Franklin Templeton Decouples AI Core to Enable Zero-Downtime Model Swaps

Pit Stops, Not Rebuilds: Modular AI Infrastructure Under Performance Pressure & Regulatory Scrutiny — Ai4

Ai4

Treating foundation models and orchestration as disposable components governed by rigid compliance contracts allows asset managers to upgrade AI capabilities continuously without risking regulatory breaches or engineering teardowns.

AI-assisted summary of Pit Stops, Not Rebuilds: Modular AI Infrastructure Under Performance Pressure & Regulatory Scrutiny — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published by the Ai4 channel on YouTube on August 14, 2026, this presentation features Max Gokhman, Head of AI and Digital Asset Solutions at Franklin Templeton. The session description details a practical playbook for building modular, agentic AI harnesses within regulated enterprises, focusing on separating durable governance and compliance layers from rapidly evolving components like models and orchestration frameworks.

## TL;DR Using the evolution of motorsport pit stops as an architectural metaphor, Max Gokhman outlines how enterprises can achieve continuous, zero-downtime AI updates without breaching strict regulatory boundaries. He proposes decoupling the durable chassis—comprising core intellectual property, governance, audit trails, and standardized evaluations—from hot-swappable components like foundation models and data feeds via explicit coupling contracts. Gokhman contrasts a real-world hardcoded copilot project that suffered costly refactoring delays against a modular agentic analyst deployment that seamlessly incorporated multi-model routing in days. He concludes by explaining how agentic workflows resolve historical friction between fundamental portfolio managers and quantitative tools while establishing an enduring competitive moat.

## Atmosphere & style The session has an engaging, crisp, and high-velocity tone, propelled by Gokhman's dual passion for track racing and enterprise systems engineering. His speaking style is polished, confident, and direct, effortlessly translating complex architectural concepts like standardized harnesses and dynamic human-in-the-loop oversight through vivid analogies like Formula 1's Drag Reduction System. The atmosphere feels like an insider technical debrief from the racing paddock, effectively balancing the competitive urgency of rapid AI iteration with the disciplined risk management required in financial services.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=4s) In regulated industries facing continuous 24/7/365 operations, the target for enterprise AI maintenance and model upgrades is '0.0 seconds' of downtime, surpassing even Formula 1's 1.8-second pit stops. - [00:04](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=4s) The dilemma between freezing a tech stack for compliance and constantly chasing new frameworks is an architectural problem, not a governance problem. - [02:56](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=176s) A modular AI architecture decouples the durable chassis (governance, risk limits, auditability, client objectives) from consumable, hot-swappable components (foundation models, data streams, orchestration frameworks). - [02:56](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=176s) Couplings between the chassis and components must be treated as formal software contracts defining inputs, outputs, latency, and compliance requirements. - [05:04](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=304s) Regulatory constraints must be directly hardcoded into software logic to prevent autonomous agents from discovering and exploiting compliance loopholes. - [05:04](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=304s) Standardizing evaluation benchmarks across engineering, legal, cybersecurity, and compliance allows new models or frameworks to be deployed without renegotiating compliance approvals. - [05:04](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=304s) Autonomous agents provide telemetry and point-in-time flight recorder replay, giving regulators completely transparent audit trails of observed data and reasoning steps. - [05:04](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=304s) Human-in-the-loop oversight acts like aerodynamic downforce: it should be minimized on simple straightaways to maximize speed and concentrated selectively on complex, high-risk operational turns. - [08:50](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=530s) Tightly 'welding' dependencies into early copilot systems caused a 3-month teardown at Franklin Templeton, whereas modular harnesses allowed intelligent model routing to be deployed in days. - [08:50](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=530s) Enterprises should adopt a 'partner for a season, not for life' philosophy with external AI vendors, re-evaluating technology partnerships annually as capabilities evolve. - [11:43](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=703s) While switching between comparable LLMs is straightforward, transitioning to fundamentally different architectures like world models requires engineering entirely different harnesses. - [11:43](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=703s) Autonomous agents bridge the historic 'quantamental' divide by enabling fundamental analysts to leverage complex quantitative tools through natural language without requiring deep mathematical backgrounds.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=4s) From 67 Seconds to 0.0: The Motorsport Metaphor and Enterprise AI

Max Gokhman introduces himself with a dual identity: on weekends, he races cars at Laguna Seca, while during the week, he leads the build-out for Franklin Templeton's artificial intelligence and digital asset solutions. His background spans running one of the earliest pure AI hedge funds and serving as deputy chief investment officer for Franklin Templeton’s multi-asset team. Drawing on motorsport, Gokhman contrasts the 67-second pit stop of a 1950 Indianapolis 500 racecar with McLaren's record-setting 1.8-second Formula 1 pit stop in 2023. The dramatic speed improvement was not achieved because racecars became simpler—modern race machines are orders of magnitude more complex—but because the pit stop evolved into an intricately choreographed routine where every critical part is designed to be modular.

In financial services, where the rise of asset tokenization and blockchain rails pushes markets toward 24/7/365 continuous operation, even a 1.8-second stop is unacceptable. The operational goal is "0.0": performing maintenance, model upgrades, and component swaps without ever taking the system offline. Gokhman notes that many enterprise leaders fall into a false dichotomy: they either freeze their tech stack under the guise of compliance and safety, or they frantically chase every emerging framework. Freezing a stack guarantees obsolescence—locking into a foundation model from 18 months prior leaves a firm "completely at the back of the grid." Conversely, constantly jumping between new frameworks is like "driving on wet cement because you've never let it dry." Gokhman emphasizes that resolving this dilemma is not a governance problem, but an architectural one.

- **1950 Indy Car Pit Stop**: 67 seconds. - **2023 McLaren F1 Pit Stop**: 1.8 seconds. - **Enterprise Target**: 0.0 seconds (zero-downtime hot swapping during live operations).

---

[02:56](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=176s) The Core Architecture: Durable Chassis vs. Modular Components

To build an AI architecture capable of zero-downtime pit stops, Gokhman separates the system into two distinct layers: the durable "chassis" and the modular "components," joined together by standardized "couplings." The chassis represents the firm's core intellectual property, compliance standards, and business identity. It houses non-negotiable elements such as firm-wide governance policies, risk limits, client investment objectives, point-in-time audit logging, human-in-the-loop triggers, and standard evaluation suites (evals).

On the other side of the boundary are the components: foundational large language models, live data feeds, and orchestration frameworks. These are consumable elements that must be upgraded at the cutting-edge pace of the market rather than the slower pace of internal enterprise release cycles. The bridge between the chassis and components is the coupling—a formal software contract defining inputs, outputs, and governance constraints. When couplings and evals remain standardized, swapping one model or orchestration layer for another becomes seamless.

```mermaid flowchart LR accTitle: Modular AI Harness Architecture subgraph Chassis["Durable Chassis (Core IP)"] G["Governance & Risk Limits"] A["Point-in-Time Audit & Replay"] H["Human Oversight Triggers"] E["Standardized Eval Harness"] end

subgraph Couplings["Couplings (Contracts)"] C1["Input/Output Specifications"] C2["Governance Encodings"] end

subgraph Components["Modular Components (Hot-Swappable)"] M["Foundation Models"] D["Market & Alternative Data Feeds"] O["Orchestration Frameworks"] end

Chassis <--> Couplings <--> Components ```

- **Chassis Assets**: Proprietary investment IP, risk parameters, regulatory guardrails, point-in-time audit logs, human intervention rules, standardized evals. - **Coupling Contracts**: Explicit technical specifications governing inputs, outputs, latency, and compliance requirements. - **Modular Components**: Foundation models, external vendor data streams, agent orchestration engines.

---

[05:04](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=304s) Staying Fast: Standardized Evals, Flight Recorder Replays, and Human Downforce

Gokhman details the operational mechanics that allow an enterprise to move quickly without breaching regulatory boundaries. First, constraints must be encoded directly into code rather than left to ambiguous guidelines. In motorsport, teams routinely comb through regulations looking for loopholes to exploit, but enterprises cannot afford to have autonomous agents discovering and exploiting loopholes in risk or compliance policies. Second, evaluations must be standardized across engineering, legal, cybersecurity, and compliance teams. By securing upfront stakeholder alignment on a single evaluation suite, any candidate component that passes the benchmark can be plugged directly into production without having to re-litigate compliance approvals from scratch.

Third, systems require complete telemetry and replay capabilities, mirroring the post-session debriefs where a racing driver and engineer analyze brake pressure, steering angle, and tire contact patches on the tarmac. In asset management, regulators require precise visibility into decision-making. If an investigator asks a human trader whether they acted on material non-public information (MNPI), the human reflexively denies it. An autonomous agent with a point-in-time audit trail, however, "technically isn't worried about going to jail or getting fined" and can supply a fully transparent record of exactly what data it observed, when it observed it, and how it reasoned.

Finally, Gokhman uses aerodynamics to explain the role of human-in-the-loop oversight. In race engineering, aerodynamic downforce creates drag that slows a car on straightaways, but provides the essential grip required to keep it from flying off the track in high-speed corners. Formula 1 introduced the Drag Reduction System (DRS) to shed downforce on straights while restoring it in corners. Similarly, human oversight should not bottleneck simple, high-throughput computational tasks; instead, human judgment must be concentrated selectively on nuanced, complex, high-risk, or ethically sensitive operational turns.

---

[08:50](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=530s) Tales from the Paddock: The Welded Co-Pilot vs. The Modular Agentic Analyst

Gokhman contrasts two real-world enterprise deployments at Franklin Templeton to illustrate the difference between hardcoded systems and modular harnesses. The first project was an early client-facing copilot. The team tightly integrated and "welded" all dependencies—including the third-party orchestration layer—directly into the application. When an external vendor released a markedly superior orchestration framework, the team was forced to completely dismantle the software chassis, rebuild internal connectors, and re-engineer the pipeline. That refactoring effort consumed three months and brought the team dangerously close to missing their launch deadline.

Having learned from that failure, the team designed their subsequent project—an autonomous agentic research analyst—with full modularity from day one. Every sub-component was treated as a hot-swappable element connected to the core investment engine via explicit contracts. When a frontier AI lab introduced a top-tier model with steep API pricing, the team quickly deployed an intelligent router: complex, high-value tasks were routed to the expensive frontier model, while routine background tasks were directed to a cheaper commodity model. Because governance contracts and eval harnesses were already established, the change was executed in days rather than months.

- **Mistake ("Welding")**: Hardcoding dependencies and orchestration into the copilot chassis resulted in a 3-month teardown and near-missed deadline. - **Success ("Pit Stop")**: Building a modular agentic analyst enabled the team to insert a multi-model smart router in days to optimize frontier model costs. - **Core Takeaways**: 1. Separate core governance from modular components. 2. Treat couplings as rigid contracts—never weld components to the chassis. 3. "Partner for a season, not for life," re-evaluating external technology partners annually just as racing teams do.

---

[11:43](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=703s) Q&A: Interoperability, Harness Types, and Fundamental Active Management

During the audience Q&A, Gokhman clarifies how Franklin Templeton approaches vendor interoperability, model architectures, and the long-term evolution of asset management:

```mermaid flowchart TD accTitle: Human-Agent Active Management Loop A["Human Portfolio Managers & Analysts"] -->|"Market intelligence, qualitative insights, hypothesis generation"| B["Agentic Systems & Quant Tools"] B -->|"Deep contextual computation, screening, automated synthesis"| A A -->|"Reinforcement & institutional investment logic"| C["Proprietary Knowledge Chassis"] C -->|"Governance & evaluation boundaries"| B ```

An audience member asks about the overhead of maintaining cross-vendor interoperability. Gokhman responds that standardized evaluation suites make it easy to benchmark competing vendors (e.g., orchestration framework A versus framework B) on identical criteria. This abstraction also allows the firm to deploy different component configurations across different use cases, much like using a single engine architecture across both a rally car and a GT track car while swapping the suspension and tires.

Addressing a question about model swappability and compatibility, Gokhman notes that swapping models is like matching a tire to a wheel rim: while moving between comparable LLMs (such as switching between GPT-4o and Claude 3.5 Sonnet) is straightforward, moving from an LLM to an embodied or world model requires building an entirely different harness. A world model harness prioritizes visual, spatial, and physics data while consuming unstructured text at a secondary layer, whereas an LLM harness does the opposite.

Asked about the future of investment management, Gokhman reflects on the post-2008 Global Financial Crisis (GFC) trend of "quantamental" investing, where firms attempted to pair quantitative analysts with fundamental portfolio managers. That initiative largely struggled across the industry due to deep cultural and methodological divides. Autonomous agents resolve this friction by acting as an intuitive interface: fundamental analysts can interact with agents in plain language without needing to understand the underlying quantitative mechanics, similar to how a driver relies on anti-lock brakes (ABS) without knowing the low-level mechanical engineering. Gokhman concludes that Franklin Templeton's competitive moat rests on combining human intelligence gathering with AI processing power to continuously expand the firm's proprietary investment IP across equities, fixed income, and private credit.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=4s) completely at the back of the grid. — Max Gokhman

> [00:04](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=4s) driving on wet cement because you've never let it dry. — Max Gokhman

> [05:04](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=304s) technically isn't worried about going to jail or getting fined — Max Gokhman

> [08:50](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=530s) Partner for a season, not for life, — Max Gokhman

## Mentioned Max Gokhman, Franklin Templeton, Laguna Seca, Indianapolis 500, McLaren, Formula 1, GPT-4o, Claude 3.5 Sonnet, Drag Reduction System (DRS), Ai4

## Video description Max Gokhman - Head of Artificial Intelligence and Digital Asset Solutions, Franklin Templeton

In motorsport, no team rebuilds the car mid-race — they engineer it so key parts can be swapped or adjusted in seconds without compromising safety or speed. Enterprises in regulated industries like investment management face the same challenge with AI: model capabilities are evolving monthly and peers keep competition fierce, but decision quality and compliance cannot be compromised. This session presents a practical playbook for building agentic harnesses that evolve at the speed of technological change — separating the durable layers like governance, client constraints, and auditability from the modular ones such as models, data streams, and orchestration frameworks. Attendees will leave with concrete frameworks, vendor-selection criteria, and lessons from building agentic systems inside one of the world's largest asset managers. If your AI roadmap changes every quarter but your scrutineers don't, this talk will help tune your chassis to perform within track limits.

[Source](https://www.youtube.com/watch?v=4V5ikAAMdHg)

中文简报

富兰克林邓普顿解耦AI核心，实现零停机模型替换 将基础模型与编排系统视作受严格合规契约约束的一次性组件，使资产管理机构能够在不面临合规违规风险或工程推倒重构的前提下，持续升级AI能力。 Max Gokhman以赛车进站维修的演变为架构隐喻，阐述了企业如何在不逾越严格监管边界的前提下实现持续、零停机的AI升级。他提出通过明确的耦合契约，将由核心知识产权、治理、审计追踪和标准化评估构成的持久“底盘”，与基础模型及数据源等支持热插拔的组件进行解耦。Gokhman对比了一个因硬编码而承受高昂重构延误代价的真实Copilot项目，以及一个在数天内无缝接入多模型路由的模块化智能体分析师部署案例。他最后总结指出，智能体工作流既化解了基本面投资组合经理与量化工具之间长久以来的摩擦，同时也建立起了持久的竞争壁垒。 富兰克林邓普顿（Franklin Templeton）的 Max Gokhman 指出，面对 24/7 全天候运转的金融市场，企业级 AI 维护与模型迭代的目标是“0.0 秒”停机；在技术栈固化合规与盲目追赶新框架之间的矛盾本质上是架构问题，而非治理问题。 (0:04) 其核心方案是将耐用的“底盘”（核心治理、风控限额、投资目标与审计日志）与可热插拔的“组件”（基础模型、数据流、编排框架）解耦，并通过定义严格输入输出与合规约束的技术契约相连接。 (2:56) 合规约束必须直接以代码形式硬编码以防 Agent 钻空子，跨部门标准化的评测基准与“飞行记录仪”点位回放可大幅简化审批流程；人机协同应如赛车下压力般运作，在平直跑道上降阻提速，仅在高风险拐点集中介入。 (5:04) 以实际经验为例，早期硬编码第三方编排层的 Copilot 因重构耗时 3 个月险些违期，而模块化的 Agent 分析师仅用数天便上线多模型智能路由优化了成本，验证了外部技术伙伴应“按赛季合作而非绑定终身”的策略。 (8:50)

Reading summary

Max Gokhman of Franklin Templeton argues that enterprise AI requires an architectural shift from monolithic codebases to modular 'chassis and component' harnesses, enabling zero-downtime upgrades and cost optimization without compromising regulatory compliance. Durable Chassis Versus Hot-Swappable Components (2:56) Enterprise technology leaders often face a false choice: freeze their software stack for compliance and risk rapid obsolescence, or constantly chase new frameworks and operate on unstable foundations. Drawing on motorsport evolution—from 67-second pit stops in 1950 to 1.8-second Formula 1 stops in 2023—Gokhman asserts that 24/7/365 tokenized financial rails require an operational target of '0.0' downtime, hot-swapping models during live production. Achieving this requires separating the system into a durable core chassis and modular components connected via standardized couplings. The durable chassis houses the enterprise's proprietary intellectual property, firm-wide risk limits, regulatory guardrails, point-in-time audit logs, human intervention triggers, and evaluation suites. Foundational models, orchestration layers, and live vendor data feeds represent modular, consumable components. By enforcing rigid input, output, and governance contracts at the coupling layer, firms can replace frontier components at market speed without disrupting core business rules. Telemetry, Standardized Evals, and Dynamic Human Downforce (5:04) Operating at speed under regulatory scrutiny requires converting governance guidelines into hardcoded software constraints, preventing autonomous agents from discovering unintended policy loopholes. Gokhman advocates aligning engineering, legal, cybersecurity, and compliance teams around a single, upfront evaluation suite. Any candidate component that passes this standardized benchmark can deploy into production without restarting interdepartmental approval cycles from scratch. Accountability depends on comprehensive telemetry and replay capabilities. While human traders may reflexively deny acting on material non-public information (MNPI), autonomous agents provide immutable point-in-time audit trails recording exact data inputs and intermediate reasoning steps. Human oversight is structured analogously to aerodynamic downforce: rather than introducing drag across simple, high-throughput workflows, human intervention functions like Formula 1's Drag Reduction System, applying concentrated judgment strictly to complex, high-risk, or ethically sensitive edge cases. Practical Implementation: Welded Copilot Versus Modular Routing (8:50) Franklin Templeton's internal engineering illustrates the performance difference between hardcoded and modular systems. An early client-facing copilot tightly integrated external orchestration frameworks directly into the application chassis. When an external vendor released an upgraded framework, integrating it required a three-month code dismantling and reconstruction effort that nearly caused the team to miss its deployment deadline. In contrast, the firm designed its subsequent autonomous research analyst with modular couplings from inception. When an AI vendor launched a premier foundation model with high API pricing, engineers integrated an intelligent multi-model router within days. Routine background computation was directed to low-cost commodity models while complex queries routed to the expensive frontier model. Gokhman recommends treating external technology relationships as seasonal partnerships re-evaluated annually rather than permanent commitments. Architectural Boundaries and Bridging the Quantamental Divide (11:43) While swapping between similar large language models is relatively direct, Gokhman notes that cross-paradigm transitions require entirely separate harness architectures. Migrating from an LLM to a world or embodied model demands rebuilding the underlying software infrastructure, because world models prioritize visual, spatial, and physics inputs while treating unstructured text as a secondary data layer. Gokhman also reports that autonomous agents resolve long-standing friction in 'quantamental' investing, which struggled across the asset management industry after 2008 due to cultural and methodological rifts between quantitative researchers and fundamental portfolio managers. Natural-language agent interfaces allow fundamental analysts to execute advanced quantitative screening and contextual synthesis without needing low-level mathematical engineering knowledge. Gokhman concludes that an asset manager's durable competitive moat does not stem from external foundation models or vendor tooling, but from expanding proprietary investment logic and human intelligence within the chassis across asset classes.

精炼阅读摘要

富兰克林邓普顿（Franklin Templeton）人工智能与数字资产解决方案负责人 Max Gokhman 指出，面对技术迭代与监管审查，企业无需全盘推倒重建系统；通过将核心治理底盘与外部模型组件解耦，构建类似赛车维修站的模块化架构，可在零停机状态下实现即插即用的敏捷升级。 架构解耦：持久底盘与可替换组件 (2:56) Gokhman 引用赛车运动的演进指出，从1950年印第安纳波利斯500赛车的67秒进站，到2023年迈凯伦车队1.8秒的世界纪录，速度的突破源自极度精细的模块化编排而非机械系统的简化。在资产代币化与区块链推动金融市场迈向全年无休（24/7/365）运行的背景下，企业运营目标必须达到“0.0秒”，即在生产系统持续运行的同时完成模型与组件热插拔。他强调，因合规担忧而冻结技术栈会导致系统迅速过时，而盲目跟风频繁更换框架则如同在未干的水泥地上驾车；化解这一矛盾的关键在于架构分层而非行政管控。 该模块化架构由三部分构成：第一是“持久底盘”（Chassis），沉淀企业核心知识产权、合规政策、风险限额、时点回溯审计日志、人工介入规则及标准化评估套件；第二是作为消耗品的“模块化组件”（Components），包括前沿大语言模型、外部数据流和编排框架，须紧跟市场节奏快速更替；第三是连接二者的“联轴器”（Couplings）软件契约，严格限定输入输出规范与治理边界。只要契约保持稳定，替换底层组件便不会扰动核心业务底座。 合规代码化：时点审计与DRS协同治理 (5:04) 为在合规红线内维持敏捷，Gokhman 提出必须将风控与合规要求直接编写为代码约束，防范自主智能体像赛车手钻研赛事规程那样搜寻合规漏洞。企业应在工程、法务、网络安全与合规团队间确立统一的标准化评估基准，候选组件只要通过测试即可直接部署，无需逐次重复跨部门审批流程。此外，系统须具备类似赛车数据遥测的“飞行记录仪”回放能力。面对内幕信息（MNPI）合规质询，人类交易员往往下意识否认，而具备精确时点审计追踪的智能体在技术上既不担心罚款也不畏惧诉讼，能够客观还原决策瞬间观察到的完整数据切片与推理链路。 针对人机协同（Human-in-the-loop），Gokhman 借用赛车空气动力学的下压力与减阻系统（DRS）进行说明：气动下压力虽在直道上产生拖拽阻力，却能在高危弯道提供防止车辆失控的抓地力。因此，企业的人工监督不应演变为常规高通量计算的流程瓶颈，而应通过机制设计，选择性聚焦于高风险、高复杂度及伦理敏感的关键业务决策拐点。 工程实证：焊接式系统与即插即用智能体 (8:50) Gokhman 通过富兰克林邓普顿的两次实操对比了紧耦合与模块化架构的实际效能。在早期的客户端副驾驶（Copilot）项目中，开发团队将第三方编排工具等依赖项直接“焊接”嵌入系统。当市场上出现更优异的编排框架时，团队被迫将整套架构拆解重构，耗费整整3个月时间重写内部连接器与数据管线，险些导致项目逾期上线。 基于前次教训，团队在后续构建自主智能体研究分析师时全面采用模块化方案，将每个子模块设计为通过明确契约接入投资底盘的热插拔单元。当前沿人工智能实验室发布高定价的顶级模型时，团队仅用数天时间便部署了智能多模型路由器，将复杂的高价值分析分流给前沿模型，常规任务则交由低成本商品模型处理。Gokhman 据此建议企业将外部技术供应商视为“单赛季合作伙伴”，如同赛车队每年重新评估技术供应商一样保持审慎评估，切忌终身绑定。 评估基准抽象与量化基本面的融合 (11:43) 在跨供应商互操作性方面，统一的标准化评估套件使得企业能够以同一套基准横向测评不同供应商的组件表现，实现类似单套发动机架构灵活适配拉力赛车与场地赛车的跨场景复用。但在跨模型形态迁移上存在物理边界：在大语言模型之间切换如同更换轮毂轮胎般直接，但若转向具身模型或世界模型，则必须重新构建系统线束，因为世界模型以视觉、空间和物理规律为核心输入层，对非结构化文本的依赖处于次席，这与语言模型的数据优先级截然相反。 谈及资管行业演进，Gokhman 回顾了2008年金融危机后兴起的“量化基本面”（Quantamental）投资潮。该模式此前因量化分析师与基本面研究员之间深层的文化与方法论鸿沟而推进艰难。他认为自主智能体正在充当自然语言抽象接口，基本面分析师无需掌握量化工程底层原理即可借助智能体调动算力，如同驾驶员依赖防抱死制动系统（ABS）而无需通晓其机电细节，从而有效消除了两大投资阵营的协作摩擦。 Gokhman 认为资管机构的护城河不在于特定模型组件，而在于结合人类情报与智能体算力，持续沉淀股票、固收和私募信贷领域的专有投资知识底盘；然而该架构的长期稳健性完全取决于企业能否在一开始便建立严密完备的标准化评估合约，当面对跨模态等非对称技术范式演进时，模块化线束本身仍存在重新工程化的适应成本。

Detailed notes — zh

## Background 本内容由 Ai4 频道于 2026 年 8 月 14 日发布在 YouTube 平台上。页面描述指出，富兰克林邓普顿（Franklin Templeton）人工智能与数字资产解决方案负责人 Max Gokhman 借鉴赛车中快速进站调优的理念，分享了受监管行业在面对技术迭代与合规双重压力时，如何通过解耦持久层与模块化组件构建企业级 Agent 基础设施的实战框架与落地经验。 ## TL;DR 演讲通过赛车进站比喻，深入探讨了金融机构在面对 AI 技术飞速更迭与严格监管审查时应采取的系统架构策略。演讲者指出，企业应将基础设施解耦为固化的核心持久底盘（涵盖治理、风险控制、时点审计与评估集）与可热插拔的模块化组件（如 LLM、数据流与编排框架），通过标准化的耦合契约消除对特定供应商的强依赖。随后结合富兰克林邓普顿的真实案例，展示了从硬编码焊接架构耗时数月改造，转向模块化路由器数天平稳上线的跃升，并进一步阐释了利用智能体作为自然语言交互层化解量化与基本面分析师长期隔阂、构建专属投资护城河的人机协同范式。 ## Atmosphere & style 整场演讲节奏紧凑且富有激情，演讲者凭借周末赛车手与资管科技高管的双重身份，自如地在硬核技术工程与赛道极限博弈的比喻之间穿梭。现场叙事生动而务实，既有反思早期将系统硬编码“焊死”底盘导致交付延误的自嘲与坦率，又有谈及将智能体作为 F1 赛车“减阻系统（DRS）”及 ABS 刹车隐喻时的自信与透彻。整场分享充满工程极客与金融实战派兼具的冷静务实感，令听众仿佛置身于飞驰赛车的维修站，既感受到技术狂飙的推背感，又体会到底层架构稳如磐石的控制力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=4s) 面对全天候交易与监管压力，资管企业落地 AI 的终极目标是实现“0.0 秒进站”，即在业务高速飞驰中完成架构升级而不停机。 - [01:41](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=101s) 必须将系统解耦为持久的“底盘”（治理、审计、业务逻辑与评测）与可热插拔的“组件”（模型、数据流、编排层），避免技术栈锁死或频繁推倒重来。 - [01:41](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=101s) 通过“耦合接口契约”严格规范输入、输出与治理指标，以机器可执行的标准彻底消除对特定大模型与供应商的强依赖。 - [05:04](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=304s) 将合规、法务与安全红线一次性编码固化到底层逻辑中，杜绝智能体钻空子，同时为工程团队赋予合规框架内快速迭代的自由。 - [05:04](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=304s) 建立跨部门统一的标准化评估套件，外部新模型或新工具只要跑通固定评测即可快速接入，无需反复进行准入协商。 - [05:04](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=304s) 引入高精度遥测与时点数据回放机制，完全记录 AI 在决策时刻的信息输入与推理逻辑，从而极大满足金融监管审查要求。 - [05:04](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=304s) 人机协同如同赛车的“下压力与 DRS”：在纯计算环节自动化全开释放极速，仅在复杂语义或伦理灰色等高难度“弯道”触发人工裁决。 - [08:50](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=530s) 硬编码依赖的旧版 Copilot 替换编排引擎耗时长达 3 个月，而模块化 Agent 分析师通过智能路由器仅用数天即平稳上线更强模型。 - [10:33](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=633s) 树立“以赛季为周期合作，而非托付终生”的供应商选型原则，始终保留根据性能与性价比动态更换技术栈的主动权。 - [13:06](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=786s) 同类大语言模型切换成本低，但引入世界模型等跨范式架构时无法直接热插拔，必须预先为其量身构建专用的数据与运行支架（Harness）。 - [15:04](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=904s) AI 智能体作为自然语言交互层，成功化解了长期困扰资管机构的“基本面 vs 量化”文化冲突，让投资经理无需懂代码即可调用顶尖量化工具。 - [15:04](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=904s) 未来资管机构的壁垒在于人机共生：人类负责深入一线捕捉非共识机会并将投资直觉注入系统，持续沉淀独有的专有知识资产。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=4s) 赛车进站与 AI 架构：从 67 秒到 0.0 秒的极致追求 演讲者 Max Gokhman 以一种双重身份开场：“我是谁？你们为什么要在意我？”他在周末是一名赛车手，屏幕上展示着他在拉古纳塞卡赛道（Laguna Seca）飞驰的身影；而在工作日，他担任富兰克林邓普顿（Franklin Templeton）人工智能与数字资产解决方案负责人，负责推动内部及客户的 AI 流程与多智能体（Multi-Agent）框架落地。在此之前，他曾担任多资产团队的副首席投资官（Deputy CIO），并创办过最早的纯 AI 对冲基金之一。 随后，Max 将话题引向赛车运动的核心机制——进站维修（Pit Stop）。1950 年印第安纳波利斯 500（Indy Car）赛车进站一次需要耗费整整 67 秒；而到了 2023 年，迈凯伦车队（McLaren）创造了惊人的 1.8 秒进站世界纪录。赛车变简单了吗？并没有，现代赛车的复杂程度呈指数级提升。之所以能实现超高速进站，是因为整个过程是一场经过精密编排的舞蹈，赛车上的每一个核心部件都是高度模块化的。尤其在金融与资产管理领域，随着代币化和区块链技术的普及，交易将演进为全天候（24/7/365）不间断运行。因此，金融企业面对的终极目标不是迈凯伦的 1.8 秒，而是“0.0 秒”——必须在赛车飞驰的同时完成进站调优，绝不能停下引擎。 - **1950 年印第安纳波利斯赛车进站时间**：67 秒。 - **2023 年迈凯伦 F1 车队进站时间**：1.8 秒。 - **全天候交易环境下的终极目标**：0.0 秒，即不停车完成架构升级与维护。 --- ### [01:41](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=101s) 底盘与组件：解耦持久层与可插拔层 面对日新月异的技术，许多企业陷入了两难困境：要么选择“冻结技术栈”（Freeze stack），认为锁定并全面验证后的流程最为稳妥负责；要么盲目“追逐每一个新框架”（Chase every framework）。Max 直言这两种极端都是灾难。如果 18 个月前锁定了某个基础大模型，现在就会直接跌落到“发车格的最底端”；而如果频繁更换技术，就如同“在未干的水泥地上驾车”，注定无法前行。这绝不是一个合规治理问题，而是一个纯粹的“系统架构问题”。 破解该难题的关键在于将 AI 基础设施解构为两大核心维度： 1. **底盘（The Chassis）**：这是企业的核心持久层与独有知识产权（IP），包含合规治理体系、风险边界设定、客户目标定义、针对智能体的时间点（Point-in-time）审计机制、人机协同（Human-in-the-loop）切入点以及评估测试集（Evals）。 2. **组件（The Components）**：如同赛车上的轮胎、前鼻翼或空气动力学套件，属于可快速更换的高速升级层，涵盖基础大模型（Foundational models）、数据源（Data feeds）以及编排框架（Orchestration frameworks）。 连接底盘与组件的关键媒介被称为**“耦合接口”（Couplings）**。耦合接口本质上是一份严格的“协议契约”（Contract），清晰定义了组件的输入、输出以及治理约束指标。通过固定的评估基准与输入输出规范，团队可以在行业前沿发生变革时，像更换赛车轮胎一样瞬间替换大模型或编排层。 ```mermaid flowchart TD accTitle: Modular AI Architecture: Chassis vs Components accDescription: Architectural diagram separating durable enterprise chassis from hot-swappable components via coupling contracts. subgraph ChassisGroup["底盘系统（Chassis - 核心持久 IP）"] Gov["合规治理与风险控制"] Audit["时间点审计与评估集 (Evals)"] HumanOversight["人机协同与干预规则"] end subgraph CouplingLayer["耦合接口契约（Couplings / Contracts）"] Contracts["输入 / 输出 / 治理要求规范"] end subgraph ComponentsGroup["模块化组件（Components - 热插拔层）"] FM["基础大模型 (LLMs)"] Data["实时数据流 (Data Feeds)"] Orch["编排框架 (Orchestrators)"] end ChassisGroup <--> CouplingLayer CouplingLayer <--> ComponentsGroup ``` - **持久底盘的要素**：治理规则、风险阈值、客户业务逻辑、智能体时点审计、标准化评估标准。 - **可插拔组件的要素**：基础大语言模型、外部数据流、Agent 编排引擎。 - **耦合接口的核心作用**：确立机器可执行的规范契约，彻底消除对特定供应商的强依赖。 --- ### [05:04](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=304s) 速度与合规并存：编码约束、统一评估、数据回放与人类在环 为了在监管压力下依然保持高速迭代，Max 提出了四项关键工程实践： 1. **规则一次性编码（Encode constraints once）**：在 F1 赛车中，车队热衷于钻研规则漏洞以提高车速；但在金融合规中，绝不能让 AI 智能体去寻找治理规则的空子。所有合规、法务、信息安全与技术红线必须统一固化为底层代码逻辑，让相关监管审查团队完全放心，从而赋予工程团队快速前进的自由。 2. **标准化评估（Standardized evals）**：与法务、合规及网安部门提前达成共识，形成一套通用的准入评估套件。当市场上出现新组件时，只要能通过这套统一测试，就能直接接入生产系统，无需就准入机制重新展开冗长的各部门拉锯谈判。 3. **精准数据回放（Telemetry & Point-in-time Replay）**：赛车手赛后会与工程师调取遥测数据，精确分析在弯道踩刹车的踏板压力或转向角度是否导致轮胎与路面失去抓地力。智能体也必须具备这种能力——明确记录其在做出决策的特定时间点“究竟获取了哪些信息”、“为何采取该行动”。相比于人类分析师可能对内幕交易（MNPI）等问题遮遮掩掩，AI 智能体“不担心坐牢或被罚款”，只要有完善的回放机制，它会彻底如实交待决策链路，这反倒极大地取悦了金融监管机构。 4. **人类作为下压力与 DRS（Humans as Downforce）**：在空气动力学中，下压力（Downforce）会在直道上降低最高极速，但在高速过弯时能将赛车死死按在地面以防失控。人类在环机制正是 AI 系统的下压力：不要让合规团队去审查每一个繁重的基础计算步骤；应引入类似 F1“减阻系统”（DRS）的机制——纯粹的逻辑计算环节关闭人工介入以释放算力极速，唯有在遭遇复杂语义、边缘案例或伦理灰色地带等高难度“弯道”时，再迅速介入人工裁决。 - **遥测分析对比**：从分析赛车刹车压力与轮胎附着力，延伸至复盘智能体在微秒级交易或决策节点的信息状态。 - **下压力与 DRS 哲学**：直道计算自动化拉满，弯道复杂决策引入人工兜底，兼顾系统流速与安全边界。 --- ### [08:50](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=530s) 实战案例对比：焊接架构的惨痛代价 vs 模块化进站的敏捷迁移 为了证明架构解耦的实战价值，Max 对比了富兰克林邓普顿团队先后经历的两个真实开发案例： **反面案例：强行“焊接”的 Copilot 产品** 团队构建的首个面向客户的 Copilot 辅助产品曾令全员振奋。然而，他们当时将整个技术栈“焊死”（Welded）在了一起——将所有系统依赖项硬编码到底层架构中，甚至将整个智能体编排层焊在了底盘上。不久后，市面上出现了一款性能大幅领先的第三方编排框架。为了完成替换，团队不得不把系统整个拆散、破坏底盘，重新手动连线所有接插件。这次升级足足耗费了 **3 个月时间**，整个项目险些错过预定的上线交付期。 **正面案例：模块化打造的智能体分析师（Agentic Analyst）** 吸取教训后，团队从设计第一天起就贯彻了模块化思路。底盘内嵌富兰克林邓普顿沉淀的核心投资策略与业务认知，外部组件全部设计为“热插拔”。当一家前沿模型厂商推出了一款性能强悍但定价极具侵略性的前沿大模型时，团队迅速完成评估，判定其高价值特性仅需用于特定复杂推理环节，其余日常任务完全可以分流给其他性价比更高的供应商模型。由于输入输出契约与治理规则早已固化，工程师仅用几天时间便搭建了一个智能模型路由器（Smart Router）。这项比当年更换编排层复杂得多的架构改造，仅仅耗时**数天**便平稳上线。 | 架构维度 | 焊接架构（旧版 Copilot 案例） | 模块化架构（Agentic Analyst 案例） | |:--- |:--- |:--- | | **集成方式** | 依赖项硬编码，编排层与底盘一体焊接 | 建立解耦耦合接口契约，组件全部热插拔 | | **面临变故** | 出现更优异的第三方编排引擎 | 出现高性能但高成本的前沿大模型 | | **改造方案** | 彻底解构底盘，重写并重组全部连接线 | 引入智能路由器（Smart Router）按需动态分流 | | **改造耗时** | 耗时 **3 个月**，险些延误交付上线 | 仅耗时 **数天**，在飞驰中完成升级调优 | --- ### [10:33](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=633s) 核心法则与现场问答：短期合作、语义理解与供应商互操作性 Max 总结了三条核心落地原则： 1. **彻底分离治理逻辑与可插拔组件**：底盘代表核心资产与永续 IP，组件代表消费品。 2. **拒绝将任何部件焊死在底盘上**：必须通过标准化的耦合契约（Couplings）维持各层独立性。 3. **以赛季为周期选择技术伙伴，而非托付终生（Partner for a season, not for life）**：赛车队每年都会根据战绩与技术进步重新评估技术赞助商与供应商；在技术迭代极其迅猛的 AI 时代，企业同样必须保留随时根据性能更换底层工具的主动权。 在演讲尾声的现场提问环节，观众针对系统设计细节提出了针对性问题： - **关于业务场景的语义理解模块（Semantic Understanding）**： 有观众询问团队是否构建了针对不同细分用例的语义理解组件。Max 给予了绝对肯定的回答，指出“上下文即是一切”（Context is everything），针对特定业务语义的理解模块是整套系统必不可少的关键切片。 - **关于跨供应商的可移植性与互操作性（Interoperability）**： 有观众追问在频繁更换底层供应商时，系统如何维持敏捷度，其带来的工程开销是否可控。Max 解释道，这正是构建“耦合接口与统一评估套件”的初衷所在。通过预先跑通一套标准评测，团队不仅能迅速判断供应商 A 与供应商 B 是否达标、差异何在，甚至能够横向对比多款产品。更重要的是，这使得团队能够为不同业务场景选用最适配的组合配置——“并非每一辆赛车都需要装载相同的轮胎；如果你打造的是一辆拉力赛车（Rally car），所选配的硬件自然会截然不同”。 ### [13:06](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=786s) 模块化模型替换与适配支架（Harness）设计 在问答互动环节，观众借用车轮改装的案例提出了一个核心疑问：如果更换模型就像将 16 英寸轮毂换成 18 英寸轮毂，该如何设计底层架构才能确保整体流程不产生刚性依赖？Max Gokhman 赞赏了这个比喻，并顺势用赛车工程学进一步拆解——如果把模型比作轮胎，那么工程团队首先要确保“模型必须能够严密匹配轮毂尺寸”。正如你绝不可能把拉力赛轮胎强行装到一级方程式（F1）赛车上一样，企业在进行 AI 模型替换时，必须建立在同类规格与应用场景相匹配的前提之下。 Max Gokhman 举例说明，在同属大型语言模型（LLM）的梯队中，将 GPT-4 / GPT-4o 替换为 Claude 3.5 Sonnet 是相对平滑且对等的；但如果业务场景需要的是一个能够理解物理与时空因果的“世界模型（World Model）”，盲目套用大语言模型就会彻底失效。在这种情况下，团队的核心任务是先针对世界模型构建专属的“运行支架（Harness）”。一旦这个底层支架建立完毕，团队便可以在其规范内自由插拔与替换不同的世界模型。不同的支架在数据流动与处理权重上存在本质差异：例如世界模型虽然同样会接收非结构化文本，但其利用层级更偏底层基础；而大语言模型虽可辅助处理视觉数据，其侧重点与处理模式却完全不同。 ```mermaid flowchart TD accTitle: Modular Harness and Model Compatibility subgraph UseCaseLayer["应用场景定义"] UC1["语言推理与生成场景"] UC2["环境仿真与因果预测场景"] end subgraph HarnessLayer["专用适配支架 (Harness)"] H1["LLM 专用支架\n(重文本/轻视觉交互)"] H2["世界模型专用支架\n(重物理时空因果/轻文本)"] end subgraph SwappableModels["可插拔模型层 (Tires on Rims)"] M1["GPT-4 / GPT-4o"] M2["Claude 3.5 Sonnet"] M3["专用世界模型 A"] M4["专用世界模型 B"] end UC1 --> H1 UC2 --> H2 H1 <--> M1 H1 <--> M2 H2 <--> M3 H2 <--> M4 ``` - **同规格热插拔**：同类通用大语言模型（如 GPT 系列与 Claude 系列）之间的切换成本极低，只要输入输出接口和业务逻辑解耦即可快速完成。 - **跨范式重构**：从文本 LLM 切换到世界模型等新型架构时，无法直接替换，必须从头针对其数据流和计算特性设计专属 Harness。 - **差异化数据处理**：不同模型支架对多模态数据的依赖程度各异，世界模型将文本置于更基础的辅助层，而 LLM 则将视觉信息作为多模态扩展。 --- ### [15:04](https://www.youtube.com/watch?v=4V5ikAAMdHg&t=904s) 人机协同：智能体如何重塑主动投资与“量化基本面”壁垒 当被问及资产管理行业与投资决策如何拥抱这种模块化技术演进时，Max Gokhman 明确表示，即便自己过去曾亲自管理过多套完全量化的黑盒策略，他也坚信未来的投资界绝不会变成一个“毫无人类参与的纯机器世界”。人类在投资体系中依然占据核心生态位，但核心逻辑在于人类必须与 AI 展开深度共生与协作。他回顾了 2008 年全球金融危机（GFC）之后行业内部轰轰烈烈的“量化基本面（Quantamental）”投资浪潮——当时资管机构极力促成量化分析师与基本面分析师联手，但由于两种人群在思维模型、决策逻辑和认知语境上的巨大鸿沟，这种尝试在绝大多数机构中都因严重的文化冲突而宣告失败。 AI 智能体（Agents）的出现彻底打破了这种协作僵局。如今在富兰克林邓普顿（Franklin Templeton），基本面分析师无需学习复杂的编程或数学建模，只需使用自然语言与智能体对话。底层的智能体可以并发执行极其精密复杂的量化模型与数据情境化分析，而前台的投资经理则完全不必关心其底层算法的每一行具体实现。Max 形象地打了个比方：“你不需要懂得防抱死制动系统（ABS）的具体机械原理，也同样知道只要踩下刹车踏板车就能稳稳停住。” ```mermaid flowchart LR accTitle: Human-Agent Collaborative Investment Workflow subgraph Humans["人类基本面分析师"] H_Insight["实地调研与前沿洞察"] H_Teach["传授投资直觉与决策逻辑"] end subgraph Agents["AI 智能体中间层"] A_NLP["自然语言交互界面"] A_Context["数据多维情境化分析"] end subgraph Backend["后台系统与护城河"] B_Quant["复杂量化与风控引擎"] B_Moat["专有投资知识库 (Proprietary Moat)"] end H_Insight --> A_NLP A_NLP --> B_Quant B_Quant --> A_Context A_Context --> H_Insight H_Teach --> B_Moat ``` 在这种全新的工作流中，AI 负责繁重的数据挖掘、交叉比对和量化运算，人类投研人员则深入一线获取独家情报、捕捉非共识机会，并向 AI 解释“自己为何会做出这种判断”。通过不断将人类顶尖投资经理的思考逻辑注入模型，资管公司正在建立起真正具备排他性的专有知识库（Proprietary Knowledge Stack）。这种护城河不再仅仅局限于预测下一季度的财报数字，而是提炼出对资产定价最本质的判断力——究竟什么才是一笔优质的股票、优质的债券或优质的私募信贷投资。这也正是富兰克林邓普顿在主动投资管理领域训练 AI 并构建长期竞争壁垒的核心路径。 - **化解历史文化冲突**：智能体作为自然语言翻译层，使基本面投研人员能够无缝调度后台顶级量化工具，消除了十多年来“量化 vs 基本面”的协作断层。 - **ABS 防抱死制动隐喻**：投资人员只需关注业务输入与决策输出，繁复的底层算力与模型调度被彻底封装在智能体之后。 - **构建内生知识护城河**：人类的核心价值转移至“训练模型理解投资直觉”，沉淀出涵盖股票、固定收益与私募信贷（Private Credit）的专有投资方法论。 ## Video description Max Gokhman - Head of Artificial Intelligence and Digital Asset Solutions, Franklin Templeton In motorsport, no team rebuilds the car mid-race — they engineer it so key parts can be swapped or adjusted in seconds without compromising safety or speed. Enterprises in regulated industries like investment management face the same challenge with AI: model capabilities are evolving monthly and peers keep competition fierce, but decision quality and compliance cannot be compromised. This session presents a practical playbook for building agentic harnesses that evolve at the speed of technological change — separating the durable layers like governance, client constraints, and auditability from the modular ones such as models, data streams, and orchestration frameworks. Attendees will leave with concrete frameworks, vendor-selection criteria, and lessons from building agentic systems inside one of the world's largest asset managers. If your AI roadmap changes every quarter but your scrutineers don't, this talk will help tune your chassis to perform within track limits. [Source](https://www.youtube.com/watch?v=4V5ikAAMdHg)

Source quotes

[0:04] Max Gokhman: completely at the back of the grid. [0:04] Max Gokhman: driving on wet cement because you've never let it dry. [5:04] Max Gokhman: technically isn't worried about going to jail or getting fined [8:50] Max Gokhman: Partner for a season, not for life,

Original transcript

[0:04] Hi everyone. Welcome to the session. Um who  am I? Why should you care? Um on weekends, [0:11] I race cars. That's me at Laguna Seca. During the  day, I lead the build-out for Franklin Templeton's [0:18] AI and digital asset solutions. So, I'm  helping both our clients and ourselves [0:23] lead with AI processes and multi-agent frameworks.  Before that, I was deputy chief investment officer [0:30] for our multi-asset team. I also started  one of the first pure AI hedge funds. [0:36] And I've done a couple other relevant things.  So, hopefully that gives you context. Now, [0:42] let's talk about pit stops. In 1950, it took 67  seconds for an Indy car to go through a pit stop. [0:55] In 2023, McLaren had a 1.8 second pit  stop. Did the cars get simpler? No. [1:05] They were incredibly more complex, but the  entire pit stop was a choreographed dance [1:11] where every component was modular. This is the  core of what we're going to be talking about. Now, [1:18] the reality is, especially in finance, especially  with the advent of tokenization and blockchain [1:25] technology, we're going to be trading 24/7  365. So, the target we're shooting for [1:32] isn't McLaren's 1.8, it's 0.0. We need to do  a pit stop without actually stopping the car. [1:41] So, a lot of enterprises think that they have a  choice. They can freeze their technology stack, [1:48] or they can chase every single framework.  Now, it feels very safe, very responsible [1:55] to freeze your stack. We're fitting a process that  is robust. We've got everything validated. Now, [2:04] we can keep going. Well, the problem with that is  the pace of technology will make it obsolete. What [2:09] would happen if you locked into a foundational  model 18 months ago? You would be completely [2:16] at the back of the grid. And conversely, if you're  chasing every single framework, it's kind of like [2:23] driving on a wet cement because you've never  let it dry. And because of that, you are also [2:30] going to fail. So, really, what you need to do is  consider that this is not a governance problem. [2:39] This is an architecture problem. You can't  change You can't chase every single framework. [2:46] And you also can't freeze your stack. So, what  can we do? Well, let's think about everything [2:56] as two components. And you're right, this  is the slide to take a picture of if you're [2:59] going to take a picture of anything. First, the  chassis. That's durable. That is your IP. So, [3:06] things like governance. What is governance  for your business? Things like risk limits. [3:12] How do you define risk within your enterprise?  What are the guardrails that you can't go past? [3:17] What are your clients' objectives? What  do they need? How do you think about that? [3:24] How do you do your audits? And audits need to be  point in time for an agent. We'll talk about that [3:29] a little bit more later, but especially in my  industry, that's critical with agentic systems. [3:35] How do you keep the human in the loop, and when  does the human stay in the loop? What are your [3:40] evals? Evals are another topic we'll come back to  shortly, but that is a critical component of how [3:45] you can swap things in and out. And then on the  other side are your components, you know, tires, [3:51] nose pieces, aero, the things that are designed  to be upgradeable, that are designed to be swapped [3:57] quickly. Your foundational models, your  data feeds, your orchestration frameworks, [4:04] those are the things that you want to be  able to swap quickly as the industry changes, [4:10] not at your pace, but at the pace of the  cutting edge. How do we think about this? [4:17] Components are supposed to be connected to your  chassis through couplings. Couplings are really [4:26] effectively a contract. Think about  writing down what are the inputs, [4:32] what are the outputs, and what are the  requirements for governing that component? So, [4:39] how do I govern my foundational model? What  does it take in? What does it give? How does it [4:46] work with an orchestration layer? These are all  abstractions we can think of, write them down, [4:51] and then if we have the same set of eval, the  same set of requirements, always the same set [4:57] of inputs and outputs, swapping one component for  another becomes very straightforward. So, let's go [5:04] a little bit further now. What lets you stay fast?  And this is where we're going to double click on [5:10] some of those ideas. Number one, you encode your  constraints once. In Formula 1, it's very common [5:17] to take a very loose interpretation of the rule  book, to find loopholes. You don't want your [5:24] agents finding loopholes in your governance. You  want to make sure that everything is encoded in [5:31] the code. Those requirements are absolutely core  to keeping your compliance, your cybersecurity, [5:40] your tech team, and your legal team happy,  which is what's going to help you move faster. [5:47] Now, standardizing evaluations is also really  important, as I mentioned. When we talk about [5:53] standard evals, it means we worked with all of  those other groups, the legal, the compliance, [6:01] cybersecurity. We said, "This is what we all agree  needs to pass for this component to be accepted." [6:11] And if they all agree, we now have one eval. If  the new component passes that eval, we can plug it [6:17] in without re-litigating the entire process every  single time. And then there's replay. So, this [6:26] is something that I find is actually very similar  between my racing career and my investing career. [6:32] Where after every race, I get out of my car and  me and my race engineer look at all the data [6:38] that my car collected on the track. Every single  point. How much brake pressure was I applying? [6:44] Did I turn the steering wheel a little too hard  and lost loose contact patch with with the tarmac? [6:50] Similarly, we need to understand why the agents  did what they did. What did they know at that [6:57] particular point in time? This is actually for  regulated industry incredibly important point. [7:02] Because if I ask a human analyst, "Hey, did  you trade on material non-public information?" [7:08] They'll most likely say, "Of course not."  But how do I know if that's actually true? [7:13] With a good governance system for an agent, the  agent technically isn't worried about going to [7:19] jail or getting fined. So, the agent will answer  honestly. And so, the agent will be able to tell [7:25] you exactly what it knew, when it knew, and why it  did what it did. And then we can iterate, improve, [7:31] and make regulators far happier than they probably  thought they could be when dealing with an AI. [7:38] And then lastly, humans are the a This is again  a bit of a racing analogy, but downforce is what [7:45] keeps the car down to the ground, really  low. And downforce actually slows you [7:52] in a straight line, but it helps you avoid  flying out going into a corner. Very similarly, [7:59] humans can slow down the agentic process when  it's just doing pure computation, but when we're [8:06] thinking about all the complexities and nuance,  that's where humans are really important. So, [8:12] what do I mean by that? It's putting the human  oversight at a point where things get nuanced and [8:19] complex. Don't have humans looking at every single  part of the process. If you guys are familiar with [8:25] now an old Formula 1 concept of DRS or drag  reduction system, that actually lowered the [8:31] downforce on straights and put it back on in  the corners. So, if it's a simple computation, [8:38] you don't need a human in the loop. If there's  nuance, there's questions, there's that [8:44] ethics perhaps that we need to address, that's  where human oversight is really critical. [8:50] So, now I want to walk through two examples and  I'll start with a failure as you can see on the [8:55] image here. So, we had our first co-pilot that we  built, the first client-facing product that I was [9:01] really excited about it. Um we welded everything.  When I say welded, and remember, welding versus [9:11] connecting, welding meant we hardcoded all  of the dependencies. And that included an [9:17] orchestration layer. And then a third party came  out with a much better orchestration layer. Well, [9:23] we had to take everything apart. We had  to deconstruct the chassis. We had to [9:29] rewire all of the connectors cuz we didn't have  those contracts, we didn't have those couplings. [9:36] And that took 3 months. We  almost missed our due date. [9:41] Now, we learned our lesson. There was another  situation, and this was a pit stop. We had now [9:49] a new tool, an agentic analyst that we built, and  we built it to be modular from day one. Everything [9:56] that could be a component was hot swappable to our  core chassis of our core investment knowledge. So, [10:04] when we had a frontier model come out with pricing  that was really aggressive, we said, "Okay, well, [10:10] there's certain things that we do need to pay for  with this frontier model, but certain other things [10:15] where we can use a cheaper model from another  vendor." And so, we just built a smart router. [10:19] And because we had the governance locked in,  because we built that contract, that coupling [10:26] from the get-go, this took days instead of months.  And it was actually a much more complex change. [10:33] So, with that, three things that I want you to  take away here, and then we'll have time for some [10:40] questions. Cuz I'm moving quite fast, as you would  expect from a race car driver. One, separate your [10:47] governance from what is swappable. Remember, the  chassis is your core IP. That is what you want to [10:55] keep. That is what's robust. The components,  those are all the things you can swap in. [11:05] The couplings are what lets you do that. Those are  the contracts. Don't weld anything to the chassis. [11:14] Lastly, partner for a season, not for life. If  you think about most racing teams, they choose [11:24] their partners for with technology, and they  reevaluate them every year. Same thing is true [11:29] at the pace in which we're moving with AI.  If you have the right framework with modular [11:34] components, you're able to do that. That is  why we built modular harnesses. So, with that, [11:43] happy to open it up to questions. Are you doing  anything with a module for semantic understanding [11:50] of each of your use cases? We we are, yes.  And and that's a really critical component. [11:56] Um context is everything, so that is something we  absolutely need to do. When talking about changing [12:03] partners, being able to adapt to different  partners, how do you think or uh operate about [12:09] the portability or the interoperability across  vendors? Like, how is your system agile enough [12:14] to move from one to another? Does it require  too much effort? Do you think that overhead is [12:19] useful or or manageable? So, it it that that's  kind of the whole point of the the coupling idea, [12:26] right? Is if we build a standard set of evals, so  we can run them once and we can say, "This is an [12:32] eval for, say, an orchestration layer." Does it  work with vendor A? Does it work with vendor B? [12:39] Do they all need the same criteria? Where is  one lacking versus another? We can actually [12:43] not only swap them, but we can also compare them  to each other. And that, by the way, another thing [12:49] that that can mean is we can run different  components, different systems for different [12:57] use cases, which is also really important.  Not every single, you know, car, if you will, [13:02] needs the same set of tires. If we're building  a rally car, though, it may still have the [13:06] same engine as a GT car. So, some components  stay the same, some components we swap out. [13:14] You spoke about replacing a model by another  one at any time, but it's not a video edit. [13:21] Like, switching a 16-in each wheel by a 18-in  one. And what what if we do we come on to make use [13:29] to make your process not depend on [13:37] That's great great analogy by the way. Um [13:41] it's you need to make sure that your model is  going to be I mean using the wheel approach [13:45] I'd say make sure your model fits fits your rim,  right? If we're thinking of the model as a tire, [13:49] right? So, you know, put you can't  put rally tires on a Formula 1 car. [13:55] And that is going to be the case here as well. So,  if I'm moving let's to give you a real example, [14:00] if I'm say switching between GPT-4 6 and Sonnet 5,  that's pretty comparable. But if I'm if I need a [14:09] world model and I'm using a large language model,  that is just not going to be the same thing cuz [14:13] actually what I need to build then is the harness  for a world model. Once I build that harness, [14:20] I may be able to swap different world models in  and out. The way I would expand a little bit of [14:25] on what you're saying is we need to think about  how we can build the harnesses really well from [14:32] the get-go for a vote for that use case and then  we can use that harness to swap between different [14:38] components that fit that specific harnesses need.  So, world model harnesses will look very different [14:43] from large language model harnesses. Of course, as  we build the harness components themselves, we may [14:49] be able to swap some of those in and out. Like a  world model may still use unstructured text data [14:54] just like a large language model would. It just  would use it at a lower level and you know, [14:59] similarly a large language model would use some  visual data, but it wouldn't use it as much. [15:04] Thank you for sharing. I want to ask a  high-level question giving your experience [15:08] in asset management and also investment. How do  you see asset management industry move forward [15:13] with this technology or modular approach? Like  especially with investment. Thank you. Sure. So, [15:20] at a very high level, I don't think we're going to  in a human-less world. And I say this having run [15:25] several fully quant strategies, black box  strategies. Um I think humans have a place [15:31] in the investment process. However, humans need to  work together with AI. Now, here's an interesting [15:38] thing. About, you know, post-GFC, like 2010-ish,  there was this huge move towards what was called [15:46] quantum mental investing, and I was a part of  that as well, where we would take quantitative [15:52] investors and fundamental investors and make them  work together. That, generally speaking, did not [15:57] work well, because quants think very differently  from fundamental investors. And that created a lot [16:06] of cultural issues at most places that were  one and then try to bring the other one in. [16:11] However, with agents, fundamental investors,  like a lot of the folks we have at Franklin, [16:17] are able to just talk to the agents. So,  even though the agents may be running really [16:22] sophisticated quantitative processes on the back  end, the humans that they're talking to don't need [16:27] to necessarily understand the full specificity  of the process. So, using a car analogy, [16:33] you don't need to know how your anti-lock  braking system works in order to know that [16:37] your car will be able to stop when you push  the brake. And that is a big change. So, [16:45] we will get to a world, I believe, fairly  soon. We're building it right now at Franklin, [16:50] where humans and agents are really  working together, truly collaboratively, [16:55] and the agents are running a lot of sophisticated  analysis, they're contextualizing data, [16:59] the humans are out in the world gathering  intelligence, coming up with more of the [17:05] innovation and the insights that are unique, and  they're training the models to understand why [17:12] they're thinking the way they are. And I think  that's how we actually grow that proprietary [17:17] knowledge stack. So, the way we're thinking about  expanding our mode as an asset management company, [17:23] especially one that's so focused  on fundamental active management, [17:26] is growing the proprietary information that we  have and not just the information that's relevant [17:32] for the next earning cycle, but how we actually  invest. What makes an attractive bond or equity [17:40] or private credit investment? That's what we're  working with our humans training the AI on. [17:48] All right, let's give Max a round  of applause. Thank you so much.