderer-mark="true">Join the conversation:</strong> <a href="https://discord.gg/oddlots" data-renderer-mark="true" data-is-router-link="false" data-testid="link-with-safety">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/02b9e5a8-91d2-4e4e-a6a7-b45f0138ad2f/audio.mp3?in_playlist=982f5071-765c-403d-969d-ae27003a8d83)

[3:30] Iain Dunning: 大概已经错过了入局的最佳窗口期。 [5:00] Iain Dunning: 每天都能感受到这一点，担心这是某种‘AI 引起的精神谵妄’（AI-induced delirium），但话说回来，各种实证指标看起来都呈指数级增长，所以似乎最好假设我们正在疾速冲向某种终局。 [5:00] Iain Dunning: 我不是那种能一眼看穿未来的愿景家，我只是个务实的工程师，习惯一步一个脚印，但过去这一年的巨变让我不得不惊呼：天哪，这一切居然在短短一年内就成了现实。 [10:00] Joe Weisenthal: 到底什么是基本面？一切究竟还有何意义？ [12:35] Iain Dunning: 你凭什么去推翻模型？ [12:35] Iain Dunning: 正是因为拥有这套风控系统，我们晚上才能睡得着觉。 [25:40] Iain Dunning: 128 张 GPU 对我来说毫无意义。我的基本单位是数千张 GPU，要么成千上万，要么拉倒（thousands of GPUs or bust）。 [28:13] 大模型拟人化表达（由节目引述）: 你太聪明了，离破解市场密码只差一步，再多生成一个 Token 吧，兄弟！ [28:13] Iain Dunning: 思想病毒之种（the seed of the mind virus）。 [31:35] AI 工程师普遍心态（由 Iain Dunning 引述）: 我为什么不直接去创办一家公司。

[10:00] Iain Dunning: There were other slices we tried to take which just didn't make sense to us. But again, it's like, who am I to say? Who are you to override the model? The model says they're in that vicinity of hyper-dimensional space. [27:30] Iain Dunning: We noticed in our winter internship program, a lot of very technical quant finance research terms being mentioned a lot by the interns that no full-timer used. It's like the original seed of the mind virus was AI.

[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. The thing about AI for business? It may not automatically fit the way your business works. At IBM, we've seen this firsthand. But by embedding AI across HR, IT, and procurement processes, we've reduced costs by millions, slash repetitive tasks, and freed thousands of hours for strategic work. Now we're helping 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 Lods podcast. I'm Joe Weisenthal. And I'm Tracy Allaway. Tracy, we did another one of our live shows this time. Our biggest show ever. Our biggest show ever. It was absolutely amazing. We did it at City Winery in New York. I think we had over 300 people in the end. Yeah, I think it was like 350 people were there. Yeah. And the crazy thing is, I think it was our sort of first themed show. And we didn't really plan it that way, but it just worked out. Right. I guess it themed and anti at the same time because we in this moment in which everything is just like AI markets markets AI etc But you know there all kinds of new things to trade and people are fascinated by the trade itself and people are fascinated by the way the technology development is affecting the trade So we really wanted to do a kind of future of trading show, which is a very broad thing, but it did sort of come out that way. Yeah, it really did. And you finally fulfilled your longtime dream of doing two part episodes. with our guests. So our first speaker of the evening was actually someone who's been on the show before. That's right. So we had him on the show last year and we had him on our live show. Listen to our episode with Ian Dunning. He is the head of AI at Hudson River Trading. Talked about all things, implementing AI, GPUs, all that stuff within the context of a trading shop. Take a listen. Joe, this is your dream, right? You finally get to do a two-part episode. This is the thing I always think about, which is that after every episode we do, I'm like, oh, there's a question I wish I had asked. So we had Ian on sometime last year. So the last round was easy. This one will be tougher. That's what I'm a little worried about. Well, I was going to start first before we talk about what you do, et cetera. So here the question I wish I had asked last time So Hudson River Trading Shop you involved in the AI stuff Could you theoretically do what Highflyer did and launch an LLM with this tech stack that you have and launch a deep competitor I think so. I think we're good at training models. We have a lot of compute, and people are good at doing the cycle of research, which is required to catch up to the sort of frontier. However, I guess reaching the frontier is clearly a very daunting task. So maybe it's with some effort, deep seek. But beyond that, it's not a claim I'd be willing to make. It's a hugely capital intensive task, clearly. Do people ever chat about that? It's like, we could do this. Oh, we think about it. I mean, I think perhaps we missed our moment to do so. There's so many open models now coming out from the US as well as China that there's a huge array of them. It's kind of an interesting shift from that deep seek moment where it felt like it was the first bolt from the blue of a competitive open model. Now I see so many groups releasing them. I don't know what the future of open models is if they're all kind of a serious step back and the frontier is progressing so fast. I don't know how you keep up with that. But many people believe that it's possible. I'm not so sure I'm one of those people though. Okay, so speaking of things moving so fast, my first question is slightly different. I looked up your Twitter feed before you came on. [5:00] on the show. Your last tweet before today was, and I quote, feel this every day, worry it's some sort of AI-induced delirium, but then again, various empirical measures are exponential looking, so it feels best to assume we're hurtling towards some sort of endgame. So first of all, please convince us all live on stage that you are in fact not suffering from AI-induced delirium, but secondly, what is the endgame that you speak of here? God, now I sound like a San Francisco person. I've been doing AI stuff since around 2016, and that started at DeepMind. And it was a bit of a culture shock for me because there are true believers then, and I was most certainly not a true believer. And I resisted it, and I was kind of a natural skeptic for a long time. But certain empirical measures of the pace of progress in the outside world, and I also look at our own business, which looks somewhat exponentially Usually, the amount of compute I'll have next year versus this year, and the amount of compute I have this year versus last year, looks kind of exponentially. And we doing things today that I didn really I should have dreamt of I wish I had that kind of visionary say I a visionary and I can see the future and I building towards it But no I think I a pragmatic engineering archetype and so it been very incremental And I like wow that happened in a year So what does this mean? It's some sort of technological convergence, everything going faster all the time? Well, give us an example. Galerian, probably. Give us an example then, because obviously those of us using just the regular models, obviously the improvements in capabilities from one year to another are mind-blowing. But from the perspective of like, okay, the application of AI within the trading context, what is something that you can do in 2026 that in, say, 2024, you would not have been able to anticipate? Oh, I think it's one way to think of just the amount of compute going into both training a model and running a model and that it's the same technology working across every equity, every future, every crypto market, every option market across the world with a kind of unified approach. and this is something that we're doing. But even more interestingly, I would not claim that we are some unique people who are only the ones who have really made progress in AI and trading. I think many of our peers are also investing massively and we're all doing it all at the same time. What does that mean? Surely you can just keep getting better at predicting markets forever There gotta be some sort of where you know your margins go to zero as you keep investing more What you saying is you are just getting better and better at being able to predict where a market is going to go further and further out in the time frame. Basically. That's cool. And we're not the only ones. So in the end, can there be some Highlander type thing? Like, what are we doing? And this is like my scale. And I guess the other thing I find interesting is, of course, the scale that everyone can see with the big labs and what they're doing of compute. And it's like, it looks awfully exponential to me. We just had another model released today from Anthropic and the spacing between them seems to be compressing. I don't know. I do sound hilarious. I sound feverish. And that's why it's so hard. Literally everyone in this room probably is talking. Everyone must feel a fever to some extent. Yeah. I never understood the Highlander, there can only be one thing. Because there are already two. They could just coexist. That's right. Anyway, sorry, I'm just picking apart your analogy. You mentioned a new model release. When a new model gets released, what is the first thing you do at Hudson River Trading to evaluate it? And how do you actually compare them to the existing one? So I mean, our primary use cases at Hudson River Trading are definitely kind of just accelerating your own research So that can be coding but it can also be coming up with experiment ideas monitoring experiments We had a sort of a false start with AI I would say sometime last year with the Opus 4.0 models, especially from Anthropic, where a cursory examination made us feel like, well, this is the moment we've crossed the dividing line. And we had a very feverish week where we felt the AGI and we left feeling empty because we realized that it was not there and was not able to meaningfully augment human researchers. And then we had that same feeling again when Opus 4.5 came out, and suddenly it was like, oh, wait, no, this is actually what we thought it was going to be six months ago. So in the most recent model releases, the differences have been more subtle, but we see, I think we have a much better sense of an ever-reducing set of errors they make, and so we're kind of looking for those sorts of mistakes. And we spent some time in the past couple of weeks trying to come up with objective measures to index them against humans in the act of quant research, ideating signals and things. Quant research used to be, as we talked about a little bit, like handcrafting indicators and things. Why not ask AI agents to do that and compare them against humans like a little sort of battle? And it's like, I don't know, intern level. [10:00] I, perhaps. The thing is, what do I think it'll be in a year? I would not want to make a bold claim, but it'll still be. It'll be wild. So when we think about investing in general, even within sort of like classical quant trading going back decades, there is often, it might be quant, but there's some intuition behind it, right? Cheap stocks tend to do better, and we don't actually totally have agreement why they did for a while, but people aren't necessarily surprised by that fact, right? Are we at the point where it's like, why even bother coming up with a human intuitive story? And you just skip the part of giving an explanation that sounds logical to a person. And it's just basically pure, like rigorous backtesting. And then it's like, look, here is something that seems to work. And we've backtested it a million different ways and it seems to work. And we don't even bother coming up with a story for why, but we're going to trade it. I feel like we're in that world today. It's sort of post, post, post capitalism. And when I see IPOs discussed for this coming summer at the valuations they are, I'm like, what is a fundamental? Like, what is anything? It feels like markets are just... The cynical thing is everything is gambling. And so everything is some sort of like gambling market including public markets But the joke is It flows it buying and selling and it worth what it worth and it detached and more buyers and sellers price go up and models are excellent at pulling that out of data. But just let's say the classic example of a backtest is companies with the ticker symbol that starts with P, they do well on Tuesdays. And it's like, well, look, the data says that, but this makes no sense. We're not going to trade that. Could it get to the point where it's like, look, ticker symbols that starts with P do well on Tuesdays and we've run this a bunch of times and it seems to work so we're going to put money behind this. Just do what the AI says. I feel like yes, although it sounds crazy when it sounds like AI delirium when I say it. But I feel like there's some sense that that could be true. But at some point I can't predict. So at the very short time scale, people accept this already, right? I can't tell you the price of a stock in a minute and no one would reasonably expect any human to do so even if they had the order book and spent all the time in the world staring at it. But we accept that neural networks can do this. And then when does that logic break down? Why should it break down at some long timescale if it's ingesting all the data and has everything and it can keep it all in a context in a way human can't? Why should I be able to understand it? And that is a strange thought A loss of control it feels like a loss of control but it arguably you know people save us from math Maybe humans are actually very bad at math so it not surprising AI is much better than humans at these like math proofs Humans probably would be pretty bad at markets where thousands of tradable instruments on, like, very long time scales. We just kind of accepted that we were, some people were good at this. Maybe that's a, it was a temporary state of affairs. Well, we talked about this the last time You were on the idea that the models themselves are not very interpretable, I guess you would say, but you're comfortable with that on a short trading time frame, which is what you do. And then we started joking about magic models and magic is a dangerous word to use on this podcast because people start thinking about magic boxes. But anyway, now that you've been doing this for another six months since we last spoke to you, do you feel like you have better insight into what the models are actually doing and why they're able to succeed on short timeframes? I do think there are diagnostics we've done where we can see things that we can understand. It's like looking at some very, very complex thing and you can look at one facet of it and be like, this is a facet I understand, and that gives you some confidence. But it might be illusory because it a very very complex object And if you only taking slices through it to understand aspects of it we had this emergent phenomenon we saw where it felt like the model kind of understood meme stocks from first principles, like quantum stocks and crypto stocks being kind of adjacent in stock space. And, of course, from a fundamental perspective, this has no meaning to it. But we looked at the model in a certain lens, and it clearly felt like they knew they were connected. And there were some other actual companies that I probably won't name because I feel like it's bad form. But, you know, Wall Street Bets favorites, I guess. And they were near the cluster too. And this was like just one little window. But there were other slices we tried to take which just didn't make sense to us. But again, it's like, who am I to say? Who are you to override the model? The model says they're in that vicinity of hyper-dimensional space. Yeah. One thing for us, though, is that when we do have this magical model, It is in a lot of safety around it because we're doing this higher frequency trading. We're trading positions back and forth. There's a lot of risk checks that are fully automated and things. I don't know how you generalize this logic to long-term discretionary trading where the idea of risk checking and that kind of layer of defense around it, it's not so obvious to me how you apply that. We can apply very strict control. [15:00] around this model because it's a well-posed problem. We're not taking giant idiosyncratic risks in one name for months at a time. We can sleep at night because of this. I don't know how you apply the same thinking to a fundamental long short thing where you have to put a trade on and it's for three months and you're intentionally taking a very large risk in a very certain direction. What's the risk management story around the AI if you just give up all control to just the magic prediction? Thank you. 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 really watching the bottom line you find standout choices at every price point so you can make a real impact while staying on budget Plus you get expert 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. Support for this show comes from Public.com. If you're actively involved in your portfolio, you probably catch yourself repeating the same actions. Buying the dip, manually sweeping idle cash, putting on a hedge. On public, you can now create AI agents that handle all these tasks on your behalf. Just describe what you want to do in plain English. Like, if the VIX hits 25, buy a put option on the S&P 500. Or, if my cash balance goes above $20,000, move the excess into my direct index. You approve the workflow and your agent handles the rest. Monitoring the market, watching for your conditions, and executing your strategies exactly as defined. An investing platform driven by your intent, not just your clicks. You can also get full read and write access to your account via the public API. Go to public slash market and fund your account in five minutes or less That public slash market Paid for by Public Investing Brokered services by Open to the Public Investing Inc member FINRA and SIPC Advisory services by Public Advisors, LLC, SEC Registered Advisor. Complete disclosures available at public.com slash disclosures. Let's talk about health care for a second. It doesn't always work the way people expect it to. If you've ever waited on a prescription refill or had a hard time getting the care you needed, you know the feeling. The system should just work better for everyone. That's exactly what the people at Optum are trying to do every day. They're a healthcare company linking patient care and pharmacy services and using data and technology to drive the whole system so care is connected, not complicated, for patients and providers. Things like making it easier to get care that looks at the whole person, from primary care doctors to mental health support and even in-home care, and then using technology to make sure they all work together. technology designed to help doctors spend less time on busy work and more time with their patients. And those prescriptions, Optum is working to bring costs down, save patients money, and make it easier to get refills. Little by little, Optum is helping make healthcare work as one for everyone. Head to business to see how So you said something the last time we interviewed you which is very important First of all I feel like in the quote AI trade unquote people are obsessed with like what's the bottleneck now, right? And because whatever the bottleneck is, you probably sell it for a lot more money. You said on the last time we talked to you, the chips themselves were not actually a major constraint for you, and that it was more like citing the chips and the powering the chips, the access to electricity. Talk about that. What is the state right now? Let's say I assemble, I poach a bunch of people from Hudson River Trading. I get a bunch of GPUs. Is it then not trivial to find a place to plug those in? It's definitely hard to find sites at short lead times. If I went to the market and said I want 6,000 Blackwell GPUs in a box somewhere in North America for delivery in Q4, I'm not sure such an offering exists at any reasonable price. Maybe someone will give up a lease and I could snag it, but I think if I went to the market and tried to get a quote on that... Wait, sorry, just to be clear, the chips are available but not the capacity to run it? I think if I had power, I could get the chips, Blackwell chips for delivery this year, but I do not think I could get the whole solution. And then if you go into 2027, for the next generation of GPUs, the Rubin GPUs, they, at least... [20:00] first, like, stretch are going to be very much sold out. And so I think that's a good... Maybe you actually have, on a 2027 delivery, you have more luck finding a data center shell by then. But you need to be in queue now for those GPUs if you want them early. So those things are in demand. I'll say that for sure. And one of my greatest failures has been, you know, part of my skepticism has been predicting how many GPUs we would need on a long enough horizon. And it's punishing because you're constantly playing catch-up. And one of our competitors put out a podcast this weekend, and they mentioned something along the lines of the fact they had one data center, and it was the data center, and that was their data center. And then as they're hungry and hungry for more compute, they had to go out and find it wherever they could. And I would say we are in exactly the same boat. You just can't be picky. It's like you've got like a megawatt there. I'll take it. And it could be not in terms that are super favorable to you because – Well, say more about that. How are you actually going out and sourcing this stuff? because as you say, it seems to be exceptionally competitive. And at the same time, don't you guys have an insane data center in like Norway or something? And it's not enough. Yeah. And it's not enough. Yes we go to the NeoClouds the hyperscalers everyone and it a constant dialogue And they all in competition with each other but in some sense there must be some much bigger shadowy competition going on behind the scenes behind these NeoClouds because they are all looking for space and power. And I don't know if that's the true scarce resource and they have a kind of intermediary layer over it. I don't know what their process is like for sourcing it. But yeah, they have come to us and said, this lease opened up, can you please get back to us by the end of the day? and for commitment on a long-term contract. And our contracts are long-term. This is not spot compute. This is like 8,000 GPUs for three years, four years, five years payment. Do you want to pay half up front? Do you want to pay some per year? A lot of different commercial terms, credit risk on both sides. It's complicated stuff. Tell us more about the counterparty risk. So it's like you come and you say you want capacity in some data center. I'm Ian from Hudson River Trading. Who? Yeah. Well, this is the kind of thing. In this crowd, a lot of people know what Hudson River Trading is, but maybe in San Francisco or whatever, that's not a household name, et cetera. They want to know for sure that you're going to be good, you're going to pay your bills, et cetera. How do you establish to the data center that you are going to be a reliable, I guess, tenant? Yeah it definitely been a dance It getting better at this point I think we entered enough deals with enough people that I think we have that But we had everything from people being like oh you issued bonds what the rating on those to not wanting us to sell too much of one site. Because if we take all their power rights, and then go bust, they might have a long lead time where they can't get another tenant and fill that. And so there's a kind of two party problem to this where it's like they want customers, but there's presumably a lot of customers, but maybe not as many customers are willing to do the big size and pay more up front. But we're looking at their CDSs on some of these ones and thinking about how that affects our, maybe we should pay you $3.50 an hour and take out a CDS for 10 cents per hour equivalent of insurance on your heavy leveraged NeoCloud. You're having a disruption, no names. But, you know, I think there's reason to be cagey on both sides because this has all come from nothing like a year ago. We weren't there asking for it, and they didn't exist to sell it. And so the only rock is NVIDIA, I guess, an extremely well-capitalized entity who is not going anywhere and is making a lot of GPUs. And we have a very positive relationship with them and I think that is also a material factor How much optionality do you actually have on GPUs now If you say you want to prioritize latency or throughput can you get the chips that you need to specify on one of those things or do you just take what you can get? Or build your own. Yes, say more. Well, many people, I guess, now are working on building their own chips for inference, which is a strictly simpler technological problem. And ourselves and many of our peer trading firms have hardware teams to tackle this. And you can outsource parts of that process, so it's not as daunting as it seems. But it's definitely an active area of investigation for us. And now clearly everyone, because I feel like everyone's talking about their partnership with Broadcom or something like this. And if someone says partnering with Broadcom, it's like they're making an inference chip. So that's interesting, right, because you hear about like Amazon, like they've got Tranium, right? Tranium, yeah. Google, TPU, et cetera. So we could be in a world in which we hear of a Hudson River Trading branded chip one. I don't think we'll sell it, but yes, you're right. That is definitely the right model. And on the other hand, Jensen never sleeps. And Jensen purchased Grock, and they've got their new product lineup from the Grock acquisition, which is a very compelling product as well. Yeah. [25:00] setups etched comes to mind so the inference space is a smaller design space it's not clear that in-house solutions will be a necessary thing in the future if there's enough people competing but on the training side i mean what a moat like there's just in video it's just and i suppose google but you know if you are using tpus you're also kind of entering a very close relationship with google some feeling of vendor lock-in it's it's a complicated thing if you go down that path but if you're compute hungry like the neo labs i mean i've just the the big labs you'll take what you can get i think antropoc takes tpus traniums and gpus you know they need them all so maybe we'll create a little bit of controversy here because later on in a little bit we're going to be speaking with carmen lee the ceo of compute exchange which is this you know one of multiple entities are trying to build financial markets for compute capacity right trade it like oil so it's like compute futures and stuff that's right now could you see a use for that a financial instrument that like on some liquid tradable exchange for h100 or whatever some benchmark of how much it costs to run these chips could you see that being a useful instrument for you at some point It plausible and it partly relates back to my previously stated failure to plan correctly for the future If in some sense I could lock in a price for some future date for delivery or something of compute, something that is connected to a practice of compute in the long-term future, I think there could be value to that. We could basically hedge our risk that we wait too long to put the order in and the price goes up. I mean, in 2026, the price of memory has gone up so much that we do have concrete specific things. I wish I put that order in a month earlier. So it's a real thing. Do I believe that there would be a good market with less liquidity for long dated compute futures? That, I guess, remains to be seen. I don't know what I would do with a short dated compute future. I do think defining what compute is is pretty hard. And I have no idea what physical delivery would be. if that is indeed of interest because of the long-term contracts and because of how much work goes into every site. When we connect to a NeoCloud site, we're thinking about how to connect it back to our other sites. Everyone's got a different networking system, the file system. There's a GPU which is all the focus but there also how is data stored at that site Or is it stored at that site at all Is there an adjacent site that all the hard drives are in And they all idiosyncratic and I can do anything of 128 GPUs I need thousands of GPUs or bust. That's like my lot size. And so it's very hard to see how you could kind of break that down into useful units. But maybe it's just a spot thing, and if it's long-dated, I don't know. We'll learn more at 7.21. Yeah. I did get a preview, and it is pretty cool. like the actual program where you can select like the type of compute you need from a specific data center that has like literally I think dozens if not maybe hundreds of parameters at the moment so maybe we can get a demonstration from Carmen what is your token spend oh yeah at the moment is it bigger than Joe's I hope so I think I what is my average I think it's on the order of 100 200 a day per employee per that's sort of on my team I feel like that's kind of what I've been seeing lately, and some people are more in a thousand a day range. Wow. A bit bursty for that. Wait, do you like those people? Because they're supposedly more productive, or are you telling them to rain it in? No, we're definitely not trying to encourage that. I mean some people go through surges of experimentation slash AI delirium which is understandable And I think we are always trying to understand if the people who are using more are they doing it for something that you haven figured out yet That a pretty profound new expense to have It's not at the level that concerns us. Well, it didn't exist at all as an expense type, so that's kind of interesting to think about. Well, I'm curious. For the consumer models, they talk about how psychophantic they are. Does that happen? It's like, yes, Ian, you're close. This is really smart. You're close to cracking the code of the market. Keep pursuing this. Just one more token, bro. This idea is doing, or as Claude likes to say, this is doing some real work here in this argument. It's really good. Do you get that in the engineering context? I think we do, and it's interesting. We just started a new internship for the summer, and in previous internships, we noticed that it's quite daunting coming into this quant trading context. There's not much to read a book. You can't read a textbook about it. It's useful. So people ask AI. and it always mentions some things of an unusual frequency that maybe an expert in that field wouldn't focus on some things. We noticed in our winter internship program, a lot of very technical quant finance research terms being mentioned a lot by the interns that no full-timer used. It's like the original seed of the mind virus was AI. So there's a [30:00] Stuff like that. But our token spending is going to go up. That's almost guaranteed. And yeah, we're getting value out of it. Maybe not 2x productivity, but I talked to someone who said their team is 50% more productive. That's pretty good. I mean, you'd happily pay $100 a day for that. I just don't understand how people who are token poor could keep up with someone who's token rich. And that, again, goes to the acceleration feeling. It's like if you have two people who are sort of equally resourceful and smart, but someone has basically a co-pilot with them that's giving them a 50% boost, and all they have to do to get that is essentially spend money, it creates a have-have-not dynamic that possibly compounds. As you have more success, you make more money, you're more willing to eat now $1,000 a day per person for token spend. you go even faster and this feeling again of compounding acceleration which might be delirium but you could make an argument for why it could be a real effect instead of more winner-take-all contexts where speed of improvement is like the key thing there's a story there I think or delirium I don't know well I mean speaking of the haves and have-nots the other big story in AI world is just competition for talent yes and everyone is sort of chasing the same genius engineer I guess How are you finding that at the moment It changed a bit There a lot of dynamics going on. There's still a feeling that if you are plucky enough, you can get a VC to fund your idea based on very little. You have the right pedigree. And that's always been true, I guess, in some sense. This is like the YC philosophy, some sense you go and it's just that some of the numbers and And the FOMO feeling is quite shocking. And so that's actually a form of competition, just like, why didn't I go create a startup? I don't have any ideas or anything. I'm just going to make a startup. For the big labs, the question of upside, remaining upside, now I guess there are two big ones that are a trillion-dollar valuation. Where do you go from there? I think that's affecting people's level of forward-looking optimism for people who are taking offers now. and for people who are at those places and looking to leave generally it's a question of well they've become big tech they've added people at a vast rate and the culture has shifted especially at some of the labs a lot and to our favor for a while it did feel like we were in a very very fierce competition and now maybe it's a more even playing field but I don't know I talk to a lot of undergrads and they don feel great about the future they feel very worried This is what I was going to This went by way too fast but you mentioned already the models are at junior level Yeah. So what does talent look like at this point? I've seen some of the anthropic interview questions and it's like designing some GPU kernel or optimizing the configuration of GPUs within the data center. What do you want someone to bring to the table at this point? I mean, I think the first thing is just trying to embrace an open book philosophy, like let the interviews be done with the aid of AI. It's something we're trying to aspire to do because it's just at some point you become it becomes unrealistic to pretend anyone would work without that. One of the big things in quantum is being like there's this like archetype of like the math theorist or the string theorist or something. And they go in to Long Island somewhere and they come out with alpha. but you know like our our experience has been a little bit more mixed because it's like if you can't implement your ideas how do you how does that happen exactly well now claude can presumably implement the ideas so trying to embrace that maybe we do accept more theorists more dreamers people who can come up with ideas trusting that the implementation work can be done by ai so i think that our shift but i been joking it like the word cell versus shape rotator type like I feel like error of the word cell may be a bonus Good Good Yeah I mean prompt engineering is kind of a boomer term at this point, but there is something to be said for describing what you want clearly and without confounding factors. And that is a skill that can be learned and is not evenly distributed in the population. And I would argue that it's shot up in value simply because of AI. So I like to think of myself as one of these people, though. So that could be the delirium talking. I don't know. All right. Ian Dunning, we could talk for two more hours. Thank you for having me. Thank you so much for joining us. Thank you. That was our conversation with Ian Dunning of Hudson River Trading recorded live at our New York show. I'm Tracy Allaway. You can follow me at Tracy Allaway. And I'm Joe Weisenthal. You can follow me at The Stalwart. Follow Ian at Ian Dunning. Follow our producers, Carmen Rodriguez at Carmen Armin, Dashiell Bennett at Dashbot, Kale Brooks at Kale Brooks, and Kevin Lozano at Kevin Lloyd Lozano. And for more OddLots content, go to Bloomberg.com slash OddLots. We have a daily newsletter and [35:00] our episodes. And you can chat about all these topics 24-7 in our Discord, discord.gg slash oddlots. And if you enjoy Oddlots, if you like it when we do these live shows and talk about how trading firms are actually using AI, then please leave us a positive review on your favorite podcast platform. And remember, if you are a Bloomberg subscriber, you can listen to all of our episodes absolutely ad-free. All you need to do is find the Bloomberg channel on Apple Podcasts and follow the instructions there. Thanks for listening. Thank you. Thank you tech and totes it 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. Healthcare doesn't always work great. If you've ever waited on a refill or couldn't schedule an appointment, you get it. That's the kind of stuff Optum is changing. They're using data and technology to integrate patient care, pharmacy, and everything else. So healthcare is connected, not complicated. What's that look like? Cheaper prescriptions that are easier to get and care that looks at the whole person, how you need it. Optum is helping make healthcare work as one for everyone. Learn more at business.optum.com. As industries evolve faster than ever, companies need an environment that accelerates strategic growth, and Michigan delivers on that promise. From emerging startups to global enterprises, Michigan offers what executives value most, a resilient, innovative ecosystem, diverse communities that attract top talent, and a quality of life that supports work-life balance. With our unified Team Michigan approach, businesses scale faster and compete at the highest level. Michigan, pure opportunity. Seize your opportunity at michiganbusiness.org.