Two Sigma AI Chief Schuster Restricts LLMs to Coding, Text Extraction
Ai4
Decades of proprietary data and distributed infrastructure create quantitative investing moats that large language models and lone operators cannot replicate.
## Background
Published on YouTube by the Ai4 channel on August 29, 2025, this fireside chat features Mike Schuster, Managing Director and Head of the AI Core Team at Two Sigma, in conversation with Gregor Schubert, Assistant Professor of Finance at UCLA Anderson. According to the page description, the discussion focuses on navigating AI hype to identify impactful tools in investment management, the necessity of programming and human expertise, and establishing a realistic perspective on LLM capabilities.
## TL;DR
Mike Schuster shares how Two Sigma transitioned deep neural networks, vector embeddings, and reinforcement learning into high-impact quantitative production systems. While large language models significantly boost developer productivity and streamline unstructured text extraction from corporate earnings and Fed speeches, direct market prediction remains constrained by severe data scarcity and low signal-to-noise ratios. Schuster argues that true competitive moats rely on decades of proprietary data and specialized institutional scale rather than off-the-shelf generative AI wrappers. Ultimately, high-performing AI teams require foundational programming mastery across C++ and Python, scientific common sense, and a research culture that embraces ambiguity and experimental failure.
## Atmosphere & style
The conversation carries a grounded, measured, and intellectually rigorous tone that methodically deconstructs speculative AI hype. Schubert’s structured academic framing pairs seamlessly with Schuster’s calm, battle-tested engineering pragmatism, creating a candid and collegial fireside exchange. Listening to the session feels like an insider look into the realities of production machine learning, characterized by quiet confidence, technical realism, and a refreshing emphasis on scientific patience over industry sensationalism.
## Key takeaways
- [01:10](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=70s) Integrating modern deep learning and reinforcement learning into quantitative finance enables firms to systematically replace operational bottlenecks with GPU-accelerated neural networks and automated decision pipelines.
- [03:50](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=230s) Within Two Sigma's Core AI stack, vector embeddings serve as pervasive foundational primitives and reinforcement learning drives decision tasks, while LLMs function primarily as research and coding assistants rather than direct market prediction engines.
- [06:08](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=368s) LLMs deliver immediate developer productivity gains by automating codebase audits, identifying low-level hardware and memory optimization targets (such as SIMD/AVX instructions), and synthesizing comprehensive test suites.
- [08:28](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=508s) Enterprise deployment of generative AI in finance demands strict data isolation to safeguard intellectual property, alongside mandatory human verification to prevent subtle hallucinations from corrupting production systems.
- [10:38](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=638s) A quantitative firm's core moat relies on decades of accumulated proprietary data, computing scale, and deep institutional expertise across hundreds of PhDs—assets that cannot be replicated by consumer-grade LLM wrappers.
- [10:38](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=638s) Quantitative finance faces a unique paradox of holding massive overall datasets while constantly struggling with acute data scarcity when testing specific daily forecasting hypotheses (e.g., ~250 trading closes per year).
- [14:27](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=867s) A primary high-leverage application of LLMs in finance is unstructured text feature extraction, reducing the time needed to parse, clean, and structure decades of corporate earnings calls and central bank speeches from weeks to roughly an hour.
- [17:54](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1074s) Top-tier AI engineering demands full-stack proficiency across Python, C++, and Java applied to multi-week distributed training pipelines, guided by 'scientific common sense' and cross-team collaboration.
- [21:43](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1303s) While lean startups possess agile execution speeds, unseating established quantitative incumbents is exceptionally difficult due to massive barriers in historical data, tooling complexity, and institutional knowledge.
- [24:12](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1452s) Transitioning from academic settings to industrial AI requires normalizing failure and managing prolonged ambiguity, as sustained risk-taking is essential for novel algorithmic breakthroughs.
- [28:38](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1718s) The greatest near-term impact for LLMs lies in practical workflow improvements, such as replacing multi-tier customer support queues with zero-wait automated agents and streamlining legacy user interfaces.
- [32:32](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1952s) Building persistent memory and automated context-gathering into AI systems is the critical next frontier to remove the friction of users repeatedly re-explaining background information.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=4s) Background and Career Journey: From Speech Recognition to Google
Gregor Schubert opens the conversation with a brief overview of Mike Schuster's trajectory in computer science and machine learning. Mike grew up in Germany, earned his PhD in Japan, and spent a long tenure at Google working on core technologies like recommendation systems and machine translation before transitioning to quantitative investment firm Two Sigma in 2018. Adding color to the biographical sketch, Mike clarifies that his foundational entry point into the machine learning domain was speech recognition. His doctoral thesis in Japan centered heavily on speech systems, which led him to work at Nuance Communications and subsequently Google during an era when modern deep learning paradigms were just beginning to revolutionize natural language and audio processing.
- **Academic background:** PhD completed in Japan focusing on speech recognition research.
- **Industry experience:** Nuance Communications, followed by 12 years as a research scientist at Google (focusing on speech recognition, machine translation, and recommendation systems).
- **Two Sigma:** Joined in 2018 as Managing Director and Head of the Core AI Team.
---
[01:10](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=70s) Establishing Core AI at Two Sigma and the Financial Sector Shift
Reflecting on his transition from Big Tech to quantitative finance, Mike notes that at Google and earlier firms, research scientists operated with substantial autonomy, often tackling open-ended technological questions. Moving from the West Coast to the East Coast in 2018 to join Two Sigma, he encountered an open mandate when inquiring about his core objectives: "they said well, you know, you figure it out." Rather than viewing this ambiguity as a roadblock, Mike welcomed the creative freedom to formulate a bottom-up roadmap, assemble a core team, and scan the entire enterprise for high-value applications where modern deep learning tools could replace or augment existing approaches.
Mike and his team systematically explored business units across Two Sigma to identify bottlenecks ripe for GPU acceleration, deep neural networks, and modern frameworks such as TensorFlow and PyTorch. Over the course of seven years, the Core AI team expanded beyond traditional supervised modeling into complex automation pipelines and reinforcement learning, which was just gaining significant traction in the late 2010s. This integration has reshaped how the firm approaches automated decision-making and alpha prediction.
```mermaid
flowchart LR
accTitle: Two Sigma Core AI Integration Workflow
subgraph S1["Exploration"]
A["Identify Internal Business Bottlenecks"] --> B["Select Modern ML Frameworks<br/>(TensorFlow, PyTorch, GPUs)"]
end
subgraph S2["Implementation"]
B --> C["Core Predictive Modeling & Embeddings"]
B --> D["Reinforcement Learning Pipelines"]
end
subgraph S3["Production"]
C --> E["Alpha Generation & Automation"]
D --> E
end
```
---
[03:50](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=230s) The Core AI Tech Stack: Embeddings, Reinforcement Learning, and Chatbots
When asked about the concrete components of Two Sigma's modern AI toolkit, Mike breaks down how technologies like vector embeddings, reinforcement learning (RL), and conversational interfaces are deployed. While Two Sigma has leveraged machine learning for its entire 25-year history, early approaches relied on simpler statistical formulations. Today, neural networks with vast parameter counts train on terabytes of financial data utilizing distributed GPU clusters. Within this framework, embeddings—which Mike characterizes plainly as internal vector representations within a neural network—are fundamental building blocks used across virtually all quantitative models.
```
+-------------------------------------------------------------------------------+
| AI Technology Usage in Core AI |
+--------------------------+----------------------------------------------------+
| Technology | Status & Primary Application |
+--------------------------+----------------------------------------------------+
| Vector Embeddings | Core primitive; pervasive across all multi-terabyte|
| | neural network models and representations. |
+--------------------------+----------------------------------------------------+
| Reinforcement Learning | Built from scratch at Two Sigma; scaled up into |
| | substantial production success for decision tasks. |
+--------------------------+----------------------------------------------------+
| Chatbots & LLMs | Deployed for research assistance and developer |
| | productivity; not used as direct prediction engines.|
+--------------------------+----------------------------------------------------+
```
Reinforcement learning represented a greenfield initiative when Mike's team first introduced it to the firm. Starting from minimal proof-of-concept experiments, RL has since matured into a highly successful core competency for sequential decision-making and policy optimization. Conversely, conversational chatbots and generative sentence-completion tools are primarily utilized for helper utilities and exploratory research rather than direct market modeling within the Core AI team.
---
[06:08](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=368s) Developer Productivity and LLM-Assisted Software Engineering
Addressing the broader economic debate surrounding generative AI productivity gains, Mike shares his hands-on experience using modern large language models to modernize software infrastructure. When joining Two Sigma, Mike had developed an extensive, bespoke C++ library. To evaluate LLM capabilities firsthand, he spent an hour setting up an LLM-assisted workflow to audit the codebase for optimization targets. The model rapidly returned high-leverage recommendations that Mike recognized as sound engineering improvements, but which normally would have required weeks of tedious manual profiling:
- **SIMD/AVX vectorization:** Identifying low-level optimization sites where Advanced Vector Extensions (AVX) instructions could accelerate numerical computations.
- **Memory management:** Recommending modernized memory allocation patterns and cache layout refactors.
- **Automated test synthesis:** Generating comprehensive test suites and unit harnesses before applying structural changes.
While human domain knowledge remains essential to direct the refactor and prevent the adoption of faulty suggestions, nearly every engineering team across Two Sigma now integrates coding assistants into their daily workflows to streamline development velocity and eliminate boilerplate overhead.
---
[08:28](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=508s) Safeguards, Security, and Governance Around Enterprise AI
Because financial institutions operate under stringent regulatory mandates and handle highly sensitive proprietary strategies, deploying large language models requires strict security boundaries. Two Sigma's earliest hurdle—even prior to the widespread public release of ChatGPT—was ensuring enterprise data isolation so internal IP and code never leak out to public frontier models.
Beyond network-level security and local hosting considerations, strict governance protocols govern model execution. Two Sigma prohibits unconstrained feedback mechanisms, such as feeding an LLM's output directly back into an autonomous optimization loop without human oversight. Because small hallucinations can trigger critical downstream errors in quantitative finance, engineering standards require teams to verify generated outputs "in 50 different ways" before deploying them to production pipelines.
[10:38](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=638s) Institutional Moats and the Financial Data Dilemma
Two Sigma operates at a massive institutional scale, employing roughly 1,800 people across diverse domains including asset management, real estate, and venture capital. With that scale comes 25 years of accumulated institutional knowledge and a research bench of 300 to 400 PhDs who have spent decades tackling intricate quantitative problems. Mike Schuster emphasizes that this depth of collective experience cannot simply be replicated by an individual working from home with a large language model. While modern AI tooling provides widespread efficiency gains, the internet often fosters an illusion that complex real-world challenges are easy to solve. In practice, the multifaceted problems Two Sigma tackles are far more complex than generative AI wrappers can handle alone. Gregor Schubert notes that his academic research aligns with this view: companies capturing substantial stock-market productivity gains from generative AI are those possessing deep preexisting infrastructure and proprietary data assets.
```mermaid
flowchart TD
accTitle: Two Sigma Institutional Advantage vs LLM Capabilities
subgraph Moat ["Two Sigma Core Moat"]
A["25 Years of Proprietary Data"]
B["300-400 STEM PhDs"]
C["High-Scale Computing Infrastructure"]
end
subgraph Limitations ["The Financial Reality"]
D["Small Sample Sizes<br/>(e.g., 250 Daily Data Points/Year)"]
E["High Signal-to-Noise Ratio"]
F["Multi-Year Statistical Validation"]
end
Moat --> G["Robust Model Deployment"]
Limitations --> G
```
However, data in quantitative finance presents a unique paradox. Two Sigma has collected and purchased enormous quantities of data over a quarter-century—often holding far too much data to process fully, similar to major tech companies. At the exact same time, financial research constantly suffers from acute data scarcity when testing specific hypotheses. For instance, attempting to forecast Apple’s closing stock price one day in advance yields only about 250 trading data points per year. At that sample rate, rigorous statistical validation requires observing a model's out-of-sample performance over multiple years to prove true predictive power—a timeline firms rarely have the luxury to sit through.
- Two Sigma employs approximately 1,800 people and maintains a quantitative research staff of 300 to 400 PhDs.
- Daily asset forecasting operates under severe statistical constraints (e.g., Apple stock has only ~250 trading day closes per year).
- Empirical research confirms that generative AI productivity boosts accrue primarily to firms with proprietary datasets and mature internal infrastructure.
---
[14:27](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=867s) Practical Machine Learning: From Neural Nets to LLM Feature Extraction
When Schuster first arrived at Two Sigma seven years prior, a prevailing sentiment among some practitioners was that "neural networks don't work" in quantitative finance. Treating that skepticism with technical rigor, Schuster questioned the premise, demonstrating that success depends entirely on how networks are architected, regularized, trained, and evaluated. Over the subsequent seven years, deep neural networks and reinforcement learning—the algorithmic foundation behind DeepMind's AlphaGo—have established widespread, high-value utility across the entire firm.
```mermaid
flowchart LR
accTitle: LLM Feature Extraction Workflow in Quantitative Research
subgraph Extraction ["1. Unstructured Data Extraction"]
A["20 Years of Fed Speeches & Earnings Calls"] --> B["LLM Ingestion & Transcription<br/>(Completed in 1 Hour vs 1 Week)"]
end
subgraph Analysis ["2. Signal Generation"]
B --> C["Feature Extraction<br/>(Keyword Density & Policy Signals)"]
C --> D["Macro Predictive Models<br/>(Interest Rate Shift Forecasts)"]
end
```
Addressing modern LLMs, Schuster illustrates how they dramatically accelerate quantitative workflows through unstructured feature extraction. Rather than spending a week having an engineer manually aggregate, clean, and transcribe twenty years of Federal Reserve Chair speeches, a researcher can prompt an LLM to assemble and structure the corpus in roughly an hour. The quantitative pipeline can then parse semantic signals—such as the frequency of specific hawkish or dovish phrases—to model potential interest rate movements following Fed meetings. Schubert corroborates that text-based feature extraction from earnings calls and central bank addresses has quickly emerged as the primary mainstream use case for LLMs across academic finance.
- Schuster overcame initial internal skepticism regarding deep learning; neural networks and reinforcement learning now operate firm-wide.
- LLM feature extraction reduces unstructured data preparation cycles from weeks to roughly an hour.
- Large language models excel at filtering noise and parsing dense, non-standardized text corpora such as corporate earnings transcripts and macroeconomic speeches.
---
[17:54](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1074s) Navigating the AI Talent Market and Essential Engineering Skills
In light of soaring global demand and sensational media headlines reporting "$100 million salaries" for top AI researchers, Schuster offers a grounded perspective on hiring. Over the past decade, the talent pendulum has repeatedly swung between big tech and quantitative finance. While the current zeitgeist strongly favors pure tech labs, Schuster advises people to "relax a little," noting that industry hype cycles inevitably rebalance. Finding exceptional talent has always been difficult because the role demands a rare convergence of hard technical capabilities, architectural intuition, and collaborative temperament.
On the technical side, candidates need strong programming foundations—specifically across Python, C++, and Java—alongside proficiency in deep learning frameworks like PyTorch and TensorFlow. While STEM disciplines (computer science, mathematics, physics) are typical, candidates from non-STEM backgrounds like English are welcomed if they possess elite coding chops. More critically, researchers must master "complex systems"—architectures managing terabytes of streaming data and millions of parameters that require days or weeks to train, where a single subtle bug can invalidate the entire pipeline. Schuster compares holding the mental model of these interconnected systems to having the composition of an entire symphony in one's head. Alongside technical aptitude, two other traits are mandatory: "scientific common sense" (the disciplined pragmatism to solve the actual problem directly rather than overcomplicating it) and strong soft skills to ensure transparent communication in a high-stakes, competitive industry.
- Core technical toolkit requirements: Python, C++, Java, PyTorch, and TensorFlow applied across large-scale distributed systems.
- High-parameter models require mental mastery over complex systems where single coding errors break multi-week training runs.
- Success depends on "scientific common sense"—practical, direct problem solving—combined with cross-team communication and collaborative culture.
[21:43](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1303s) Collaborative Systems and the Moat of Scale
As technological architectures grow increasingly complex, individual contributors can no longer build and manage entire pipelines alone. Modern quantitative and AI systems demand tight teamwork to set up unit tests, optimize computational efficiency, and maintain collaborative feedback loops. Mike Schuster notes that while finding top-tier talent from elite universities is achievable, finding candidates who combine strong technical chops with genuine collaborative instincts is challenging. Gregor Schubert asks whether the rising baseline complexity of internal tooling, data pipelines, and infrastructure creates high barriers to entry that favor massive, established funds over newer, leaner startups.
```mermaid
flowchart TD
accTitle: Big Fund Advantages vs Startup Dynamics
subgraph Established["Scale & Incumbency Advantages"]
D["Decades of Data & Infrastructure"]
E["Hundreds of Seasoned Engineers"]
K["Institutional Knowledge"]
end
subgraph Startups["Startup Trade-Offs"]
S["Agile, Fast Movement"]
B["High Barrier to Complex Replication"]
end
D & E & K -->|High Barrier to Entry| B
```
Schuster acknowledges that while a small team can occasionally build a specialized firm from scratch, competing directly with established incumbents remains exceptionally difficult. An organization with decades of institutional knowledge, massive historical data troves, and hundreds of experienced researchers possesses an enormous structural moat. While smaller startups have the advantage of being able to move faster without bureaucratic drag, dislodging a major fund is akin to attempting to replace an automotive giant overnight.
- Building modern quantitative systems requires collaborative engineering practices, including shared unit testing and code optimization.
- Two Sigma's moat rests on roughly 25 years of institutional experience, terabytes of accumulated proprietary data, and around 400 specialized practitioners.
- Schuster compares the difficulty of unseating an established fund to a new market entrant trying to "replace Toyota tomorrow."
---
[24:12](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1452s) Bridging the Academic Gap and Normalizing Research Failure
Schubert raises the question of how academic AI research translates to complex production environments, pointing out that academic papers often rely on isolated toy datasets explored by small teams. Schuster explains that although Two Sigma hires heavily from PhD programs and emphasizes scientific rigor, recent graduates often struggle with the transition from structured academic success to open-ended industrial research. In university programs, top students are conditioned to complete six-week courses and consistently receive top marks without experiencing substantial setbacks. In a quantitative industrial setting, reality is far messier, and genuine breakthroughs require dealing with persistent ambiguity and frequent dead ends.
```mermaid
flowchart LR
accTitle: Research Mindset vs Academic Conditioning
subgraph Academia["Academic Conditioning"]
A1["Structured Problem"] --> A2["Short Timeframe"]
A2 --> A3["Guaranteed Success / High Grade"]
end
subgraph Industry["Applied AI Research"]
I1["High Uncertainty"] --> I2["Frequent Failure & Exploration"]
I2 --> I3["Iterative Breakthrough"]
end
```
When Schuster joined the firm, he observed that new researchers were overly fearful of failing. He actively encouraged them to embrace risk, reminding them that "if you never fail... you're not taking enough risk." He stresses that while long consecutive multi-year slumps must be avoided, occasional failure is inherent to the scientific method. For instance, when the team initially explored reinforcement learning, standard gaming benchmarks provided little guidance on how to properly construct financial evaluation and test sets, requiring prolonged experimentation under high uncertainty.
- High-performing PhD graduates often arrive with little prior experience with failure due to structured academic pipelines.
- Healthy research cultures require tolerating failure as an inevitable byproduct of ambitious risk-taking.
- Applying paradigms like reinforcement learning to finance introduces non-standard implementation challenges, such as designing entirely new testing and validation frameworks.
---
[28:38](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1718s) Cross-Team Culture and Practical LLM Opportunities
When asked about potential cultural friction between open-ended research teams and fast-paced trading desks or portfolio managers, Schuster emphasizes organizational autonomy and mutual respect. Rather than dictating workflows across different divisions, each team operates with the methods best suited to their operational domain. Looking forward toward the broader AI landscape, Schuster highlights that the immediate priority is moving past superficial hype so that individuals and organizations genuinely internalize everyday model usage.
Rather than fixating on speculative claims, Schuster sees immense near-term value in practical, process-oriented applications. Key opportunities include replacing painful customer service queues with comprehensive, zero-wait AI agents that possess the combined knowledge of an entire support organization, as well as fixing unintuitive web interfaces and optimizing legacy workflows. Schubert concurs, highlighting that incorporating persistent memory and automated context-gathering will eliminate the friction of repeatedly re-explaining background details across user interactions.
- Cross-functional harmony inside a quantitative firm relies on mutual autonomy rather than imposing rigid, top-down research doctrines across diverse teams.
- True enterprise AI adoption requires employees to deeply integrate models into daily work rather than merely observing the technology from afar.
- Key impact areas highlighted for LLMs include context-rich automated customer support, user interface (UI) redesign, and workflow optimization.
[32:32](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1952s) Persistent Context and Session Wrap-Up
Reflecting on the trajectory of AI systems and their usability, the discussion concludes with a look at the friction caused by stateless interactions. Rather than users "having to constantly remind them and having to input lots of context," developing systems capable of retaining memory and background across interactions will mark a critical leap forward. Removing this repetitive cognitive load is poised to make AI tools substantially more practical and effective across everyday personal use cases as well as professional interactions with enterprises.
With those closing thoughts on context retention and practical utility, the speakers wrap up the fireside chat, exchanging mutual thanks before the session concludes to applause from the audience.
## Notable quotes
> [01:10](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=70s) they said well, you know, you figure it out. — Mike Schuster
> [08:28](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=508s) in 50 different ways — Mike Schuster
> [14:27](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=867s) neural networks don't work — Mike Schuster
> [17:54](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1074s) $100 million salaries
> [17:54](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1074s) relax a little — Mike Schuster
> [17:54](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1074s) complex systems — Mike Schuster
> [17:54](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1074s) scientific common sense — Mike Schuster
> [21:43](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1303s) replace Toyota tomorrow — Mike Schuster
> [24:12](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1452s) if you never fail... you're not taking enough risk. — Mike Schuster
> [32:32](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1952s) having to constantly remind them and having to input lots of context
## Mentioned
Mike Schuster, Gregor Schubert, Two Sigma, Google, Nuance Communications, UCLA Anderson School of Management, Ai4, ChatGPT, DeepMind, AlphaGo, Apple, Federal Reserve, Toyota, TensorFlow, PyTorch, Python, C++, Java
## Video description
Mike Schuster, Managing Director (Head of AI Core Team), Two Sigma
Gregor Schubert, Assistant Professor of Finance, UCLA Anderson School of Management
Abstracts
Mike Schuster, Head of the AI Core Team at Two Sigma, will share how the firm is approaching the emergence of new powerful AI models; including how to navigate the AI hype to identify and implement the tools that will deliver impact in investment management. Through this discussion, Mike will share the lessons he’s learned on navigating the AI landscape from an extensive career working at the cutting-edge of AI and Machine Learning; including the importance of programming skills, the necessity of human expertise in AI integration, and the need for a realistic understanding of what the latest LLMs can and cannot do.
Two Sigma AI 负责人 Schuster：大模型仅限用于编程与文本提取
数十年的专有数据与分布式基础设施构筑了量化投资护城河，大语言模型与单打独斗者均无法复制。
Mike Schuster 分享了 Two Sigma 如何将深度神经网络、向量嵌入与强化学习转化为高影响力的量化生产系统。尽管大语言模型显著提升了开发者的工作效率，并简化了从企业财报和美联储演讲中提取非结构化文本的流程，但直接进行市场预测仍受到严重数据稀缺和低信噪比的制约。Schuster 认为，真正的竞争护城河依赖于数十年积累的专有数据和专业化的机构规模，而非现成的生成式 AI 套壳产品。最终，高水平的 AI 团队需要精通 C++ 和 Python 的底层编程功底、具备科学常识，以及包容模糊性与实验失败的研究文化。
在 Two Sigma 的核心 AI 技术栈中，向量嵌入（Vector Embeddings）已成为基础核心原语，强化学习（RL）被深度用于驱动自动化决策流程，而大语言模型主要定位为研究与编程辅助工具而非直接的市场预测引擎。 (3:50)
大语言模型为量化软件工程带来了即时生产力提升，能够自动化完成代码库审计、识别底层 SIMD/AVX 硬件向量化与内存优化点，并快速合成测试套件。 (6:08)
量化投资面临独特的“数据悖论”，即宏观数据总量庞大但在特定日度预测假设下（如单只股票每年仅约 250 个收盘点）面临严重的数据稀缺，这使得机构 25 年积累的私有数据与几百名博士的专业经验构成消费级 AI 无法复制的护城河。 (10:38)
大语言模型在量化金融中最高杠杆的实际应用是非结构化文本特征提取，能将解析和清洗数十年的美联储主席演讲及财报电话会议语料的时间从数周大幅缩短至约一小时。 (14:27)
Reading summary
Two Sigma Core AI head Mike Schuster argues that meaningful artificial intelligence impact in quantitative finance stems from deep institutional infrastructure, complex systems engineering, and disciplined experimentation rather than generative AI hype. While large language models accelerate programming and unstructured data extraction, direct financial prediction and decision-making still rely on specialized architectures, proprietary data, and rigorous governance.
Deploying Modern ML and Reinforcement Learning (1:10)
When Mike Schuster joined Two Sigma in 2018 after 12 years at Google, he was tasked with identifying where modern deep learning could augment quantitative methods. Overcoming initial internal skepticism that neural networks were ineffective in finance, his team demonstrated that success relies on appropriate network architecture, regularization, and training. Over seven years, the Core AI team integrated distributed GPU clusters to train neural networks on multi-terabyte datasets, establishing vector embeddings as ubiquitous building blocks across quantitative models.
Schuster's team also built reinforcement learning capabilities from scratch. Although the team had to design novel testing and validation frameworks due to a lack of standard financial benchmarks, reinforcement learning matured into a core competency for sequential decision-making and policy optimization. In contrast, generative chatbots and large language models are utilized for research assistance and developer tooling rather than as direct market prediction engines.
LLM Utilities, Acceleration, and Enterprise Governance (6:08)
Within software development and research pipelines, LLMs deliver major operational efficiencies. Schuster evaluated LLMs on a bespoke C++ codebase, finding that the models rapidly pinpointed high-leverage optimization targets—including SIMD/AVX vectorization and modernized memory management—while generating unit test harnesses that would typically require weeks of manual profiling. In quantitative research, LLMs accelerate unstructured feature extraction; ingesting and structuring twenty years of Federal Reserve speeches or corporate earnings calls takes roughly an hour instead of a week, allowing researchers to rapidly model macro sentiment and policy signals.
Deploying these tools in quantitative finance nevertheless demands strict security and governance. Two Sigma enforces data isolation to protect proprietary code and IP from leaking to commercial frontier models. Furthermore, the firm prohibits unconstrained autonomous feedback loops without human oversight; because small hallucinations can cause critical downstream errors, outputs must be thoroughly verified before entering production pipelines.
Institutional Moats and Financial Data Realities (10:38)
Schuster emphasizes that an individual using public generative AI tools cannot replicate the capabilities of an established quantitative investment firm. Two Sigma’s structural advantage relies on roughly 25 years of proprietary data, specialized computing scale, and a research staff of 300 to 400 STEM PhDs. Schuster compares replacing an established quantitative fund to attempting to "replace Toyota tomorrow." Gregor Schubert corroborates that empirical research shows productivity gains from generative AI accrue primarily to institutions with mature internal infrastructure and proprietary assets.
This institutional scale operates alongside a unique data paradox. Although the firm holds vast quantities of data, testing specific market hypotheses often involves acute data scarcity. Forecasting daily stock closes yields only about 250 trading data points per year for a single asset, requiring years of out-of-sample observation to statistically validate genuine predictive power.
Systems Engineering, Talent, and Research Culture (17:54)
Addressing high-profile reports of extreme AI compensation packages, Schuster notes that talent competition between Big Tech and finance routinely fluctuates across market cycles. Core engineering roles require proficiency in Python, C++, and Java alongside frameworks like PyTorch and TensorFlow. However, Schuster stresses that elite candidates must also possess the architectural intuition to manage complex, distributed systems with multi-week training runs, paired with "scientific common sense" and collaborative skills.
Schuster also highlights a cultural gap among recent PhD graduates entering industrial research. Conditioned by academic programs to expect predictable success over short timeframes, new researchers often fear failure. Because applying cutting-edge techniques to finance involves persistent ambiguity, Schuster encourages calculated risk-taking, warning that an absence of failure indicates insufficient ambition.
Schuster and Schubert conclude that the immediate frontier for enterprise AI lies in practical workflow integration rather than speculative hype. Sustainable impact will come from automating support workflows, refining user interfaces, and equipping systems with persistent context retention so users no longer need to repeatedly re-enter background information.
## Background
本内容由 YouTube 频道 Ai4 于 2025 年 8 月 29 日发布。视频记录了 Two Sigma 董事总经理兼 Core AI 负责人 Mike Schuster 与 UCLA 安德森管理学院金融学助理教授 Gregor Schubert 的对谈，围绕量化投资中如何穿透 AI 炒作、落地现代机器学习模型与大语言模型展开深入讨论。
## TL;DR
Mike Schuster 结合其在 Google 及量化先锋 Two Sigma 的多年研发经验，系统剖析了现代 AI 与深度学习在量化投资中的真实落地场景。他指出，大模型与代码助手能极大加速特征提取与底层 C++ 代码重构，但量化投资的核心壁垒依然建立在 25 年积累的专有数据、深厚工程架构与系统化协作之上，绝非单纯调用 LLM 即可颠覆。对谈还探讨了金融数据的两极困境、顶尖 AI 人才的‘系统级’与‘科学常识’要求，以及在工业界推行‘容忍失败’的科研文化的重要性。
## Atmosphere & style
整场对话节奏紧凑明快、务实接地气，褪去了当前科技界常见的浮躁狂热，充满资深工业界科学家的冷静与严谨。Gregor Schubert 从学术与金融机制切入抛出尖锐问题，而 Mike Schuster 则以举重若轻的幽默和坦诚一一拆解，对话在平实生动的工程实战比喻（如构思交响乐、汽车制造壁垒）中展开，给人一种拨开技术迷雾、回归第一性原理与长期主义价值的清醒感。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=4s) Mike Schuster 从早期语音识别与 Google 机器翻译的长期研发经验出发，于 2018 年加入 Two Sigma 组建 Core AI 团队，全面推进量化投资向现代深度学习与 GPU 计算栈的转型。
- [03:50](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=230s) Two Sigma 的 AI 技术栈将 Embeddings 作为底层通用特征表达，并在特定量化决策场景中落地了强化学习（RL），但明确将 Chatbot 定位为研究辅助工具而非核心量化投资引擎。
- [06:27](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=387s) LLM 在代码工程辅助上表现出极高价值，能快速针对大型 C++ 代码库提出向量化优化（AVX）与内存管理建议并生成单元测试，大幅压缩重构周期。
- [08:28](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=508s) 量化金融领域对大模型设立了严格的安全护栏：严禁核心代码与数据外泄，禁止让大模型输出以无监督闭环方式自我优化，且所有生成结果必须进行全方位交叉验证。
- [10:38](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=638s) 单纯的 LLM 无法颠覆成熟量化机构的壁垒，Two Sigma 依赖 25 年积累的数据管道、交易策略及数百名博士构筑的系统工程经验构成了实质性护城河。
- [12:11](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=731s) 金融量化面临独特的数据两极悖论：一方面拥有海量高噪声数据难以高效利用，另一方面在特定关键预测场景下（如次日收盘价验证）有效统计样本极度匮乏且需漫长时间回测。
- [13:56](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=836s) LLM 极大简化了非结构化文本的特征提取流程，例如将美联储过去 20 年发言纪要的清洗、转录与关联分析耗时从一周压缩至一小时。
- [17:54](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1074s) 量化 AI 团队招聘核心看重对 TB 级与百万参数级复杂系统的全局推演掌控力，以及务实解决问题的“科学常识”与跨团队沟通协作精神。
- [25:18](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1518s) 学术研究常局限于确定性小数据集，工业界量化系统要求研究员适应高噪声环境并建立“从不失败说明承担风险不够”的容错试错文化。
- [30:10](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1810s) 未来 AI 落地应摆脱单纯的概念炒作，重点突破客服知识库整合、极简 UI 工作流重塑，以及无需用户反复提供背景的长效上下文记忆机制。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=4s) 嘉宾背景介绍：从语音识别、Google 到量化先锋 Two Sigma
对谈在主持人 Gregor Schubert（加州大学洛杉矶分校安德森管理学院金融学助理教授）的简短开场中展开。由于时间紧凑，Gregor 直奔主题，介绍了本场嘉宾 Mike Schuster 的履历。Mike 现任量化巨头 Two Sigma 的董事总经理兼 Core AI 团队负责人。Gregor 用极为凝练的一句话概括了他的早年经历：在德国长大，随后赴日本攻读博士学位，之后加入 Google 工作长达 12 年，深入研发过推荐系统和机器翻译，并于 2018 年正式加入 Two Sigma。
Mike 欣然认同了这段“极简人生时间线”，并特别补充了自己职业生涯中不可或缺的基石——语音识别技术。他在攻读博士学位期间就专注于语音识别领域，早期曾在该领域的知名公司 Nuance 工作，随后在 Google 长期深耕。正是语音识别这一技术路径，成为了他迈入机器学习与现代人工智能前沿的起点。
- **教育与职业轨迹**：德国成长 $\to$ 日本获博士学位（主攻语音识别） $\to$ Nuance $\to$ Google（12年，推荐系统、机器翻译、语音识别） $\to$ 2018 年加入 Two Sigma
- **核心职能**：Two Sigma 董事总经理（Managing Director），领导 Core AI 核心团队
---
### [01:10](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=70s) 建立 Core AI 团队：在量化投资中落地现代机器学习与强化学习
谈及在 Two Sigma 的职责与团队定位时，Mike 对比了纯科技巨头与量化金融机构的工作生态。在 Google 期间，研究科学家（Research Scientists）往往以相对独立的研究为主；而在 2018 年，Two Sigma 内部希望将现代 AI 技术系统性地引入金融决策体系，Mike 借此契机从西海岸搬到了东海岸。当他询问管理层自己的具体任务是什么时，得到的回答出乎意料地开放——“你来弄清楚我们该做什么”（"You figure it out"）。
这种看似充满挑战的开放式指令，恰恰契合了 Mike 的工作风格。他接手了一支初期规模不大的团队，拥有充分的自主权去制定规划、探索核心痛点。团队采取了主动出击的方式，在全公司各个业务线中“扫楼式”地寻找落地场景，全面推动从传统统计学向现代深度学习栈的升级：
```mermaid
flowchart TD
accTitle: Two Sigma Core AI 落地探索流程
subgraph S1["需求发掘与技术布局"]
A["走访全公司业务线<br/>寻找痛点"] --> B["引入现代 AI 技术栈<br/>(TensorFlow / PyTorch / GPU)"]
end
subgraph S2["核心业务赋能"]
B --> C["高维特征与市场预测"]
B --> D["交易与系统流程自动化"]
B --> E["前沿强化学习 (RL) 落地"]
end
```
经过七年的持续建设，Core AI 团队将神经网络、GPU 加速计算以及 TensorFlow 和 PyTorch 等现代框架深度嵌入到了预测建模和自动化流程之中，并探索了当时刚刚兴起的强化学习（Reinforcement Learning）技术。
- **核心转型**：从传统量化模型向现代深度神经网络、大规模 GPU 并行训练体系演进
- **落地重点**：金融信号预测、复杂量化系统自动化、端到端强化学习决策
---
### [03:50](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=230s) 技术栈解构：Embeddings 的普及、强化学习的突破与对 Chatbot 的审慎态度
针对主持人关于技术栈构成的具体提问（包括 Embeddings、强化学习与 Chatbot 是否成为标配），Mike 逐一进行了拆解。他指出，Two Sigma 在长达 25 年的量化投资历史中一直在使用机器学习，早期方法相对简单，而如今的模型参数量巨大，依赖数以 TB 计的海量数据在 GPU 集群上训练完成。
谈及技术细节时，Mike 强调不同工具在量化场景下的定位截然不同：
1. **向量嵌入（Embeddings）**：在现代 AI 体系中，Embeddings 本质上就是神经网络内部的特征向量。在 Two Sigma，这早已成为随处可见的底层基础设施，支撑着所有核心模型的特征表达。
2. **强化学习（Reinforcement Learning）**：这是团队从零开始自主引入并落地的技术。最初从小规模简单算例起步，历经反复迭代与打磨，如今已在特定量化决策与执行场景中取得了相当显著的成效。
3. **聊天机器人（Chatbots）与 LLM**：虽然生成式对话系统近年来极为火爆，甚至能被宽泛定义为“自动补全句子的系统”，但 Mike 明确表示 Core AI 团队并不专注于做通用聊天机器人。这类技术在内部更多作为研究辅助或通用生产力工具存在，而非核心量化投资引擎。
---
### [06:27](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=387s) AI 辅助编程实战：重构大型 C++ 代码库与效率提升
当话题转向编程助手（Coding Assistants）对日常研发效率的实际影响时，Mike 分享了自己近期亲身体验的一个生动案例。他在剛加入 Two Sigma 时，曾亲手编写过一套功能庞大的底层 C++ 代码库。为了测试大语言模型（LLM）究竟能带来多少实质性帮助，他花了大约一小时配置好辅助工具，向模型抛出核心问题：“这个库中有什么可以优化的？”
LLM 迅速给出了一整套高质量的重构建议，其精准度令 Mike 大为惊叹：
```mermaid
flowchart LR
accTitle: LLM 辅助大型 C++ 代码库优化工作流
subgraph Step1["LLM 宏观代码审查"]
A["输入大型 C++ 代码库"] --> B["生成优化清单<br/>- 内存管理优化<br/>- AVX 指令集矢量化<br/>- 数据格式改造"]
end
subgraph Step2["交互式落地推进"]
B --> C["优先级排序与方案拆解"]
C --> D["自动生成针对性单测 (Unit Tests)"]
D --> E["开发者把关与代码落地"]
end
```
这些建议包括调整底层数据结构、利用 AVX 指令集进行向量化计算优化，以及改进内存管理等。Mike 直言，虽然这些优化方向作为资深专家的他也完全明白可以做，但如果纯靠人工一行行重构，可能需要耗费数周时间。借助 LLM，开发者可以进一步追问具体实现路径，甚至直接要求模型“先为该模块编写单元测试”，极大压缩了探索与验证周期。目前，Core AI 团队乃至整个 Two Sigma 内部几乎所有人都在日常工作中使用 AI 编程工具，显著提升了工作流效率并简化了复杂工程开发。
- **关键优化维度**：底层数据格式转换、AVX（高级矢量扩展）指令集调用、精细化内存管理优化
- **效率倍增器**：快速搭建测试脚手架（Unit Tests）、协助理清重构优先级，将“数周的手工工作量”大幅压缩
---
### [08:28](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=508s) 金融级安全红线与护栏：验证机制与竞争优势的思考
针对量化投资这一对“精确度与严谨性”要求极高的特殊行业，主持人提问 Two Sigma 如何在内部为 LLM 设立安全护栏。Mike 透露，甚至早在 ChatGPT 引起全球轰动之前，团队在探索大模型时遇到的首要也是最大的难关就是**数据安全性**。
在金融合规与监管的硬性约束下，公司内部核心数据与核心代码绝对不能外泄。在此之上，Two Sigma 制定了极为严格的模型使用规范：
> “你可以用它来写代码，但绝对不允许把 LLM 的输出结果直接以闭环方式喂回系统让它自我优化——因为谁也不知道在无监督的循环迭代中会发生什么。”
除了防止失控的死循环优化外，Mike 强调了结果验证的极端重要性：对 AI 生成的任何输出，团队必须“用 50 种不同的方式进行交叉验证”（"check that it does the right thing... in 50 different ways"），确保万无一失。
在对谈最后，主持人引出深层思考：随着生成式 AI 与前沿机器学习逐步成为量化投资的标准工具箱，像 Two Sigma 这样提早数年布局先机机构所拥有的先发优势是否会被行业迅速抹平？还是说深厚的人才储备与技术积淀将始终构成不可替代的复利壁垒？Mike 随后就此展开了进一步阐述。
### [10:38](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=638s) 护城河与幻象：大模型无法单枪匹马颠覆量化投资
Mike Schuster 介绍了 Two Sigma 的基本体量：全公司拥有约 1,800 名员工，业务覆盖资产管理、房地产投资和风险投资等多个领域。作为量化投资界的领军机构之一，公司积累了 25 年的深厚底蕴，汇聚了三四百位长期深耕该领域的博士。他直言，这种数十年来凝聚的集体智慧与工程经验，绝不是某个个人在家里打开电脑随便跑跑就能轻易超越的。
谈及外界对生成式 AI 的追捧，Mike Schuster 指出，虽然新技术赋予所有人平等的提速机会，但“有了 LLM 就能突然搞定 Two Sigma 过去 25 年所做的一切”纯属无稽之谈。现实世界中的复杂问题远超普通人的想象，无论是金融交易还是其他行业皆是如此。他批评互联网往往给人一种虚假的安全感，让人误以为许多底层困难已被解决，但只要真正卷起袖子去触碰现实业务，就会发现水深得多。
- **Two Sigma 人才储备**：约 1,800 名员工，拥有 300 至 400 名长期致力于量化研究与算法构建的博士。
- **经验壁垒**：历经 25 年积累的数据管道、交易策略与系统工程体系，构成了对抗单纯算法炒作的核心壁垒。
---
### [12:11](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=731s) 金融数据的两极困境：海量噪声与极度匮乏
Gregor Schubert 结合自己在学术界的研究指出，能从生成式 AI 提升生产力并转化为股价增长的公司，往往不是单靠使用工具的能力，而是依赖已有的软硬件基础设施以及**专有数据资产**（proprietary data assets）。Mike Schuster 对此深表赞同，并阐述了金融领域独特的数据悖论：一方面公司 25 年来采买并积累了海量数据，甚至像科技巨头一样多到难以全盘高效利用；但另一方面，在金融的特定关键场景下，**用于验证假设的有效数据往往严重不足**。
为了形象地说明这一困境，Mike Schuster 举了预测苹果公司（Apple）收盘价的例子：如果目标是提前一天预测次日收盘价，一年中只有约 250 个交易日（即 250 个数据样本）。从统计学角度来看，要在如此小样本、强噪声的环境下验证一个模型到底是真的抓住了Alpha还是纯属运气，需要长达数年的回测与实盘检验。在瞬息万变的商业环境中，团队根本无法耗费数年去等待结论，因此数据缺乏与稀疏性是金融量化必须面对的巨大瓶颈。
```mermaid
flowchart TD
accTitle: 金融数据两极困境示意图
subgraph DataSurplus["数据过剩维度（无结构海量噪音）"]
A["25年持续买入与积累的数据"] --> B["体量堪比科技巨头"]
B --> C["噪音极大 / 难以直接喂入模型"]
end
subgraph DataScarcity["数据匮乏维度（有效验证样本极少）"]
D["例如：预测次日收盘价"] --> E["年均仅约 250 个交易日数据点"]
E --> F["统计显著性验证往往需要数年时间跨度"]
end
```
---
### [13:56](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=836s) 从破除“神经网络无用论”到 LLM 的特征提取实验
回顾自己在 Two Sigma 的历程，Mike Schuster 提到他刚加入时，金融圈内普遍流传着“神经网络在金融里不管用”的根深蒂固偏见。作为 AI 领域的资深专家，他当时深感荒谬：“不管用到底是什么意思？”他认为关键在于如何根据问题设定训练方式、测试流程以及构建正确的架构。经过团队七年多的深耕，神经网络已经证明了其巨大价值，并在公司内部全面开花；同样，像战胜人类围棋大师的 AlphaGo 所使用的强化学习（Reinforcement Learning）技术，也逐步被引入到了核心业务流水线中。
对于当前热门的 LLM，Mike Schuster 勾勒了一个极具代表性的工作流场景：假设要预测美联储（Fed）议息会议后的利率变动，传统流程下研究员或工程师需要耗费整整一周去收集、清洗美联储主席过去 20 年的发言纪要并转录文本；而如今利用 LLM，一个人在一小时内就能完成全套自动化数据拉取、转录并初步提取高频词模式与利率波动之间的关联线索。Gregor Schubert 补充道，这类非结构化文本的特征提取（Feature Extraction）如今已成为学术界和量化界分析财报电话会议（Earnings Calls）与央行发言的最主要利器，极大降低了从庞杂噪声中萃取有效信号的门槛。
- **神经网络的平反**：从 7 年前被主流怀疑，到如今成为全公司范围内的标配武器。
- **强化学习落地**：借鉴 AlphaGo 等突破性进展，在复杂决策与交易执行上展开长周期探索。
- **LLM 赋能研究流水线**：将美联储 20 年发言数据收集清洗的耗时从**一周缩短至一小时左右**。
---
### [17:54](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1074s) AI 人才争夺：硬核系统工程、科学常识与团队协作
针对近期科技圈动辄开出数千万甚至上亿美元天价年薪挖角顶级 AI 研究员的热潮，Gregor Schubert 询问量化基金是否也面临极度严峻的人才招聘困难。Mike Schuster 表现得相当淡定，他笑称这种风向在过去十几年里经历过多次摇摆——时而由金融转向科技，时而由科技回流金融。目前科技圈风头正盛，但他劝大家“放松一点”（relax a little），狂热的技术炒作可能明年就会发生变化；而事实上，**寻找真正兼具软硬实力的顶尖人才一直都极度困难**。
在硬技能方面，除了扎实的计算机、数学、物理等 STEM 基础（即便学英语等文科，只要代码过硬也可胜任）以及熟练使用 TensorFlow、PyTorch、Python、C++ 和 Java 之外，他最为强调的是对**“复杂系统”（Complex Systems）的掌控感**。核心交易系统常常涉及数以百万人为调整的参数、海量 TB 级数据并需连续训练数天甚至数周，任何一行代码的细微偏差都会引发系统崩溃。这种对复杂系统在脑海中全局推演的能力，宛如作曲家构思一整部大型交响乐。此外，他还提出了“科学常识”（scientific common sense），即保持务实、不搞虚架子、直奔问题本质去解决的能力。在软实力上，沟通与协作至关重要——尽管金融行业外部竞争残酷，但若在公司内部搞孤岛与内耗，庞大而精密的量化系统将根本无法平稳运转。
| 评估维度 | 核心要求与典型特征 | 关键价值 |
|:--- |:--- |:--- |
| **硬技能 (Hard Skills)** | 深度掌握 Python / C++ / Java / PyTorch / TensorFlow；具备数理或工程推演底子。 | 具备将数学理论无缝转化为高效实盘代码的落地能力。 |
| **系统思维 (System Sense)** | 能在大脑中构建并运转参数达数百万、数据达 TB 级的复杂模型架构，犹如构思交响乐。 | 避免在训练长达数天的庞大系统中因微小错误而导致整体崩溃。 |
| **科学常识 (Common Sense)** | 极其务实接地气，专注于使用最直接有效的方法解决实际痛点，拒绝花架子。 | 避免过度工程化，在成本与算法效果之间求取最优解。 |
| **软技能 (Soft Skills)** | 跨部门清晰沟通、强协作精神、抵制恶性内耗。 | 支撑大规模跨领域团队共同维护高复杂度工程。 |
### [21:43](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1303s) 系统的复杂性壁垒与大机构的护城河
量化与 AI 研发系统的复杂性正在呈指数级上升，个人单打独斗完成端到端任务的时代已经彻底过去。Mike Schuster 指出，这种演变与科技及其他行业的历程如出一辙，如今必须依靠高度协作的团队才能推进项目——从编写单元测试、优化执行效率，到团队内部持续探讨改进方案。然而，要在顶级名校毕业的顶尖人才中找到既具备卓越技术能力又擅长团队协作的人，往往极具挑战性。虽然这种对人才协作特质的要求多年未变，但底层系统的复杂度却已不可同日而语。
面对这种基础设施和工程复杂性，Gregor Schubert 提出疑问：这是否意味着量化投资行业将进一步走向寡头垄断，只有拥有几十年历史、积累海量数据并能负担庞大内部工具链的大型基金才能生存，还是说新兴 AI 工具反倒会成为一种“平民化”力量，促成更多初创基金的诞生？Mike 回应称，虽然初创团队凭借精简的人数在某些方向上行动更快，但大机构的护城河依然难以撼动。他以汽车行业作比：“你不可能明天就跑去把丰田（Toyota）给取代了，这非常困难。”
- **团队协作与工程实践**：现代量化系统要求深入细致的分工，涵盖单元测试构建、性能调优及持续架构评审。
- **历史沉淀与规模效应**：大型机构拥有长达 25 年的实战经验、数十年积累的 PB 级（Terabytes/Petabytes）数据，以及数百位经验丰富的长期从业人员，构筑了极高的竞争壁垒。
- **初创与巨头的动态平衡**：小型团队依靠敏捷性可以探索特定利基市场，但在系统化基础设施和全面竞争上面临巨大天然劣势。
```mermaid
flowchart TD
accTitle: Big Fund Moat vs Startup Agility
subgraph BigFund["大型成熟量化机构（如 Two Sigma）"]
A["25 年历史经验与行业认知"] --> D["坚实壁垒与高执行效率"]
B["数十年积累的专有海量数据集"] --> D
C["数百名资深研究员与工程团队协作"] --> D
end
subgraph Startups["新兴初创团队"]
E["精简团队架构"] --> F["决策链路短、探索速度快"]
F -.->|难以正面突破| D
end
```
---
### [25:18](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1518s) 学术象牙塔到工业界：拥抱不确定性与“失败”的价值
学术界的实验室研究通常由两三个人在小型“玩具数据集”（toy dataset）上验证算法，证明其在单一特定场景下的有效性。当来自顶尖高校的博士生进入工业界时，往往会遭遇巨大的认知冲击。Mike 观察到，这些精英人才在过往的求学经历中几乎从未经历过真正的失败——一门课程上 4 到 6 周，拿个 A 就顺利通关。但在量化工业界的真实复杂系统中，现实远比书本严苛得多，遭遇挫折和失败是家常便饭。
为了打破新员工对失败的恐惧，Mike 在团队中极力推行健康的试错文化。他常对团队强调：“如果你从不失败，那就说明你承担的风险还不够，你没有在做正确的研究（if you never fail, you're not taking enough risk）。”真正的科研必须伴随着巨大的不确定性，研究人员在一开始可能只有一个模糊的方向。例如在探索如何将强化学习（Reinforcement Learning）引入金融场景时，团队就曾因金融环境与最初游戏环境的底层逻辑完全不同，耗费了大量时间摸索如何正确构建测试集。
- **高校与工业界脱节**：学术研究往往局限于简化环境与确定性考核，而真实系统涉及极度复杂的实现挑战与噪声数据。
- **风险与容错机制**：鼓励阶段性试错以换取突破性进展，将“失败”视作科研探索的必然组成部分（但避免连续多年毫无产出）。
- **面对未知的研究心态**：在缺乏现成范式时（如金融时序下的强化学习验证），必须具备耐受模糊性和自主搭建评估体系的能力。
---
### [28:38](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1718s) 基金内部的研究文化与跨团队协同
在对冲基金这种高度关注盈亏与执行结果的环境中，研究团队推行的这种容忍失败、拥抱未知的“科研心态”是否会与交易员或投资组合经理（PM）产生文化冲突？针对 Gregor 的提问，Mike 表示任何大型组织内部不同角色之间存在思维差异都是正常的。每个职能团队都有自己深耕多年的经验与做事方式，关键在于建立相互尊重与包容的协作机制。
Mike 阐述了他的管理哲学：尊重各部门既有的专业方法与路径，研究团队做自己的探索与尝试，既不强行指令其他团队该如何操作，也不受制于单一的教条。这种求同存异、各司其职的开放态度，不仅确保了科研创新团队拥有足够的探索自由度，也保证了整套机构复杂体系的高效运转。
---
### [30:10](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1810s) 超越炒作：LLM 的实际落地与未来应用图景
谈及对前沿 AI 技术的展望（特别是最新一代大模型的演进），Mike 直言，当前最紧迫的不是空谈概念，而是让人们真正“内化并深度使用这些工具”。大众在新闻媒体上看到了大量关于 AI 的炒作，但即便在顶尖科技公司内部，真正深入探索并挖掘模型潜能的人依然不够。Mike 对 AI 的未来持非常积极乐观的态度，认为其将在多个实际流程中释放巨大生产力。
在具体落地场景上，Mike 举了两个极具现实痛点的方向：
1. **智能客户服务系统**：颠覆传统冗长且体验极差的呼叫中心。未来的系统应无需排队等待，直接汇聚整个客服知识库的所有经验，不再需要将用户反复转接五六次还给不出正确答案。
2. **直观的 UI 与业务流程优化**：重塑互联网上大量设计极其晦涩难用的网页与交互界面，利用 AI 大幅简化操作链路与工作流。
Gregor 对此深表赞同，并补充指出，大模型的真正威力在于“个性化与定制化”以及“上下文记忆能力的引入”。当 AI 能够在交互中自动携带用户历史背景与特定情境、不再需要重复解释上下文时，人机协作的效率与体验将迎来质的飞跃。
### [32:32](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1952s) 上下文记忆与 AI 系统的未来实用性及对谈结语
讨论在展望 AI 系统的未来进化方向中走向尾声。讲者总结指出，未来的 AI 如果能够在交互中自主保持长效记忆，使用户不再需要“不断提醒它们并重复输入大量背景信息”（"having to constantly remind them and having to input lots of context"），这将会彻底改变用户体验。这种长上下文与记忆机制的完善，将极大提升各类 AI 系统在个人日常生活以及与企业级交互中的实用价值与落地深度。
对谈在此积极的展望中圆满结束，讲者与主持人互相致谢，全场响起掌声。
- **减少重复提示的必要性**：摆脱每次交互都需要重新输入背景上下文的痛点，是 AI 走向高实用性的关键。
- **跨场景的赋能价值**：长效记忆与无缝上下文理解不仅能优化个人助手体验，也能深化企业内部业务流程与人机协作的效率。
## Notable quotes
> [01:10](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=70s) You figure it out — Two Sigma Management
> [08:28](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=508s) 你可以用它来写代码，但绝对不允许把 LLM 的输出结果直接以闭环方式喂回系统让它自我优化——因为谁也不知道在无监督的循环迭代中会发生什么。 — Mike Schuster
> [08:28](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=508s) check that it does the right thing... in 50 different ways — Mike Schuster
> [10:38](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=638s) 有了 LLM 就能突然搞定 Two Sigma 过去 25 年所做的一切 — Mike Schuster
> [13:56](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=836s) 不管用到底是什么意思？ — Mike Schuster
> [21:43](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1303s) 你不可能明天就跑去把丰田（Toyota）给取代了，这非常困难。 — Mike Schuster
> [25:18](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1518s) if you never fail, you're not taking enough risk — Mike Schuster
> [32:32](https://www.youtube.com/watch?v=E5GO5_u9UBQ&t=1952s) having to constantly remind them and having to input lots of context — Mike Schuster
## Mentioned
Mike Schuster, Gregor Schubert, Two Sigma, UCLA Anderson School of Management, Google, Nuance, TensorFlow, PyTorch, C++, Python, Java, AVX, ChatGPT, Apple, Fed, AlphaGo, Toyota
## Video description
Mike Schuster, Managing Director (Head of AI Core Team), Two Sigma
Gregor Schubert, Assistant Professor of Finance, UCLA Anderson School of Management
Abstracts
Mike Schuster, Head of the AI Core Team at Two Sigma, will share how the firm is approaching the emergence of new powerful AI models; including how to navigate the AI hype to identify and implement the tools that will deliver impact in investment management. Through this discussion, Mike will share the lessons he’s learned on navigating the AI landscape from an extensive career working at the cutting-edge of AI and Machine Learning; including the importance of programming skills, the necessity of human expertise in AI integration, and the need for a realistic understanding of what the latest LLMs can and cannot do.
[Source](https://www.youtube.com/watch?v=E5GO5_u9UBQ)
Source quotes
[1:10] Two Sigma Management: You figure it out
[8:28] Mike Schuster: 你可以用它来写代码，但绝对不允许把 LLM 的输出结果直接以闭环方式喂回系统让它自我优化——因为谁也不知道在无监督的循环迭代中会发生什么。
[8:28] Mike Schuster: check that it does the right thing... in 50 different ways
[10:38] Mike Schuster: 有了 LLM 就能突然搞定 Two Sigma 过去 25 年所做的一切
[13:56] Mike Schuster: 不管用到底是什么意思？
[21:43] Mike Schuster: 你不可能明天就跑去把丰田（Toyota）给取代了，这非常困难。
[25:18] Mike Schuster: if you never fail, you're not taking enough risk
[32:32] Mike Schuster: having to constantly remind them and having to input lots of context
Source quotes
[1:10] Mike Schuster: they said well, you know, you figure it out.
[8:28] Mike Schuster: in 50 different ways
[14:27] Mike Schuster: neural networks don't work
[17:54] $100 million salaries
[17:54] Mike Schuster: relax a little
[17:54] Mike Schuster: complex systems
[17:54] Mike Schuster: scientific common sense
[21:43] Mike Schuster: replace Toyota tomorrow
[24:12] Mike Schuster: if you never fail... you're not taking enough risk.
[32:32] having to constantly remind them and having to input lots of context
Original transcript
[0:04] All right. Um, thank you everyone for
[0:05] coming. Um, so we don't have a lot of
[0:08] time, so I think we'll we'll jump right
[0:10] in. I just wanted to do a a very quick
[0:12] intro to Mike, which will give you some
[0:14] idea of what we're going to be talking
[0:15] about. So Mike is the head of the core
[0:18] AI team at uh, Two Sigma. Uh, he's also
[0:21] managing director at Two Sigma. Um, I
[0:24] think the shortest possible biography
[0:25] for you is you grew up in Germany. You
[0:27] did a PhD in Japan. You ended up at
[0:29] Google eventually and did worked on
[0:31] recommendation systems and machine
[0:33] translation before joining two sigma in
[0:35] 2018. And you can fill in additional
[0:37] details uh as we talk as necessary.
[0:41] >> Uh yes. So that is a good super short um
[0:46] life lifeline I'd say. uh I also worked
[0:49] a lot on actually speech recognition
[0:51] which uh first at a company called
[0:53] nuance and then you know long at Google
[0:56] uh which was a very interesting time I
[0:58] also did my my my uh PhD partially about
[1:02] speech recognition so that's that's how
[1:03] I got into this whole field uh of
[1:05] machine learning and you know all this
[1:08] interesting technology yes
[1:10] >> great so and you've you've worked in a
[1:12] lot of different roles um and so can you
[1:14] tell us more about your current role at
[1:17] two sigma So how what does core AI mean?
[1:21] Where does that fit in terms of the
[1:22] overall organization? Like what does
[1:24] your team look like? What do you do?
[1:26] >> Right. I mean maybe I should first say
[1:28] what I did before I came to Twigma,
[1:30] right? So I was for 12 years in tech at
[1:33] Google and you know at other companies
[1:34] before that and basically was you know
[1:37] we we were called research scientists
[1:40] and uh we were mostly by ourselves uh of
[1:44] course working with other people and in
[1:48] 2018
[1:50] uh I had the chance to come to Twigma uh
[1:53] because the there were people there who
[1:55] wanted um AI what we call today AI or
[1:59] modern technology in a in a financial
[2:02] company and eventually I said okay let's
[2:05] try this this is uh very new I was very
[2:08] interested in it um I had done some
[2:10] stuff by myself but I really didn't know
[2:12] much about it uh at least not from their
[2:15] point of view I had had to move from the
[2:17] west coast to the east coast u and and
[2:20] then they said well um so I said you
[2:24] know what uh what are we supposed to do
[2:26] and and and they said well you know you
[2:28] figure it out right and For me that
[2:31] sounds like like maybe a little bit um
[2:34] like a a difficult thing to do but
[2:36] actually that is something that I
[2:38] personally like because I could make my
[2:40] own plans. uh we you know we inherited a
[2:44] small team we could find our own
[2:45] problems and what we started doing was
[2:48] okay let's walk around the company let's
[2:50] find places where we can use things like
[2:53] neuronet networks where we can use GPUs
[2:57] where we can use uh tools like
[2:59] TensorFlow PyTorch um where you know
[3:02] programming matters very complex systems
[3:06] possibly where we can use reinforcement
[3:08] learning that had just come out at the
[3:10] time uh in a in a in a in a bigger way
[3:14] and I had uh a bunch of experience on on
[3:18] different levels uh in in different
[3:20] technologies and you know we found one
[3:23] problem after the next where we could
[3:25] apply this at to sigma and that is what
[3:27] we have been basically been doing for
[3:29] seven years now uh and you know we can
[3:31] go into more details but it's partially
[3:33] prediction partially helping with
[3:36] automation
[3:38] uh you things like reinforcement
[3:41] learning which which is relatively new
[3:43] uh and and that has been a really good
[3:46] time I think because it's it's the right
[3:48] way for finance to go.
[3:50] >> So maybe we'll we'll stay there for a
[3:53] second. So can you tell us a little bit
[3:54] more about like what technologies
[3:55] actually you ended up finding useful in
[3:58] terms of what you're what the the text
[4:01] now looks like like do you use a lot of
[4:03] embeddings? Is it is it reinforcement
[4:06] learning for prediction mostly? uh are
[4:09] chat bots now a thing that that that
[4:11] you've started using in the company like
[4:13] which technologies are actually being
[4:15] used?
[4:15] >> Yeah, I mean so just these three
[4:17] examples chat bots uh embeddings and
[4:20] what else did you say? Okay. Uh
[4:22] >> enforcement learning or
[4:23] >> enforcement learning. Okay. So um these
[4:25] are of course like very different
[4:27] topics. Um so let's start with the
[4:29] simplest one embeddings. Uh when I came
[4:32] uh so two sigma has done machine
[4:35] learning and what we call today AI
[4:37] really for for the whole time of it its
[4:40] existence for 25 years. uh it was very
[4:44] simple in the beginning uh and it became
[4:46] more and more complex and uh so for me
[4:50] that meant first you know let's use
[4:52] neural networks more uh today we have
[4:55] models you know with with lots of
[4:57] parameters that are trained on on
[4:59] terabytes of data on on GPUs and things
[5:01] like that and what we call embeddings or
[5:04] what people call embeddings in the AI
[5:06] field is is basically just a vector of
[5:09] one of inside of those neural networks
[5:12] and yes of Of course, we're using them
[5:13] all the time for everything. This is
[5:15] this is a very basic thing. So about the
[5:17] other two examples. So reinforcement
[5:19] learning uh we introduced this at two
[5:22] sigma and uh it was we really uh we
[5:25] started from scratch. Uh we tried very
[5:29] simple examples first and eventually it
[5:31] became this this bigger thing and and we
[5:33] have we're having quite some success
[5:34] with it today. Um chat bots that's of
[5:38] course something really recent. uh you
[5:41] know there are probably some that are
[5:43] used in research and we use it for
[5:45] various you know helper functions and
[5:47] things like that uh but you know not
[5:50] this is not something we do in my team.
[5:52] I mean what is exactly a chatbot? A
[5:53] chatbot could be you know just a um a
[5:57] system that finishes a sentence for
[5:59] example could be called a chatbot but uh
[6:02] I mean they're probably in use but I you
[6:05] know not in we haven't done that in my
[6:06] team much. um to the degree that you're
[6:09] not using them in terms of the like
[6:11] internal products that you're building
[6:13] um do you use sort of coding assistant
[6:16] tools as part of your team and have you
[6:18] sort of seen impacts of that? Um we were
[6:20] on a panel earlier where we talked about
[6:23] uh the impact on the economy like have
[6:24] you seen some of the productivity
[6:25] improvements that people talk about?
[6:27] >> Uh yes. So um in fact I tried it out
[6:31] myself just a little while ago and uh I
[6:34] was blown away how useful and good it
[6:37] can be. That doesn't mean you can trust
[6:40] everything but you know so when I came
[6:43] to to Sigma I had written this larger
[6:45] C++ library for a lot of stuff and uh
[6:49] and you know not um I I wanted to see
[6:53] okay what can an LLM help me to improve
[6:56] on this library and I just you know I I
[6:58] set it up took me an hour or so and then
[7:01] I asked the question what can I improve
[7:03] in this library and I it came back with
[7:05] a whole bunch of uh suggestions pretty
[7:07] quickly.
[7:08] that basically I knew I could do, but
[7:10] also I knew it would take me probably
[7:12] weeks to do it if I would do it all by
[7:14] hand. Um, and
[7:16] >> you didn't just want to tell the chatbot
[7:17] to do it.
[7:18] >> Well, okay. So, wait. Right. So, first
[7:20] it comes up with a list list of
[7:22] suggestions. It tells you which one are
[7:24] the probably the more useful ones. It's
[7:25] things like, okay, you should change
[7:27] this format. You can maybe use uh AVX
[7:30] commands here. You can maybe improve
[7:33] your memory management there. Stuff like
[7:35] that, right? And because it was a lot a
[7:38] fairly large library, um this is
[7:40] difficult stuff to do. And um the next
[7:42] step would then be, okay, how do I do
[7:44] this exactly? And I can basically ask
[7:47] the uh the the the LLM, hey um what
[7:51] would you do first? Please write me a
[7:53] test for this. Um you know, and and do
[7:56] things like that that really helped me
[7:58] to be more productive. Uh so I didn't do
[8:01] it eventually. I know I could do it. I
[8:03] just didn't have the time. But I know
[8:04] that people in my team um and many
[8:07] people around the company
[8:09] almost everybody's using using these
[8:11] tools now because it really speeds up
[8:13] your workflow. Um and and that means you
[8:17] know more efficiency, more simplicity
[8:20] maybe. Uh but you know of course you
[8:22] have to be careful with what it actually
[8:25] produces and you have to check that it
[8:26] does the right thing. Um given that
[8:29] you're in a an industry where sort of uh
[8:32] doing things precisely and doing them
[8:34] well is very important. Um do you have
[8:36] some sort of additional safeguards on
[8:38] LMUS internally in some ways or do you
[8:41] use your own opensource models?
[8:44] So yeah, uh initially when we started
[8:48] using LLMs and that was probably even
[8:50] before chat GPT came out because there
[8:52] were other ones that weren't that famous
[8:54] but uh the biggest problem was to make
[8:57] it secure
[8:59] >> and because of course the information
[9:01] that is in the company can't go out and
[9:04] uh this is something that is uh probably
[9:06] true for for almost every company but
[9:08] specifically for finance because of
[9:10] because of regulations and other things
[9:12] and so that was the big the first big
[9:14] hurdle. Um then the next step is of
[9:17] course you have to have some rules of
[9:20] what you can actually do with with the
[9:22] LLM. So, okay, you can use it for
[9:24] coding, but you maybe you're not allowed
[9:27] to feed the output of the LLM right back
[9:30] in and have it, you know, optimize in a
[9:32] loop because who knows what is going to
[9:35] happen, right? Maybe they, you know, I
[9:37] know we have some rules like that. I I
[9:39] don't know them now exactly but there
[9:40] are a bunch of rules that are important
[9:43] and exactly as you say uh it's very very
[9:46] important to really check what comes out
[9:49] and make make sure you know in 50
[9:52] different ways that you're getting the
[9:54] right answer
[9:56] >> and so given that generative AI but also
[10:00] AI more generally sort of is becoming
[10:02] more and more part of the standard
[10:05] toolkit in some ways of quantitative
[10:07] investing Um, how do you think about
[10:10] sort of the advantage that some of the
[10:12] companies that I think jumped into this
[10:14] field a bit earlier than others like Two
[10:15] Sigma have in the space? Is that
[10:18] advantage eroding over time? Like is
[10:20] this something that becomes where it
[10:21] becomes you have to invest more and more
[10:23] over time in order to sort of stay ahead
[10:25] of the game or is is it just having
[10:28] advantage in terms of people skills that
[10:30] is always complimentary even as new
[10:32] tools come out? Like how do you think
[10:33] about sort of the advantage there that
[10:35] companies like Two Sigma have? Uh so two
[10:38] sigma has a total of roughly,800
[10:41] employees and you know we do a lot of
[10:43] different things real estate and venture
[10:45] fund and you know all kinds of things.
[10:48] >> Um so it's a relatively big and
[10:51] traditional uh place in in the asset
[10:54] management uh world and with that comes
[10:58] 25 years of experience you know and
[11:00] three 400 PhDs um who have worked on
[11:03] these things uh for a long time. So that
[11:07] is a lot of collective experience and um
[11:12] that is I would say not easy to beat
[11:17] with somebody who's just sitting at home
[11:19] let's say as an example right um sure
[11:22] there there are things that you can do
[11:25] but I would I would actually say
[11:26] everybody has the same advantage and
[11:29] everybody will be able to speed some
[11:31] things up. It's um
[11:35] it's not true that
[11:38] suddenly
[11:39] you can do everything that a company
[11:42] like Two Sigma does you can do with an
[11:44] LLM. I mean the the the real problems
[11:47] are much much more complex than than
[11:49] people realize. I mean that's not true
[11:51] just for finance but for almost anything
[11:54] we see today and especially
[11:57] um I I think the internet
[12:00] um gives us the impression that many
[12:02] things are a lot easier than they than
[12:04] they really are once you start looking
[12:06] at them and that's almost true for I
[12:08] mean for so many problems and and and
[12:10] issues.
[12:11] >> Yeah, I think this this also aligns with
[12:13] some of my own research. So I have a I
[12:15] have some research on sort of looking at
[12:16] what companies uh see greater increase
[12:19] in their stock prices when um they have
[12:22] productivity benefits from generative
[12:24] AI. And what think one thing we found is
[12:26] that it's not just having some sort of
[12:28] ability to use these tools is that you
[12:30] need to have some existing
[12:32] infrastructure and in particular
[12:34] proprietary data assets that is very
[12:36] valuable for companies. Um and that's
[12:38] really where a lot of the benefit comes
[12:39] from.
[12:40] >> Yes. So
[12:43] uh right there there's definitely the
[12:46] data aspect. I mean uh in our case uh we
[12:50] have collected and bought data uh for
[12:53] for 25 years really there's so much data
[12:56] that it's uh often
[12:59] you know similar to tech companies
[13:00] there's way too much data to actually
[13:02] use uh in a certain way but also
[13:06] in another problem with the data is that
[13:10] specifically in finance there's often
[13:12] not enough data uh to to really test
[13:16] your hypothesis correctly. I mean, you
[13:18] know, I always give this example. If you
[13:20] would try to predict the closing price
[13:22] of Apple
[13:23] um a day in advance, right? Um you only
[13:27] have 250 data points a year. So, I mean,
[13:31] how good does your model need to be and
[13:34] how long do you need to test it until
[13:36] you actually know that you have
[13:37] something that is good? And uh if you
[13:40] know about statistics uh then you will
[13:43] find out that's probably going to take
[13:44] multiple years depending on how good
[13:46] your model is and uh often we don't have
[13:49] the time to wait for these kinds of
[13:51] things right so um yeah data is
[13:53] definitely a huge issue
[13:56] >> and so you talked about earlier about
[13:59] how um when you first arrived at two
[14:01] sigma you were sort of going around and
[14:04] looking for opportunities to apply some
[14:06] of this toolkit that you were you were
[14:08] bringing Um, can you maybe I don't know
[14:10] if there are some examples that you can
[14:11] talk about about sort of like where
[14:13] things worked out well and you you
[14:16] actually like were able to use AI in a
[14:18] in a way that was surprising maybe to
[14:19] people uh who hadn't thought about it
[14:22] beforehand or or you can also talk about
[14:24] examples where it didn't work where you
[14:25] you you anticipated it might work. Well,
[14:27] I mean I have some examples that worked
[14:29] well and uh those are uh neuronet
[14:33] networks for example, you know, in the
[14:35] beginning when when I came and I talked
[14:38] to some people and they said the general
[14:40] opinion is that neural networks don't
[14:42] work and and you know I go like what
[14:44] does that mean? I mean it's it it just
[14:47] didn't make sense uh on on a deeper
[14:50] level. uh because it depends on of
[14:52] course what you do with them, how you
[14:53] train them, how what you actually uh how
[14:56] you set up your training and testing and
[14:57] and all these things and after now seven
[15:01] years we can say there were certainly
[15:05] a lot of value in using more neural
[15:08] networks in the right way and that is
[15:09] not only in our team that is happening
[15:11] all across the company. So it's I mean
[15:14] there you go. This is this is definitely
[15:16] a good example. Another good example for
[15:18] us was reinforcement learning. Um I'm
[15:21] I'm not going to go into the details but
[15:23] this is also a very interesting
[15:24] technique that um you know that is the
[15:27] technique that that won basically Alph
[15:29] Go and and u
[15:33] for those those of you who are familiar
[15:34] with the with the games right um this
[15:37] was you know it took a long time but it
[15:40] was very interesting for us uh to see
[15:42] that. So now you know a hypothetical
[15:44] example
[15:46] that actually I don't even know if
[15:47] somebody is doing this but people have
[15:49] talked about it um where we use an LLM.
[15:53] So let's say you would want to predict
[15:56] um the interest rate changes that that
[15:59] are happening after a Fed meeting. Um so
[16:03] you could get first you need all the
[16:05] data that um that the fed chair um you
[16:10] know that that the fetch chair speaks.
[16:13] So, and the first step is you ask the
[16:15] LLM or you ask an engineer, please find
[16:17] me the last 20 years of fetch speeches,
[16:19] transcribe them. Um, and you know, today
[16:22] you can very easily do this by yourself
[16:24] with an with an LLM. And then the next
[16:27] step would be, okay, please build me
[16:28] some kind of model. What are the
[16:30] interest rates? Okay, are there
[16:32] connections? You know, if he says often
[16:35] certain words, maybe there's a higher
[16:37] chance that the interest rates will go
[16:38] up. And this kind of these kind of um ex
[16:43] well not experiments but but but
[16:45] workflows you can actually go through um
[16:47] very easily with LLMs and uh and that is
[16:52] what you know I what is a currently for
[16:56] me a hypothetical example because people
[16:58] have talked about it we haven't done
[16:59] this ourselves exactly but um here you
[17:02] see how useful this can be compared to
[17:05] engineers you know when you go like okay
[17:07] it's going to take me a week to find all
[17:09] the data, right? You can do it maybe in
[17:11] one hour.
[17:12] >> Yeah, that's that's a really interesting
[17:14] example. One that um I think is used a
[17:16] lot in finance research outside of funds
[17:18] nowadays, sort of in academic finance
[17:20] research, is a lot of feature extraction
[17:22] from earnings calls, Fed speeches, and
[17:24] so on. That's sort of one of the key use
[17:25] cases that people spend time on. Yeah, I
[17:28] mean there's there's especially for
[17:30] feature extraction, it's it's super
[17:32] useful because
[17:34] uh it is very e much easier to to scan
[17:38] large unstructured data sets that um
[17:42] that are hard to process otherwise,
[17:45] right? Because you don't have the right
[17:46] format. There's all this noise in there.
[17:48] And you know, LM will help you with
[17:51] these uh with these problems a lot
[17:53] better.
[17:54] So given that there's sort of this big
[17:57] boom in terms of demand for AI
[17:59] researchers like we we read in the news
[18:01] about the you know hundred million
[18:02] dollar salaries um has it become harder
[18:06] to hire for this sort of skill set like
[18:08] you were mentioning that that you know
[18:10] your experience in in AI research was
[18:13] important for being able to identify the
[18:15] right use cases. Is it harder to find
[18:16] someone like you uh in the market now?
[18:19] um you know I don't know it's it it has
[18:22] been changing over the last decade or so
[18:25] multiple times away from tech and
[18:28] towards finance and then away from
[18:31] finance and towards tech and right now
[18:33] it seems a little bit it's towards the
[18:35] tech world is somehow more important
[18:39] it's more interesting more more future
[18:42] is there right and um what I usually
[18:46] tell people is to relax a little it you
[18:48] know what has happening right now in
[18:50] tech maybe that will not happen tomorrow
[18:52] or next year
[18:54] uh so I'd say it has always been
[18:58] difficult to find actually the right and
[19:01] good people it's uh because there's so
[19:04] many qualities that that you know we
[19:06] want that uh that not that often really
[19:11] come together and maybe if you want to
[19:13] we can talk a little bit about
[19:15] >> yes what are those qualities
[19:16] >> right about the qualities I mean uh in
[19:19] terms of
[19:21] um hard skills,
[19:25] you know, programming, computer science,
[19:27] it it's usually good if you have a
[19:29] science background in in some kind of
[19:31] sense. Uh maybe math, maybe physics, um
[19:34] it's usually usually STEM, but it
[19:37] doesn't have to be STEM, right? We have
[19:38] people who studied English as well,
[19:40] right? And um but you need to be able to
[19:42] program. So that's a that's a key thing.
[19:45] Um if you are familiar with tools like
[19:48] TensorFlow or PyTorch and and C++ and
[19:52] Java and and Python
[19:55] and you have actual experience with
[19:57] these with this that is um that is a big
[20:00] plus. Um
[20:02] and you know it all comes together
[20:04] because eventually what especially our
[20:08] group is is particularly good at is
[20:11] complex systems. That's how I describe
[20:13] this. Um so we have systems that that
[20:17] run through you know terabytes of data
[20:19] with millions of parameters that takes
[20:21] days to train or weeks to train in some
[20:23] cases and you know anything you change
[20:27] if you make one mistake the whole thing
[20:29] could not work. So you have to have a
[20:31] good feeling for complexity. You have to
[20:34] have it somewhere in your head. You have
[20:36] to and you know and that that knowledge
[20:39] can come really from anywhere. If you
[20:41] probably if you can write a symphony and
[20:43] you have it in your head that is that is
[20:44] maybe similar kind of skill that you
[20:46] need. So that is for I would say the
[20:50] hard skills. Um maybe another one to add
[20:54] is you know I always call it um
[20:56] scientific common sense. You know just
[20:59] we need to we need to be practical. We
[21:01] need to solve a problem and um and you
[21:04] know some people are just better at
[21:06] being practical and and and doing what
[21:10] is necessary to solve exactly the
[21:12] problem and not doing something else. So
[21:14] and then there are soft skills which uh
[21:17] are also extremely important uh for us
[21:20] at least um and those are you have to be
[21:25] able to communicate with your neighbors
[21:27] and your bosses and your you know all
[21:29] the people around you. You have to be
[21:32] collaborative. I mean finance the
[21:34] problem is in finance is is very
[21:36] competitive you know for for companies
[21:38] and it sometimes is also very
[21:39] competitive within companies. But the
[21:41] problem is that the systems are so
[21:43] complex and becoming more and more
[21:46] complex that basically you cannot do
[21:48] everything by yourself anymore. And this
[21:50] is something that has happened in tech
[21:52] in in many other areas as well. You have
[21:54] to have a group of people who work
[21:56] together to get something done. That
[21:58] means you know I don't know setting up
[22:00] unit tests, setting up um you know how
[22:03] to make it the most efficient, discuss
[22:06] in a group what to improve and so on.
[22:08] And um this is sometimes hard to find
[22:12] when you have really excellent people
[22:14] from excellent universities uh that can
[22:17] also be collaborative. Um so it's that
[22:22] is what we're basically looking for and
[22:24] that really hasn't changed much in the
[22:26] last uh I would say well since since I
[22:29] came it hasn't really changed much uh
[22:31] with the exception of that the systems
[22:34] became more complex. So the kind of
[22:37] complex systems that you're describing
[22:39] it sounds like it would make it a lot
[22:41] harder for people to set up for instance
[22:43] a new fund that would try to replicate a
[22:45] similar structure. Does this mean that
[22:47] we're going to see in some sense more
[22:48] concentration more larger funds that can
[22:50] afford sort of the infrastructure back
[22:52] end for all of this and that have been
[22:54] around for a while and got to collect
[22:55] all the data and and build up this
[22:57] internal tooling or or is there some
[23:00] sort of democratizing equalizing force
[23:03] here where where some of these new tools
[23:05] are actually make it make it easier to
[23:06] have startup hunts.
[23:08] >> Yeah, I think we we touched on that
[23:10] earlier. I mean just today I heard a
[23:12] great talk by by you know these two guys
[23:14] here who had set up their own thing and
[23:16] it's uh uh you know and they're probably
[23:19] not,800 people right so it's just just a
[23:22] few people but um so it can definitely
[23:24] be done but uh you know
[23:29] it is still going to be true that there
[23:31] are there's 25 years of experience there
[23:34] there are you know terabytes of data
[23:36] collected over decades
[23:39] Um there are 400 people who have been
[23:44] doing this for a long time. It's very
[23:47] hard to compete with that. Uh you know
[23:50] it's it's not I wouldn't say it's
[23:52] impossible. You have other advantages as
[23:54] a as a smaller place. Maybe you can move
[23:55] a little faster in in certain ways but
[23:58] it's going to be hard. Uh that is and it
[24:01] you know it's it's similar for I don't
[24:04] know a car company or any other big
[24:06] company. you know, you can't just go and
[24:08] and replace Toyota tomorrow. This is
[24:10] this is going to be difficult. Um,
[24:12] >> that's a I think that's a really good
[24:14] point and and I see this a lot. So, I
[24:16] liked your point about scientific common
[24:17] sense or also what you said earlier
[24:19] about sort of just there's certain
[24:21] scientific methods people need to
[24:23] follow, but also there's broader
[24:24] organizational complex systems
[24:26] management that's involved in making
[24:27] this all run. And it's not just all at a
[24:30] at an individual researcher level. It's
[24:32] it's about making the whole organization
[24:34] function together. That's actually kind
[24:35] of hard. Um because I think this this is
[24:37] true outside of uh outside of finance
[24:40] like in all organizations like the
[24:43] implementing AI becomes much more of a
[24:44] broader organizational issue. It's not
[24:46] just everyone individually uses an AI
[24:48] tool or or implements an AI model in
[24:50] some ways. Um does this sort of this
[24:53] difference in organizational complexity
[24:56] between uh what you're what you're doing
[24:59] and maybe what an indivi individual
[25:01] researcher is doing in a lab. Does that
[25:03] make it harder for aca more academic
[25:06] research to be relevant? Like what does
[25:09] the pass through look like there in in
[25:10] terms of people coming up with new ideas
[25:12] in an academic setting? Like do you read
[25:14] the papers that are coming out? Um do
[25:17] you is it useful?
[25:18] >> Um yeah so we hire a lot of people from
[25:22] uh from PhD programs right and out of
[25:25] from academic programs as as well. and
[25:29] we are
[25:31] trying to be very scientific about
[25:34] everything. That's that's really I think
[25:36] the power of at least our company right
[25:38] that we um
[25:41] yeah I was I was trying to think where
[25:43] to go with this but uh what was your
[25:47] initial question was was just to to what
[25:50] degree um the fact that uh research in
[25:53] the lab is oftentimes done by like two
[25:56] people working with a toy data set and
[25:58] showing that their model can do one
[26:00] particular thing well in one particular
[26:02] setting. To what degree does that limit
[26:04] the ability to make use of this research
[26:06] in your setting where you you're working
[26:07] with much more complex systems and you
[26:10] have much more of a
[26:12] many more implementation challenges than
[26:14] individual researcher has.
[26:15] >> Yeah. I mean maybe one thing I can
[26:17] mention is that uh what especially
[26:20] people who who have gone through you
[26:23] know maybe high school and then
[26:24] undergrad and then grad and PhD um in
[26:28] most cases there has not been a lot of
[26:31] failure.
[26:33] Um so people you know they do a class
[26:35] and within 30 days or six weeks you're
[26:37] done with the class you get an A you
[26:39] move on right that's often what people
[26:41] expect and especially the very good
[26:43] people that's what their life has been.
[26:45] So when they start at at two sigma what
[26:48] often happens is reality is much more
[26:51] difficult than anything they have seen
[26:53] before and uh there's there could be a
[26:57] lot of failure and I usually try to
[27:00] encourage people you know I remember
[27:01] when I when I came in the beginning
[27:03] people were very afraid of failing and I
[27:07] told them you know if you never fail
[27:09] this you you're not taking enough risk
[27:11] you need to you know you're not doing
[27:13] the right
[27:14] Uh and I think that helped a little bit.
[27:17] Um you know failure is fine. I mean you
[27:20] shouldn't fail all the time in a row,
[27:21] right? Not multiple years in a row. But
[27:23] once in a while this is this is what
[27:25] research exactly is. Um what is what I
[27:28] also found really interesting is that
[27:30] you need to have you need to have this
[27:33] research mindset that
[27:36] uh it is not clear in the beginning
[27:39] where where this re research is going to
[27:42] take you. you have a slight idea. Oh, I
[27:45] want to use let's say as an example
[27:46] reinforcement learning somehow. Okay,
[27:49] how are we getting there? And I remember
[27:53] in the beginning we had issues um like
[27:56] okay how do we set up a test set for
[28:00] reinforcement learning because you know
[28:02] you this is not how the original these
[28:05] games actually worked. it was set up in
[28:07] a very different way and um it took us a
[28:11] while to get to that point and this is
[28:13] something that that people you know that
[28:16] this uncertainty that you need to have
[28:19] uh to make the right steps in that
[28:21] research is something that is um that we
[28:24] are trying to you know really nurture as
[28:26] much as possible and probably in in my
[28:29] team maybe the most uh and that is I
[28:33] think absolutely necessary. So
[28:38] [laughter]
[28:38] >> all right and okay so and so you're
[28:41] you're nurturing this this research
[28:43] mindset is that are there ever any sort
[28:46] of culture clashes with other parts of
[28:49] the fund in some ways like is that is
[28:51] that the same mindset that also makes
[28:53] someone successful at being a trader or
[28:56] a like portfolio manager in the setting
[28:58] or is there or
[29:01] >> of the frictions that sometimes arise
[29:02] from running sort of a research minded
[29:04] organization inside a hedge fund.
[29:05] >> Well, I mean uh so first of all, I don't
[29:08] know exactly what is happening in in
[29:10] many other teams, right? uh but I would
[29:14] say it's probably my guess is it's
[29:16] similar like in every other company
[29:20] uh that there are people who think a
[29:22] little bit differently and um sometimes
[29:26] obviously we run into these kinds of
[29:28] things but we um you know my at least my
[29:33] approach has been so far is you know
[29:36] everybody has their own way of doing
[29:37] things and that's fine they have a lot
[29:39] of experience with with what they're
[29:41] thing um I I will accept it um and then
[29:47] you know but we may do it a little bit
[29:50] differently and that usually works out
[29:52] works out well. I'm not we are not
[29:54] trying to tell other people exactly what
[29:56] they are supposed to do or vice versa.
[29:58] >> Um and I think maybe that's a general
[30:00] concept that that needs needs to apply
[30:03] in in I mean in generally in companies
[30:06] or or in organizations or whatever you
[30:08] want to call it. Okay, I think we're
[30:10] almost out of time. Um, but before we
[30:12] close out, I just want to ask for a
[30:15] somewhat personal uh assessment of what
[30:19] are you sort of excited about in terms
[30:21] of new technologies coming out? So, we
[30:22] had GPT5 come out last week like uh what
[30:26] what aspects of either generative AI or
[30:29] AI do you think is going to be really
[30:31] exciting over the next couple of years,
[30:32] either professionally or personally, and
[30:34] and where you you'd really like to see
[30:36] go?
[30:37] >> Uh, okay. That's a good question. I
[30:40] mean, there's so many use cases that I
[30:42] would like LLMs really to be used more.
[30:45] So, I mean, number one is maybe I want
[30:48] people to really use this stuff. A lot
[30:50] of people talk about it and see it in
[30:52] the news and, you know, but what is
[30:54] really important is that people
[30:56] internalize using it and and seeing what
[30:59] you can do with it. And I'm not sure
[31:02] enough people are actually doing that um
[31:04] even within the company. And as an
[31:06] example of what I would like to see is
[31:09] um for example customer service as an
[31:12] example. I mean the it's it would be
[31:14] great to have a system where you you
[31:17] call and you don't have to wait and you
[31:19] talk to a system that has the knowledge
[31:21] of all customer service agents right
[31:23] there without you know being handed over
[31:27] six times to to another person and and
[31:29] then not maybe getting the right answer
[31:31] in some cases. So um another case would
[31:36] be um there are so many you know non
[31:41] unintuitive web pages out there. So for
[31:43] UI I think it could really help to
[31:46] optimize UI and processes in general a
[31:49] lot lot more and I'm sure that is going
[31:51] to happen. Um and you know I personally
[31:54] actually I'm I think pretty rosy about
[31:58] all this stuff compared to you know some
[32:02] of the the people you know some some of
[32:05] the things that we see in the internet
[32:06] or even heard here here today. So I
[32:10] think there's a lot of positive um
[32:13] energy and a lot of positive for for us
[32:16] will come out of this.
[32:18] >> Yeah I I would have said something
[32:20] similar. I think it's like the the
[32:22] personalization and customization
[32:24] element and adding memory in a useful
[32:26] way into these systems such that they
[32:28] already have the context for what
[32:31] they're talking to you about without you
[32:32] having to constantly remind them and
[32:34] having to input lots of context. I think
[32:36] it's going to make a lot of these
[32:37] systems much more useful both in our
[32:39] personal lives and our interactions with
[32:41] companies. Um, okay.
[32:42] >> Right.
[32:43] >> Thank you so much.
[32:44] >> Okay. Thanks a lot.
[32:47] Okay. [applause]
