Jane Street Makes Deep Learning Top Driver of Systematic Trading

Jane Street将深度学习作为系统化交易的核心驱动力 现成机器学习模型因极端噪声在金融市场失效，迫使量化交易机构构建定制化的基础科学架构，这些架构现已主导其核心策略。 在与Grant Sanderson的一场坦诚圆桌讨论中，Jane Street的工程师和研究人员通过杂货店管理库存和合理定价的比喻，揭开了做市业务的神秘面纱。他们概述了扁平化、去个人英雄主义的工程文化，重点介绍了诸如共享集群算力的实时拍卖等独特内部机制，以及横跨基础设施与研究的流动性职位头衔。讨论小组成员详细阐述了公司二十年来从基础电子表格到深度学习的演变过程——深度学习现已成为其系统化交易的最大单一驱动力，并强调高噪声的金融数据需要前沿的基础科学研究，而非套用现成模型。小组成员还分享了各自多元的职业发展路径，并赞扬了互助且高留存率的工作环境，同事们在其中如同大学同学攻克难题般协同合作。 机器学习已迅速演变成为Jane Street系统化交易的最大单一驱动力，推动其业务从早期的电子表格模型转向前沿深度学习与学术级基础科学研究。 (7:57) Jane Street并未依赖自上而下的管理层硬件配额，而是通过动态的实时内部拍卖引擎在整个计算集群中分配GPU和CPU资源，所有团队均可查看任务和效率指标。 (3:33) 相较于仅具备狭窄领域专业背景的求职者，该公司优先聘用适应能力强、学习速度快的解题型人才，极高的长期员工留存率使这一策略具备可持续性。 (18:04) 内部诸如“交易员”和“研究员”等职位头衔主要是历史命名巧合所致，并非严格的企业层级制度，技术岗位实际上处于涵盖工程、系统开发与建模的流动连续谱系之中。 (10:03)

Jane Street engineers and researchers explain to Grant Sanderson how their market-making operations rely on an academic, low-ego culture, internal market mechanisms, and foundational machine learning rather than off-the-shelf algorithms. Collaborative Culture and Internal Resource Markets (0:00) Jane Street maintains an intentionally flat, low-ego environment designed to prevent status signaling and groupthink. Group deliberations regularly begin with unanchored formats where participants state or record views independently before debating evidence, allowing perspectives to shift fluidly. Formal management exists for long-term project coordination, but daily communication cross-cuts functional teams without hierarchical gating. This decentralized ethos governs infrastructure allocation. Rather than relying on bureaucratic quotas, the firm operates a real-time internal auction for its shared compute cluster, the hive. Researchers submit dynamic bids reflecting their projects' expected value. Workloads and metrics like model flops utilization remain visible across teams, empowering any colleague to pause malfunctioning runs or assist in debugging. While an absence of strict safeguards once enabled a researcher to unintentionally alter bids across the entire cluster, operational controls are adjusted iteratively to maintain experimental freedom. Market Making and Custom Machine Learning (5:51) Panelists liken market making to a retail grocery store: instead of forcing market participants to trade directly with disparate counterparties, the firm holds inventory, prices instruments fairly, and provides continuous liquidity for assets like exchange-traded funds. While external observers often assume quantitative trading involves feeding market data into standard, off-the-shelf deep learning models, panelists emphasize that extreme signal noise in financial data makes this ineffective. Addressing this noise requires treating quantitative modeling as basic science and conducting frontier research with custom architectures. Team members note that modern machine learning has evolved from early experimental efforts into the primary driver of systematic trading at the firm, with researchers perceiving their progress to be at the base of a steep S-curve. Fluid Roles and Twenty Years of Evolution (10:03) Internal job titles are largely historical accidents, departing sharply from traditional Wall Street divisions of manual execution desks and isolated research silos. Technical roles exist across a spectrum comprising developer tooling, trading-adjacent infrastructure, and quantitative research. Desk neighbors frequently share technical tasks regardless of whether their formal designation is trader or researcher. Technological tools have evolved in roughly five-year cycles, transitioning from simple human spread-checking in Microsoft Excel spreadsheets to cluster-based linear models, gradient-boosted decision trees, deep neural networks, and modern transformer architectures. Despite this technical overhaul, long-tenured staff note that the firm's core personnel and culture have remained consistent over two decades. Hiring Adaptable Generalists for Long Tenures (16:15) Hiring focuses on general problem-solving ability, intellectual curiosity, and adaptability over specialized domain familiarity. Candidate assessments favor probability puzzles and games over resume recall or machine learning course credentials. Long employee retention justifies this strategy, making it viable to invest significant time retraining staff as technological requirements shift. Panelists describe non-linear paths into the firm, including a pure math student who completed a multi-year PhD leave of absence mid-career, and an engineer recruited in 2005 from an OCaml hobbyist mailing list. High retention is attributed to an environment resembling college study groups, where colleagues actively support one another, exercise substantial personal agency, and organize informal chalk talks on advanced mathematical and physics topics. The discussion demystifies modern quantitative trading by illustrating that sustained performance stems from scientific problem-solving, custom research architectures, and high-trust peer collaboration rather than black-box algorithms or rigid corporate structures.

Jane Street 团队在与 3Blue1Brown 的访谈中指出，公司依靠去中心化、低自负（low-ego）的协作文化与内部算力竞价机制，已将机器学习发展为其系统化交易的核心驱动力；其研究更接近应对高噪声金融数据的基础科学探索，而非单纯套用现成模型。 去中心化文化与算力实时竞价 (0:00) Jane Street 内部刻意剥离了等级观念与自负倾向。团队会议常采用防锚定机制，参会者在听取他人观点前先写下独立判断，随后通过贝叶斯式探讨促成观点自然收敛。跨业务沟通不设层级壁垒，工程师与交易员可直接跨组交流。这种扁平化模式也直接延伸至底层系统，公司并没有采用自上而下的行政审批分配硬件，而是为内部集群“hive”搭建了实时竞价引擎。研究员需评估实验与模拟任务的实际业务价值来进行动态竞价。 全公司对集群资源形成共有监督氛围。所有任务对员工透明开放，如果有人发现某个作业持续报错或算力利用率（MFU）过低，可主动暂停该任务或协助排障。团队早期甚至保持广泛的全局读写权限，曾有一位研究员在 Jupyter notebook 中调整参数时遗漏参数，意外修改了集群上全部任务的竞价。公司并未因此施加僵化的事前官僚限制，而是依据具体故障模式迭代完善权限与工具。 做市逻辑与机器学习的科研范式 (5:51) 在解释做市商角色时，团队将其类比为“股票领域的杂货店”。正如杂货店建立库存以避免消费者直接前往各个农场采购一样，做市商为 ETF 等金融工具持续提供双向流动性并承担定价风险，确保衍生份额与底层资产价格锚定。随着市场预测需求演进，现代机器学习成为了定价与库存管理的关键手段。受访团队强调，外界常误以为量化机构只需租用大型 GPU 集群运行现成开源模型即可获利，但金融数据具备极高的噪声，这导致简单套用通用架构彻底失效。 因此，该机构将机器学习视作前沿基础科学研究，专注于定制底层架构以解析微弱信号。团队成员表示，机器学习已从实验性项目迅速崛起为驱动其系统化交易的最大单一力量。整个团队认为当前的机器学习应用仅处于陡峭“S 曲线”的底部，未来仍有广阔的发展空间。 模糊角色分工与二十年技术迭代 (10:03) 与传统华尔街机构中“大批交易员手动录入订单、辅以小规模独立研究组”的割裂结构不同，Jane Street 的岗位边界呈现连续的光谱分布。所谓“交易员”的称谓更多是历史沿革，其日常职责往往高度偏向定量研究；许多员工甚至不清楚隔壁同事在系统名录中的官方头衔是交易员还是研究员。技术人员沿统一轴线分布：包括维护代码审查工具与 PB 级市场数据管线的底层基建工程师、精通交易逻辑的系统开发人员，以及训练长周期神经网络的量化研究人员。 在公司任职二十年的 Craig 回顾指出，早期交易主要依赖人工观察两只相关证券的价差，并在 Microsoft Excel 表格中直接运行交易系统；当时在集群上记录数据并跑通线性模型已被视为尖端技术。二十年间，技术栈大约每五年发生一次重大更替，依次走过了手动规则、线性回归、梯度提升决策树（GBDT）、深度神经网络直至现代 Transformer 架构的演化历程。尽管技术手段彻底重塑，但核心团队稳定，工程导向与低摩擦的协作氛围始终未变。 高留存率背后的招聘与学术氛围 (14:35) 量化金融行业普遍存在频繁跳槽现象，而团队成员将本公司的高留存率归因于高自主权与类似大学探讨课业（problem sets）的同侪氛围。同事间互相关注彼此的长远成长，员工倾向于以每日的实际收获、自主空间和智力挑战来决定去留，而非预设固定工作年限。基于这种极高的人才留存预期，公司的招聘哲学始终定位于挑选聪明、务实且适应能力强的问题解决者，而非仅具备特定课程背景的技术专才，因为公司有充分的时间对员工进行长期的机器学习再培训。 员工的入职轨迹与日常工作同样兼具多样性与浓厚学术色彩。例如 Nitya 原本立志从事纯数学研究，通过实习加入后，公司甚至支持她请假数年完成博士学位再重返全职岗位；Alok 在实习阶段便将原本被视为纯理论的 Huber M 估计量等稳健统计学方法运用于真实异化数据处理；而 2005 年通过 OCaml 邮件列表偶然求职加入的 Craig，至今仍保留着极客好奇心。在办公区的小会议室里，员工经常自发组织关于统计力学、配分函数、费曼图乃至图指数化（e 的图次方）等前沿物理与数学白板研讨，展现出高度智力驱动的极客文化。 访谈通过解构做市机制、算力市场与技术跃迁，展现了 Jane Street 将学术科研习惯融入高频交易的工程实践。不过，对于面对极端噪声数据时定制化深度学习模型相较于传统统计套利策略的超额收益持续性与鲁棒性界限，讨论中并未展开更多定量论证。

## Background 该视频由知名量化交易公司 Jane Street 于 2026 年 8 月 21 日发布在 YouTube 官方频道。内容为知名数学科普频道 3Blue1Brown 的创作者 Grant Sanderson 造访 Jane Street 纽约办公室，与四位在职员工（Nitya、In Young、Craig 与 Alok）展开圆桌对谈，探讨公司的核心业务、算力集群的实时拍卖机制、20 年来的技术架构演进，以及员工多元的个人职业轨迹。 ## TL;DR 视频深入揭秘了 Jane Street 的技术运作与去中心化文化：它将自身定义为提供流动性的“股票杂货铺”，如今深度学习已成为其系统化交易的最大单一驱动力，内部甚至通过全员透明的“实时算力拍卖”来调配集群资源。员工回顾了技术栈从 2005 年依赖 Excel 电子表格到如今现代前沿深度学习与基础理论研究的二十年变迁。此外，受访者还分享了各自从攻读纯数博士、在 OCaml 邮件列表偶然求职到加入公司的真实故事，展现了公司偏好选拔并长期投资于具备极高适应力的“通才”的招聘哲学。 ## Atmosphere & style 整场对谈呈现出浓厚的学术沙龙与极客智趣氛围，没有传统华尔街金融机构的浮夸或官僚感，取而代之的是纯粹探索真理的低自我（low-ego）沟通。Grant 的提问敏锐且直击关键，受访者之间的互动自然松弛、默契十足，并不时穿插诸如面试扔筹码、白板即兴推导统计物理与费曼图等生动趣事。观众在聆听过程中能直观感受到顶尖智力高密度碰撞时的兴奋感，以及一种兼具工程师浪漫与理性求知欲的独特企业文化。 ## Bulletin - [03:33](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=213s) Jane Street 内部并不依赖传统科层审批，而是通过内部实时算力拍卖系统调配全公司 GPU 集群，由工程师和研究员自主评估作业商业价值并进行动态竞价。 - [07:57](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=477s) 团队纠正外界对量化交易的刻板印象：金融数据信噪比极低导致开源现成模型基本失效，公司正经历机器学习爆发式技术 S 曲线的起步期，深度学习已成为驱动其系统化交易的最大单一引擎。 - [12:44](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=764s) Jane Street 过去 20 年经历了核心建模工具每 5 年一次的大迭代，交易体系从 2005 年依赖 Excel 电子表格和简单线性回归演进为深度神经网络及梯度提升树。 - [19:02](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=1142s) 公司将极低的人才流失率与招聘策略深度绑定，优先选拔具备极强适应能力的通才，从而使得投入长达一年的前沿技术再培训在商业回报上具备可行性。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=0s) Jane Street 推崇低自我的真理探索文化，采用扁平的“单一大团队”组织架构，鼓励员工跨组自由脑暴与直接沟通，初期拥抱适度混乱以激发创意。 - [03:33](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=213s) 算力集群采用全透明实时拍卖机制，员工可全局查看作业并直接介入干预异常任务或低下 MFU，形成独特的基层有机治理模式。 - [05:51](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=351s) 公司用“股票的杂货铺”比喻其做市商业务本质，核心在于通过计算资产公允价值、双向挂单与撮合流动性，解决跨国及复杂资产的配置便利性。 - [07:57](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=477s) 现代做市商的预测逻辑类似杂货超市预估节日供需，必须进行大量原创性底层基础科学研究，现成开源模型无法胜任极低信噪比的金融数据。 - [10:03](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=603s) 公司内部“交易员”与“研究员”边界高度模糊且大多从事量化研究，从底层基础设施到交易系统再到长期深度学习模型呈现出高度流动的连续光谱。 - [12:44](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=764s) 二十年技术变迁剧烈，交易系统从早期人工观察 Excel 电子表格价差演化为现代庞大的神经网络架构，而核心技术范式几乎每五年彻底重塑一次。 - [14:35](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=875s) 极高的人才留存率源于低内耗协作与高自主权，员工之间形成了如同大学室友般彼此全心投入对方成功的同侪支持文化。 - [16:15](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=975s) 公司制度具备极高包容度与弹性，甚至支持员工在中期暂时离职攻读数学博士学位并在毕业后重返核心岗位。 - [19:02](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=1142s) 面试设计刻意规避记忆型题目与特定课程套路，通过高压博弈游戏和陌生数学难题重点考察候选人在未知情境中的逻辑推演与适应能力。 - [21:29](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=1289s) 大学阶段看似远离应用的抽象数学理论（如 Huber M 估计量等稳健估计理论）在处理金融现实中的畸变和高噪声数据时发挥了核心作用。 - [24:28](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=1468s) 2005 年元老员工 Craig 因在 OCaml 邮件列表看到帖子而入职，源于其纯粹出于个人趣味用小众函数式语言折腾黑白棋与分形几何的极客精神。 - [26:17](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=1577s) 办公区常年保持浓厚的学术探索风气，员工闲暇时通过即兴白板讲座讨论配分函数、费曼图乃至抽象图论等高深理论物理课题。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=0s) Jane Street 的低自我沟通文化与“单一大团队”架构 Grant Sanderson 首先向来自 Jane Street 的员工们抛出核心问题：Jane Street 到底是一家怎样的公司？抛开表面的量化金融标签，其独特的公司文化到底体现在哪里，在技术上又究竟在做什么？ 针对文化的独特性，团队成员指出 Jane Street 呈现出一种“最美妙的怪异”。在公司日常交流中，人们几乎看不到争强好胜的“自我（ego）”，没有人试图在对话中证明自己比别人聪明或单纯为了辩赢一个观点。相反，交流的出发点完全是为了探寻真理。员工们对未知事物极其开放，大家在踏入任何一场对话时，甚至都抱着“自己的观点在离开时可能会被改变”的心态。例如在会议讨论中，团队经常会让所有人先亮出各自未经他人影响的初始想法，充分阐述后大家再分享吸收他人意见后的后续想法；令人惊讶的是，最终大家的观点往往会走向高度收敛。这种习惯甚至在员工与大学同学聚会时显得格格不入，同事们不仅带有深厚的贝叶斯思维习惯，还流行着诸如“那并非不合理（That's not unreasonable）”等极具辨识度的“Jane Street 语录（Jane Streetisms）”。 除了讨论氛围，Jane Street 最大的组织特征是“整家公司如同一体（there's kind of just one company）”。与传统大型企业中复杂的科层汇报链条不同，这里的团队之间不存在“先让我的主管去找你的主管预约时间”的官僚作风。任何员工如果对某个技术或业务产生疑问，都可以直接走到另一个小组的办公桌前打招呼：“嘿，我正在思考一件事，你们有人能抽出哪怕一分钟脑暴一下吗？我觉得你们在这方面肯定很有经验。”当 Grant 质疑这种去中心化的模式难以规模化、容易演化为纯粹的混乱时，团队承认这确实伴随一定程度的混沌，但在项目初期激发想法的轻量级阶段，拥抱这种混乱远比过早建立层级约束更有效；只有当确定要启动多人的长期大项目时，正式的项目负责人和组织结构才会介入。 > “在大家展开对话时，目的并不是为了满足虚荣心，或是非要证明某个观点、向人炫耀自己有多聪明；如果是那样的话会非常累人。真正的做法是以一种纯粹觉得某个话题很有意思的态度去切入，通过把事情摊开聊透来接近事实真相。” —— Nitya --- ### [03:33](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=213s) 算力集群的实时竞价系统：集体责任与有机治理 这种“拥抱混乱”与扁平信任的组织理念，在 Jane Street 的技术基础设施管理上得到了极其戏剧化的体现。与大型科技公司内部通过繁杂的官僚审批和芯片额度指标来调配 GPU/算力不同，Jane Street 直接为全公司的计算集群运行了一个“实时算力拍卖系统（real-time auction for our compute）”。所有的计算资源完全向全员开放，没有设置死板的护栏（guardrails）。每位工程师或研究员的任务，是自主而现实地评估自己所提交的作业（job）相对于全公司其他任务的商业或研究价值，并据此给出竞价。 这种机制培育了一种强大的集体责任感： - 算力系统的运行状态对全员透明，所有人都可以看到当前运行的作业。 - 如果有人离开工位时提交的任务报错或异常循环，其他人路过看到后不会置之不理，而是会主动帮忙暂停任务并留言提醒。 - 如果某位同事注意到某个任务的模型算力利用率（MFU, Model Flops Utilization）极其低下，他可以直接走过去沟通：“嘿，我看到你跑的这个作业 MFU 表现非常糟糕，到底发生什么了？”问题往往在基层对话中几分钟就得到解决，无需上报管理层。 - 整个集群系统甚至曾一度是“全员全局可读写（globally read-write）”的。团队笑谈过往的离谱轶事：曾有一位毫无戒心的研究员在 Jupyter Notebook 中调整自己任务的出价时，不小心漏传了参数，导致该命令瞬间重置了集群上所有任务的出价。这种粗糙而强大的系统往往是随着需求自然演化修复，而不是一开始就设置一堆限制性条框。 ```mermaid flowchart TD accTitle: Jane Street 算力拍卖与协同机制 accDescr: 展示研究员提交算力任务、参与实时竞价，以及全员透明监督与协作纠错的闭环流程。 subgraph Cluster["统一算力集群 (Hive Cluster)"] Auction["实时拍卖机制 (Real-time Auction)<br/>按作业预估价值自由竞价出资"] Jobs["运行作业 (Running Jobs)<br/>对全员透明全局可见"] Auction --> Jobs end User["研究员 / 工程师"] -->|"自主评估作业价值并出价"| Auction Jobs -->|"暴露指标 (如 MFU、报错率)"| Monitor["全员自由观测与干预"] Monitor -->|"发现异常直接线下走动沟通"| User Monitor -->|"紧急情况下暂停或协助调整"| Jobs ``` --- ### [05:51](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=351s) 从“股票的杂货铺”到前沿“AI 实验室” 对于外部世界而言，Jane Street 的业务性质始终蒙着一层神秘面纱。Grant 指出，外界对它的认知通常在“宽客金融（Quantitative Finance）”这种模糊套话与涉及“商业机密”之间摇摆。团队回顾了创始人 Sandor 曾经给出的官方拗口定义——“通用型金融服务流动性提供商（generalist financial services liquidity provider）”，但大家公认这种说法对普通人毫无助益。 为了向大学生和公众通俗解释，团队经常使用“股票杂货铺（grocery store for stocks）”的比喻来描述其做市商（market making firm）的核心基因： 1. **聚合便利性**：如果没有杂货铺，消费者买苹果必须去果园，买牛奶必须去奶牛场；做市商就像杂货铺一样，在市场上随时挂出买价和卖价，使买家和卖家能够随时完成兑付。 2. **公平合理定价**：杂货店如果把一个苹果标价 10 美元就无人问津，做市商同样必须精准计算资产价值，保持极具竞争力的买卖价差。 3. **跨境与复杂流动性转换**：以交易所交易基金（ETF）为例，如果一个身在美国的普通投资者想要配置中国或海外新兴市场的股票，直接跨国购买极其困难，而买入对应的 ETF 则是最高效的途径。做市商的核心工作之一就是确保 ETF 价格与底层一篮子股票公允价值紧密锚定，并在两端提供流动性。 > “我们核心业务之一——这在很多年前绝对是无可替代的核心，但我认为至今仍驱动着我们交易哲学的——就是成为一家做市商机构。我对大家打的比方就是：它就像是一家‘股票杂货铺’。” —— In Young 然而 Grant 敏锐地指出，近年来从 Jane Street 传出的技术基调越来越不像一家传统的交易公司，反而与前沿的“AI 实验室”别无二致。对此，团队解释道，杂货铺与 AI 实验室之间看似遥远的鸿沟，在深层预测逻辑上其实是相通的：就如同一家现代杂货超市必须极其精准地预测万圣节前夕需要储备多少南瓜灯一样，金融市场的定价与撮合背后，本质是对海量多模态数据进行前沿的统计学习与深度表征，深度学习和大规模计算早已成为驱动其系统化交易最强劲的核心引擎。 ### [07:57](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=477s) 预测供需与打破对做市商的刻板印象 做市商的核心工作之一在于预测供需关系以提供流动性，比如根据季节或事件（如万圣节）提前准备库存，满足未来市场的交易需求。然而，公众和外部求职者对这类金融业务常常存在根深蒂固的误解。许多人以为这只是一些套利机制或依赖传统资产管理机构来维系平衡，但现实完全由高度复杂的预测模型所驱动。 ```mermaid flowchart LR accTitle: 做市商预测与流动性提供流程 accDescr: 展示从预测未来供需到通过机器学习建模，最终向市场提供流动性的流程闭环。 subgraph S1["市场预测阶段"] A["市场周期与事件<br/>(例如万圣节等需求)"] --> B["预估未来供需缺口"] end subgraph S2["策略执行阶段"] B --> C["机器学习预测模型"] C --> D["建立并持有合适库存"] D --> E["向市场提供即时流动性"] end ``` 谈及行业误区时，一位受访者指出，外界不仅严重低估了 Jane Street 内部机器学习应用的计算规模，更低估了其工作的原创性与前沿性： > “I think there's a misconception not just about the scale of machine learning that we're doing, but the novelty of it. Like I think it's like kind of easy to imagine that mostly what we're doing is we're taking some like models off the shelf and maybe we like bought a bunch of GPUs and then we're like running it. And that like basically doesn't work at all.” 在金融数据领域，信噪比极低，各类数据充斥着巨量噪声，直接套用开源现成模型根本行不通。为此，团队不得不开展大量底层基础科学研究（Basic Science），从第一性原理探索前沿理论，并紧跟甚至推动学术界的研究边界。虽然外界对这一转变仍有信息差，但在过去几年里，Jane Street 内部的机器学习实践经历了爆炸式增长，正如受访者形容的那样，目前才刚刚站在技术跃升“S 曲线的起点”，而机器学习早已成为“量化系统化交易的最大单一驱动力”（the single largest driver of our systematic trading）。 - 外界常误以为金融交易仅仅是采购一堆 GPU 并运行开源标准模型，实则由于金融数据极高噪声的特性，必须进行大量的原创性基础科学研究。 - 过去几年内，机器学习技术的落地迎来了巨大突破，目前内部正处于这波爆发式技术 S 曲线的起步阶段。 --- ### [10:03](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=603s) 职能分工的真实形态：打破“交易员打字下单”的幻想 当银行或其他传统交易机构的人加入 Jane Street 时，往往会大吃一惊——这里的组织氛围更接近科技巨头或学术机构，而非刻板印象中的华尔街金融机构。最大的误解之一来自于团队规模：外部观察者看到公司拥有一大群“交易员”（Trader），而专门设立的“研究团队”（Research Group）规模却相对较小，便脑补出数以百计的人在彭博终端（Bloomberg）上手工敲键盘挂单的场景。 事实上，这种职务名称纯粹是历史遗留问题。在 Jane Street，“交易员”这个职位在很大程度上本身就是从事量化研究工作的，整个公司的系统化程度远超外界想象。甚至在座的员工内部开玩笑说，随便挑出一位同事，大家往往都说不清对方在花名册上的正式头衔究竟是“研究员”还是“交易员”。公司内部并不依赖僵化的层级划分，而是一种更具弹性的流动态网络： ```mermaid flowchart TD accTitle: Jane Street 内部工程与研究光谱 accDescr: 展示从基础设施工程到前沿研究算法的连续职能光谱。 A["核心基础设施工程师<br/>(以代码审查、分支管理、PB级市场数据处理为主)"] B["交易系统软件工程师<br/>(精通代码且深入理解交易机制)"] C["短期视角量化交易员<br/>(关注具体证券、日常系统表现与即时交易)"] D["长期视角量化研究员<br/>(提交集群计算任务、训练深度学习模型与开发工具)"] A <--> B B <--> C C <--> D ``` - **基础设施工程师**：服务于其他程序员，搭建代码审查流程、分支管理系统，以及负责摄取、清洗与管理高达 PB（Petabytes）级别的海量市场行情数据。 - **交易软件工程师**：主攻软件开发，同时对交易业务和市场机制有极深刻的理解。 - **交易与研究的光谱两端**：一端更聚焦于具体证券的实时微观结构和每日交易系统表现；另一端则提交大规模计算集群任务，训练更长周期迭代的机器学习模型，二者之间存在广泛的重叠和流动空间。 --- ### [12:44](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=764s) 二十载技术变迁：从 Excel 电子表格到神经网络 Craig 作为在 Jane Street 工作了 20 年的元老级员工，回顾了过去二十年技术栈的沧桑巨变。在 2000 年代初期，如果团队能够把交易日志记录下来，并在小型计算集群上训练一个线性模型，大家就已经觉得“非常高大上”（pretty fancy pants）了。当时大量交易是直观且简单的：只要懂得两种证券的基本结构，肉眼比对一下价格，就能意识到定价出现偏差，随后手动下单促成均值回归。 > “There were literally trading systems that would run in an Excel spreadsheet, you know, to like do these kinds of things and you could just look at the trades and be like, 'Oh, yeah, that was a good trade.' You didn't need any like statistics or things like that, you know.” 那时的许多交易系统甚至直接搭建在 Excel 电子表格里运行，根本不需要复杂的统计学模型辅助决策。但今天，建模工作已彻底转向前沿的机器学习技术，涵盖深度神经网络、梯度提升树（GBTs）以及各类复杂的现代架构。Craig 指出，几乎每隔五年左右，公司所倚赖的核心技术范式就会经历一次彻底的迭代。然而令人称奇的是，尽管底层代码和算力架构翻天覆地，公司的核心文化却始终如一，许多创立之初的核心成员至今依然活跃在业务一线。 - **2005 年前后**：依赖 Excel 运行部分交易系统，只要理解资产运作逻辑，仅凭肉眼盯盘就能识别定价偏差，模型仅限于简单的集群线性回归。 - **现代技术架构**：完全演变为以现代深度学习为主导，深度集成神经网络、GBTs 等复杂算法体系。 - **迭代周期**：核心建模工具与算法技术大致每 5 年就会发生一次重大技术迭代。 --- ### [14:35](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=875s) 为什么 Streeter 愿意长期留任：无摩擦的同侪文化与高自主权 在人员流动极为频繁的金融与科技行业，Jane Street 却维持了极其罕见的员工高留存率。当主持人询问为何员工能在这里扎根如此之久时，团队成员解释道，关键在于人与人之间深度的情感链接与协作范式。在很多公司，“友好”可能仅仅是一种职场礼节，但在 Jane Street，许多同事在现实生活中也是彼此真正的挚友。 > “I think that people help each other out and collaborate not just cuz they're like nice people and they're friendly, which is, you know, really great, but because like they feel like really invested in your success. Like, you know, when you were in college and you like had a roommate or a friend and like they wanted you to succeed on the problem set like almost maybe sometimes more than you did.” 这种彼此全心投入对方成功的文化，就像大学时代通宵达旦、甚至比你本人更希望你解开难题的室友一样。这种氛围极大地削减了跨领域探索时的协作摩擦力。员工每天早上走进办公室，不仅是去面对充满智力挑战的硬核问题，更是去和值得信赖的伙伴并肩作战。大家并不会在一入职就设定“我要在这里呆满十年”的僵化目标，而是每一天都在自然地评估：在这里是否能持续学到新知、是否拥有足够的个人自主权（agency），以及整个探索过程是否充满乐趣。正是这种极低的人际内耗与高密度的智力碰撞，构筑了 Jane Street 最核心的文化护城河。 ### [16:15](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=975s) 从数学学术圈到 Jane Street：Nithya 的博士休假之路 在探讨外部其他职业机会时，与会者谈到留在 Jane Street 的切身感受：在这里学习新知识的速度随着工作年限的增长反而越来越快，这种持续成长的状态令人十分满意。随后，Grant Sanderson 将话题引向大家的入职经历，首先询问 Nithya 在加入 Jane Street 之前的背景与职业轨迹。Nithya 回忆道，自己原本坚定地认为学术数学才是自己的唯一归宿，家人和朋友虽然全力支持她走数学家之路，但也建议她至少去尝试一次教书以外的工作，看看外面的世界究竟如何。于是她申请了 Jane Street 的暑期实习，没想到彻底爱上了这里。经过两轮暑期实习后，她顺理成章地接下了全职 Offer。 更与众不同的是，Nithya 的职业路径并没有按部就班地留在交易室内。在 Jane Street 工作了几年之后，她选择暂时离开公司去攻读数学博士学位（PhD），毕业后又重新回到了 Jane Street。Grant 好奇这种“工作后请假读博再回归”的做法在公司是否普遍，Nithya 解释道，虽然像她这样在中途请长假读博的人极少，但 Jane Street 在满足员工个人独特发展需求上展现了极高的灵活性，公司非常愿意为员工的具体情况量身定制支持方案。 > “I think I'm one of the very few people who like went on like pseudo leave for a few years to get a PhD in the middle of my career.” —— Nithya - **初衷转变**：原本立志终生从事纯数学研究，经亲友劝说抱着“尝试一次非教职工作”的心态申请实习，结果完全被实际工作吸引。 - **职业轨迹**：首度实习 $\rightarrow$ 次年二次实习 $\rightarrow$ 全职入职工作数年 $\rightarrow$ 休假攻读数学博士 $\rightarrow$ 博士毕业重返 Jane Street。 - **公司文化弹性**：尽管中期读博并非普遍惯例，但公司对于优秀员工的个性化诉求展现出极大的包容度与制度弹性。 --- ### [18:04](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=1084s) 扔筹码、追飞机与“二级快乐”：独特的现场面试体验 谈及当年的面试经历，Nithya 笑着分享了一段戏剧性十足的往事。面试最初由概率论谜题和各类综合解题测试组成，电话面试非常有趣。到了现场面试环节，她原本以为会考查更深奥的高等数学理论，结果却发现面试几乎全程都在“玩各种博弈游戏”。这让她在现场感到巨大压力，但事后回想起来却又觉得十分刺激，完全是一种典型的“二级快乐”（Type II Fun，即身处其中时备受煎熬，事后回味却妙趣横生）。 整个面试日的终局更是充满戏剧性。Nithya 当天的最后一场面试是与 Sandor（Jane Street 核心创始人之一）进行，两人当时甚至在对着墙壁扔筹码玩概率博弈游戏。面到一半时，两名招聘人员突然猛地推门冲进房间大喊：“她要赶不上飞机了！她要误机了！”Nithya 甚至还没反应过来，就被一路急匆匆地推出了走廊塞上开往机场的交通工具。当时她心想自己这下肯定搞砸了面试，没想到最终顺利拿到了录用通知。 ```mermaid flowchart LR accTitle: 面试流程经历 accDescr: Nithya 经历的 Jane Street 典型面试历程 subgraph Round1["初筛阶段"] A["电话面试<br/>概率谜题 & 杂项解题"] end subgraph Round2["现场阶段"] B["现场博弈测试<br/>高压游戏模拟"] C["终面互动<br/>与合伙人扔筹码博弈"] end subgraph Result["戏剧收尾"] D["HR冲入抢人赶飞机"] E["斩获全职录用"] end A --> B B --> C C --> D D --> E ``` --- ### [19:02](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=1142s) 招聘核心策略：寻找具备适应能力的通才与长期人才投资 针对 Grant 提出的“过去五年面试标准与考查内容有何演变”的提问，团队深入剖析了 Jane Street 的用人哲学。面试的核心标准多年来始终保持一致：寻找聪明、勤奋、讲理、具备极强适应能力且渴望学习新事物的人。尽管如今掌握机器学习（ML）技能确实是一个加分项，但面试题目依然刻意避开那些“只要你在大学上过某门特定课程就能直接套用解法”的既有知识题，转而采用开放、新颖的数学谜题，以此观察候选人在面对陌生情境时的现场思维推演与沟通碰撞。 团队进一步指出，这种注重“通才适应力”的策略与 Jane Street 极低的人才流失率和长期留任文化深度绑定。如果在一家平均任期仅有三年的公司，花费整整一年时间对员工进行全方位的机器学习转型再培训是极其不划算的，企业更倾向于直接去市场上高薪挖人；但因为 Jane Street 的员工普遍具有很长的在职生命周期，为有潜力的员工投入长周期的新技能重塑便具有了极高的 ROI。即便在短短一小时的面试中很难精确预判候选人一到五年后的产出上限，但历史证明，五年前擅长学习和应用新工具的人，面对如今工作要求的剧烈演进依然能快速上手。 > “The bulk of our strategy has always been to hire people who are just like smart and hardworking and reasonable and adaptable and like to learn new things and then you can just do the new thing.” —— 团队成员 > “It wouldn't be worth your time if the median time until somebody left was in 3 years to spend a year retooling them on ML. You'd rather just go and like hire somebody, right? But if you're pretty sure that most of these people are going to be here for a long time then it like makes more sense to invest in that kind of thing.” —— 团队成员 - **题目设计原则**：弱化特定课程工具的记忆依赖，侧重于未见过的数学难题与逻辑推演过程。 - **通才（Generalist）优势**：技术栈与工作重心随时代不断迭代，优秀的学习底色能让员工持续完成自我技术更新。 - **人才重训经济学**：高留存率构成了企业内部长期技术重构与培训投资的底层商业逻辑。 --- ### [21:29](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=1289s) Alok 的入职历程：当抽象数学理论照进现实交易建模 随后，Alok 分享了自己在 2018 年夏天申请定量研究实习的经历。大学时他同时修读计算机科学（CS）与数学，当时内心对两者的定位截然不同：CS 是为了日后能找一份糊口的工作，而数学纯粹出于个人智趣，他甚至预设未来的工作只会用到编程，能碰上数学只能算“额外惊喜”，绝非求职的硬性门槛。在他当时的认知里，真正能同时融合数学与计算机的领域只有两个：一个是广义的机器学习与数据科学，另一个则是充满神秘感的金融行业。当时他甚至揣度金融公司里大概坐着一帮整天用高斯过程（Gaussian process）等复杂模型试图解释一切的理论学者。 然而，在 Jane Street 度过一个暑假后，Alok 惊讶地发现量化交易在本质上与前沿数据科学极其相似，学校里那些看似极其抽象、远离实用的象牙塔理论，在实际交易建模与数据清洗中竟然找到了极其硬核的落脚点。 > “Like I was actually kind of shocked at how well a lot of the other stuff that I've been studying kind of transferred or like ideas I had learned in college classes that seemed pretty abstract and like theoretical actually found like pretty concrete applications in just like analyzing the modeling we were doing and things like that.” —— Alok 当 Grant 邀请他给出一个具体例子时，Alok 提到了他在统计课上学到的 Huber M 估计量（M-estimators）理论。这套由 Peter Huber 提出的统计思想原本看起来像是纯粹的理论构造——往往只在渐近理论或宇宙级规模的假设下才成立。但当面对现实金融世界中充斥着噪声、缺失与畸变的异常数据（malformed data）时，这类鲁棒统计学工具展现出了不可替代的强大现实威力，让年轻时的他深感理论与工程结合的奇妙。 ### [24:17](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=1457s) 稳健估计量与实习项目 实习生在项目期间致力于将理论模型落地为具体的业务实践。在统计学与建模中，存在一类被称为稳健估计量（Robust Estimators）的外部大类，当时的研究重点正是如何将这种数学理论应用到具体的建模难题中，从而提高模型在非理想分布或噪声环境下的稳定性。 ### [24:28](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=1468s) Craig 的传奇入职经历：2005 年的 OCaml 邮件列表 主持人 Grant Sanderson 转向 Craig，提到两人过去曾多次相遇，因此非常真诚地问起 Craig 当年究竟是如何加入 Jane Street 的。Craig 笑着清了清嗓子，坦言自己的入职故事确实非同寻常。那时的 Jane Street 规模极小，同事们后来甚至开玩笑形容他像是“被狼养大的野孩子”（raised by wolves）。在当今的量化圈，绝大多数人要么毕业于顶尖名校，要么在中学时就通过奥数夏令营等圈子互相熟识，而 Craig 的经历在当时显得十分随意，或者从某种意义上说，极其精准。 Craig 透露，自己这一辈子其实基本只申请过这一份正式工作。当时他已经准备开启攻读博士学位的学术道路，制作简历纯粹只是为了练练手。然而，他偶然在 OCaml 编程语言的邮件列表（mailing list）上看到了一则招聘启事，便顺手投出了简历。2005 年时的招聘流程极为简短，放在今天甚至难以想象。正如 Craig 自嘲所言：“The process was very different back then. Like I don't think I would get hired now. Um, they only did like one phone interview, you know, and then I I came in in person because free trip to New York, why not, you know?” 在接到一轮电话面试后，他抱着“顺便免费去纽约旅游一趟”的心态参加了现场面试，结果对现场遇到的团队成员印象极佳，随即决定接受 Offer。 ```mermaid flowchart LR accTitle: Craig 加入 Jane Street 的历程 accDescr: 流程图展示了 Craig 在 2005 年如何从学术准备走向全职量化交易公司的独特路线。 subgraph S1["学术准备"] A["计划攻读博士学位"] B["练习制作简历"] end subgraph S2["偶然契机"] C["浏览 OCaml 邮件列表"] D["发现招聘帖子并投递"] end subgraph S3["早期简易流程"] E["单轮电话面试"] F["纽约现场面试 (免费差旅)"] G["加入团队 (2005 年入职)"] end A --> B B --> C C --> D D --> E E --> F F --> G ``` - **学术转向**：Craig 原本并未打算进入金融量化界，攻读博士学位是其第一选择。 - **面试流程精简**：相比现代量化金融行业动辄多轮严格的技术考核，2005 年的招聘仅包含一轮电话面试和一次现场面试。 ### [25:38](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=1538s) 对小众语言 OCaml 的黑客情结与趣味编程 Grant 追溯起技术细节，好奇 Craig 当初为何会关注并深入研究 OCaml 这门在当时相当小众的函数式编程语言。Craig 解释道，最根本的原因在于自己骨子里就是一个广泛意义上的“书呆子/硬核技术迷”（nerd）。对于他而言，写代码并不是为了严肃的商业目的，而是用计算机去尝试各种好玩的新鲜事。 在日常折腾业余项目时，Craig 会默认选用 OCaml。无论是想在屏幕上绘制分形几何图形（fractals），还是为了探究博弈人工智能而编写一个黑白棋（Othello）对战机器人，OCaml 都是他顺手拈来的第一工具。正是这种单纯出于个人兴趣的技术探索与积累，最终促使他点开了那个改变职业轨迹的邮件列表帖子。 ### [26:17](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=1577s) 办公室内的即兴物理讲座与学术探讨 话题随后延伸到团队成员在工作之余的休闲活动。在 Jane Street 的办公室文化中，经常能看到大家在闲暇时聚在一起交流高深学术问题。例如 Alok 经常会在旁边的会议室里即兴为大家开设小型数学或物理讲座，大家把这种探讨当作一种纯粹的放松和社交方式。 Grant 和 Craig 聊起了 Alok 最近举办的一次即兴讲授。那次讨论从统计力学（Statistical Mechanics）切入，涉及配分函数（Partition Functions），进而探讨到一种深层的审美哲学：只要你的思维视野拓展得足够广阔，你几乎可以欣赏并理解这世上的任何事物。话题随后自然延伸到了费曼图（Feynman Diagrams）与统计力学中配分函数的紧密联系，甚至是更为抽象的数学构架——比如“图的指数”（e to the power of graphs）。Grant 打趣道，这题目听起来简直就是 3Blue1Brown 频道的经典“标题党”素材，引得全场大笑。 - **跨学科交流**：探讨统计物理中的配分函数与量子场论中费曼图之间的深层数学关联。 - **独特的办公室氛围**：即兴的白板讲座涵盖抽象代数、图论与理论物理，形成了浓厚的学术探索风气。 ### [27:05](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=1625s) 访谈总结与致谢 在对话的尾声，Grant 对这次座谈表达了诚挚的谢意。他提到自己虽然此前与 Kenji、Nithin 等 Jane Street 成员有过接触，但外界对这家公司的真实运转模式依然抱有许多未解的疑问。这次面对面的深入探讨不仅解答了大量核心疑惑，也让外界能够真实看到在算法交易、系统架构、前沿机器学习背后活生生的人与独特的工程师文化。正如 Grant 总结所言：“I think this helped clarify things. I think it'll also help just make it a lot more real for whoever might be watching this.” 整个 roundtable 讨论在轻松愉悦的氛围中圆满结束。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=0s) 在大家展开对话时，目的并不是为了满足虚荣心，或是非要证明某个观点、向人炫耀自己有多聪明；如果是那样的话会非常累人。 — Nitya > [05:51](https://www.youtube.com/watch?v=zPLc3jjHbnU&t=351s) 我们核心业务之一——这在很多年前绝对是无可替代的核心，但我认为至今仍驱动着我们交易哲学的——就是成为一家做市商机构。我对大家打的比方就是：它就像是一家‘股票杂货铺’。 — In Young ## Mentioned Jane Street, Grant Sanderson, 3Blue1Brown, Nitya, In Young, Craig, Alok, Sandor, Kenji, Nithin, Peter Huber, OCaml, Excel, Jupyter Notebook, Bloomberg ## Video description We are longtime fans of Grant Sanderson and his YouTube channel, 3Blue1Brown, at Jane Street. So when we had a chance to sit down with Grant for a roundtable conversation at our New York office, we jumped at it! In this video, Streeters Nitya, In Young, Craig, and Alok cover what Jane Street actually does, why we run a real-time auction for our compute cluster, and how the work has changed over 20 years – from trading systems running in Excel spreadsheets in 2005 to deep learning now being the single largest driver of our systematic trading. They also shared their own paths to Jane Street, from spotting a job on the OCaml mailing list to taking a mid-career detour for a PhD. [Source](https://www.youtube.com/watch?v=zPLc3jjHbnU)

[0:00] Nitya: 在大家展开对话时，目的并不是为了满足虚荣心，或是非要证明某个观点、向人炫耀自己有多聪明；如果是那样的话会非常累人。 [5:51] In Young: 我们核心业务之一——这在很多年前绝对是无可替代的核心，但我认为至今仍驱动着我们交易哲学的——就是成为一家做市商机构。我对大家打的比方就是：它就像是一家‘股票杂货铺’。

[18:04] The bulk of our strategy has always been to hire people who are just like smart and hardworking and reasonable and adaptable and like to learn new things and then you can just do the new thing.

[0:00] Okay, one of the things I I want to do, [0:01] thank you for taking the time, is just [0:03] try to get a feel for what the Jane [0:04] Street culture even is and also just [0:07] technically what do you actually do? [0:09] What are the salient features? Right, [0:11] what comes out is like you want to [0:12] distill a sense of what Jane Street [0:14] culture actually is as it maybe differs [0:16] from other places, what pops to mind? [0:20] >> Jane Street is really weird, but in like [0:21] the best possible way. [0:22] >> Okay. [0:23] >> Like truly when people are um having [0:26] conversations, like the goal isn't to [0:29] like have ego or like to prove a point [0:31] or kind of show people how smart you are [0:32] or something. That'd be very exhausting. [0:33] It's really approaching like a topic [0:35] that's kind of interesting [0:37] uh and trying to arrive at a truth at [0:39] the truth by kind of like talking [0:41] through things and um I think people are [0:42] very just like receptive to things that [0:45] they didn't know before and um I don't [0:47] know, it's like pretty cool if you can [0:48] like approach any conversation and [0:50] expect to have your view change um [0:53] at the end of it. [0:54] >> Is there a time that I guess any of you [0:56] like had your mind changed in one of [0:57] these interactions or vice versa where [0:59] you're like having an interaction and [1:00] you're like change someone's mind that [1:02] felt very low ego uh [1:04] that like exemplifies what you're [1:06] talking about here? [1:07] >> I feel like it would more it would be [1:08] like more anomalous to find a time when [1:10] that didn't happen. Like I think like a [1:11] lot of meetings like you know, you know, [1:13] people like come in with some some [1:14] thoughts and often we'll ask people to [1:16] like share their thoughts before they [1:17] hear everyone else's just to like see [1:18] what everyone is starting with and talk [1:19] about it and then maybe at the end we'll [1:21] like share what our like after thoughts [1:22] are based on hearing everyone's and [1:24] usually they've converged. It's like [1:25] very It's like very nice to see [1:26] actually. It's like wow. I think it's [1:28] like a pretty unique um way of [1:31] conversing um and I feel like I know [1:35] that it's unique from just like going [1:37] back and talking to people from college [1:39] and they're like wow the way that you [1:41] talk is pretty strange. [1:42] >> [laughter] [1:43] >> Like the way you talk about interactions [1:45] here or you've like adopted a certain [1:47] Bayesian habit that like [1:49] >> Oh, there's like a certain Bayesian [1:50] habit. There's also like Jane [1:51] Streetisms, you know, people will say [1:53] things like oh, that's not unreasonable. [1:54] Um [1:55] >> [laughter] [1:55] >> I don't know. I I think there's just [1:56] like phrases you just hear a lot here [1:58] and it's just absolutely wild outside of [2:00] Jane Street. [2:01] >> My impression is that we have much more [2:03] of like [2:04] there's kind of just one company. [2:06] >> Mhm. [2:07] >> Um [2:08] and it doesn't matter so much like which [2:10] group you're in or something. So, people [2:12] talk across groups a lot without going [2:15] through [2:16] uh like their manager talks to somebody [2:18] else's manager kind of things to set it [2:20] up. You just kind of walk over to [2:22] another group and you say, "Hey, I was [2:23] thinking about something. Does anybody [2:24] have a minute to just spitball on it?" [2:26] Like it seems like something that you [2:27] guys would know something about and [2:30] you know, so it's it's lots of very [2:32] organic [2:33] >> That seems hugely important, but it also [2:35] seems very hard to scale, right? I you [2:37] know, if I try to picture why is it that [2:38] other organizations evolve kind of [2:40] hierarchically and it is you're going to [2:42] manage manage manager is that and maybe [2:44] I'm naive here, it just seems like chaos [2:46] otherwise. [2:47] >> It is a little bit. [2:48] >> Okay. Okay. And so is there a kind of an [2:50] embracing the chaos of that because of [2:52] the the benefits of [2:54] cross expertise communication? [2:55] >> Yeah, a little bit, but I mean there are [2:57] like you know, people have managers, [2:59] right? And there are like kind of team [3:01] leads or project leads or you know who [3:03] to ask about certain kinds of topics. Um [3:05] so, there is some structure and like [3:07] organization to it. But I think that [3:09] that kicks in where, you know, if you're [3:11] going to start like a month-long project [3:13] with multiple people, now okay, let's be [3:15] organized about it, right? Um but if [3:17] it's just kind of like you're going to [3:18] walk over to try to come up with some [3:20] ideas or something, that's at the like a [3:23] small enough scope where [3:25] just the chaos seems like the better way [3:27] to do it, I guess. [3:28] >> Got you. Got you. Um man, I have many [3:30] more questions there, but I want to say [3:32] >> on to that. [3:33] >> Yeah, yeah. I want [3:34] >> really emblematic of this chaos is we [3:36] run like a real-time auction for our [3:38] compute. Uh it's like our compute [3:39] cluster [3:40] is available for everybody and [3:43] you know, like your task would be [3:45] you know, a task of realistically [3:46] thinking about how to value kind of like [3:49] this job um that you're running versus [3:51] all the other things and um there's not [3:53] really any guardrails. [3:55] >> Wait, that's so interesting. So as [3:56] whereas like other company large [3:58] compute, there's a whole like [3:59] bureaucratic like shifting around chips [4:02] kind of game to decide like what [4:03] allocation they get, it's just straight [4:05] up auction on and like anyone can kind [4:07] of bid for the compute timer. And it's [4:09] funny, I almost would have said that [4:10] that's a case of organization. We've got [4:12] an auction system that somebody [4:13] implemented. [4:14] >> I was I mean I do think there are like [4:16] advantages cuz I think like people like [4:17] really feel collectively responsible [4:19] like in a good way feel like everything [4:20] everyone is doing. Like I feel like if I [4:22] really don't leave my desk or you know [4:23] some job running and then it's like [4:25] doing something horrible like bumping [4:26] into errors like it's not on me to be [4:28] like checking every like one minute like [4:29] I do feel like I do feel more [4:30] responsible for my jobs than everyone [4:32] else's but like someone else might see [4:33] it and like you know be like oh I [4:34] noticed there's an error let me just [4:35] like pause it for you until you come [4:37] back and I think that's really nice and [4:38] people feel like that sense like we're [4:40] like one you know one team even outside [4:42] the local team I'm on. [4:43] >> Yeah, that's so fascinating. [4:45] >> I mean you'll you'll kind of like the [4:46] people who are actually I don't know [4:47] running these jobs like people have [4:50] access to kind of the jobs that are [4:51] running and if you kind of see something [4:53] that's a little off you can just go and [4:54] talk to them and say hey I see that you [4:56] know like you're running this job and [4:58] this MFU is just terrible like what's [4:59] going on [5:01] and you can have a conversation and a [5:03] lot of like things can be sorted out [5:04] without it like going up the ranks or [5:06] something and I think it just allows [5:08] things to run much faster. [5:09] >> I think it's globally read right too [5:11] right? If like you can just nuke [5:12] somebody else's job anytime. [5:14] >> Really? [5:15] >> [laughter] [5:15] >> Yeah, there there was an instance where [5:18] a unwitting researcher ran a command in [5:21] a Jupiter notebook um [5:23] uh of adjusting the bids on his job. [5:28] He accidentally ran the command without [5:30] any arguments which meant that he [5:32] changed the bids of every single job [5:34] running on the hive. [5:35] Um [5:36] >> And he had access to do that? [5:37] >> that was not great. I think we fixed [5:38] that but like [5:40] it is kind of like a you know we'll [5:42] build a system as it kind of makes sense [5:44] and kind of change it organically as it [5:46] fits rather than coming up with like a [5:48] structure um necessarily that's [5:49] restrictive from the get-go. [5:51] >> curious how you describe to outsiders [5:53] what Jane Street does, cuz you've got [5:55] the [5:56] quantitative finance just like hand-wave [5:58] sort of description, then you've got a [6:00] level of specificity that is like trade [6:02] secrets and kind of obviously can't talk [6:03] about, but there's a lot of room in [6:05] between for actually clarifying for you [6:08] know, people outside of finance like me, [6:11] what what do you do? What is the actual [6:14] uh organization pointing towards? [6:16] >> Yeah, I guess there was a a phrase that [6:18] I think Sandor came up with uh which is [6:20] kind of a mouthful of uh Jane Street [6:22] being a [6:23] uh generalist financial services [6:26] liquidity provider. [6:27] >> I will be honest, that has so far not [6:29] added clarity. [6:31] >> When we have like college students in [6:32] the office, I give a lot of intro talks [6:33] and one thing I say a lot is like, you [6:35] know, I think one of Jane Street's core [6:37] businesses that was really really really [6:38] our core business, you know, many years [6:40] ago, but I think still drives like our [6:42] philosophy around trading is being a [6:43] market making firm. [6:44] >> Okay. [6:44] >> Which is, you know, what I what I tell [6:46] folks is it's like a grocery store for [6:47] stocks. Um not literally, but you know, [6:50] I feel like, you know, grocery stores [6:51] are pretty magical. Like if you had to [6:52] like go, you know, buy all your [6:53] groceries and you had to like go to a [6:55] farm to buy your apples and go to a [6:57] dairy farm to buy, you know, milk, you'd [6:59] be pretty sad. And the grocery store has [7:01] to think about how to price things [7:02] fairly, right? They can like charge [7:03] like, you know, $10 for an apple if they [7:05] want, but no one is going to buy their [7:06] apples. And I think that that's kind of [7:07] like a pretty similar principle. Like [7:09] you know, we do do the business of like [7:11] kind of pricing ETFs relative to stocks [7:12] and kind of like, you know, if they [7:13] trade out of line, but I think the [7:15] different way to like conceptualize that [7:16] is that like, you know, if I am like, [7:18] you know, sitting in the United States [7:19] and I want to like invest in, you know, [7:22] some companies in like China or I want [7:23] to invest in some companies in, you [7:25] know, some foreign market, the best way [7:26] that I can do that is by buying an ETF. [7:29] >> So that's I mean, that's one category of [7:30] answering the like what Jane Street [7:32] does, which I'm sure doesn't scratch the [7:33] surface like on just like the trading [7:35] half. But a lot of the tone that I also [7:37] hear much more recently is more it's a a [7:40] more similar to if I'm talking like an [7:41] AI lab. [7:42] >> Yeah, I think um maybe the gap between [7:45] AI lab and grocery store as metaphors is [7:47] like a little bit hard to [7:49] maybe one thing that would help there is [7:51] um if you think about it with a grocery [7:52] store, they need to know to have a lot [7:54] of jack-o'-lantern pumpkins for, you [7:57] know, um Halloween and and these kinds [8:00] of things, right? So, there's a lot of [8:02] like anticipating supply and demand and [8:04] then you can kind of see how that's a [8:05] prediction problem that could be [8:07] somewhat complicated, right? Um [8:10] so that of where, you know, the machine [8:13] learning and stuff like that kind of [8:14] comes in. You want to have sort of the [8:16] inventory to provide liquidity to [8:19] you know, what's upcoming and what [8:20] people will want. Um [8:22] >> This tracks with I think a a conception [8:24] that because, you know, in theory if [8:26] there's value to be added from these [8:27] instruments like ETFs just as an [8:29] example, you could have like a [8:30] management firm around it that like is [8:32] deliberately trying to keep keep things [8:34] in line. But then if it's in a [8:35] market-driven way, it's some other [8:37] participant in the market that does [8:38] that. Is there anything that's [8:39] misunderstood? Like if if you see in in [8:42] so far as there's, I don't know, [8:43] discussion in popular press or just like [8:46] uh interviews that you have with [8:48] candidates coming in, are there [8:49] misconceptions where someone thought [8:51] about it one way and you realize, oh, [8:53] actually, that's just that just doesn't [8:54] match with what you see on the inside? [8:56] >> I think there's a misconception not just [8:58] about the scale of machine learning that [8:59] we're doing, but the novelty of it. Like [9:01] I think it's like kind of easy to [9:02] imagine that mostly what we're doing is [9:03] we're taking some like models off the [9:05] shelf and maybe we like bought a bunch [9:06] of GPUs and then we're like running it. [9:07] And that like basically doesn't work at [9:08] all. Um in in financial domains, I think [9:11] the problems that we're working on are [9:12] much much much noisier. And so we've had [9:14] to do a there's a lot more like really I [9:16] think it surprised me even a lot more [9:17] like basic science that we're doing. Um [9:19] like, you know, where we're really [9:20] trying to understand fundamentals and [9:21] working on the frontier of what is [9:23] known, you know, in the research [9:24] community at large. [9:25] >> Mhm. [9:26] >> that surprising that the misconceptions [9:28] exist just because um things have ramped [9:30] up really really quickly um over like [9:33] the last couple years. Mhm. Um but yeah, [9:35] I don't know. It's just been kind of a [9:36] incredible ride. [9:38] >> So, you felt the meaningful shift just [9:39] in the last couple of years for like [9:41] what the what the actual work feels [9:43] like? [9:43] >> Honestly, just kind of like the amount [9:44] of success that we've seen with um some [9:46] of the ML efforts. It kind of feels like [9:48] we are just kind of like at the [9:50] beginning of the S curve. And um but [9:52] it's also just like uh very meaningful [9:54] input to our trading. Um and I would say [9:57] it's like, you know, the single largest [10:00] driver of um the systematic trading that [10:02] we have. [10:03] >> Yeah, I think there's a lot of [10:04] misconceptions, but they're kind of like [10:06] different coming from different [10:07] directions, you know? [10:09] People at other trading firms or like [10:11] banks or stuff like that will generally [10:12] assume it's going to be more similar to [10:15] these other institutions than it [10:16] actually is. Um so, often if we hire [10:19] somebody from one of these places, [10:21] they'll be surprised that it's maybe a [10:23] little bit more like tech or academia in [10:26] some ways or some something kind of like [10:27] that [10:28] um compared to what they were used to. [10:31] >> People see that we have like a lot of [10:33] traders and then like a research group [10:36] that's meaningfully smaller and assume [10:37] that means like the majority of our [10:39] trading is being done by people manually [10:41] entering orders into like a a Bloomberg [10:44] or something. But like mostly this is [10:46] historical accident in naming and like [10:49] the the role that we call trader is like [10:51] in many ways largely a research one. Um [10:54] and [10:55] like the firm is way more systematic [10:57] than I think people tend to assume. [10:59] >> Right. Yeah, pop quiz, which of the [11:00] people here appear in the corporate [11:01] directory as a researcher or a trader, [11:03] right? Like [11:04] >> [laughter] [11:05] >> I'm not sure. [11:05] >> I was like it's just generally like [11:07] you're like thinking of your coworker [11:08] like, "I'm actually not sure what their [11:09] title is." [11:10] >> Yeah, there are sort of like, you know, [11:12] teams and some kind of structure to this [11:14] even though not everyone's exactly sure [11:16] which team they're on or you spend what [11:18] some time on one and some time on the [11:19] other, that kind of thing. [11:21] Um [11:21] but yeah, broadly there's kind of like [11:23] specialization [11:25] at the level of [11:26] um [11:27] people who mostly code but understand, [11:30] you know, the trading and stuff pretty [11:32] well. Then there's people who uh code [11:35] but mostly with other coders as their [11:37] customers, right? And so they're like a [11:39] little bit of a step further removed [11:41] from the trading and stuff like that. [11:42] >> This is like internal tools. [11:44] >> Yeah, internal tools for like managing [11:47] all of our [11:49] I guess branches, you know, [11:51] and who's code reviewed what and is it [11:53] ready to be released and various kinds [11:55] of infrastructure for [11:57] dealing with all the market data. It [11:58] turns out there's actually a lot of [12:00] data. [12:01] Um [12:02] petabytes, so [12:04] that's a big task. [12:06] Um [12:07] And then there's kind of like a a loose [12:09] um [12:11] spectrum between sort of like traders [12:14] and researchers with a lot of in [12:16] between. [12:17] Sort of having to do with how much you [12:19] think about like particular securities, [12:22] what is happening today in particular [12:24] trading systems [12:25] as opposed you know, that on the trading [12:27] end as opposed to a little bit more like [12:30] you know, launch off some compute job to [12:32] do some machine learning stuff to train [12:34] some model or build a tool for doing [12:36] that, you know, that would eventually [12:37] like feed its way into these trading [12:40] systems on kind of like a longer time [12:42] cycle. [12:44] >> Got you. Now, this is maybe while I'm [12:46] talking to you Craig, like [12:47] you've been around for 20 years. [12:49] In that time, how are things different [12:52] compared to then? And I have some [12:55] suspicions for what the answer might be, [12:57] but as you look over the last 20 years [12:59] or so, like what is the difference? And [13:01] let's maybe start like technically and [13:02] then I'll ask culturally. Like [13:03] technically for what the actual work [13:06] looks like back then compared to now. [13:08] >> Yeah, it's it's very very different. I [13:10] mean, the the work um [13:12] when I started, you felt pretty fancy [13:14] pants if it was like, oh, we've actually [13:16] logged data, you know, and we can now [13:19] like train a linear model on a compute [13:22] cluster, you know, like, wow. Uh [13:25] and so that was kind of considered [13:28] you know, the more advanced stuff we [13:29] were doing as opposed to um [13:32] in a way, a lot of it back then was [13:34] stuff that you could just kind of look [13:36] at two securities and if you knew how [13:37] they work and thought about it a bit, [13:39] you would be like, "Oh, that's out of [13:40] line. We should trade them to be in [13:41] line." [13:42] Um and so, there were literally uh [13:44] trading systems [13:46] that uh would run in an Excel [13:48] spreadsheet, you know, to like do these [13:50] kinds of things and you could just look [13:51] at the trades and be like, "Oh, yeah, [13:53] that was a good trade." You didn't need [13:54] any like statistics or things like that, [13:55] you know. [13:56] Um and so, the challenges for that kind [13:59] of stuff were totally different. Um at [14:02] this point, uh [14:04] you know, most of the modeling type of [14:06] work really is machine learning kind of [14:07] stuff. Um neural nets increasingly [14:11] recently, you know, GBTs and stuff also, [14:14] just all kinds of things. And we hit [14:16] many, you know, points in between. So, I [14:18] would say like every [14:20] five years or so, really the techniques [14:22] have changed like quite a lot. [14:25] And yet kind of strangely, it feels um [14:28] similar as far as like kind of the [14:31] culture and stuff like that. Like a lot [14:32] of the original people are still around. [14:35] >> Why is it that people stick around so [14:36] long, right? Like what what is it that [14:38] makes people stay that seems to not [14:40] happen at other places? [14:42] >> I think like earlier you'd asked like [14:44] them about like, you know, what makes [14:45] Jane Street Jane Street. And I think one [14:46] thing that I've really felt is like the [14:48] people like like there are lots of [14:50] workplace settings that are super nice, [14:51] that are super friendly, where like your [14:52] colleagues are super friendly to you. Um [14:54] I mean, Jane Street is also one of [14:55] those. But I think more than that, like [14:57] a lot of the my colleagues, like [14:58] especially even new ones I've met and on [15:00] teams I'm on too recently, are like my [15:01] real life friends. Um I think that like [15:03] people help each other out and [15:05] collaborate not just cuz they're like [15:06] nice people and they're friendly, which [15:08] is, you know, really great, but because [15:09] like they feel like really invested in [15:11] your success. Like, you know, when you [15:12] were in college and you like had a [15:13] roommate or a friend and like they [15:14] wanted you to succeed on the problem set [15:16] like almost maybe sometimes more than [15:17] you did. And I think that like I really [15:19] feel that from a lot lot people around [15:20] me that they are like as invested in my [15:22] success as I am. Um and they really feel [15:25] that like for me as a person and not [15:26] just cuz like they're a really nice [15:27] person or that's like kind of a norm at [15:28] the company. I think it has become a [15:30] norm at the company for people to help [15:31] each other out. But I think that also [15:33] means just like everything feels like [15:34] kind of like really low friction. It [15:36] feels really like fun to go to go to [15:37] work. I'm like, "Oh, I'm going to go to [15:38] work. I'm going to meet my friends. I'm [15:39] going to work on really cool problems." [15:41] Like if, you know, things are really [15:43] tough and challenging and high friction, [15:44] especially as they can be sometimes when [15:46] you are moving into new areas, there's [15:47] going to be a lot of support and people [15:49] are going to be like with me doing it. [15:50] And I think that really, I don't know, [15:51] for me that's [15:53] that that that makes a big difference. I [15:54] think for a lot of people I talked to it [15:55] seems like a really important input. [15:57] >> I mean, from a personal point like I [15:59] feel like it's not like I came into I [16:01] came into Jane Street thinking, "Oh, [16:02] like I'm going to work here for 10 [16:04] years." Um it's kind of been like an [16:06] everyday reevaluation of like do I feel [16:08] like I'm learning a lot here? Am I [16:09] having fun? Like um you know, [16:12] are the conversations that I'm having, [16:14] like the amount of agency that I have, [16:15] like is it, you know, better um versus [16:18] like um comparable other opportunities [16:20] and [16:20] I just kind of feel like like the pace [16:22] at which I feel like I've learned new [16:25] things um has kind of like increased [16:27] over the years, which is kind of um [16:28] shocking. And I feel like that's like a [16:30] pretty nice feeling to have. [16:32] >> Mhm. [16:33] What one of the things I was a little [16:34] bit curious to ask was some of the story [16:36] around joining Jane Street. And I might [16:38] start actually, Nithya, with you. Are [16:40] you able to tell me like what were you [16:42] doing before you joined Jane Street and [16:44] kind of what was the trajectory to come [16:45] in? [16:46] >> Yeah, like I think I personally had like [16:48] a I don't know, maybe a normal [16:50] trajectory for a person here, but it [16:52] felt like pretty unique. Like I really [16:53] thought I wanted to be a mathematician. [16:55] I was like really like set. And my [16:57] family and my friends were like, "You [16:58] know, this like math thing is super [16:59] great. We would really love for you to [17:00] be a mathematician. Like maybe you [17:01] should like try like once like working [17:03] at like a job that's like not teaching [17:04] math just like once. Like see how you [17:05] like it. You never know." And I applied [17:07] for an internship and I got it um at [17:10] Jane Street and I came and I like really [17:11] really loved it. I was like super [17:13] surprised. Then I came back the next [17:14] summer and then I came back full-time. [17:16] >> Okay, that's So the the there kind of [17:18] speaks volumes. [17:18] >> [laughter] [17:19] >> But your path was even less inertial [17:20] than that, right? Cuz you kind of went [17:22] back to academia and then came [17:24] >> Yeah, so I after I came, I worked at [17:25] Jane Street for a couple of years, and [17:27] then I did go on leave and get my PhD [17:28] and then come back. And that was like a [17:30] super lovely experience. [17:32] >> Do you like how normal is that? You you [17:34] took a little bit of a leave to do the [17:35] PhD. Is that a little bit of an [17:38] exception or are there other Jane [17:39] Streeters that have a similar um leave [17:42] for PhD return? [17:43] >> I think I'm one of the very few people [17:45] who did the particular thing that I did, [17:47] but I think there are like lots of [17:48] people at Jane Street who've had their [17:50] own unique particular experiences, and I [17:52] think I don't know, I feel like people [17:53] at Jane Street have been super flexible [17:55] about making, you know, people's [17:56] individual circumstances possible. I I [17:58] think I'm one of the very few people who [18:00] like went on like pseudo leave for a few [18:02] years to get a PhD in the middle of my [18:03] career. [18:04] >> What was the What was the interview [18:05] process like? [18:06] >> Oh, the interview process was like super [18:08] fun. Uh I was asked like a mix of like [18:11] probability questions and like kind of [18:13] more like miscellaneous problem-solving [18:14] questions. I think the phone interviews [18:15] are really fun. When I came to the [18:17] onsite interviews, I think it was like [18:19] very different than what I was [18:20] expecting. I think I was like kind of [18:21] expecting to do like more like even more [18:22] advanced math. And then I think that [18:24] onsite interview was like playing a [18:25] bunch of games, um which I found, you [18:27] know, very fun, but also in the moment [18:29] very stressful. I think I was like not [18:31] really prepared. I didn't have a lot of [18:32] friends who'd gone through the interview [18:33] process. I think it was like I don't [18:34] know if people call it type two fun like [18:36] in hindsight like, "Wow, that was such [18:37] an interesting interview process." In [18:39] the moment, I don't think I would've [18:40] told you that. [18:41] >> Yeah, I got you. [18:41] >> I I really remember that I had like a [18:43] lot of interviews scheduled during my [18:44] day, and my final interview was with [18:46] with Shander, and we were like throwing [18:47] chips at the wall. Um and uh during that [18:50] interview midway through, um two [18:51] recruiters ran in and said, "She's going [18:52] to miss She's going to miss her flight. [18:53] And she's going to miss her flight." And [18:54] then I was like shoved out of the [18:55] hallway [18:56] >> [laughter] [18:56] >> onto an airplane. And I was like, "Wow, [18:58] I really messed up and failed the [18:59] interview." [19:00] Um but it seemed to go okay for me. [19:02] >> Is there anything when you're like [19:04] interviewing people that you feel has [19:05] changed uh kind of the the things that [19:07] you were looking for, the ways that you [19:08] would interview, you know, 5 years ago [19:10] versus now? [19:12] >> The bulk of our strategy has always been [19:14] to hire people who are just like smart [19:16] and hardworking and reasonable and [19:18] adaptable and like to learn new things [19:20] and then you can just do the new thing. [19:22] Like now for sure like machine learning [19:23] skills are um [19:25] a plus. [19:27] But I think [19:28] mostly it hasn't changed a super ton [19:31] just because we're mostly looking for [19:33] these very general kind of skills. So we [19:35] deliberately try to like [19:37] kind of have like math puzzles but not [19:39] like something where if you took the [19:41] right class you would have the tool and [19:43] then it's easy but more just something [19:44] to talk about and think through that's [19:46] kind of [19:47] maybe a little bit novel to them or [19:48] something like that. [19:49] >> Which is not easy to do, right? Like I [19:51] think it's very easy to like hire the [19:53] person who's smart and generalist and [19:55] like can adapt to a new thing but way [19:57] easier said than done. [19:58] >> Yeah, and I think also it relates to [20:00] your previous question about like [20:01] turnover and how much people tend to [20:03] stay here. It wouldn't be worth your [20:04] time if the median time until somebody [20:07] left was in 3 years to spend a year [20:09] retooling them on ML. You'd rather just [20:11] go and like hire somebody, right? Um but [20:14] if you're pretty sure that most of these [20:15] people are going to be here for a long [20:16] time then it like makes more sense to [20:18] invest in that kind of thing. [20:20] >> Yeah, a really tricky thing about [20:21] interviewing um is uh especially [20:23] interviewing for a job where you don't [20:26] come in with the full skills that needed [20:28] to do the job is that you're [20:30] basically trying to evaluate like [20:32] what will a person's like skills and [20:35] output look like a year from now or 5 [20:37] years from now. Um [20:39] and [20:41] yeah, this is really hard to do in like [20:42] an hour or like multiple 1-hour sessions [20:45] or something like that. Like how do you [20:46] how do you tell if you can like, you [20:48] know, train a person over the course of [20:49] months to be highly effective at [20:51] something. [20:52] Um [20:54] and there's no there's no punchline [20:55] here. I don't have like oh yeah, we we [20:57] did it. We did it with this and no, we [20:58] didn't. Oh, we like we like try we try [21:00] to do the sticks. We like try to, you [21:01] know, [21:02] the strategy of like [21:04] find people who can like learn to do a [21:07] bunch of things and um basically find [21:09] like generalists who are really [21:11] adaptable has worked really well and I [21:13] think it's also one of the ways that [21:15] we've managed to adapt to like new [21:17] demands of the job quickly because like [21:19] if someone was good at learning and [21:20] applying new skills 5 years ago, even if [21:23] the skill set that we, you know, are [21:25] demanding from them has changed quite a [21:26] bit now, they're probably still pretty [21:27] good at that. [21:29] >> And it makes sense that that would kind [21:30] of line up with the kind of people who [21:31] just want to learn. The that seems to be [21:33] like what what makes the career a little [21:35] bit more fulfilling. Alok, uh kind of [21:38] same question for you on the origin [21:40] stories. Like where were you before Jane [21:41] Street? What was the trajectory into it? [21:43] What was the interview process like? [21:45] >> Yeah, so I I also came in through a [21:48] relatively like typical pipeline. I [21:50] guess I applied to the the research [21:51] internship for summer of 2018. [21:55] At the time I was studying math and [21:56] computer science, but it was always kind [21:57] of like, I don't know, my general [21:59] assumption was like computer science is [22:01] the one you do to like get a job and [22:03] then math is like [22:05] quite interesting, but like I don't know [22:06] if it uses that in like [22:07] >> Was that the motive for like joining CS [22:09] with it? Was like kind of a vocational [22:11] one? [22:12] >> Not exactly. Like I also did just enjoy [22:14] CS for its own sake, but I had kind of [22:16] imagined in my mind like, "Okay, like [22:19] the jobs I'm considering are the jobs [22:20] that will like use the CS part of my [22:22] background. Like if I get to do any [22:23] math, that's like a fun bonus, but like [22:26] not really a a criteria I can set." Um [22:29] but like I I guess even at the time it [22:31] was [22:31] kind of clear in my mind that there were [22:34] roughly two fields that cared about like [22:37] both these things and one of them was [22:38] like broadly machine learning and data [22:40] science and the other was like whatever [22:42] they were doing over in finance. Um [22:44] wasn't really sure at the time. My [22:46] expectation was like, "I don't know. [22:47] There's like surely there's a bunch of [22:49] people like sitting around with their [22:50] Gaussian brain models that like, you [22:52] know, explain everything." Um [22:54] or at least explain everything in like [22:55] this very specific domain. Um I guess [22:57] after after spending a summer there I [22:59] realized there was like [23:01] in many it's more similar to data [23:04] science elsewhere that I give it credit [23:06] for. Like I was actually kind of shocked [23:08] at how well a lot of the other stuff [23:10] that I've been studying kind of [23:12] transferred or like [23:13] ideas I had learned in college classes [23:15] that seemed pretty abstract and like [23:16] theoretical actually found like pretty [23:18] concrete applications in just like [23:20] analyzing the modeling we were doing and [23:22] things like that. So that was that was a [23:23] really interesting surprise. I was [23:25] really kind of glad to to learn that. [23:27] >> Examples are always hard, but does an [23:29] example pop into mind of that abstract [23:31] way of thinking that you as the young [23:33] math and CS student didn't think would [23:35] pop in and suddenly it did? [23:37] >> Sure. So one thing that I had studied [23:38] just like the semester before in a in a [23:40] stats class I was taking was like the [23:42] theory of N-estimators, which is this [23:45] idea from a guy named Huber in like the [23:48] 1900s and it seemed like I don't know [23:51] the kind of thing where [23:53] the theoretical knots researchers like [23:56] come up with something that is like [23:57] asymptotically correct or only if you're [23:59] working at a galactic scale and things. [24:01] Like I don't know, there are a lot of [24:02] cases like this, but uh turns out this [24:04] is actually just like a pretty [24:06] concretely useful thing when you're [24:07] trying to deal with like [24:09] malformed data or things like that. Like [24:11] it it just [24:12] sometimes it just is is better. Like it [24:15] just using these [24:17] this kind of outer class of like robust [24:19] estimators. So this is like what I I [24:20] think I worked on during the internship [24:22] actually. Like try to apply this theory [24:24] to a concrete like modeling problem. [24:26] >> Oh, that's good. That's good. [24:28] Now Craig, I feel like you and I have [24:29] crossed paths a number of times in the [24:30] past. And so this is not just a [24:32] cookie-cutter question. It's actually [24:33] sincere. I I don't know. How did you [24:35] come to Jane Street? [24:37] What was that? Cuz I think this [24:39] >> [clears throat] [24:39] >> might be a slightly different answer [24:40] from [24:41] >> Yeah, it is a little bit different [24:42] [laughter] because [24:43] the company was pretty small and like [24:45] people sort of joked that I was raised [24:46] by wolves or something like that because [24:48] [laughter] [24:48] you know, it seems like you know, most [24:50] people these days have come from some [24:52] decent school and like you know, Uh, [24:55] they knew each other from math camp or [24:57] or you know something like that. For me [24:58] it was pretty um, [25:00] random. [25:01] In fact, I think [25:03] or you could call this a very not random [25:05] in a way. I only applied to one job [25:08] ever basically. I just I saw a posting [25:11] on the OCaml mailing list and I had made [25:13] a resume just kind of for practice but [25:15] mostly I was planning to start my PhD [25:17] and then [25:19] you know, I just thought, well, I'll [25:20] send it in, you know, and the process [25:22] was very different back then. Like I [25:23] don't think I would get hired now. Um, [25:25] they only did like one phone interview, [25:27] you know, and then I I came in in person [25:29] because free trip to New York, why not, [25:31] you know? And then I just end up liking [25:32] the people I met and stuff like that and [25:34] so I took the job and that was in like [25:37] 2005 and here I am. [25:38] >> Let me back up a little bit. So you're [25:40] you're following OCaml a little bit. Are [25:42] you able to just describe one, what that [25:44] is and then uh, like why were you into [25:47] it? Why are you like looking at the jobs [25:49] postings on this kind of obscure [25:51] language? [25:51] >> Oh yeah, I mean just nerd broadly, [25:52] right? That's a baseline. So whenever I [25:54] >> it for? Like what were the first [25:56] projects [25:56] >> projects, you know? Like I want to like [25:58] graph a fractal or something or like, [26:00] you know, when you just kind of like use [26:01] a computer to do fun things, I would by [26:04] default [26:05] use OCaml, you know, made a little uh, a [26:07] fellow bot, you know, if you made that [26:09] if you know that board game like [26:10] >> Oh yeah, yeah. [26:10] >> just to kind of see how do you make bots [26:12] for games, you know? Um, so it's just [26:14] stuff like that, nothing serious, you [26:16] know? [26:17] >> What do y'all do for fun? [26:20] >> I would call your your physics lessons [26:22] in the office hanging out. Like like [26:24] either the power of graphs or something [26:26] or what is all this stuff? [26:27] >> Okay, yeah, every so often I give like [26:28] an ad hoc like math or physics talk in a [26:31] in a side room. [26:33] >> Uh, [26:33] >> What was the most recent ad hoc Alec [26:35] talk? [26:36] >> Uh, something about system mechanics and [26:38] like we're talking about some partition [26:40] functions and then we ended up on the [26:41] topic of like how you could appreciate [26:43] anything basically uh, as long as you [26:46] you know, expand your mind enough. [26:48] >> It was Feynman diagrams, wasn't it? [26:49] >> Um, yeah, so we ended up so that then we [26:52] were talking about how Feynman diagrams [26:54] kind of relate to partition functions in [26:55] statistical mechanics and like they're [26:56] kind of they're very, very closely [26:58] linked to things. Um but but now you get [27:00] into things like Yeah, e to the power of [27:03] graphs and it starts getting a little a [27:04] little odd. [27:05] >> Yeah. Yeah, no. That's just such classic [27:08] like Three Blue and Brown clickbait on [27:10] like [laughter] [27:10] e to the power of a graph [27:12] >> Yeah, you should do that one. [27:13] >> No, no, that would be fun. That actually [27:14] would be fun. I mean, honestly, that was [27:16] extremely enlightening, I think. Just [27:18] I've I've chatted with you know, like [27:19] Kenji I guess crossed your path before [27:21] and Nithin we've talked before, but even [27:22] then I feel like there's a lot of um [27:24] question marks I have about Jane Street. [27:25] I think this helped clarify things. I [27:27] think it'll also help just make it a lot [27:29] more real for whoever might be watching [27:31] this. So, let me just say thank you so [27:33] much for taking the time. I really [27:34] enjoyed this. [27:37] >> [music]