Man Group Deploys AI Quants as Token Consumption Soars 86-Fold

Token消耗激增86倍：Man Group部署AI量化模型实盘交易 随着智能体工作负载激增并将Token预算转化为直接运营驱动力，Man Group正积极利用自主生成的AI量化模型交易客户资金。 Man Group高管Gary Collier和Tushara Fernando详细阐述了该对冲基金如何在主观交易与量化交易部门整合生成式AI和智能体工作流。讨论涵盖了多智能体系统的部署——该系统能够自主研究、编写代码并回测定量交易模型（其中已有15至20个模型正在积极管理客户资金），以及为这些模型供给所需的多层数据架构。此外，他们还探讨了自1月以来内部Token消耗暴增86倍对组织带来的影响、员工角色从手动编程者向系统编排者的转变，以及该机构在严监管市场环境中维持严格风险管理、可审计性与知识产权控制的具体举措。 全球大型对冲基金英仕曼集团（Man Group）正积极将生成式 AI 深度融入全谱系主动资管业务，从单纯的统计预测建模转向赋能员工、多模态数据合成与智能体（Agent）工作流。 (5:25) 英仕曼集团耗时 18 个月研发的多智能体系统已成功自主完成文献扫描、提出经济假设、编写代码及回测；目前已有 15 至 20 个完全由 AI 构思的量化策略通过投资委员会审核并实盘运作。 (14:14) 在数据架构上，该机构将数据分为高频 Tick 数据、另类非结构化数据以及沉淀内部方法的制度化知识层；实践表明对数据进行预标注与元数据结构化的投资回报率显著高于单纯微调基础模型。 (20:00) 自年初至 2026 年中，英仕曼集团的 Token 消耗量激增 86 倍，技术、财务与人力等多部门均在广泛应用；为防止中心化瓶颈并鼓励创新，机构已将 Token 预算下放至各业务单元自主调配。 (30:00)

Executives at Man Group report that operationalizing generative AI and multi-agent workflows has driven an 86-fold surge in token consumption, expanding automated quantitative strategy generation while augmenting discretionary fundamental research. Autonomous Strategy Generation and Research Augmentation (14:14) Across its trading desks, Man Group utilizes generative AI to ingest massive unstructured datasets and automate quantitative research pipelines. For discretionary managers, specialized agents monitor alternative data, podcasts, and research reports asynchronously—such as synthesizing a cloud executive's technical interview to identify data center interconnect capacity as a key hardware bottleneck—delivering real-time alerts without replacing human judgment. In systematic quantitative research, the firm runs an end-to-end multi-agent platform developed over 18 months. Agents scan academic literature, formulate economic hypotheses, write code, run backtests, and critique results. CTO Gary Collier confirmed that 15 to 20 quantitative trading models conceived entirely by this AI pipeline have passed human investment committee review and are actively managing client capital. Tiered Data Infrastructure and Auditable Models (20:00) Deploying AI within a regulated fiduciary environment demands strict data structuring and explainability. Man Group processes nearly a terabyte of raw exchange tick data daily across three architectural tiers: high-frequency market data, unstructured alternative datasets, and an institutional context layer that encodes internal processes and proprietary terminology. Head of Data and AI Tushara Fernando emphasized that pre-tagging data with rich semantic metadata yields a higher return on investment than fine-tuning raw base models. To prevent the black-box risks common in consumer AI, agents are programmed to state a plain-English investment thesis before generating any code or executing models. Because Man Group trades across medium- and long-term horizons spanning days to months rather than sub-millisecond high-frequency windows, all trade rationales remain fully auditable and defensible to risk committees and regulators. Managing an 86-Fold Surge in Tokens (30:00) Between January and mid-2026, Man Group experienced an 86-fold increase in token spend as agentic harnesses allowed non-technical staff across finance, operations, and human resources to automate complex, multi-step tasks equivalent to up to 16 hours of human labor. Rather than restricting use through centralized automated classifiers, the firm federated explicit token budgets to individual business units, allowing department heads to reallocate operational budgets for additional tokens. Internal analysis of user behavior led the firm to educate its roughly 1,800 employees on context window management and prompt efficiency, including executing version-control commands outside inference loops to save tokens. As autonomous background agents drive an increasing share of usage, the firm faces a broader shift from tracking individual human consumption to budgeting for automated business workflows. Conductor Talent and Preserving Investment Alpha (40:00) The rapid democratization of coding tools has shifted hiring criteria across the firm. Rather than prioritizing technologists focused on manual code debugging, Man Group looks for strategic 'conductors' capable of orchestrating complex multi-agent workflows. The firm deploys platform engineers directly onto trading floors alongside investment professionals, codifying standard research processes into firm-wide AI playbooks while allowing fundamental portfolio managers to retain the proprietary 10% of their workflow that represents their personal edge. Addressing whether AI will commoditize market edge into a race to the bottom, Fernando argued that algorithms and datasets alone do not produce alpha. Because individual quantitative features rapidly decay into generic risk factors, sustainable edge continues to rely on the firm's broader ecosystem, including proprietary broker relationships, specialized data pipelines, and execution infrastructure. While Man Group has expanded AI usage without resorting to strict token austerity, asset managers face ongoing uncertainty regarding catastrophic agent errors and determining when exponential token spending must directly offset headcount or translate into measurable alpha.

全球大型对冲基金英仕曼集团（Man Group）高管指出，该机构的Token消耗量在数月内激增86倍，生成式AI与多智能体系统已从实验性研究深入至自主量化建模与实盘资金管理。 多智能体量化研究体系与实盘落地 (14:14) 英仕曼集团首席技术官Gary Collier透露，公司已开发自主量化研究系统逾18个月。该系统依托多智能体架构运行：前端智能体负责阅读金融文献、学术论文与标注数据集并提炼经济假说，后续代码智能体自动编写逻辑、抓取市场数据并回测，最后由评估智能体对结果进行批判与交叉验证。截至访谈时，已有15至20个完全由AI构想的量化模型通过了人工投资委员会的审查与验证，正实际管理着客户资产。 在主观基本面投资方面，AI则充当信息倍增器。例如，系统曾自动转录并提炼大型云服务商技术主管在小众播客中的深度观点，帮助主观投资经理厘清AI硬件产业链的瓶颈不仅在于GPU，更在于大型数据中心的占地空间以及设施间的高速互联网络。目前最主要的瓶颈并非单纯的算力或数据，而是人类委员会消化、审核并治理海量AI生成策略的处理能力。 分层数据架构与上下文治理 (20:00) 英仕曼集团数据与AI主管Tushara Fernando将底层数据架构分为三层：第一层是全球交易所每日产生近1TB的原始逐笔高频行情数据；第二层是结构混乱的另类及非结构化文本，需建立统一知识层来映射代码、行业与公司；第三层则是涵盖机构专属语言与回测规范的“上下文层”。机构实践发现，直接向基础前沿大模型灌入信用卡等未加工数据难以捕捉微观差异，相比耗费资源做全量微调，对数据添加清晰标签、元数据描述与业务规则语义层能带来更高的投资回报率。 由于对冲基金受严格的受托人责任与合规要求约束，系统设计严禁出现消费级AI常见的误操作或不可控错误。因为持仓周期通常介于数天、数周至数月，而非亚毫秒级高频交易，AI智能体被硬性要求在编写任何策略代码前，必须先用清晰通顺的英文完整写明投资逻辑与信号假设，以确保所有决策均能接受风险委员会与监管机构的严格审计。 Token消耗激增86倍与自主预算转移 (25:00) 自2026年1月至年中，英仕曼集团整体Token支出激增86倍，这一增长并非仅局限于量化技术团队，还迅速扩散至财务、运营与人力资源等部门。Fernando援引SWE-bench与METR等评测基准指出，智能体可自主执行的复杂任务周期大约每7个月翻倍一次，目前已能处理相当于人类16小时工作量的多步骤任务，促使员工直接利用自动化脚手架构建整套应用。 在管理约1,800名不同技术背景员工的消耗时，管理层并未采取严格的Token紧缩政策或强制流量路由，而是将Token预算分发至各业务部门，允许主管自行平衡内部营运支出。技术人员也自发总结降本技巧，例如在智能体循环外部拦截并处理基础Git版本控制指令，避免将冗长输出回灌至上下文窗口中浪费Token。同时，Token支出的主体正逐步从人工提示词交互，转向全天候在后台自主运行的自动化智能体工作流。 人机协同重塑人才定位与阿尔法护城河 (40:00) 随着代码编写门槛大幅降低，英仕曼集团对技术人才的考核标准正在发生改变。传统专注于代码底层纠错的工程师需要转向成为多智能体系统的“指挥家”与架构师。同时，非结构化数据的处理能力使系统化策略得以提前切入过往高度依赖人工协商的非流动性市场（如结构化信贷或小众新兴市场）。在机构协作上，该机构长期推行工程师与主观交易员同场办公的模式；基本面投资人通常乐意将约90%的通用研报梳理与回测流程沉淀为机构共享的AI手册，但仍会保留最核心10%的个人直觉壁垒。 面对行业内“AI阿尔法是否会因技术普及而被迅速套利抹平”的质疑，Fernando强调，单一的大模型或开源数据本身无法形成持久的超额收益。AI挖掘出的表层因子会极快演变为公开的通用风险因子，真正的超额收益依旧根植于包括券商网络、订单执行架构、专有数据管线以及深度回测体系在内的完整机构生态系统。 英仕曼集团的实践表明，金融机构落地AI的核心痛点正从基础算力与模型微调，转移至业务上下文标注、合规可解释性以及人类对海量自主策略的治理吞吐量上；随着Token支出呈指数级攀升且逐渐由自主智能体主导，市场与管理层未来势必将密切关注此类支出能否明确转化为经风险调整后的超额收益或人员成本的实质优化。

## Background 本期内容来自彭博社（Bloomberg）旗下的知名播客节目《Odd Lots》，于 2026 年 7 月 9 日发布。节目简介指出，在经历了高频交易与智能投顾等多轮技术洗礼后，本期特邀全球最大上市对冲基金之一英仕曼集团（Man Group）的首席技术官 Gary Collier 与数据与 AI 主管 Tushara Fernando，深入探讨该机构如何在严守风控的同时将 AI 落地到量化与主观投资中，并剖析其今年 Token 支出激增 86 倍背后的创新应用与安全挑战。 ## TL;DR 节目全面复盘了英仕曼集团（Man Group）在生成式 AI 与多智能体系统（Agentic Workflows）领域的实战经验与治理框架。对冲基金正在从单纯由算法预测特征，转向全面构建能让 AI 自主构思、验证并运行实盘量化模型的“元系统”，目前已有 15 至 20 个完全由 AI 生成的策略在实盘中受托管理资本。面对自 1 月以来暴增 86 倍的 Token 支出，英仕曼并未急于通过自动化路由削减成本，而是推行全员透明化教育与预算联邦制，让各部门像管理劳动力一样自主调度 AI 工具。此外，嘉宾强调数据清洗与机构语义层才是量化超额收益（Alpha）的核心壁垒，模型能力唯有与基础设施、合规沙箱及前线部署工程师（FDE）深度结合，才能真正撬动投资生产力并筑牢护城河。 ## Atmosphere & style 对话在两位主持人 Joe Weisenthal 与 Tracy Alloway 轻松幽默且极具穿透力的提问中展开，金融一线实操的敏锐度与技术极客的严谨求实碰撞出密集的思维火花。两位嘉宾既坦诚揭示了员工用最贵模型查天气、挑午餐的真实日常窘境，又以顶尖量化机构的高维视角冷静拆解了 Token 预算核算、多智能体闭环与信托责任的边界，全程毫无空洞的技术布道。听感既充实又接地气，充满行业一线打破盲目炒作、直击业务本质与资本逻辑的清醒与纵深感。 ## Bulletin - `30:00` 全球顶尖上市对冲基金英仕曼集团（Man Group）自年初以来 Token 消耗量激增 86 倍，技术应用已从研发部门全面蔓延至财务、运营及人力资源等全业务线。 - `13:16` 英仕曼已实现多智能体自主端到端构建策略的闭环，目前已有约 15 至 20 个完全由 AI 自主构思、生成信号并完成回测验证的量化模型通过人类投资委员会审核并投入实盘交易。 - `26:00` 在成本管控上，英仕曼放弃了行业通用的自动化意图路由技术，转而推行预算联邦制与全员成本透明化教育，以此解决非技术人员在长上下文窗口中混合低效琐事调用的资源浪费难题。 - `23:45` 机构指出微调开源模型投资回报有限，寻找 Alpha 的根本燃料在于高质量数据与跨数据源的统一语义层，前沿基座大模型更多在代码工程与工具生成中发挥核心优势。 ## Key takeaways - `01:46` 桥水与英仕曼的实践表明，直接让大模型推荐股票是死胡同，开源基座模型结合机构专有数据微调或构建统一语义层，在特定金融任务上的表现已超越通用前沿闭源模型。 - `05:40` 英仕曼技术路线具有极强自主性，从硬件、网络设施、数据管道到交易和中后台运营平台均全栈自研，并将数据与机器学习职能全面升级为赋能全员的“数据与 AI 团队”。 - `10:00` 大模型在主观基本面选股中充当多模态信息综合器，能够异步处理技术播客和另类数据等非结构化信息，捕捉传统财报电话会之外的产业链瓶颈与隐性投资逻辑。 - `13:16` 量化投资正从“编写交易策略”跃升为“将构建系统化策略的过程本身系统化”，由多智能体分工协同完成阅读学术论文、生成假设、编写回测代码及模型审计的完整闭环。 - `18:27` 当前对冲基金应用 AI 的最大瓶颈并非算力或数据，而是在飞速扩张的技术边界中保持组织认知对焦，并建立起足以履行信托责任的自治智能体安全沙箱与评估机制。 - `21:28` 机构构建了涵盖市场逐笔 Tick、非结构化另类数据以及“机构上下文与知识”的三层数据架构，核心是通过自然语言元数据让 AI 理解现实业务含义并学会专有研究规范。 - `23:45` 任务属性决定技术路线侧重：工程与代码开发重度依赖最顶尖的前沿基底模型，而 Alpha 研究与信号挖掘的胜负手完全取决于底层专有数据的清洗与语义加工质量。 - `26:00` 英仕曼实行 Token 预算联邦制，将配额下放至各业务部门灵活调剂，并公开个人账单与开销分布以开展全员教育，反向倒逼员工掌握 Prompt 压缩与上下文节流技巧。 - `31:40` 随着任务评测基准 METR 显示的自主任务时长每 7 个月翻番，人机协作范式由传统的局部单元测试辅助，进化为智能体独立承接长达 16 小时的端到端应用功能构建。 - `35:48` 受监管要求约束，英仕曼坚决杜绝黑盒逻辑，要求智能体在生成任何实盘回测代码前，必须先以清晰的自然语言输出投资逻辑与信号假设，确保全链路可解释与可审计。 - `40:00` 生成式 AI 的文本与合约解析能力打破了非结构化市场的准入门槛，使得资产证券化信贷、场外交易等传统难以获取规整价格的边缘领域也能被纳入多资产量化框架。 - `41:15` 全员招聘门槛被 AI 重塑，技术人员正从深陷底层逐行 Debug 的执行者，转变为统筹、编排与监督多个专业自主智能体协同作战的“智能体指挥官”。 - `43:45` AI 极大降低了执行代码的边际成本，团队核心瓶颈从“如何实现功能”转向高阶“顶层规划”，即系统架构设计与跨部门工作流重塑。 - `45:28` 暴增的 Token 支出本质上正在从传统 IT 软件采购演变为购买“自主智能体工时”，其自主跨部门调用的特性给机构内部的成本核算与归属带来了全新治理命题。 - `47:15` 面对全市场同质化模型竞争，超额收益的护城河不在于单点算法，而在于连接全球复杂市场的执行闭环、清算网络、自研回测算力环境以及专有非公开数据的系统生态。 - `50:00` 英仕曼早已常态化推行“前线部署工程师（FDE）”机制，工程人员与宽客无缝并肩作战，通过将公用流程固化为标准化 AI Playbooks 来平衡全员知识沉淀与明星投资人的核心 Alpha 保密。 ## Detailed walkthrough ### `01:46` AI 在投资中的迷思：从粗浅预测到技术范式迭代 节目伊始，Joe Weisenthal 与 Tracy Alloway 展开了一场带着几分幽默的探讨。Joe 坦言自己现在对 AI 最着迷的切入点是它在投资领域的具体落地，而不是过去那种简单粗暴的使用方式。当 ChatGPT 刚问世时，大众最直接的反应往往是直接问模型“我该买哪只股票”，或者试图借此预测通胀走势以在预测市场上套利，但实践早已证明，这种指望模型直接给出致富代码的粗糙方式完全是一条死胡同。这引发了一个更具深度的行业疑问：投资行业历来不乏技术革命的洗礼，从早期的系统化量化投资、高频交易（HFT），再到前些年声势浩大的智能投顾（Robo-advisors），眼下的生成式 AI 究竟只是传统机器学习技术的一次温和迭代，还是一场正在重构投资本质的范式转移？ Joe 借由量化投资的演进史指出，海量数据运算一直以来都是量化交易的根基。例如确立“价值股跑赢高估值股票”或“动量策略有效”等经验法则，本质上依赖强大的算力与数据清洗来证实这些现象的存在；然而哪怕这些策略行之有效，学界与业界对其深层成因依然争议不断。这与当前的 AI 大模型展现出的特质惊人地相似：模型能够识别出令人叹为观止的复杂模式，却难以用人类的日常语言解释其底层的推导机理。Tracy 对此深表赞同，并指出如今所有金融峰会的核心话题无一例外都围绕着数据——如何获取大规模、高洁净度、且最好具备排他性的“专有数据”（proprietary data）。 Joe 还分享了一个极具代表性的行业新案例：就在录音前一周（2024年7月初），全球头号对冲基金桥水（Bridgewater）发布了一篇颇具启发性的研究论文。桥水利用自身的专有金融数据，对开源大模型 Qwen（通义千问）进行了微调，用于精准辨别金融资讯中究竟哪些内容才真正具备“新闻价值”（newsworthy）。研究结果表明，这种“开源基座模型 + 专有数据”的定制组合，在特定金融任务上的表现甚至击败了当时美国最先进的前沿商业模型。这一发现直指金融机构当下最焦虑的核心：当全行业都能接入顶尖大模型时，驱动超额收益（Alpha）的核心壁垒到底从何而来？ - 行业过往技术迭代轨迹：智能投顾（Robo-advisors）$\to$ 系统化/量化投资（Systematic Quants）$\to$ 高频交易（HFT）$\to$ 生成式 AI 与大语言模型。 - 桥水实验核心结论：针对特定金融新闻有效性筛选，开源模型经专有数据微调后，性能超越通用顶尖闭源前沿模型。 ### `05:40` 走进英仕曼集团：全谱系资管的技术底座与角色定义 为了拆解顶尖机构如何真正把 AI 嵌入投资系统，节目邀请到了全球最大上市对冲基金之一——英仕曼集团（Man Group）的首席技术官（CTO）Gary Collier，以及数据与 AI 主管（Head of Data and AI）Tushara Fernando。Gary 首先对英仕曼集团的业务版图进行了拆解，将其定义为一家“全谱系主动型资产管理公司”（Full-spectrum Active Asset Manager）。所谓全谱系，是指其业务纵横交错：横跨另类投资与只做多（Long only）策略、公开市场与私募股权信贷、系统化量化交易与传统基本面主观投资（Fundamental Discretionary），并拥有专门的解决方案业务将上述能力打包成客制化投资组合。 这种业务的“全谱系”特性也直接决定了 CTO 职能的广度与深度。Gary 强调，英仕曼在技术路线上有着极强的自主性（very opinionated），技术栈从底层硬件服务器的选择、网络设备配置，一直向上延伸到终端用户使用的软件界面；在数据链路和投资流程上，从定制的数据源抓取、投研框架构建、算法交易系统，再到中后台的运营作业平台，几乎所有核心环节都是高度自主研发的。 ```mermaid flowchart TD accTitle: 英仕曼集团全谱系技术与资产管理架构 accDescr: 展示英仕曼集团从底层基础设施、自研技术链条到上层多元资产业务与AI赋能的整体架构。 subgraph Infra["底层基础与技术栈 (CTO 统一把控)"] HW["定制服务器与低延迟网络设备"] DataFeed["自研数据管道 (市场/另类/专有知识)"] CorePlatform["投研框架、交易系统与中后台平台"] HW --> DataFeed --> CorePlatform end subgraph Business["全谱系投资业务应用"] direction LR Quant["量化系统化投资 (Quants)"] Discretionary["基本面主观投资 (PMs/分析师)"] Private["私募与长仓/另类策略"] end subgraph AIEngine["Data & AI 团队赋能"] ML["传统机器学习 (回归/预测)"] GenAI["生成式 AI (知识提取/工具构建)"] end CorePlatform --> Business DataFeed --> AIEngine AIEngine --> Business ``` Tushara 随后阐述了数据与 AI 团队的权责，其团队既要统筹市场行情数据与另类数据，更要将公司积累的内部知识与上下文注入 AI 模型之中，进而把先进的技术能力赋能给宽客、量化研究员以及基本面的投资组合经理。Tracy 特别问及 Tushara 职务头衔的变化——此前为“数据与机器学习主管”（Head of Data and Machine Learning），如今更名为“数据与 AI 主管”。Tushara 解释道，在传统量化基金语境下，“机器学习”通常意味着运用线性回归、神经网络等统计数学模型来预测特定的特征变量（features）；而随着生成式 AI 的爆发，技术的核心已经跃升为“对人的全方位赋能”（enablement of people），它不仅帮助团队处理预测任务，更让研究员和投资经理能够创造出以前根本无法构建的新工具。 - 英仕曼集团业务维度：涵盖量化系统化、基本面主观、另类投资、公募多头、公开与私募市场。 - 技术建设特点：全栈自研覆盖从硬件、网络、数据接入、投研框架到中后台运营平台。 - 职能定位演进：从以统计特征预测为主的“Machine Learning”，转变为涵盖生成式工具与全员创造力赋能的“AI”。 ### `07:37` 交易室屏幕上的新常态：生成式 AI 如何嵌入工作流 当话题转入真实的日常运作时，Tracy 抛出了一个最具现场感的问题：如果今天外人走进英仕曼的办公室，站在交易员或基金经理的身后，观察他们电脑屏幕上的操作，与 2022 年生成式 AI 爆发前相比究竟发生了什么肉眼可见的变化？ Gary 回应称，如今走过英仕曼的交易大厅，无论是交易岗位、量化研究岗位还是基本面主观投资团队，员工的电脑桌面上最直观的改变就是出现了一批崭新的独立窗口和专用 AI 工具。就在前一周，Gary 与几位基本面股票分析师交流时，分析师们甚至主动感慨道，现在只要站在基本面投资部的楼层环顾四周，几乎每一台显示器上都常驻着一个以 AI 为核心的工作窗口。“在英仕曼，没有任何一个岗位是不受 AI 影响的，”Gary 总结道，AI 并不是以某种单一的形式颠覆一切，而是根据每个职能的具体痛点，全面渗透、赋能并增强（augmenting）了公司内部的每一个角色。 ### `10:00` 赋能传统选股经理：多模态信息综合与播客中的 Alpha 谈到大型语言模型如何落地，曼氏集团（Man Group）首先以典型的传统“只做多”（Long-only）股票投资经理（PM）为例进行了阐释。传统模式下，一位投资经理通常要同时跟踪 50 只股票，他们需要阅读财报、券商研报、另类数据乃至收听播客。然而，人类每天的精力和时间极其有限，单凭肉眼根本不可能穷尽所有信息。引入 AI 后，最大的改变在于它能够跨越多模态（音频播客、券商研究、另类数据流等），将海量非结构化信息异步提炼为对投资组合切实有意义的洞察。投资经理甚至只需去喝杯咖啡或休息一夜，AI 智能体就能在后台监控全网动态，精准梳理出发生实质变化的边际信息并呈递上来。 Tushara Fernando 分享了一个近期具体案例：当时一位负责跟踪“AI 主题交易”的投资经理正试图理清产业链的核心瓶颈究竟在何处——是 GPU 供应、先进封装，还是其他环节？在一家大型超大规模云厂商（Hyperscaler）工程主管参与的技术播客中，该主管提到：虽然训练模型对 GPU 的需求越来越大，但当前更大的制约在于“数据中心资源极度短缺”，市场上越来越难找到物理规模足够大、能容纳这些庞大训练集群的算力中心。由此可以推导出两个投资结论：不仅 GPU 稀缺，未来各大超大规模数据中心之间的高性能网络互连技术也将迎来爆发式需求。这类信息来自技术一线主管，他们既不出席财报电话会，也不是买方基金经理日常能接触到的人群，但通过 AI 智能体对音频内容的自动转录与综合提炼，投资经理成功在财报之外捕捉到了独家投资逻辑。 两位主持人 Joe Weisenthal 与 Tracy Alloway 随即打趣道，看来“播客本身就是重要的 Alpha 来源”，甚至调侃在 AGI（通用人工智能）时代，最稀缺的资源恐怕就是那些“能把顶级嘉宾请到节目里深聊的主持人”。 ```mermaid flowchart TD accTitle: 传统投资经理的 AI 赋能信息流 accDescr: 展示 AI 如何跨越播客和另类数据等多种模态，为传统选股经理提炼核心投资见解。 subgraph Input["海量多模态数据源"] D1["公司财报"] D2["券商研报"] D3["行业技术播客"] D4["非结构化另类数据"] end subgraph Processing["AI 智能体处理层"] A1["多模态转录与数据清洗"] A2["跨数据源综合与边际变化提炼"] end subgraph Output["投资决策赋能"] O1["异步输出投资洞察 (如数据中心与网络瓶颈)"] O2["赋能投资经理优化组合"] end Input --> Processing Processing --> Output ``` --- ### `13:16` 从研究辅助到智能体闭环：系统化策略生成的“元系统” 随着话题深入，主持人询问曼氏集团是否正在向真正的“智能体工作流（Agentic Workflows）”迈进——即不再局限于辅助人类进行零散研究，而是由系统自主完成“构思交易策略、生成假设、回测验证乃至实盘执行”的完整链路。曼氏集团首席技术官 Gary Collier 明确指出，这绝非遥不可及的未来设想，事实上团队在此方向上已经深耕了一年半之久。在量化与系统化投资领域，如果说以往是“技术 + 量化算法 = 构建系统化交易策略”，那么引入 AI 后的核心命题就是“将构建系统化策略的过程本身系统化”（Systematizing the way that we build systematic strategies），从而赋予量化研究员巨大的生产力杠杆。 在这套端到端的流水线中，AI 智能体分工协同：第一阶段的智能体负责阅读海量学术论文并检索标注数据集，通过推理评估其中是否蕴含值得验证的真实经济学假设；一旦假设成立，下一组智能体就会自动编写底层代码封装该逻辑、提取对应的市场数据并运行回测；随后，专门的评估智能体再对前序模型的回测绩效与稳健性进行交叉审计。Gary 强调，这绝对不是仅停留在实验室里的概念验证（PoC）：截至目前，已有“大约 15 到 20 个完全由 AI 自主构思、完成信号构建与验证的量化模型”，顺利通过了人类投资委员会（Human Investment Committee）的合规与风险审查，被正式判定为“合格且适用”，直接管理着客户的真实资本在实盘中交易。 - **多智能体分工**：学术论文与数据理解 $\rightarrow$ 经济学假设生成 $\rightarrow$ 自动化代码与回测 $\rightarrow$ 独立智能体绩效评估与复核。 - **实盘落地规模**：已有 15 至 20 个完全由 AI 生成的端到端量化模型进入实盘交易。 - **治理与风控底线**：所有 AI 构思的量化策略在上线前，必须经过人类投资委员会的严格把关。 --- ### `18:27` 真正的瓶颈：在飞速扩张的技术前沿中保持人类对焦 既然播客中常有业内人士探讨自身业务的“瓶颈”，主持人借此反问曼氏集团的两位高管：在当前的技术环境下，对你们而言真正的瓶颈究竟是什么？如果可以打一个响指立刻拥有更多资源，你们最渴望的是算力还是数据？ Gary Collier 坦言，算力和数据永远是多多益善的，但它们并非最底层的瓶颈。当前机构面临的最大挑战在于：由 AI 驱动的机会窗口（Opportunity Envelope）正在以令人眼花缭乱的速度急剧膨胀，技术能力几乎每天都在迭代。从人类团队的角度来看，要时刻紧跟这种前沿技术的演进步伐极其吃力。世界上并不缺少好点子，最困难的是“如何有效筛选和过滤点子”。即便像曼氏集团这样向来对技术变革适应极快、组织敏捷度极高的头部对冲基金，在面对如此规模和速度的技术浪潮时，如何平衡战略重心、让组织认知不落后于技术边界，依然是横亘在管理团队面前最严峻的考验。 ### `20:00` 安全治理与合规底线：为何组织变革是最大瓶颈 在将生成式 AI 技术全面推向实战时，技术本身往往不是最困难的部分，真正的阻碍在于整个企业的组织变革与合规管控。作为一家受严格监管的对冲基金，Man Group 承担着沉重的信托责任（Fiduciary Duty），这决定了其部署的任何 AI 系统都必须将风险降至最低。然而，当前整个行业在模型评估（Evaluations）以及对 AI 产出结果的严格检验方面，依然缺乏标准答案，这一领域目前正处于飞速演变之中。 更深层的忧虑在于智能体（Agents）的企业级落地。在个人消费场景中，当用户遇到诸如“AI 把我的收件箱全删了”或者“AI 删光了我的照片且无法恢复”时，大家或许还能相视一笑；但在资产管理规模庞大的金融机构，类似意外是绝对不可承受的灾难。正如 Gary Collier 所指出的，如何在确保万无一失的前提下快速迭代，成为眼下最核心的瓶颈：“确保我们既能快速推进，又能绝对安全，我认为这才是真正的瓶颈所在。” - **信托责任与风险控制**：受管制的对冲基金无法承受任何因 AI 失控导致的合规事故或操作损失。 - **Agent 的安全沙箱**：行业目前对于自治型 Agent 的安全监控和行为边界尚无成熟的现成方案，必须在企业内部建立极其严格的测试与评估链路。 --- ### `21:28` 数据架构三层体系：Tick 数据、非结构化信息与机构知识 针对如何构建信息过滤与摄取管道（Ingestion Pipeline），Tushara Fernando 深入拆解了 Man Group 目前支撑 AI 运转的三层数据架构体系。不同于传统数据处理，现代对冲基金面对的数据异构程度极高，AI 若想提炼出有价值的交易信号，必须依赖针对性架构的处理。 第一层是极度规整的市场数据（Market Data），涵盖各交易所的高频 Tick 数据和订单簿（Order Books），仅全球主流交易所的逐笔 Tick 数据，Man Group 每天就要吞吐接近 1 TB；第二层是杂乱无章的另类与非结构化数据（Alternative and Unstructured Data），包括各类文本、新闻和消费记录。这层的核心在于如何对其进行结构化清洗、打上标签，并建立连接股票代码（Ticker）、行业与公司关系的统一知识图谱；第三层则是全新的架构层——“机构上下文与知识”（Institutional Knowledge）。这一层旨在将 Man Group 多年积累的研究规程、内部上下文注入模型，教 AI “说 Man Group 的内部语言”，例如教会它如何按照基金的严格规范正确运行策略回测（Backtest）。 ```mermaid flowchart TD accTitle: Man Group 数据架构与 AI 交互层 accDescr: 展示市场数据、另类数据与机构知识三层架构及其与统一语义层和任务分流的交互关系 subgraph DataLayers["底层数据三层架构"] D1["市场行情数据<br/>每日近 1TB Tick/订单簿"] D2["另类与非结构化数据<br/>杂乱文本/信用卡消费流"] D3["机构上下文与知识<br/>内部回测规范/专有流程"] end subgraph SemanticLayer["语义抽象与元数据层"] S1["自然语言描述器<br/>(赋予行列底层现实含义)"] S2["统一语义知识网络<br/>(打通标的/行业/宏观关联)"] end subgraph Downstream["下游任务分流"] T1["代码生成与工程开发<br/>(重度依赖最强基底模型)"] T2["量化 Alpha 研究与信号挖掘<br/>(重度依赖底层数据质量)"] end D1 --> SemanticLayer D2 --> S1 S1 --> S2 D3 --> SemanticLayer SemanticLayer --> T1 SemanticLayer --> T2 ``` - **海量高频吞吐**：仅基础 Tick 市场数据一项，系统每日摄入量就高达近 1 TB。 - **让 AI 讲“内部方言”**：机构知识层是区别通用大模型与专业量化助手的关键，重点在于教大模型理解专有回测流程与研究体系。 --- ### `23:45` 语义层胜过模型微调：数据质量才是 Alpha 的根本燃料 面对行业内关于“微调开源模型”还是“直接调用前沿模型（Frontier Models）”的争议，Tushara 指出，Man Group 过去确实尝试过针对特定场景的模型微调，但目前投资回报率最高（"the most bang for our buck"）的做法，实际上是前置的数据清洗、打标与元数据构建。以信用卡消费数据集为例，大模型在原生状态下虽然能看懂行列结构，却无法体会字段背后的金融与商业细微差别。 为此，工程团队通过自然语言元数据为数据集注入“现实世界描述”，直接以浅显的英文在数据字典中告诉模型：“在这一数据集中，每一行记录实际代表着一个真实的人走进了一家线下商店并购买了某件商品”。更进一步，团队投入重金打造了全公司通用的共享语义层（Semantic Layer），使得不同数据源之间截然不同的表述被统一拉平。当模型清楚字段间的映射关系后，就能在股票标的、细分板块与宏观经济之间自如穿梭，挖掘交叉信号。当被问及“最强的前沿模型”与“完美标注的数据”谁更重要时，Tushara 旗帜鲜明地表示，代码生成确实依赖最前沿的基础大模型，但对于寻找超额收益的量化研究（Alpha Research），底层数据的加工质量才是燃料，“单靠堆砌前沿模型，你将一无所获”。 - **元数据平白化**：用清晰的自然语言描述器补齐冷冰冰表格与现实经济活动之间的认知鸿沟。 - **跨数据集关联**：统一的数据语义层让模型能够将信用卡数据、财报与行业标签无缝串联，实现宏观与微观信号的自动联想。 - **任务分化法则**：编程任务重模型，Alpha 研究重数据。 --- ### `26:00` Token 预算联邦制：为什么放弃自动化路由而选择员工教育 随着全公司对 AI 资源的需求爆发，内部甚至出现了非技术人员盲目迷信顶级大模型（如随意调用最昂贵的 Claude 3 Opus）的现象。为了解决资源浪费与算力分配难题，Gary Collier 阐述了 Man Group 独特的“预算联邦制”（Budget Federation）。去年底，管理层根据不同业务团队的职能分类与预期用例构建了消耗模型，测算出总预算；而在今年，他们将预算完全下放给各个业务部门自行决策，坚持“在最合理的前提下将决策权推到最底层”。由于部门内部预算具备高度灵活性（Fungible），各团队完全可以削减其他支出挪用给 Token 采购，并在平台集成的庞大模型库中自由选择商业模型或开源权重。 有趣的是，面对当前行业热衷的“意图分类与自动低成本路由”（Prompt Routing）技术，Man Group 明确选择了“反其道而行之”——放弃由算法替人做路由，转而全力推行全员教育。Tushara 解释称，在拥有 1,700 至 1,800 名员工的 Man Group，员工的技术理解参差不齐，内部观测数据显示了大量极度业余的操作：例如有人在同一个长上下文窗口（Context Window）中，一边让模型起草复杂的投资论据（Investment Thesis），紧接着又在同一个会话里让最昂贵的模型推荐午饭去哪吃，甚至顺嘴查明天的天气。 与其用黑盒路由器掩盖问题，Man Group 选择公开每个人的 Token 账单与开销分布，对全员系统化讲解不同规格模型的特性与成本机制。这种透明化带来了意想不到的正面效果：员工不仅主动规避了上下文污染，还开始自发探索出各种极其巧妙的 Prompt 压缩与 Token 节流方案，并将这些经验反哺回公共平台。 - **预算自治与自主权**：实行联邦制预算管理，各业务线享有支配权，支持预算跨类目调剂。 - **拒绝算法路由的逻辑**：通过教育提升全员对 AI 底层机制（如 Context Window 机制、模型层级开销）的认知，而非通过自动化路由掩盖使用误区。 - **典型的使用误区**：在同一上下文窗口内混合复杂的投资分析与日常琐事（如挑选午餐地点、查询天气预报），造成极大的算力浪费。 ### `30:00` 规避无谓的 Token 消耗与惊人的 86 倍爆发式增长 对话一开始聚焦在智能体调用开销的优化技巧上。以编程智能体（Coding Agent）为例，当它需要执行一些基础的 Git 命令与版本控制系统交互时，传统的做法往往是智能体发送命令，然后把终端返回的所有结果都原封不动当作 Token 重新塞回上下文中处理。事实上，完全可以采用一些简单直观的步骤：在工具调用发起时进行拦截，直接在智能体循环（Agentic Loop）外部执行这些基础操作，而无需每次都调用 LLM 去做推理和工具解析。 尽管团队在工程层面做出了各种能效优化，但 Man Group 整体的 Token 消耗图表依然呈现出极为陡峭的上扬走势。自当年 1 月以来，公司的 Token 使用量在短短几个月内暴增了 86 倍。这一爆炸式的增长彻底超出了管理层的预期，并且这种爆发是全员性的，绝不仅限于技术或研发部门。无论是财务团队、运营团队，还是人力资源团队（People Team），都在广泛使用基于 Agent 的编程工作流。作为技术负责人，能将这种过去只有程序员才具备的能力交到非技术人员手中，无疑令人倍感兴奋。 - **Git 工具调用拦截**：避免将终端原始标准输出全部转化为 Token 输入，把无需推理的命令移至 Agentic Loop 外部处理。 - **86 倍的 Token 暴增**：自 1 月份以来消耗量飙升 86 倍，渗透到财务、运营、HR 等非传统技术职能中。 ### `31:40` 驱动飞跃的核心：更强的底层模型还是工作流容器？ 主持人提出疑问：去年底到今年初（12 月至 1 月）显然是一个关键拐点，这波陡峭增长究竟是由基础模型能力的迭代驱动的，还是由于像 Claude Code、CoWork 等高质量工作流容器（Harnesses）的出现，使人类能够超越简单的“一问一答”，真正开始操控复杂现实任务？ 技术负责人解释道，模型能力与工具容器实际上是紧密耦合的。行业中一个非常值得关注的评测基准是 METR（Model Evaluation and Threat Research），它专门评估智能体完成原本由人类处理、耗时从数分钟到数小时不等的复杂任务的能力。测试数据显示，智能体能够独立执行任务的时长每隔大约 7 个月就会翻上一倍。如今，系统已经可以承接原本需要人类专注工作 16 个小时的庞大任务。 这种质变根本性地重塑了人机协作方式与团队分工结构。团队以前是“人在回路”（Human-in-the-loop）中指挥模型干脏活累活，比如让智能体帮忙写个单元测试；而现在，智能体可以直接独立去构建应用程序的一整套核心功能，乃至从头搭建一个完整的应用。正是这种承接宏大任务的扩展能力，从根本上促成了 Token 消耗的指数级激增。 ```mermaid flowchart TD accTitle: 智能体自主协作模式的演进 accDescr: 展示从微观任务的人机回路到宏观任务自主开发的范式转变 subgraph Phase1["传统人机回路 (短时长 / 窄范围)"] A["工程师提示词介入"] --> B["LLM 针对性推理"] B --> C["局部代码交付<br/>(例如：编写单一单元测试)"] end subgraph Phase2["现代长程智能体 (跨越 16 小时任务范式)"] D["高层业务意图输入"] --> E["METR 范式扩展：自主多步骤规划"] E --> F["外部工具与环境深度交互"] F --> G["端到端功能构建<br/>(独立交付完整模块或全套应用)"] end Phase1 -.->|"能力每 7 个月翻番<br/>Token 消耗激增 86 倍"| Phase2 ``` ### `35:48` 监管与风控高压下的模型“可解释性”（Explainability） 话题随后转回金融业务的监管与合规核心：人在回路审批时，模型的可解释性该如何落实。在受严格监管的对冲基金行业中，投资组合经理（PM）或宽客（Quant）在面对风控委员会或外部监管机构审查时，绝不能用“AI 这么做，但我们不知道为什么”来推卸责任。 Man Group 强调，他们所从事的并非超高频交易（HFT）——高频领域往往构建庞大、无法洞悉的多维神经网络，黑盒输出极难解释。相反，Man Group 的投资持仓周期通常在“数天、数周到数月”之间，所有的交易决策必须具备完全的可解释性，坚决杜绝黑盒交易。在落地实践中，当智能体通过消化海量数据集或学术论文提出全新的交易假设时，它在落笔编写任何回测代码之前，第一步必须以规范、清晰的英文自然语言写明投资逻辑与信号假设，并给出明确命名，从而为全流程的审计和风控提供坚实支撑。 - **非高频交易属性**：持仓周期横跨数天至数月，天然要求决策链路能够被完全倒查和审视。 - **假设先行机制**：智能体在写代码前，必须先输出一份清晰可读的自然语言投资假设文档。 ### `38:15` 微观赋能边缘市场：降低前置研究的人力成本门槛 主持人提出一个关于未来劳动力的引申视角：AI 是否能够降低探索边缘市场的前期成本？例如，研究埃塞俄比亚证券交易所（Ethiopian Stock Exchange）等前沿市场，过去需要耗费大量昂贵的宽客和研究员精力去熟悉其交易规则、结算流程与制度背景。如果潜在利润不足以覆盖这笔前期人力成本，机会就会被直接放弃。而 AI 若能大幅压缩前期调研成本，是否就等于拓宽了基金的投资机会边界？ Man Group 确认，这一转变正在边际上以渐进的方式发生。它并不是一夜之间翻天覆地的巨变，而是由分布在公司各个角落的无数个“微观赋能”（Micro-augmentations）逐步累加而成。例如，有团队成员开发了一套结构简单的专用 AI 系统，专门用来自动阅读各类复杂金融衍生品的 PDF 说明书与技术规范文档，并将其结构化数据自动录入底层参考数据库（Reference Data Store）。正是这些在边缘场景上不断省下前置时间与人力成本的微小工具，正在一点一滴地推高整体生产力的边界。 ### `40:00` 非结构化市场中的系统化投资与 AI 的破局点 在构建系统化量化投资策略时，长期以来存在两个不可或缺的基本前提：第一是必须具备接入该金融工具进行交易的连通性（connectivity），第二是必须能够精准解析并理解该市场的资产价格。对于股票或期货等高度成熟的公开市场，这两点极为直观，量化团队只需直接接入电子订单簿（order book）即可获取清晰的买卖报价。然而，在加密货币、资产证券化信贷（securitized credit）以及许多场外交易（OTC）等欠发达或复杂的细分市场中，交易往往依赖非标准化的合同甚至口头沟通，价格的生成机制中充斥着复杂的非结构化数据，导致传统系统化模型极难切入。 生成式人工智能与大语言模型的引入打破了这一壁垒。大模型展现出了强大的非结构化数据解析与语意提取能力，使得量化团队能够在远早于以往的时间节点，系统化地理解并梳理那些模糊定价机制与合约细节。这意味着，以往因数据清洗成本过高或报价形式杂乱而被排除在系统化大门之外的资产类别，现在也能以工程化、自动化方式被纳入多资产量化投资的版图。 --- ### `41:15` 招聘标准的全面提升：从“深陷细节的技术人员”到“智能体指挥官” 面对 AI 逐步接管研究流程乃至投资经理日常工作的现状，英仕曼集团（Man Group）正在重塑其人才选拔与团队画像。首席技术官 Gary Collier 明确指出，他一直在内部极力推行一条铁律：如今进入公司的每一位新员工，不论是应聘前台的投资研发岗位，还是后台的运营支持岗位，都必须在 AI 运用能力上“拉高全公司的平均水位”（up the bar with respect to AI）。在如今公开信息与工具极度丰富的环境下，如果一个应聘者自诩聪明、充满干劲且对专业领域饱含热情，却在被问及 AI 时声称“我对其了解不多，在日常工作中也基本不用”，这在逻辑上已经自相矛盾，直接失去了竞争力。 数据与 AI 负责人 Tushara Fernando 进一步阐述了技术与投研人员心智模型的深刻转变。过去，顶级技术人员往往以“深挖细节”（in the weeds）为傲，热爱扑在底层代码里逐行调试 Bug 和处理琐碎的技术难点；然而在自动化程度极高的今天，公司真正青睐的是具备宏观战略眼光的长期思考者（long-term thinkers）。理想的员工应当乐于从具体的技术琐事中抽身，将自己定位为“这些智能体的指挥官（a kind of a conductor of these agents）”，负责端到端流程的设计与监督，像一位兼具深厚技术素养的管理者一样去编排多个智能体协同作战。 ```mermaid flowchart LR accTitle: 从传统极客到智能体指挥官的人才演进 accDescr: 展示量化技术人员从过去专注于底层代码调试，转向以顶层设计和协调 AI Agent 为主的角色转变。 subgraph Past["传统技术人员角色"] A["深陷底层执行 (In the weeds)"] --> B["手动逐行编写与调试代码"] B --> C["聚焦单点技术细节与 Bug"] end subgraph Future["AI 时代的智能体指挥官"] D["顶层架构规划与编排"] --> E["自动化端到端业务流"] E --> F["管理监督多个专业 Agent"] end Past -.->|"能力要求拉高 / 工具链进化"| Future ``` --- ### `43:45` “超级巨星”与“全员普及”：从执行转向架构规划 针对 AI 是会催生能够驾驭大批智能体的“超级巨星（Superstars）”，还是会让全员普及技能并让初级员工越级赋能（Democratization），英仕曼集团的观察是两者兼有，但在整体规模上，“全员普及”的效应更为广泛。几乎每位员工都在日常工作中频繁使用 AI 来处理调研、编写基础代码和自动化繁琐流程。相比之下，那些能够站在思维最前沿、跨越多个部门去重塑体系架构的超级创新者虽然极具影响力，但数量自然更加稀少。 这种分化带来的核心痛点在于团队工作重心的结构性迁移。当前，编写代码与构建新特性的时间成本与资金成本正在急剧暴跌，“执行（Execution）”变得极度廉价。团队面临的最大挑战不再是“怎么做”，而是“做什么”： - 核心精力必须从 80% 的时间用于执行，转向将大部分时间投入到顶层“规划（Planning）”与系统架构设计中。 - 团队必须思考：我们到底应该构建什么？系统各部分如何连接？跨职能部门的全新工作流应当如何运转？ - 最困难的往往是“把人们从习惯的工位和思维惯性中拉出来”，促使他们坐在一起共同规划全局工作流，而不是一头扎进去急着做一个单点概念验证（POC）。 --- ### `45:28` Token 支出的本质：从 IT 软件预算到“自主智能体工时” 随着英仕曼集团内部的 Token 消耗量实现了惊人的 86 倍增长，市场讨论的语境也发生了根本性逆转。主持人和管理层提到，早期行业还在讨论大模型会否冲击传统 SaaS 软件公司的估值，而如今人们已经意识到 AI 的颠覆目标不是软件，而是整个“人类劳动力市场（Human Labor）”。业界甚至开始推算 Token 消耗与员工薪酬总额的相对比例。尽管英仕曼尚未正式按 2027 或 2028 年的远期目标去制定固定的“薪资-Token 支出比率”，但公司底层的资源配置逻辑极其现实——资本与算力永远流向能产生最高经济回报的地方。 在当前的预算审批与核算过程中，英仕曼迎来了一个全新的内部治理难题：暴增的 Token 支出中，很大一部分并不是由员工手动提问消耗的，而是由后台日夜运转的自主智能体（Agents）在自动化工作流中消耗的。这些自主调用大模型的 Agent 往往横跨多个业务条线，导致费用归属模糊：这个特定的自主工作流究竟归属于哪一个业务部门？这笔 Token 账单到底该由前台算力买单还是运营线承担？管理层幽默地表示，这虽然是一个尚未完全解决的新挑战，但绝对是一个“幸福的烦恼（a great problem to have）”。 --- ### `47:15` Alpha 的护城河：从单点算法到复杂交易生态的系统博弈 针对高频交易（HFT）领域过去在微秒、纳秒级竞速中陷入“逐底竞争（race to the bottom）”的教训，主持人质疑：当全市场所有量化基金都在接入相同的大语言模型并加速挖掘信号时，AI 所带来的超额收益（Alpha）是否也会在短期内被迅速套利殆尽？对此，Fernando 给出了深层的防御逻辑：利用 AI 接入数据集、分析文本并提取特征确实变得前所未有的容易，但“能够提取特征”与“能够将特征转化为可交易的盈利组合”之间存在巨大的鸿沟。 英仕曼数十年来沉淀下来的核心壁垒并非某一个单点算法，而是一个不可复制的系统生态： - **深厚的基础设施与连接能力**：拥有直接接入全球各类复杂市场的清算通道、深度的经纪商关系（broker relationships），以及无法在公开货架上随意买到的专有非公开数据。 - **回测与计算闭环**：能够将强大的自研回测算力环境与前沿 AI 能力无缝嵌套，支持信号的高频演进与风险对冲。 - **Alpha 的动态演化**：在英仕曼集团内部，没有任何一个单独的代码仓库可以被指着说“这里就是 Alpha 所在”。Alpha 诞生于庞大系统网络各节点之间的复杂交互。 - **从 Alpha 降级为风险因子**：某些基础的数据集特征在被全市场广泛掌握并推动价格后，确实会迅速从超额收益退化为“全市场通用门槛（table stakes）”与“风险因子（risk factor）”，但顶尖量化机构的盈利能力建立在持续将跨领域异构数据集与执行端无缝缝合的整体生态之上。 ### `50:00` 前线部署工程师与组织协同架构 ```mermaid flowchart TD accTitle: Man Group的技术组织架构与业务协同 accDescr: 展示Man Group横向数据与AI平台团队与前线部署工程师协同赋能宽客及主观投资经理的结构。 subgraph 平台底层技术支持 Platform["数据与AI核心平台<br/>(Tushara Fernando团队)"] end subgraph 前线业务落地 FDE["前线部署工程师 (FDE)<br/>(长期驻扎业务一线15年)"] Quants["量化投资团队 (Quants)"] Discretionary["主观投资经理 (Fundamental PMs)"] end Platform --> FDE FDE <--> Quants FDE <--> Discretionary ``` 在探讨企业如何从文化与层级结构上最大化发挥 AI 工具潜力时，主持人提到业内正在兴起的新趋势：例如微软最新宣布成立所谓“Frontier Company”部门，旨在深入企业内部协助设计最优组织架构；而 Palantir 很早就采用前线部署工程师（Forward Deployed Engineers, FDE）模式；Anthropic 也向高盛（Goldman Sachs）等机构直接派遣工程师，以推动 Claude 等工具的深度落地。 针对外部厂商是否也向英仕曼集团（Man Group）兜售此类组织咨询服务，首席技术官 Gary Collier 笑着回应称，看到业内把这种高昂的前线工程师模式当成新鲜事物令人莞尔。实际上，Man Group 早在约 15 年前就确立了这种运作架构： - 集团内部设立专门构建大型横向底层平台的技术团队，例如 Tushara Fernando 领导的数据与 AI 平台团队。 - 与此同时，技术部门中很大一部分人员历来就是“前线部署工程师”，他们常年与量化研究员（Quants）及主观投资经理（Discretionary Managers）坐在一起并肩作战。 - Gary 常常向求职者开玩笑说，只要到公司位于泰晤士河畔的 Riverbank House 五楼办公室走一圈，让人分辨谁是工程师、谁是宽客，十有八九会猜错，因为大家屏幕上跑的代码和处理的数据几乎一模一样。这种紧密无缝的协作状态早在十多年前就已成为现实。 --- ### `52:16` 独家 Alpha 保护与组织知识沉淀的利益对齐 在资产管理机构中，AI 的威力依赖于打通和处理海量结构化与非结构化数据，但这触及了对冲基金的核心矛盾——明星投资人（Rainmakers）赖以立足和获取高额薪酬的往往是其个人独有的 Alpha 生成能力和深厚行业壁垒。主持人询问，Man Group 如何化解这种利益冲突，让投资经理心甘情愿将个人知识贡献给全公司的共享 AI 系统。 Gary Collier 指出，这在某些业务领域仍处于“探索与磨合”阶段： - **量化与系统化业务：** 长期以来就建立在高度协同和共享代码库的文化之上，知识共享没有任何阻力。 - **主观基本面投资业务：** 投资经理通常对自身投资逻辑的 90% 表现得非常开放，愿意外放并进行讨论，但唯独对那最后决定个人核心增值、最敏感的“10% 独家秘方”保持保留态度。 Tushara Fernando 进一步阐述了具体的应对与沉淀机制： - 机构层面上，绝大多数工作流程属于公共专业能力，而非直接决定回报的 10% 核心机密。 - 例如模型回测的具体方法、阅读解析投研报告的框架等，技术团队会将这些高阶工作流封装成标准化的“AI 技能（AI Skills）”和“AI 操作手册（AI Playbooks）”，沉淀进内部知识平台供全公司调用。 - 涉及具体投资策略构建和头寸判断的独特敏感细节，则允许各团队适度保留，从而在保障知识沉淀与尊重个人 Alpha 之间达成有效平衡。 --- ### `55:00` 后续复盘：86 倍 Token 增长的经济账与隐性数据捕获 在与嘉宾道别后，主持人 Joe Weisenthal 与 Tracy Alloway 对对话核心要点进行了复盘： ```mermaid flowchart LR accTitle: Token消耗飞轮与成本控制阶段 accDescr: 展示Token消耗从探索期到紧缩期的演进路线。 A["86倍Token爆发式增长<br/>(自由探索/多模型实验)"] --> B["业务指标倒逼<br/>(显性成本削减 or Alpha收入兑现)"] B --> C["Token紧缩与管控<br/>(开发内部路由/限制低效调用)"] ``` - **Token 爆炸式增长背后的经济账：** 自当年 1 月以来，Man Group 的 Token 开销激增了 86 倍。Joe 认为这极其疯狂且不可持续，最终必须反映在实打实的财务数字上——要么是明确削减了运营开支，要么是创造了新的增量收益。 - **看涨（Bullish）信号：** 令人惊讶的是，即使增长了 86 倍，Man Group 甚至还没开始自研内部智能路由机制来压低成本，仍鼓励员工自由试用各大顶尖模型进行业务实验。这反映出大型金融机构当前不仅没有陷入“Token 紧缩（Token Austerity）”，反而还在大规模容忍前期的探索式支出。 - **敲击键盘监控与知识沉淀的幽默讨论：** 针对如何让王牌投资人交出核心知识，Tracy 打趣道其实只要借助按键记录与合规监控（Keystroke Surveillance）即可搞定。像金融投资这类强监管行业，内部几乎所有操作早已全天候留痕并受监控。 - **企业是否悄然用员工行为数据训练模型：** 主持人探讨了近期 Meta 等科技巨头被爆出开始利用员工内部数据训练模型的现象。在金融圈，如果投资人为了不让自己的隐性知识变成喂养 AI 的原料，未来是否真得退回到用“纸和笔”记录策略，成为了一个值得玩味的深层次悖论。 ## Notable quotes > `01:46` 开源基座模型 + 专有数据 — Joe Weisenthal > `07:37` 在英仕曼，没有任何一个岗位是不受 AI 影响的 — Gary Collier > `13:16` 将构建系统化策略的过程本身系统化 — Gary Collier > `20:00` 确保我们既能快速推进，又能绝对安全，我认为这才是真正的瓶颈所在。 — Gary Collier > `23:45` 在这一数据集中，每一行记录实际代表着一个真实的人走进了一家线下商店并购买了某件商品 — Tushara Fernando > `23:45` 单靠堆砌前沿模型，你将一无所获 — Tushara Fernando > `26:00` 在最合理的前提下将决策权推到最底层 — Gary Collier > `41:15` 拉高全公司的平均水位 — Gary Collier > `41:15` 这些智能体的指挥官 — Tushara Fernando > `47:15` 这里就是 Alpha 所在 — Tushara Fernando ## Mentioned Joe Weisenthal, Tracy Alloway, Gary Collier, Tushara Fernando, 英仕曼集团 (Man Group), 桥水基金 (Bridgewater Associates), 通义千问 (Qwen), ChatGPT, Claude 3 Opus, Claude Code, CoWork, METR, Palantir, Anthropic, 高盛 (Goldman Sachs), 微软 (Microsoft), Meta, 埃塞俄比亚证券交易所 (Ethiopian Stock Exchange) ## Video description <p>We've gone through a number of a technological revolutions in investing, whether it was the dawn of the high frequency trading era or the introduction of robotraders. When it comes to AI, the big question that remains in the investment context is whether or not the technology will be implemented like those past tech innovations &mdash; meaning it will be integrated into the flow of the business without upending everything as we know it &mdash; or if AI will transform the very nature of investing. Right now, AI's use in investing is a mixed bag: People are excited about its potential, but several firms are still trying to figure out its value. Today, we speak with Man Group's CTO Gary Collier and Head of Data and AI Tushara Fernando about how one of the largest publicly-traded hedge funds in the world is actually implementing AI into its work. We speak with them about empowering their quants with AI tools, the challenge of integrating AI safely, and the creative ways their staff is thinking about token spending, which is up 86-fold this year.</p> <p><br>Only Bloomberg.com subscribers can get the Odd Lots newsletter in their inbox &mdash; plus unlimited access to the site and app. Sign up here: <a href="https://www.bloomberg.com/subscriptions/oddlots?in_source=oddlotspodcast" data-link-cookie="1" data-destination="https://www.bloomberg.com/subscriptions/oddlots?in_source=oddlotspodcast">bloomberg.com/subscriptions/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/ee665e22-5abb-4981-a58d-b481016ff97e/audio.mp3?in_playlist=982f5071-765c-403d-969d-ae27003a8d83)

[1:46] Joe Weisenthal: 开源基座模型 + 专有数据 [7:37] Gary Collier: 在英仕曼，没有任何一个岗位是不受 AI 影响的 [13:16] Gary Collier: 将构建系统化策略的过程本身系统化 [20:00] Gary Collier: 确保我们既能快速推进，又能绝对安全，我认为这才是真正的瓶颈所在。 [23:45] Tushara Fernando: 在这一数据集中，每一行记录实际代表着一个真实的人走进了一家线下商店并购买了某件商品 [23:45] Tushara Fernando: 单靠堆砌前沿模型，你将一无所获 [26:00] Gary Collier: 在最合理的前提下将决策权推到最底层 [41:15] Gary Collier: 拉高全公司的平均水位 [41:15] Tushara Fernando: 这些智能体的指挥官 [47:15] Tushara Fernando: 这里就是 Alpha 所在

[5:25] Tushara Fernando: Historically, machine learning in a quant firm was really about using these more traditional machine learning techniques... Whereas now with the onset of generative AI it's more than that. It's really about enablement of people to create things. [14:14] Gary Collier: There have been a bunch, I think, 15, 20 models at the last count that had gone all the way through... ideated by AI, went all the way through signal construction, the validation process, were reviewed and validated by a human investment committee and deemed fit and proper to trade our client's assets with. [20:00] Tushara Fernando: We found that if you take a data set like credit card data, for example, AI can look at that... but it doesn't really understand the nuances of it... So what we're having to do is invest in trying to add extra color, extra metadata to that information... telling AI that in plain English. [25:00] Tushara Fernando: We have a very rich data set of how people are using AI... Some of the things that we saw were just quite basic mistakes. So often people would be doing multiple tasks in the same context window... trying to figure out the best way to write an investment thesis, and then... asking Fable what the weather will be tomorrow. [30:00] Tushara Fernando: Now you can ask an agent to do a task that would take a human 16 hours. And that changes the way that you think about teams... You go from a place where you're in the loop, asking an agent to solve a problem like writing a unit test, to a problem like building an entire application in itself.

[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. 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 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. Bloomberg Audio Studios, podcasts, radio, news. Hello and welcome to another episode of the Odd Lots podcast. I'm Joe Weisenthal. And I'm Tracy Alloway. Tracy, I'm very interested in AI. No, no. Oh, really? Really, Joe? I am. That's a surprise. I'm very interested in the investment context, specifically. I mean, the actual implementation of like, how do investors use it? Because I think, obviously, just sort of substantively, incredibly important question for reasons that need no explaining. But I also think it raises very interesting sort of puzzles about what the technology is used for. And I remember when ChadGPT came out and people were asking what stock should I buy We did that prediction market episode recently And it like one thing you definitely can get much value out of is saying like which contract should I buy or what's inflation going to be so I can trade this contract? So like, but that doesn't mean that there aren't interesting ways. It just seems like, you know, the sort of the most crude version of, quote, using AI for investing is obviously a dead end deal. way. Here's the question I have. We've been through technological revolutions in investing before. Notably, we had robo-advisors. Remember that? We had systematic investing, quants. We had high-frequency trading. My big question is how much of the current use of AI is basically an iteration or an improvement on some of those kind of machine learning dynamics, let's say, versus something more substantial or more revolutionary? Is it like a minor evolution or moderate evolution or something big? Well, this came up a little bit in our conversation with Ian from Hudson River Trading. And I'm glad you brought up the quant example, because like all right big data in some sense has been part of quant since the very beginning Like how do you establish that quote value stocks outperform expensive stocks Well you just need a lot of data and computers and running that math et cetera to establish that fact. But this idea of, yes, maybe cheap stocks outperform expensive ones, momentum stocks outperform stocks with bad momentum, things that seem to be true, but we don't really know why. And there's a lot of disagreement as to the source of these things. And that really makes me think about AI because we can get these outputs from AI models that are obviously remarkable. You can recognize patterns, yes. But they can't really explain how they arrived at that pattern. And I'm very interested in where this leads us with investing and whether it's like, okay, do we get these ideas and strategies, et cetera, that work, but we can't really articulate them in plain English. I'm glad you brought up data as well, because one thing you hear at basically every finance conference nowadays is the importance of data when it comes to running LLMs and making sure that your data is actually processed, it's clean, that you have a lot of it and that you have, hopefully, proprietary data. And the people we're going to talk to have a lot of data, actually. Totally. There was just one more thing. I don't know if you caught it last week. We're recording this July 7th. [5:00] Last week, while you were on vacation, there was a very interesting paper out from Bridgewater talking about their use of proprietary data to fine-tune an open-source version of Quinn. And for a specific purpose of being able to identify what is newsworthy financial information than not, they found that the combination of open-source models plus proprietary data got them better results than the most frontier U.S. models. Yeah. That is very interesting to me. And like these are the types of things that I'm very curious about. Where does the secret sauce or the alpha actually come from? Totally. Well, yeah, especially when everyone is going to have access to like really, really intelligent models. Anyway, enough talk from us. I'm really excited. We have the perfect guests to talk about this, actually understanding the implementation of AI within the investment or asset management context. We're going to be speaking with Gary Collier, Man Group CTO, as well as Tushara Fernando, the firm's head of data and AI. So like I said, really the perfect guests. Gary and Tushara, thank you both so much for coming on OVLOTS. Great to be here. Let start with Gary or maybe we go both of you one after the other Why don you just give us a very quick description of your roles at Man Group and what specifically you do Sure I can start with that I perhaps useful to give a bit of context of the firm I describe Man Group really as this full spectrum active asset manager. So full spectrum being we cover alternatives, we cover long only, we cover public markets, we cover private markets, we cover quant systematic, cover fundamental discretionary investing, and we've got a solutions business that can bring all that together into custom client mandates. And what that means in terms of CTO role, I describe that as like full spectrum, too, because for all of the above, we're very opinionated in what every part of the tech stack looks like, right down from choice of servers, networking devices, all the way through to the end user facing pieces of software. And simply left to right, we build an awful lot of our own technology right from the custom data feeds all the way through research frameworks, trading systems and the middle and back office operating platform. So very much full spectrum role. Tushara? Yeah so on my side I look after data and AI So that everything from market data to alternative data to all of the context and knowledge that we have that we plug into our AI models Then from an AI perspective it about how do we give the great capabilities that we now have to our quants to our researchers and to our fundamental portfolio managers. Yeah, Tashara, I wanted to ask you about this, because I think your title actually used to be Head of Data and Machine Learning, and now it's Head of Data and AI. How do you think about the difference between those two terms, machine learning versus AI? That's a really good question. I think it links to one of the earlier points that you had around whether it was that we were using traditional machine learning techniques or whether we were using generative AI. And I think historically, machine learning in a quant firm was really about using these more traditional machine learning techniques that were looking at things like linear regressions. They were looking at neural nets to try to predict some feature. Whereas now with the onset of generative AI is more than that. It's really about enablement of people to create things as well as using these traditional techniques. So that's why we've changed the title of the role because it encompasses both the generative AI aspects as well as the more traditional machine learning methods Okay that makes sense So if I was sitting within a man group office today and I take you know Gary laid it out perfectly well you guys have a lot of different roles. So you have discretionary sort of traditional portfolio managers and then you have more systematic quants and all of that. But if I'm shadowing one of your traders or PMs and I'm watching what they're doing on a screen today versus what they were doing on a screen in, let's say, 2022 or something like that, What exactly has changed with the use of generative AI? What are you doing differently? I think if you were to walk across the floor, you would see different windows, different tools in use now, regardless of role, whether it be trading, quant research or discretionary investing. and chatting to a couple of discretionary analysts last week, and they were commenting to me on how everywhere you look across the discretionary floor, everybody has got an AI focus like window on their screen. And so it's affecting, improving, like augmenting pretty much every role we have in the firm, of course, in different ways, depending on what the role is. But there's no role that's unaffected by AI. What are they doing? Again, [10:00] Rather than talk about the firm Y, let's say I'm like a sort of traditional long-only stock selector, like the sort of classic thing that in our minds we often think of as an asset manager. And it's like, okay, I have now access to some of the world's most advanced models. What am I doing with them? So if you think about traditionally how a PM like that would work, they want to look across a broad set of names and they want to get access to as much data as possible for those names. So they want to look at earnings reports. They want to look at broker research. They want to look at alternative data. They want to look at podcasts. But there's only so much time in the day. There's a few hours in the day and there's 50 names. How are they going to cover them all? How do they get to the really important insight? And what AI has allowed us to do is it allowed us to access all of those different types of data, lots of different modalities, podcasts, alternative data, things like broker research reports, and synthesize them into what actually meaningful so that a PM can asynchronously get that insight They could go away for a coffee they can come in overnight and an agent has actually distilled a change that's happened on the internet and given them the insight that's meaningful to their portfolio, to their investment thesis. So just I suppose an example might be recently we had a pm that was covering the ai trade and really that one of the important things is about where's the bottleneck where's the bottleneck here and is it in the gps is it yeah and it's hard to know exactly where it is and there was a podcast cast from one of the heads of engineering from a large hyperscaler and he was saying that it was increasingly important to have more and more GPUs to train models, but data centers were becoming more scarce and it was becoming increasingly difficult to find data centers that were actually big enough to train these models. So what you could see there was a couple of things One that there GPU scarcity And the other thing is that it likely that we going to need better networking between these large data sensors in the future And that was something that came on a podcast from a head of engineering This isn someone who a PM would usually interact with They don't go to the investor calls. They're not somebody they often have access to. But through AI, the PM was able to have an AI agent transcribe that podcast and synthesize that data so that they could get better insight into that investment idea. Yeah, I feel like podcasts are an important source of alpha, Joe. Everyone should be listening to podcasts. You know, the non-joke version of that, which is one of the things people say. We had Alex Nimas on the podcast and he's like, what's going to be scarce after AGI? Yeah. People who are able to get good guests on their podcast, people who are able to get those people, you know. OK, so we're clearly talking our own book. Yes. OK, setting that aside for a second. It sounds like what most of you're doing when you describe that process is augmenting research or allowing your investors, your managers to be more efficient in their research process. There a lot of talk nowadays about agentic workflows and the actual idea that instead of having humans drive every step of a particular trade or project you have a system that can do the research it can generate ideas it can test the hypotheses and then it can even execute on them Is that something that you're sort of working towards, or is it still too far away for you guys to even be contemplating? No, that's not too far away at all. In fact, that's something we've been working on for quite some time now and if we shift focus to the quant the systematic part of the business what we've been doing now look if you take a step back and think well technology plus quant techniques that gives you the ability to build systematic strategies so what do we get if we add ai into that mix well we have the ability to think about systematizing the way that we build systematic strategies to begin with, and in effect giving a big force multiplier, big leverage multiplier to our quants, our researchers. So what we've been doing there for well over a year, now actually a year and a half, is building a system that can actually take all of those different parts of the quant research process. So the idea formation part, looking at academic papers, looking at carefully labeled data [15:00] sets reasoning about the content of those papers, the content of those data sets, are there economic hypotheses in there that could be real or at least could be worth testing out, and then having other agents build the code to encapsulate those ideas, get the right market data, run the right back tests, etc., and then further agents that evaluate the output of the prior agents. And so this is really one of the early ideas that we thought would give potentially great bang per buck. And we've been working on some time. And just to make it real, to demonstrate this is not just something that's happening in the lab, but doesn't have any real or practical consequences. There have been a bunch, I think, 15, 20 models at the last count that had gone all the way through. They started off as models that were ideated by AI, went all the way through the signal construction, the validation process, were reviewed and validated by a human investment committee and deemed fit and proper to trade our client's assets with. So this is very real It not to make believe at this point 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. Support for this show comes from public.com. If you 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 of 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.com slash market and fund your account in five minutes or less. That's public.com 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. 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 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. Yeah. Since you mentioned sort of hoovering up information that might appear on a podcast where someone identifies a particular bottleneck in their business. We are on a podcast. What is your bottleneck? What would you like to have more of right now in order to improve your process? Is it compute? Is it data? If you could snap your fingers and have more of something, what would it be? you can always use more compute and you can always use more more data but i think what the bigger problem is the world of opportunity that's afforded by the ai capabilities that we're seeing and of course being delivered all the time that that envelope is expanding like so quickly that keeping up with it from a human perspective can be like quite hard there's no sources of good ideas filtering ideas is important. But of course, whilst we're used to making lots of change at Man Group, the level of [20:00] like organizational change that we need to get involved in to put all of this stuff into effect. I think that's the bottleneck, making sure because we're a regulated business, we've got fiduciary duty. We need to make sure that the things that we're doing, the things we're deploying are done with the minimum amount of risk. And there's a lot of work in this field that really, I think the industry doesn't have firm answers to yet. I mean, evaluations, testing the results of AI processes are good and proper. That's a rapidly expanding field. Thinking about how we run agents across the business in a safe and controlled fashion. We all, I'm sure, have seen and smiled at the, well, the AI deleted my inbox or the AI deleted all my photos and I can't get them back. We can't afford to have that type of thing happen at an enterprise level. So it's making sure that we can move quickly, but safely, I think, is the bottleneck, if you like. Yeah, that makes sense. Let's get into a couple of specifics. And I want to get back to the sort of how you move forward safely. But both of you have now mentioned filtering. And that actually is precisely what I mentioned in the intro There was a paper from Bridgewater about fine a version of Quinn on their own data precisely for better filtering so that the PMs could just have more efficient use of their time to see what signal How do you build that? Is this an off-the-shelf thing? What is the tech, the model, et cetera, that you have to this ingestion pipeline? What does it consist of? Yeah, I can say that one. So I think for us, it really depends on the type of data. There's broadly three types of data that we look at. The first one is market data. That's often very structured tick data. Things like order books. We ingest every tick from most exchanges, almost a terabyte of data just from ticks per day. and then we have alternative and unstructured data that is much more messy it's much more malformed and there it's really about how we structure the data how we tag it how we connect it with each other what's the knowledge layer on top of it how do you think about the connectivity between tickers and sectors and companies How does AI actually interrogate that data in a way that is uniform What the common language between all of those data sets And then the last piece is really something that's quite new. It's our institutional knowledge. It's our context. And this is becoming a new layer in our data architecture. How do we tell AI about our processes? How do we make it speak mangroup? How do we tell it the right way to run a backtest? So those are really the three areas that we focus on. Sorry, just pushing on this point further, like all of that makes a ton of sense to me. And especially, you know, the second two are like big problems. And we know that generative AI in particular solves a lot of the unstructured data problem. You can actually get a lot of signal from, you know, yeah, a bunch of text dumped in a file. But what is, what are you building? Like, how does it work? And like, do you have to, can off the shelf frontier models, are they the best for the job? Do you do your own in-house fine tuning of open source models? Like right today what is the best tech stack for that part of the process So we have historically looked at fine for a couple of cases but where we seeing the most bang for our book at the moment is proper tagging and structuring of the data so pre What we found is that if you take a data set like credit card data, for example, AI can look at that, it can see the columns, it can see the rows, but it doesn't really understand the nuances of it. It doesn't really understand what it is. So what we're having to do is invest in trying to add extra color, extra metadata to that information. We want to have descriptors to tell it the nuances. We will say things like, when you look at this data set, each row really means that a person has gone into a shop and bought something. And you tell AI that in plain English, you give it those descriptors. And you do that for all of your data sets. And then the second piece is how do you connect lots of data sets together? They use very different language. They use very different semantics. So investing in a shared language, a shared semantic layer, a shared way of doing things is something that we've had to do so that we tag all over the column. [25:00] with this unified language so that AI can connect different data sets together. Because that's really an important thing in idea generation that knows that a field in one data set is linked to a field in another data set. It can quickly navigate between tickers and sectors for macro insight. What do you think is most important? Having the latest frontier model or a beautiful set of structured, tagged and labeled data? Yeah, I'd say it depends on the task that you're trying to do. If you're looking at a coding task, you really want the latest frontier models. But for a quant research task, I think that looking at the underlying data is what fuels alpha research. Trying to just use a frontier model will get you nowhere. So this is something that comes up in a lot of our AI deployment questions. And I don't know, maybe both of you can take this. But you have all these different teams. And I assume everybody wants really including people who don know much about AI intuitively think oh I want the strongest version of the model I want Opus 4 I want Fable whatever And I assume okay yes for like deep computational tasks or engineering problems, coding problems, yes, probably those are the best. But there are probably a lot of people who do not need anything like that at all. How do you think about the question of internal provision of token consumption and not wasting money by having people use the most advanced models, but also giving people enough green space to explore and figure out what is the maximal potential value that they can get from AI. Yeah, the economics question is an interesting one. It's an important one. What we've done there is, well, towards the end of last year, when we knew growth was likely to accelerate rapidly, we modeled what we thought the company would start to look like in terms of different classifications of users with different use cases, came up with a certain budget. And then what we've done this year is federated that out to all of the business units within the firm. I mean strong believer in pushing decision down to the lowest possible level that it makes sense to do so It allows people to be agile and use the economics that work for them and their departments Of course, budgets are fungible. If departments want to move money from some other spend and say, right, I think we should buy more tokens, then they are free to do that. And the platform that we've built, the AI platform, has a very rich, not 100% complete, but a very, very rich set of models that can be chosen, including all of the main frontier models and a number of the different open source, open weight models. Do you build a model that routes queries to the optimal sort of like cost efficient model? I know there's a lot of interest in this, the sort of classifiers and the big AI labs have them themselves, but the classifiers that route queries, is that something that you built that you have in house or how do you solve that problem? We could do that. We've chosen not to. I think that the reason is that we want people to understand the dynamics of how best to use AI and which models to use. So what we leaned on instead is education So we have a very rich data set of how people are using AI So we have great insight And some of the things that we saw were just quite basic mistakes So often people would be doing multiple tasks in the same context window They'd be trying to figure out the best way to write an investment thesis. And then they were trying to figure out where to go to lunch. And then they were trying to figure out... Asking Fable what the weather will be tomorrow. Yeah, exactly. And that is, I suppose, like a well-known problem if you're in the weeds of AI. But we have 1,700, 1,800 people at Man Group, and the technical understanding of how things work at a fundamental level is varied. So we've been really focusing on education. We're very transparent about the budgets that people have and how they're being spent. and we talk to them about the different classes of models. And it's through that that we have seen better results. And we've actually seen people finding quite creative ways to reduce token spend and contributing that back to the platform as a result of it. Wait, say more about that. Yeah. So for example, if you are using [30:00] coding agent and you want to do some basic git commands to interact with the version control system, often the coding agent will send the command. It will process the entire result that comes back from that command as tokens. Instead, there's some really simple tools and simple basic steps that you could use. Instead of calling an LLN to do inference and tool calls, it can just intercept that tool call and do it outside of the agentic loop. If I was looking at a chart of your overall token consumption, what would, I mean, I assume it's upward sloping despite some of these efficiency efforts, but like how steep is the slope at the moment? So since January, I think token consumption has gone up 86 times. Wow. So it's really, really quite, quite incredible. We were not expecting usage to be the way that it has been. And it's been across the board. It's not just been in the tech and tech adjacent departments. We have seen people in finance people in operations people in the people team using agentic coding workflows And as a technologist that just super exciting to give this new capability this new power to people who haven't been able to use it before. Well, this actually gets to a question that I've been wondering about, and it definitely feels like December and January were just like a very pivotal period. And from your perspective, that sharp inflection point up, how much was it driven by the capability of the models themselves, whether going from an Opus 4.5 to 4.6 and beyond, or this sort of discovery of these very high quality harnesses like a Cloud Code or Co-Work or whatever. I think that's what it's called, that really allows someone to do things with AI that are beyond asking questions and actually manipulate real work. The model or the harness, which would you describe as the key driver of that huge acceleration? I think the two are coupled. I think one of the interesting benchmarks to look at for this is a benchmark called meter. and what that tries to do is it looks at tasks that humans would do for different time periods from a couple of minutes to many hours And what we seeing is that every seven months or so the amount of time that an agentic workflow can go away and do a task is doubling. So now you can ask an agent to do a task that would take a human 16 hours. And that changes the way that you think about teams. It changes the way that you think about interacting with these agents, you go from a place where you're in the loop, you're asking an agent to solve a problem like writing a unit test to a problem like building an entire feature of an application or an entire application in itself. So I think it's the scalability that has allowed larger tasks to be completed that has resulted in larger token usage. 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. Support for the 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. [35:00] 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.com slash market and fund your account in five minutes or less. That's public.com 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. 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 create smarter business ibm i want to go back to i guess the oversight question And we all know that finance is a highly regulated environment And you've touched on this earlier, but you're still approving a lot of these model outputs through humans. So there's some oversight there. And I would assume that when you're approving, if a human is approving a new model or an output, that they have to understand what's actually coming out of it, right? There has to be some explainability there. If I'm a PM or, I don't know, a quant sitting in front of a risk management committee or a regulator, what does explainability actually look like? And how am I translating the model outputs into something that is understandable and also, I guess, defensible? yeah you're right explainability is super important to us and to be clear the sort of business that we're in is not the high frequency trading business where people are constructing huge neural nets and looking through like multi-dimensional spaces and the output not being like at all insurvitable we not in that space i mean our trading and holding period horizons are typically days to weeks to months So all of our trading decisions are ultimately explainable And we never want to be in a position where we don know why that trade happened The AI did it. And going back to some of the things that we talked about earlier, take the example of the AI coming up with brand new trading hypotheses based on what it's seen in terms of content of a data set, what it's seen in the contents of an academic paper and the model will go as far the system the agent will go as far as naming and writing the investment hypothesis it's one of the first things it does before it moves on to writing codes it's giving us a real english description of what it thinks the rationale for the trading signal is i'm curious you know obviously we haven't even gotten to the question of what is the future of labor and the humans in the loop and how many humans in the loop will we need in the future. But one thing I'm curious about as a way to ask this question differently, has AI allowed you to look at markets that you wouldn't have had the bandwidth before So for example let take the Ethiopian Stock Exchange or something like that There a certain amount of human labor that would be required to gain any familiarity with it whatsoever setting aside everything else And no matter how much potential profit there is in a, say, frontier market, there is a minimum amount that's going to cost. And that might take some investment opportunities off the board because the potential profit isn't big enough to justify the spend to getting up to speed. This strikes me as something that AI could potentially help with or solve for of creating a new opportunity set by reducing some of the upfront human labor costs. Is that something that you think about or have seen specifically so far in terms of mangrove? I think it's likely happening incrementally at the margins. margins. If we start to sum up all of the different micro-augmentations that we see across the firm, for example, and here's a related one, someone had built a relatively simple AI system to take data from complex instruments, PDFs, specifications, and automatically populate our reference data store with that. So yes, I think it's absolutely happening, but it's the sum of a [40:00] lots of different parts across the firm. I think what we're seeing as well in the systematic space is that you need a few prerequisites to build a systematic strategy. You need to have some connectivity to trade the instrument and you need to be able to understand what the price of a market is. Those are two real fundamental things. And for developed markets, that's really easy. You go and look at the order book, but for less developed markets, things like crypto securitized credit they're often harder to connect to they're they may be traded more verbally or the contracts are complicated and it's difficult to understand the price where there's these unstructured data nuances in the derivation of that price we've seen that AI allows us to think about accessing that market in a systematic way earlier than we could before. Just going back to the labor market side of things, I guess, you know, if I'm a PM at Man Group and more and more of my job is using AI for research or even basically supervising agents that you know I maybe help develop What does that mean for what you looking for in terms of talent Are you looking for engineers who can tweak these models? Are you looking for more traditional investors who I guess have stronger intuition about how these things might play out or some sort of combination of characteristics? What do you look for now? I think very fair to saying this is like something I've been making a strong case for that everyone we hire now into the firm should up the bar with respect to AI regardless of the role that they're coming in to do and I think that applies as much in the operation space as it does in the front office space. What does that mean? So what like someone like okay I'm capable of upping the bar with respect to ad like what does that say more about okay in the recruiting process what does that person look like that that means being as familiar with the technology as you could reasonably expect a person to be given the wealth of information that out there that when i when i hiring people you know what sort of people do i want i I I want bright people I want people who are motivated to get things done and are passionate about the subject matter, their field of expertise. And I think it's very hard to fulfill all of those criteria, particularly nowadays, and say, well, AI, I don't know much about it. I don't really use it as part of my job. Yeah, I think that the other piece that you want is somebody who's a bit more of a long-term thinker, somebody who really wants to automate a process from end to end. They are happy not to be in the weeds, in the loop. And when you think about technologists, that's actually quite difficult. People love being in the weeds. They love debugging issues, getting into the nitty gritty. But actually, fundamentally, you want to level up. You want to be a kind of a conductor of these agents. You want to be in charge of the end-to-end process rather than necessarily being in the weeds, almost like a manager who has a lot of technical expertise. So somebody who is thinking about things in broader strategic terms is much more valued than they used to be So this leads to the other thing I wanted to ask which is you could see the arrival of AI tools generating like two different outcomes here where you have some people who are just really, really good at using AI, and they become superstars and orchestrators of a bunch of different agents, as you put it. Or you could have this sort of democratizing effect where maybe you're a junior employee with not as much experience, but now you can automate a bunch of tasks. You can learn from AI, you can use it for research and things like that. What are you seeing more of at the moment, the superstar dynamic or the democratization of skill sets throughout Man Group? I think we're seeing both, genuinely both. I think given the size of the firm, though, we're seeing more of the latter. I mean, as I said at the start of the chat, almost, or everybody's using it day to day. But examples of a huge, genuine, right at the cutting edge of thinking, I think they're naturally more rare, but there's a fair few of them even saying that. They often cut across multiple teams as well, and that is hard because you need to move from a space where you spend most of your time executing to a space [45:00] where you spend most of your time planning. The time to execute is just going down and down. It's becoming cheaper and cheaper to do that. You can build code. You can build features very quickly. So the focus really needs to be on what should we build? How does it connect together? And what is the process that we want to develop across multiple teams? And it's very difficult sometimes to take people out of their seat and get them to collaborate and plan a workflow. rather than just going and trying to build a proof of concept and execute on the idea. Let's talk more about the 86x increase in token spend. If we were having this conversation back in February or January, a lot of the chat would have been very much about like, what does this mean for legacy software companies? Because that's when the big software company software sell off was, you know, really was quite intense. But I feel like this conversation in July is becoming like, no, these companies are not going after the world of software. They're going after the world of labor. And you see a lot of these conversations like people talk about the ratio of token spend to employee salaries etc And that that is the TAM that it like all human labor And maybe we not going to get there for a while I kind of hope not. But is token spend part of your tech budget? Or is it like something that is a true line item that's distinct, that's more on par with labor? And when you think about Man Group in 2027, 2028, do you talk about expected ratios of salaries to token spend? no we generally haven't started having that conversation yet i mean i would guess at some point uh it will come and the company's set up in such a way that we want to direct resources to where they produce the best economic outcomes for us so i'm sure that time will come one of the interesting things in the token budgeting process i suppose is that what we're increasingly seeing is the the spend is actually not by people is by agents and the agents relate to workflows and who owns the workflows is it this department is it that department and that's a new problem for us one that we haven solved yet but it a great problem to have i brought up earlier the sort of machine learning parallel I guess And I know you guys don do a lot of high frequency trading, but we've certainly seen a dynamic in HFT where everyone's competing and it's sort of a race to the bottom where I can't even remember where we're at in terms of micrometers. Yeah, in terms of second increments or time increments, but it's a sort of race to the bottom dynamic. Would you expect AI driven alpha to get sort of competed or arbitraged away relatively quickly as everyone seems to be hopping on the same bandwagon? Or are there certain advantages? You know, we talked about data, for instance, scale, perhaps that you would expect to stick around for some time? Yeah, I'd say it goes back to your question around where is the alpha. And what is true is that it is easier for people to onboard data sets, analyze them and build features. But that doesn't mean that they can trade them. We've been doing this for decades. So what we have are capabilities to actually access these markets. We have the relationships with the brokers We have access to data that isn just available off the shelf So it putting all of those things together putting those expertise together the access to markets all of the data that we have the rich market data and then giving AI access to the capabilities such as backtesting compute It's this whole network, this whole ecosystem that together, I think, drives alpha. There isn't one code repository in my group that I can point out and say that's where the alpha is. It's really this network of different systems that interact with each other. So I think that definitely some features of data sets will become table stakes. They go from being alpha to being a risk factor, but because everybody has them and it moves the market. but it isn't the only way that we make money. We are able to connect different data sets, different ways of doing things together, and then actually trade on those signals. Gary, I want to go back to something you said earlier when we were talking about bottlenecks, and it's like, sure, everybody wants more data, everyone wants more compute, no one would complain about these things, but where the rubber meets the road is like, does the institution have [50:00] the capacity to actually, maybe from a cultural standpoint, a sort of hierarchy standpoint, to actually get the most out of these tools. And this is clearly a very hot area. And so, for example, just last week, Microsoft announced a new thing, which they're calling the Frontier Company, which is basically a new like sort of subdivision that seems to want to specifically solve this problem, go into an organization and figure out this optimal structure. And, And arguably, this is even like what a company like Palantir is trying to do, which is the forward deployed engineers. We know about Claude sending or Anthropik sending engineers inside Goldman Sachs to really leverage. I hate using that word because it's so cliche, but yes, leverage these tools. What specifically are you seeing happening on that front? Do you have third party companies who are coming to you and saying, look, we can work with you to find what is the org structure of the future for Man Group such that it's getting the most out of these tools for a long time? Yeah I did smile at the multi dollar forward deployed engineer division that you just mentioned Partly because I mean that the way we been set up internally here for about I think 15 years We're very big on platforms, and we've got a bunch of teams, and Tashara runs one of those, that build out these big cross-cutting elements of platform technology, in his case, the data and AI platform. But a big part of the tech team are already, and have been for a decade and a half, forward deployed engineers, sitting with our quants, sitting with our discretionary managers. And one of my jokes often to people I'd be interviewing to come join the team is like, right, let's go walk across the fifth floor here in Ribbank House. And I want you to tell me who the engineers are and who the quants are. And I bet you're going to get it wrong because what you'll see is very similar stuff on that screen. And that holds just as true today as it did a decade ago. One last question, but another thing I'm curious about in the investment context. So we know that AI works when there's a big pool of data and that it can pull in the unstructured data and the structured data, et cetera, and communicate across them. In a entity such as man group are there any alignment issues in which you know if i have a subject matter expertise that generates alpha this might be why i have a seat at an organization or i might be a rainmaker you know you hear this at all kinds of different firms where compensation is linked to someone's like deep expertise in some area. Do you think about alignment so that the firm, the franchise man group is actually capturing some of the expertise and data of the superstar? How do you get them to sort of, I guess, contribute as much information as possible to this thing that requires a lot of data and information? I think in some areas that's still a little bit of a work in progress. I'm saying in some areas because in others, notably the systematic, the quant areas of the business, this highly collaborative approach and shared code bases, it's just been the way that those areas have worked for a long time. Now I not saying that you walk across the discretionary part of the floor and all of the fundamental investors are going to be quite so free and open I talked to a bunch of them about their processes and open with 90 percent But there the 10 percent Well you know this is where my personal value add lies I'm not so comfortable talking about about that. But even that said, there's a very decent and genuine amounts of collaboration there as well. And it's only when you get perhaps to the very sensitive areas that people might be a little bit more reluctant to talk. Yeah, I think the high level, there's a huge amount of shared workflows. If you think about the way that we backtest, the way that you read an investment report, these are all workflows that are people's expertise, but they're not necessarily the 10 percent that produces the returns. so those areas are encapsulated in ai playbooks ai skills that we put in our knowledge platform and they can be used across the floor whereas some of the particulars around how the strategy works and the investment pieces are somewhat held back in certain cases that makes sense all right gary and tishara thank you so much for coming on odd lots really appreciate your taking your time and talk about where you're at. Thank you. My pleasure. Thanks a lot. [55:00] you know what i think was really interesting about that conversation in part is like there is so much to figure out still right i mean just the basic token budget and like where stuff gets allocated like 86x in since january it's pretty crazy well this is the other thing i was thinking like 86 times growth in token usage like at some point that needs to show up in another concrete number whether it's expense reduction or revenue generation income generation and i don't know how much like leeway there is for that no and you figure like right now so you have this 86x explosion but as they say they're still in the moment where they haven't gotten to we want to build like an internal router to minimize this. So even with the 86, they're still in the phase where it's okay, you know, figure out what model you want to use and experiment, et cetera, which to my mind is actually kind of bullish. If you think about it for token spend that you can grow 86X and it still doesn get you to the point where like oh we got to really like clamp down on this Like maybe you know who knows like what the point is for a lot of firms that are just starting out where they actually have to start imposing some token austerity I guess we'll find out at some point. But the other thing that stood out to me, you know, you asked the question about how do you get how do you get workers to give up their own secret sauce that basically is responsible for them having a job in the first place? keystroke surveillance solves all of this, right? Yeah, you know, one of those articles, I have to say. You don't need them to offer it up. You just take all the data that they're producing. I have to say there was some story that came out a while back about Ometa is going to start training its models on its own. And I was like, they weren't doing this already. I was actually really surprised that this wasn't already. Especially in finance and investment, which is a highly regulated industry already, and I'm sure is monitoring pretty much everything anyway. Yeah, I kind of assumed that all of these companies are really building models or are they using all of the data that their own employees were generating. But yeah, you have to, it's like, oh does the person does the Rainmaker suddenly just start writing everything down on pen and paper et cetera They not implicitly uploading all of their knowledge to the AI I think that a pretty interesting question in itself. All right. Clearly, lots of interesting questions, but shall we leave it there for now? Let's leave it there. Okay. This has been another episode of the All Thoughts Podcast. I'm Tracy Alloway. You can follow me at Tracy Alloway. And I'm Joe Wazenthal. You can follow me at The Stalwart. Follow our producers, Carmen Rodriguez at Carmen Armand Daschle 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 on all of 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 talk about how finance firms are actually implementing 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. 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. 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. [60:00] Learn more at the Hartford.com slash risk mitigation. Policies provided by Hartford Fire Insurance Company and its property and casualty affiliates, Hartford, Connecticut. 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.