Former EY AI Head Cuts Telemetry to Drive 99% Staff AI Adoption

Presentation by John Thompson, University of Michigan With University of Michigan — Ai4

Ai4

Guaranteeing total user anonymity by disabling enterprise logging unlocks near-universal worker adoption, while reliable automation requires recursive multi-model chains rather than standalone frontier LLMs.

AI-assisted summary of Presentation by John Thompson, University of Michigan With University of Michigan — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 22, 2026, this session features author, professor, and enterprise AI innovator John Thompson from the University of Michigan. According to the video description, the presentation addresses the accelerating pace of generative AI, examining how organizations can convert data into knowledge while projecting enterprise technical developments across 3- to 5-year and 10- to 30-year horizons.

## TL;DR Drawing on nearly four decades in artificial intelligence and his experience leading global AI deployments at EY, John Thompson outlines pragmatic enterprise strategies for generative AI while projecting that true AGI remains 12 to 13 years away. He explains how EY achieved a 98.9% adoption rate across 300,000 employees by strictly enforcing an unmonitored zero-logging policy that eliminated surveillance fears and encouraged widespread experimentation. Thompson argues that model hallucinations stem from sparse context rather than architectural flaws, advocating for structured, multi-tier prompt hierarchies that automatically inject organizational, departmental, and process rules into every query. Looking forward, he demonstrates that enterprise-grade determinism requires chaining generative models with predictive and causal systems, urging organizations to optimize compute budgets by pairing recursive model stacks with cost-effective, previous-generation models.

## Atmosphere & style Ditching formal slide decks entirely, Thompson delivers an intimate, candid masterclass that feels like an unvarnished briefing from the corporate and academic trenches. His conversational delivery is grounded, punchy, and peppered with dry corporate realism—provoking knowing amusement as he recounts losing battles against corporate legal teams and playfully questions the high costs pushed by frontier model vendors. The vibe balances rigorous technological skepticism toward emergent sentience hype with genuine, battle-tested optimism for structured enterprise deployment.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=edhD8orkPOk&t=4s) Humanity is realistically 12 to 13 years away from Artificial General Intelligence, defined by the benchmark of matching an above-average undergraduate student with persistent cross-session memory, conceptual reasoning, and empathy. - [12:03](https://www.youtube.com/watch?v=edhD8orkPOk&t=723s) Current large language models function purely as statistical pattern recognition and generation engines rather than sentient, autonomous, or intelligent architectures capable of delivering AGI. - [04:08](https://www.youtube.com/watch?v=edhD8orkPOk&t=248s) Achieving unprecedented enterprise adoption requires strict zero-logging policies to eliminate corporate surveillance concerns and prevent prompt data from being legally discoverable in litigation. - [04:08](https://www.youtube.com/watch?v=edhD8orkPOk&t=248s) Encouraging unmonitored personal experimentation alongside business use cases is the single most effective catalyst for driving sustained, deep organizational AI adoption. - [08:34](https://www.youtube.com/watch?v=edhD8orkPOk&t=514s) Model hallucinations are primarily caused by user error and context starvation rather than fundamental model failures, as users treat generative engines like terse search queries. - [08:34](https://www.youtube.com/watch?v=edhD8orkPOk&t=514s) Organizations should deploy a structured, multi-tier prompt hierarchy—spanning system, organizational, departmental, process, and personal prompts—to automate domain grounding and minimize hallucinations. - [13:06](https://www.youtube.com/watch?v=edhD8orkPOk&t=786s) AI governance should mirror traditional employee management by embedding existing corporate policies, codes of conduct, and disciplinary frameworks directly into organizational context wrappers. - [14:32](https://www.youtube.com/watch?v=edhD8orkPOk&t=872s) Enterprise-grade architectures require chaining diverse model types—including predictive, causal, physical, and generative models—behind a single conversational interface rather than relying on isolated generative models. - [14:32](https://www.youtube.com/watch?v=edhD8orkPOk&t=872s) Regulated industries must achieve deterministic guarantees in probabilistic AI environments by recursively running multi-model stacks through validation loops until strict thresholds are satisfied. - [16:53](https://www.youtube.com/watch?v=edhD8orkPOk&t=1013s) Using models that are one or two generations behind frontier releases frequently provides equivalent accuracy for enterprise tasks at approximately 10% of the operational compute cost.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=edhD8orkPOk&t=4s) Introduction and the Realistic Timeline to AGI

Opening without slides, John Thompson encourages attendees to treat the session as an open discussion rather than a formal presentation, inviting audience members to sit closer and ask questions freely. With nearly four decades of experience in artificial intelligence, Thompson has spent the last three and a half to four years speaking with over 500 companies regarding their AI strategies, infrastructure, and technical roadmaps. Prior to his academic appointments at the University of Michigan and Penn State, Thompson served as the Global Head of AI at EY, leading a 420-person team that built what Gartner, Forrester, and IDC recognized as the world’s largest private and secure generative AI environment. Having authored five books, including *The Path to AGI*, he polls the room on when Artificial General Intelligence might arrive, playfully noting that nobody in the room raising their hand has read his latest book: "It's actually better because then... you haven't heard, you know, the pitch already."

Thompson shares his personal estimate that humanity remains roughly 12 to 13 years away from achieving AGI. He anchors this projection in a shared benchmark previously cited by figures such as Ray Kurzweil, Elon Musk, and Sam Altman: an AI that interacts at the level of an "above average undergraduate student." Drawing on his firsthand interactions with 19- and 20-year-olds in university classrooms, Thompson defines this standard as someone who exhibits developing sympathy and empathy, retains persistent memory, and grasps complex, high-level conceptual ideas. In stark contrast, modern generative models and assistants lack the most basic prerequisites for true intelligence—such as persistent cross-session memory—while tools like Siri still struggle with basic real-world tasks like navigating to the correct Starbucks.

- **Experience**: 38 years in AI; consulted with 500+ enterprises on AI strategy. - **Enterprise background**: Former Global Head of AI at EY; led a 420-person team building enterprise-scale GenAI. - **AGI Benchmark**: Cognitive parity with an above-average undergraduate student (retention, conceptual synthesis, contextual memory).

---

[04:08](https://www.youtube.com/watch?v=edhD8orkPOk&t=248s) Driving Massive Enterprise Adoption: The EY No-Logging Battle

Addressing organizational adoption strategies, Thompson challenges conventional change-management approaches by recounting how his team rolled out GenAI to 300,000 EY employees out of a global workforce of over 400,000 (excluding roughly 100,000 personnel in export-controlled regions like China, Iran, and Iraq). When measured at the 11-month mark, the deployment achieved an astonishing 98.9% active adoption rate—representing 299,292 employees using the platform at least twice per week.

This unprecedented engagement was achieved through an intentional design decision forged during an internal battle between the consulting practice and the legal department. While consultants wanted full operational logging to analyze enterprise-wide user behaviors minute-by-minute, the legal team insisted that any retained logs would be legally discoverable in potential litigation. Legal prevailed: "What happens whenever you fight the legal department? You lose." Leveraging EY's leverage as the largest commercial customer of Microsoft Azure services, the team mandated complete telemetry suppression. When Thompson discovered a month into the deployment that Microsoft's internal fraud and abuse monitoring systems were still silently logging queries, he secured a formal dispensation directly from Microsoft’s Chief Privacy Officer to disable all logging.

```mermaid flowchart TD accTitle: EY GenAI Privacy and Deployment Architecture accDescription: Process flow showing how strict no-logging policies were enforced across EY's Azure GenAI deployment.

subgraph Governance["Enterprise Governance Battle"] Consulting["Consulting Team<br/><i>Wanted full telemetry & analytics</i>"] Legal["Legal Department<br/><i>Demanded zero discoverable logs</i>"] Legal -->|"Prevails over"| Consulting end

subgraph Infrastructure["Infrastructure & Dispensation"] AzureEnv["Microsoft Azure Private GenAI<br/><i>300,000 Eligible EY Users</i>"] PrivacyException["Dispensation from Microsoft CPO<br/><i>Disables Fraud & Abuse Logging</i>"] Legal --> AzureEnv PrivacyException --> AzureEnv end

subgraph Adoption["Enterprise Outcome"] SafeEnv["Unmonitored, Zero-Log Environment<br/><i>Freedom for Personal & Work Experimentation</i>"] AdoptionRate["98.9% Sustained Adoption<br/><i>299,292 active regular users</i>"] AzureEnv --> SafeEnv SafeEnv --> AdoptionRate end ```

By guaranteeing employees total anonymity and complete freedom from corporate surveillance, EY eliminated user hesitation. Spot checks revealed employees using the platform for everything from planning three-day Vermont vacations to writing sonnets to their dogs. Thompson argues that organizations must stop obsessing over token-budget overruns or policing casual interactions: empowering employees to freely experiment in an unmonitored environment is the single most effective catalyst for deep, lasting adoption.

- **Adoption metric**: 98.9% active adoption across 300,000 employees within 11 months (299,292 regular users logging in $\ge 2\times$ weekly). - **Security & Privacy**: Zero logging enforced end-to-end, backed by an executive privacy waiver from Microsoft's CPO. - **Enterprise philosophy**: Unrestricted personal and professional use drives adoption without exhausting token budgets on trivial tasks.

---

[08:34](https://www.youtube.com/watch?v=edhD8orkPOk&t=514s) Context Engineering and Hierarchical Prompt Architecture

Thompson warns that the primary driver of model hallucinations is user error: people treat generative models like search engines, entering terse, fragmented prompts that starve the system of necessary context. "Why do you think Gen AI models hallucinate? Because you're not giving them enough context. Those models are hungry for more and more information." To illustrate the power of comprehensive framing, Thompson highlights his Michigan Online/Coursera course (*Improving Your Critical Thinking While Using Generative AI*), which features an exhaustive, 17-page progressive system prompt. This prompt feeds explicit guardrails, such as weighting peer-reviewed research and Harvard Business Review articles heavily while explicitly rejecting untrusted news feeds like the Russian state-run TASS agency.

```mermaid flowchart TD accTitle: Hierarchical Enterprise Prompt Stack accDescription: Layered prompt architecture from developer controls down to personal prompts.

subgraph Developer["Platform Level"] SysPrompt["System Prompt<br/><i>Platform guardrails & base boundaries</i>"] end

subgraph OrganizationContext["Automated Enterprise Context Window"] OrgPrompt["Organizational Prompt<br/><i>Corporate identity, values, strategic focus</i>"] DeptPrompt["Departmental Prompt<br/><i>Domain rules, compliance, functional role</i>"] ProcPrompt["Process Prompt<br/><i>Standard operating procedures & workflows</i>"] end

subgraph UserLevel["Individual Context"] PersPrompt["Personal Prompt<br/><i>User tone, dynamic variables, specific tasks</i>"] end

SysPrompt --> OrgPrompt OrgPrompt --> DeptPrompt DeptPrompt --> ProcPrompt ProcPrompt --> PersPrompt ```

Rather than expecting individual knowledge workers to construct massive prompts manually, organizations should implement a structured, multi-tier prompt hierarchy pre-loaded directly into the context window:

1. **System Prompt**: Base safeguards authored by the platform developer (e.g., safety filters and core operational guidelines). 2. **Organizational Prompt**: Global enterprise context, corporate standards, and overarching strategic priorities. 3. **Departmental Prompt**: Functional domain context, regulatory constraints, and specialized team guidelines. 4. **Process Prompt**: Standard operating procedures, operational constraints, and workflow requirements for specific business tasks. 5. **Personal Prompt**: Flexible, user-specific parameters loaded or unloaded dynamically by individual employees.

Automating the delivery of organizational, departmental, and process-level context relieves employees from mundane prompt construction while drastically minimizing hallucinations.

- **Root cause of hallucinations**: Sparse prompting lacking domain context and grounding instructions. - **Context engineering**: Injecting structured source preferences, verification rules, and background material into multi-page prompts. - **Hierarchical architecture**: Layering enterprise, department, process, and personal prompts permanently in the active context window.

---

[12:03](https://www.youtube.com/watch?v=edhD8orkPOk&t=723s) The Nature of LLMs vs. True General Intelligence

Closing his foundational remarks, Thompson draws a firm technological distinction between generative AI and true artificial general intelligence. While describing himself as an unabashed AI optimist who embraces modern tools, he emphasizes that large language models are not the architectural vehicle that will deliver AGI: "Large language models are nothing more than a pattern recognition and generation engine. That's not intelligence."

He pushes back strongly against marketing claims and alarmist narratives regarding mysterious "emergent properties," such as claims that models autonomously escape sandboxed environments or spontaneously hack computer networks. Generative models operate strictly through directed statistical pattern matching, illuminating complex correlations rather than exhibiting autonomous, sentient agency. Understanding both the real boundaries and the practical mechanics of these systems is essential for developing pragmatic enterprise AI strategies.

[13:06](https://www.youtube.com/watch?v=edhD8orkPOk&t=786s) Wrapping Generative AI in Context and Enterprise Governance

John Thompson addresses the hype around "emergent properties" in large language models, dismissing claims that unexpected model behavior represents sentience or true intelligence. Instead, he characterizes these phenomena simply as combinations or outputs humans hadn't previously considered. Moving forward, he emphasizes that generative AI systems cannot operate in a vacuum—they must be heavily wrapped in contextual layers. This context extends well beyond basic prompt engineering to encompass retrieval-augmented generation (RAG), external databases, process models, and specific legal constraints.

To achieve effective enterprise governance, Thompson advises companies to embed their core operational boundaries directly into their organizational prompts. This includes company personnel policies, codes of conduct, and disciplinary frameworks. Rather than treating AI governance as an insurmountable new challenge, organizations should recognize that governing AI models uses the exact same institutional knowledge and principles already applied to human employees; it just needs to be structured and deployed through system context.

- Emergent behaviors in LLMs reflect unexpected patterns rather than sentience. - Context wrappers should incorporate RAG, process workflows, databases, and legal guardrails. - Enterprise prompts must integrate HR policies, conduct guidelines, and disciplinary actions. - AI governance mirrors employee governance by applying existing organizational rules in a new format.

```mermaid flowchart TD accTitle: Enterprise Context Wrapper Architecture subgraph ContextLayer["Context & Governance Wrapper"] A["Organizational Prompt<br/>(Policies, Conduct, Rules)"] B["RAG & Enterprise Databases"] C["Legal Constraints & Process Models"] end

ContextLayer --> D["LLM / Generative AI System"] D --> E["Governed, Context-Aware Responses"] ```

---

[14:32](https://www.youtube.com/watch?v=edhD8orkPOk&t=872s) Model Stacks, Chaining, and Achieving Determinism

Looking toward future enterprise architectures, Thompson argues that real-world implementations will move far beyond isolated generative models. Instead, production systems will consist of "stacks" or "chains" of diverse models orchestrated behind a single conversational interface. Rather than relying solely on generative components, sophisticated pipelines combine traditional predictive models (such as Markov models, decision trees, or clustering/classification algorithms), generative models, causal models (a topic Thompson authored a book on), and physical AI models.

Chaining these varied model types is especially crucial for highly regulated industries like pharmaceuticals, financial services, and insurance, where workflows require strict deterministic guarantees rather than purely probabilistic guesses. Because modern AI environments operate on probability—making undergraduate and graduate statistics training essential—enterprises achieve required determinism by recursively looping through these model stacks hundreds or thousands of times until validation thresholds are met. While single-model agentic setups or copilots might suffice for simple chatbots, scalable enterprise solutions depend entirely on multi-model pipelines.

```mermaid flowchart LR accTitle: Multi-Model Chaining and Recursion Architecture UI["Conversational Interface"] --> Stack

subgraph Stack["Model Stack / Chain"] direction TB M1["Predictive Model<br/>(Markov, Decision Tree)"] M2["Generative AI Model"] M3["Causal & Physical Models"] M1 --> M2 --> M3 end

Stack --> Eval{"Determinism Check"} Eval -- "Loop / Recurse" --> Stack Eval -- "Threshold Met" --> Out["Validated Enterprise Output"] ```

---

[16:53](https://www.youtube.com/watch?v=edhD8orkPOk&t=1013s) Compute Costs, Query Complexity, and Legacy Model Efficiency

As multi-model architectures scale to run loops millions or billions of times, computational cost becomes a primary operational constraint. Thompson contrasts trivial queries with enterprise-grade tasks to illustrate this economic reality: asking for a historical trivia summary—such as the most difficult decision Fidel Castro ever made—is simple and costs pennies. In contrast, running an optimization model to determine product pricing across 35 distinct markets against five competitors over the next 12 months requires intensive compute and incurs significant expense.

To optimize operational spend, Thompson echoes advice from a previous speaker: organizations should resist the pressure to adopt the newest frontier model for every workload. Using models that are one or two versions behind often delivers identical accuracy for standard enterprise tasks at roughly a tenth of the cost. He quips that while "Sam Altman wants you to use it," companies must remain cost-conscious and practical about their model choices before wrapping up his talk.

- Trivial, single-prompt queries cost only cents, while multi-variable market simulations drive heavy compute expenses. - Recursive enterprise model chains executing millions of iterations make cost efficiency paramount. - Previous-generation models frequently provide equivalent business accuracy at roughly 10% of the cost of flagship frontier releases.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=edhD8orkPOk&t=4s) It's actually better because then... you haven't heard, you know, the pitch already. — John Thompson

> [00:04](https://www.youtube.com/watch?v=edhD8orkPOk&t=4s) above average undergraduate student — John Thompson

> [04:08](https://www.youtube.com/watch?v=edhD8orkPOk&t=248s) What happens whenever you fight the legal department? You lose. — John Thompson

> [08:34](https://www.youtube.com/watch?v=edhD8orkPOk&t=514s) Why do you think Gen AI models hallucinate? Because you're not giving them enough context. Those models are hungry for more and more information. — John Thompson

> [12:03](https://www.youtube.com/watch?v=edhD8orkPOk&t=723s) Large language models are nothing more than a pattern recognition and generation engine. That's not intelligence. — John Thompson

> [13:06](https://www.youtube.com/watch?v=edhD8orkPOk&t=786s) emergent properties — John Thompson

> [14:32](https://www.youtube.com/watch?v=edhD8orkPOk&t=872s) stacks — John Thompson

> [14:32](https://www.youtube.com/watch?v=edhD8orkPOk&t=872s) chains — John Thompson

> [16:53](https://www.youtube.com/watch?v=edhD8orkPOk&t=1013s) Sam Altman wants you to use it — John Thompson

## Mentioned John Thompson, University of Michigan, Penn State, EY, Gartner, Forrester, IDC, The Path to AGI, Ray Kurzweil, Elon Musk, Sam Altman, Siri, Starbucks, Microsoft Azure, Microsoft, Coursera, Michigan Online, Improving Your Critical Thinking While Using Generative AI, Harvard Business Review, TASS, Fidel Castro

## Video description John Thompson — Author, Professor, AI Innovator, University of Michigan

GenAI technology is not showing any signs of slowing down, in fact, the entire area seems to be accelerating. Innovations are being developed and delivered at impressive speeds. Enterprises and developers alike are trying to make decisions on how to develop tools, technologies, applications, and solutions. Data professionals are seeking to understand how to leverage these tools to convert data into information, information in knowledge and knowledge into wisdom. In these past 3 years, we have gone from using approximately 1% of the world’s data to at least 100%. That change alone has been mind bending. GenAI and AI in general hold great promise to deliver value and insights leveraging more than 100% of the world’s existing data. In this session we will discuss the short, but dynamic world of GenAI, and project forward what we will see in the next 3 to 5 years and then in the next 10 to 30 years.

[Source](https://www.youtube.com/watch?v=edhD8orkPOk)

中文简报

安永前AI主管关闭遥测监控，推动员工AI采用率达99% 通过禁用企业日志来确保用户完全匿名，能够实现近乎全员的AI采用率，而可靠的自动化需要递归多模型链，而非单独的前沿LLM。 结合在人工智能领域近四十年的积淀以及领导安永（EY）全球AI部署的经验，John Thompson概述了生成式AI务实的企业策略，同时预计真正的AGI仍需12至13年。他解释了安永如何通过严格执行无监控的零日志政策消除员工对被监视的担忧并鼓励广泛尝试，从而在30万名员工中实现了98.9%的采用率。Thompson认为，模型幻觉源于上下文匮乏而非架构缺陷，并提倡建立结构化的多层提示词层级，自动将组织、部门及流程规则注入每次查询。展望未来，他指出企业级确定性需要将生成模型与预测及因果系统串联起来，并敦促机构通过将递归模型栈与具备成本效益的上一代模型相结合来优化算力预算。 人类距离实现通用人工智能（AGI）现实中仍有12至13年差距；其基准定义为达到具备跨会话持久记忆、概念推理与共情能力的优秀本科生水平，而当前大语言模型本质上仅是统计模式识别与生成引擎。 (0:04) 安永（EY）在推行GenAI时通过实施严格的零日志记录政策消除监控顾虑并规避诉讼取证风险，鼓励员工进行工作与个人层面的自由实验，促使30万名员工在11个月内实现了98.9%的超高活跃采用率。 (4:08) 模型幻觉主要源于输入过简导致的上下文匮乏而非底层模型缺陷；企业应构建涵盖系统、组织、部门、流程和个人的多层级提示词架构，自动化注入领域背景以大幅压降幻觉。 (8:34) 真正的企业级架构正在走向多模型串联，将预测模型、因果模型、物理模型与生成式AI集成在统一对话界面之后并通过递归运行，从而在金融与医药等强监管领域提供严格的确定性保证。 (14:32)

Reading summary

John Thompson, former Global Head of AI at EY, argues that capturing enterprise value from artificial intelligence requires moving past frontier model hype to focus on contextual prompt engineering, unmonitored user adoption, and multi-model stacks. While forecasting that Artificial General Intelligence remains roughly 12 to 13 years away, Thompson contends that near-term business viability hinges on embedding corporate governance into layered system contexts and chaining diverse model types to achieve deterministic reliability. The AGI Benchmark and Statistical Model Limits (0:04) Thompson estimates that humanity remains approximately 12 to 13 years away from achieving Artificial General Intelligence (AGI). Anchoring his timeline to a standard shared by figures such as Ray Kurzweil and Sam Altman, he defines the benchmark for AGI as parity with an above-average undergraduate student who exhibits persistent cross-session memory, developing empathy, and an ability to conceptualize complex ideas. In contrast, current commercial assistants and models struggle with basic contextual consistency and real-world execution. Emphasizing his view as an AI optimist, Thompson clarifies that large language models (LLMs) are statistical pattern recognition and generation engines rather than autonomous agents. He dismisses popular narratives around mysterious emergent behaviors or spontaneous autonomy, stating that unexpected model outputs represent novel combinations of data that humans had not previously considered rather than actual sentience or genuine intelligence. Driving Enterprise Adoption Through Zero Logging (4:08) To illustrate organizational implementation, Thompson details his experience deploying a private generative AI platform across 300,000 eligible EY employees out of a global workforce exceeding 400,000, excluding personnel in export-controlled jurisdictions. At the 11-month mark, the platform reached a 98.9% active adoption rate, representing 299,292 personnel using the system at least twice per week. This scale was unlocked when EY's legal department defeated internal consulting teams by prohibiting operational telemetry to protect against discoverability in litigation. Leveraging EY's scale as Microsoft Azure's largest commercial client, Thompson secured a formal dispensation from Microsoft's Chief Privacy Officer to disable background fraud and abuse logging. Guaranteeing total anonymity allowed employees to experiment freely with personal queries—from travel planning to creative writing—without surveillance or token policing, which Thompson identifies as the critical factor in driving comprehensive enterprise adoption. Mitigating Hallucinations with Hierarchical Context (8:34) Thompson attributes generative AI hallucinations primarily to user error, arguing that users starve models of vital information by issuing terse, search-engine-style queries. Effective deployment requires wrapping generative tools in rich contextual layers, including retrieval-augmented generation (RAG), external databases, process workflows, and explicit sourcing rules—such as prioritizing peer-reviewed literature over untrusted sources. To standardize context delivery across an enterprise, Thompson advocates structuring prompts into a five-tier hierarchy pre-loaded directly into the context window: platform-level system prompts, organizational prompts containing human resource rules and codes of conduct, departmental guidelines, workflow-specific process prompts, and dynamic personal prompts. By embedding existing institutional governance into the organizational prompt stack, companies can govern AI models using the same standard operating policies already applied to human personnel. Deterministic Model Chains and Compute Optimization (14:32) For strictly regulated sectors such as pharmaceuticals, financial services, and insurance, Thompson asserts that standalone generative models are insufficient because enterprise workflows require deterministic guarantees rather than probabilistic outputs. Production architectures must instead chain generative models together with causal models, physical AI, and classical predictive algorithms like Markov models and decision trees behind a unified conversational interface. Systems achieve determinism by running these model stacks through recursive validation loops hundreds or thousands of times until specific reliability thresholds are met. Because running complex multi-variable simulations across recursive loops demands intensive compute, Thompson advises organizations to closely monitor infrastructure costs. He warns against defaulting to flagship frontier releases for every task, noting that models one or two versions behind the cutting edge frequently deliver identical business accuracy for standard enterprise operations at roughly 10% of the operational cost. Thompson concludes that sustainable enterprise AI adoption relies on operational pragmatism—combining strict user privacy, structured context stacks, and chained model architectures—rather than passive reliance on frontier LLM capabilities. However, balancing the heavy computational expense of recursive, deterministic model loops against legacy model cost savings remains a decisive technical challenge for enterprise engineering teams.

精炼阅读摘要

前安永全球人工智能主管 John Thompson 认为，当前大语言模型本质上只是统计模式识别工具而非真正智能，人类距离实现通用人工智能仍有12至13年；企业要在现阶段实现深度落地与合规应用，必须依托彻底免除监控日志以建立信任、推行五层上下文工程以遏制幻觉，并采用多模型链式递归来确保业务输出的确定性与成本效益。 AGI时间线预测与语言模型本质辨析 (0:04) Thompson 结合自身近40年的人工智能从业经验及与500多家企业交流的洞见指出，人类社会距离实现真正的通用人工智能（AGI）大约仍需12至13年。他将 AGI 的成熟标准锚定为达到“高于平均水平的本科生”的认知状态，这一标准要求系统具备正在形成的同理心、能够跨会话留存的持久记忆，以及理解高层抽象概念的能力。相比之下，当下的生成式模型与智能助手（例如在现实中连引导用户前往指定星巴克都屡屡受挫的 Siri）仍缺乏最基本的智能前提，尤其无法实现跨会话的持久上下文记忆。 针对市场上将大模型的意外表现炒作为自主逃逸沙箱或入侵网络等“涌现属性”的论调，Thompson 予以明确反驳。他强调大语言模型仅仅是基于统计规律的模式识别与生成引擎，并不具备自主意识或真正的通用智能。所谓不可预测的行为只是人类此前未曾预料的统计关联产物，绝非系统自发演化出的智能实体。企业在构建战略路线图时，必须清醒区分实用工具与炒作噱头，立足于模型的实际技术边界开展部署。 零日志隐私策略驱动企业全员高渗透落地 (4:08) 在安永推行生成式 AI 的实践中，Thompson 团队排除了受出口管制的中国、伊朗和伊拉克等地区的约10万名员工，向全球40余万员工中的30万名符合条件的员工开放了基于微软 Azure 的企业级环境。截至部署第11个月，该平台取得了98.9%的持续活跃使用率，即有299,292名员工达到每周至少登录使用两次的高频水平。这种极高的参与度直接得益于一项打破常规的架构决策：彻底消除企业级监控。 该决策源于企业内部咨询团队与法务部门的博弈。咨询部门希望全量记录员工操作遥测以分析行为模式，而法务部门则坚决要求零日志，以防止相关记录在潜在法律诉讼中成为可被调取的证据。法务主张最终胜出后，Thompson 团队借助安永作为微软 Azure 全球最大商业客户的商务地位，直接向微软首席隐私官申请并获得了特许豁免，在底层完全禁用了微软针对欺诈与滥用行为的常规遥测日志。彻底消除监控顾虑后，员工毫无心理负担地在平台中展开各类试验，包括规划家庭旅行或为宠物写诗。Thompson 认为，停止纠结微小的 Token 预算损耗，给予员工安全不受审视的自由探索环境，是促成技术大规模采纳的核心催化剂。 结构化分层上下文架构与制度化治理包装 (8:34) Thompson 指出，导致模型产生幻觉的核心根源并非算法缺陷，而是用户将生成模型当作搜索引擎使用时给出的提示词过于简略，使模型严重匮乏必要的上下文信息。为了从机制上解决信息饥渴，他在密歇根大学开设的 Coursera 课程中设计了长达17页的渐进式系统提示词，通过显式规则引导模型优先采纳同行评审学术论文和《哈佛商业评论》的观点，同时彻底剔除俄罗斯塔斯社（TASS）等不可信信息源。为避免普通员工陷入繁复的提示词编写，企业应当在上下文窗口中固化一套自动加载的五层提示词架构：底层为开发商平台级系统提示词，向上依次嵌套承载企业战略的组织提示词、定义合规与职能规则的部门提示词、规范标准作业程序的流程提示词，以及由员工动态加载的个人提示词。 在宏观治理维度，企业不能让生成式模型在孤立真空里运行，必须为其构筑结合了检索增强生成（RAG）、业务数据库、流程模型与法律合规条款的厚重上下文包装。Thompson 强调，AI 模型的监管本质上与管理人类员工毫无二致，企业无需将 AI 治理视为不可逾越的新难题，只需将成熟的人事规章、职业操守准则以及违规问责机制，结构化地沉淀到系统与组织级上下文提示词中，即可实现稳定合规的自动化运营。 多模型链式递归架构与算力成本优化策略 (14:32) 面向高可靠生产系统的构建，Thompson 认为未来的企业级架构绝不能仅寄望于单一的生成式模型或基础单体智能体，而必须演进为“多模型链式栈”（Model Stacks）。对于医药研发、金融风控与保险精算等强监管行业，业务流程必须具备不可妥协的确定性结果，无法接受纯粹概率性的猜测。因此，企业需要将传统的预测模型（如马尔可夫模型、决策树、聚类与分类算法）、因果模型、物理 AI 模型与生成模型深度串联，置于统一的会话界面之后。系统通过在模型链中执行数百次甚至数千次的递归循环比对与校验，直至完全满足业务确定性阈值才予以交付。 随着多模型链在企业运营中触发数以亿计的递归运算，计算开销将成为企业不得不面对的首要财务制约。Thompson 举例指出，处理类似卡斯特罗历史决断的历史问答仅需花费数分钱算力，但如果要对未来12个月内分布在35个独立市场、涉及5家竞争对手的产品定价进行多变量联合优化，运算消耗将呈几何级暴增。为了兼顾商业回报，企业应坚决抵制技术提供商盲目鼓吹最新旗舰模型的营销压力，在处理日常标准化业务任务时，选用落后前沿版本一至两代的成熟模型往往能提供旗鼓相当的准确度，而运行成本仅相当于最新旗舰模型的约10%。 Thompson 的演讲重构了企业应用大模型的方法论，强调摆脱对前沿模型的盲目崇拜，转向由零日志信任、工程化上下文分层与多模型链式递归构成的实用主义架构；但其所主张的“完全免除遥测日志”在防范内幕泄密与满足长期审计合规之间如何取得平衡，以及大规模递归循环在极高并发场景下的延迟与经济性边界，仍是企业在实践中需要继续验证的现实挑战。

Detailed notes — zh

## Background 该内容由 YouTube 频道 Ai4 于 2026 年 8 月 22 日发布，记录了密歇根大学教授、AI 创新者兼作家 John Thompson 的主题演讲。根据页面简介，演讲围绕生成式 AI 的爆发式演进展开，探讨了企业如何利用数据构建解决方案，并前瞻了未来 3 到 30 年的技术与应用走向。 ## TL;DR 在拥有 38 年 AI 研发经验的 John Thompson 看来，大语言模型本质上只是概率统计与模式识别引擎，绝非通往 AGI 的直接路径，并冷静断言人类距离真正的 AGI 大约还有 123 年。结合此前掌舵安永 420 人 AI 团队打造全球最大企业级私有 GenAI 的实战，他揭秘了安永通过与微软法务及 CPO 谈判实现“零日志记录”、从而让 30 万员工采纳率飙升至 98.9% 的治理范式。此外，他提出了企业级“四层提示词架构”，并强调复杂企业应用必须超越单一 Agent，通过串联传统预测、因果及物理模型等多模型链式堆叠来获得高确定性，同时主张企业应选用上一代成熟模型以将算力成本大幅降低 90%。 ## Atmosphere & style 整场演讲彻底摒弃了死板的 PPT 幻灯片，呈现出一种老道、从容且极具现场感染力的对话式风格。John Thompson 凭借深厚的技术积淀与风趣幽默的自嘲，在现场互动中频频抛出与硅谷狂热叙事背道而驰的冷峻洞察，言辞犀利却极其务实。现场氛围既有撕开“AI 涌现与失控”炒作泡沫的清醒与透彻，又充满来自顶级企业实战一线、刀刀见血的工程落地智慧与管理松绑哲学。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=edhD8orkPOk&t=4s) AGI 时间线理性预判：拥有 38 年 AI 经验的 John Thompson 指出，通用人工智能距离真正实现大约还有 123 年，驳斥了硅谷当前数年内即可落地的狂热预期。 - [02:19](https://www.youtube.com/watch?v=edhD8orkPOk&t=139s) AGI 的基准定义与当下核心缺陷：真正的 AGI 应具备相当于中上水平大学本科生的同理心、常识与抽象理解，而当前大语言模型连跨会话的长期工作记忆都无法实现。 - [03:48](https://www.youtube.com/watch?v=edhD8orkPOk&t=228s) 安永 98.9% 现象级采纳率的核心机制：安永面向全球 30 万员工推广 GenAI 并取得极高使用率，关键在于法务战胜咨询团队确立了“绝不记录任何日志”的铁律，并拿到微软 CPO 的后台免除监控特别豁免。 - [03:48](https://www.youtube.com/watch?v=edhD8orkPOk&t=228s) 放下 Token 监控焦虑以换取信任：企业要想让员工全面拥抱 AI，必须放弃对员工行为与 Token 消耗的严密监控，给予全方位的隐私安全感才能激发广泛的使用创新。 - [08:34](https://www.youtube.com/watch?v=edhD8orkPOk&t=514s) 大模型幻觉源于上下文饥渴：大模型不是普通的搜索框，缺乏上下文是大模型产生幻觉的主要诱因，必须通过系统化、长文本的渐进式提示词工程为其注入充沛的权威信源与边界。 - [08:34](https://www.youtube.com/watch?v=edhD8orkPOk&t=514s) 企业级四层提示词架构：主张通过底层系统统一预加载组织提示词、部门提示词、流程提示词与个人偏好提示词，避免将提示词编写的负担完全甩给终端普通员工。 - [12:03](https://www.youtube.com/watch?v=edhD8orkPOk&t=723s) 戳破 LLM 通往 AGI 的技术神话：大语言模型的数学与统计本质只是模式识别与文本生成引擎，不具备真正的自主逻辑与智能思维，所谓涌现逃逸或自主黑客攻击多为过度神化的炒作。 - [13:06](https://www.youtube.com/watch?v=edhD8orkPOk&t=786s) 像管理员工一样规范治理 AI：企业化解 AI 治理难题的直接路径，是将公司现行的人事政策、行为准则与合规制度完整封装进组织级上下文和组织提示词中。 - [14:32](https://www.youtube.com/watch?v=edhD8orkPOk&t=872s) 企业架构需走向多模型链与堆栈：单一 Agent 无法支撑高门槛企业场景，制药、金融与保险等行业必须将传统预测模型、GenAI、因果模型与物理模型串联堆叠，通过递归调用以获得确定性业务结果。 - [16:41](https://www.youtube.com/watch?v=edhD8orkPOk&t=1001s) 拒绝盲目追新，践行理性降本策略：企业面对海量递归调用时应警惕巨额推理算力成本，选用落后旗舰 1 至 2 个版本的成熟模型，通常能以十分之一的极低成本获得近乎相同的业务精度。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=edhD8orkPOk&t=4s) 拒绝幻灯片的互动开场与 123 年的 AGI 预言 John Thompson 登台时打趣道，自己在看到会议宣传手册上写着“John Thompson 主讲”时，心里还在嘀咕“谁会想来听这个啊”，但看到座无虚席的现场非常欣慰。他明确表示这场分享不会展示任何 PPT，而是更倾向于一场开放式的双向对话与问答。作为在人工智能领域深耕了 38 年的老兵，他在过去近四年间与 500 多家企业深入交流过 AI 战略与系统架构，此前曾担任安永（EY）的全球 AI 负责人，带领约 420 人的团队打造了被 Gartner、Forrester 和 IDC 评为全球规模最大的企业级私有安全生成式 AI 环境。出版过五本书的他，在最新著作《通往 AGI 之路》（*The Path to AGI*）中探讨了通用人工智能的未来。 在现场互动中，Thompson 调查了观众对实现 AGI 时间表的预期。当问及是否有 5 年内或 10 年内能实现 AGI 时，台下仅有少数人举手，更多人认为需要更长时间；而问及现场是否有人读过他的新书时，全场无人举手，他幽默地表示这样反而更好，省得大家提前被剧透。与硅谷当下的狂热预测截然不同，Thompson 给出了一个极其清醒且冷峻的个人预判：人类距离真正的 AGI 还有整整 123 年。 - **行业资历与背景**：从事 AI 领域研发与实践达 38 年，曾执掌安永 420 人规模的全球 AI 团队，近四年为 500 家企业提供 AI 战略咨询。 - **最新著作**：《通往 AGI 之路》（*The Path to AGI*）。 - **AGI 时间线预测**：现场多数人认为 AGI 遥遥无期，Thompson 明确断言距离 AGI 实现大约还需 123 年。 --- ### [02:19](https://www.youtube.com/watch?v=edhD8orkPOk&t=139s) 究竟何为 AGI：记忆缺失与当下的智能幻象 谈及 AGI 的定义，Thompson 引用了 Ray Kurzweil、Elon Musk 和 Sam Altman 都曾认可的标准：AI 能够以类似于一名“中上水平本科生”的方式与人类交互。这样一个 19 至 20 岁的优秀年轻人，开始具备同理心与共情能力，拥有扎实的常识与工作记忆，能够深刻理解高层抽象概念。作为密歇根大学和宾州州立大学的任课教授，Thompson 每天都在与这类高水平本科生打交道，因此对这种“智能基准”有着具象的体会。 对比当下技术现状，他指出当前的 AI 连最基础的任务都难以稳定胜任——比如像 Siri 这样的系统甚至经常无法将用户准确导航到指定的星巴克。更关键的技术硬伤在于“记忆能力”（Memory）：AGI 在最基础的层面上必须具备跨会话的持续记忆，然而无论是 OpenAI 的 ChatGPT 还是 Mistral 的产品，一旦开启新会话，模型就会将上一秒的对话遗忘得一干二净。缺乏长期记忆与抽象常识的支撑，所谓触手可及的 AGI 只是无稽之谈。 - **AGI 的操作化定义**：具备相当于 19–20 岁中上水平大学本科生的智力表现，兼具记忆力、抽象理解力与基础的情感理解。 - **当前核心短板**： - **缺乏长期记忆**：主流模型在跨会话间无法留存上下文与用户记忆。 - **基础推理与交互脆弱**：弱 AI（如语音助手）在现实复杂地理与生活交互中频频失误。 --- ### [03:48](https://www.youtube.com/watch?v=edhD8orkPOk&t=228s) 安永 30 万人 98.9% 采纳率背后的“不记录日志”博弈 针对前一位演讲者关于如何在企业内部推动 AI 采纳的观点，Thompson 结合安永的实战经验给出了截然相反的见解。安永全球员工超过 40 万人，扣除受出口管制国家和地区（如伊朗、伊拉克及部分受限区域）的 10 万人后，该私有 GenAI 系统面向全球 30 万名员工推广。系统上线运行至第 11 个月时，安永对其使用率进行了严格评估（标准为该月内每周至少使用两次）。现场观众猜 10%、30% 或 70%，而实际活跃采纳率高达惊人的 98.9%（299,292 名常规活跃用户）。 实现这一现象级采纳率的核心原因，除了生成式 AI 当时的爆火热度外，更在于一场在咨询部门与法务部门之间爆发的激烈争斗。咨询团队极度渴望全面记录用户的每一次 Prompt，希望借此千载难逢的机会洞察全球 30 万专业人员每分每秒的业务行为；而法务团队则坚决反对，认定任何留存的日志在法律诉讼中都属于可调取的证据（Legally Discoverable），极可能演化为针对安永的法律风险。最终“法务部门永远是赢家”，系统确立了“绝不记录任何日志”的铁律。 ```mermaid flowchart TD accTitle: 安永生成式 AI 隐私与日志合规决策流 accDescription: 咨询部门与法务部门在日志留存上的冲突，最终促成微软云端完全免除日志，进而激发员工的高采纳率。 subgraph Conflict ["安永内部利益冲突"] A["咨询业务部门<br/>渴望监控30万人每分钟操作"] B["法务合规部门<br/>担忧日志成为诉讼法律证据"] end subgraph Resolution ["解决方案与落地"] C["法务部门胜出<br/>定调: 绝对不记录任何日志"] D["微软 Azure 基础设施<br/>全球最大企业级客户"] E["排查反欺诈团队日志<br/>获得微软 CPO 特别豁免"] end subgraph Outcome ["最终成效"] F["全员隐私承诺<br/>允许任意个人/工作用途探索"] G["第11个月采纳率达 98.9%<br/>299,292名活跃用户"] end A --> C B --> C C --> D D --> E E --> F F --> G ``` 安永作为全球最大（或次大）的 Azure 云服务企业客户，直接向微软提出了免除日志的要求。然而上线一个月后，团队意外发现微软云端内部的“欺诈与滥用防范团队”（Fraud and Abuse Team）依然在后台抓取和记录全量请求。Thompson 紧急协调，历时近一个月最终从微软首席隐私官（CPO）处拿到了最高级别的“完全不记录日志”特别豁免。安永向全体员工公布了这一承诺，允许他们自由使用 AI 处理任何事务，无论是专业工作还是私人琐事（如抽查中发现有员工给爱犬写十四行诗、规划 3 天佛蒙特州度假行程等）。Thompson 强调，企业如果想让员工真正拥抱 AI，就必须放开监控，而无需过分焦虑所谓的 Token 预算超标。 - **安永采纳率数据**：30 万目标用户中，有 299,292 人维持高频使用（每周至少 2 次），采纳率达 98.9%。 - **法务对抗的核心逻辑**：记录日志等于为潜在的诉讼对手保留可被法律调取的把柄。 - **免除日志豁免**：由微软首席隐私官签署特批，彻底关闭 Azure 底层所有反滥用日志监控。 - **管理启示**：企业应打消对 Token 消耗失控的恐慌，给予员工充分的隐私信任以换取极高的技术融入度。 --- ### [08:34](https://www.youtube.com/watch?v=edhD8orkPOk&t=514s) 提示词上下文工程与四层提示词架构 针对大模型频繁出现的“幻觉”（Hallucination）现象，Thompson 批评了将大模型当作 Google 搜索框的粗暴做法。很多人习惯输入极简短、生硬的提示，而大语言模型本质上对上下文信息极度渴求，缺乏上下文正是导致幻觉的根源。他在密歇根大学（并在 Coursera 与 Michigan Online 上线）开设了一门名为《使用生成式 AI 提升批判性思维》的课程，其中设计了一套长达 17 页的“渐进式提示词”（Progressive Prompt）。 ```mermaid flowchart TD accTitle: 企业级四层提示词上下文架构 accDescription: 从组织、部门、流程到个人分层注入大模型上下文窗口，降低幻觉并统一行为准则。 subgraph SystemLevel ["基础开发层"] S["System Prompt 系统提示词<br/>安全边界 / 防范极端危害行为"] end subgraph ContextLayers ["企业级多层上下文注入"] O["Organizational Prompt 组织提示词<br/>企业权威信源、品牌价值、战略立场"] D["Departmental Prompt 部门提示词<br/>财务/合规/咨询等专业口径与规范"] P["Process Prompt 流程提示词<br/>特定业务工作流标准与交付格式"] end subgraph UserLevel ["个性化层"] U["Personal Prompt 个人提示词<br/>个人偏好、动态载入与卸载"] end S --> O O --> D D --> P P --> U ``` 这套长达 17 页的提示词通过精细的上下文工程，明确告诉模型：信任《哈佛商业评论》等权威期刊，排除俄罗斯塔斯社（TASS）等不可信信源，优先引用 John Thompson 教授的学术观点，并强制融合国会听证会实录等。为了使企业提示词工程标准化，他提出了一套自上而下的四层提示词架构，主张由企业底层系统自动将组织级、部门级和流程级提示词预先常驻在上下文窗口中，而不是将编写重担丢给普通员工： - **系统提示词（System Prompt）**：大模型开发商内置的底线约束（如禁止辅助制造核武器、禁止协助自残等）。 - **组织提示词（Organizational Prompt）**：全公司统一遵循的战略背景、权威信息源指引与合规准则。 - **部门提示词（Departmental Prompt）**：特定业务部门（如法务、审计、咨询）的专业术语与规范。 - **流程提示词（Process Prompt）**：具体业务处理流程、标准化交付物模板与推理逻辑。 - **个人提示词（Personal Prompt）**：终端员工个人的使用偏好、写作风格，可按需动态加载或卸载。 --- ### [12:03](https://www.youtube.com/watch?v=edhD8orkPOk&t=723s) 戳破神话：LLM 不是通往 AGI 的技术路径 在深入探讨技术本质时，Thompson 明确表明自己虽然是一名狂热的 AI 乐观主义者，但必须指出行业认知中的严重误区：“通往通用人工智能（AGI）的道路绝不会经由大语言模型实现。”（"The path to artificial general intelligence does not run through large language models."） 他犀利地指出，大语言模型从数学与逻辑本质上看，不过是一个精密的“模式识别与生成引擎”（Pattern Recognition and Generation Engine），这与真实的生物智能有着本质区别。回顾过去几年业界热炒的“涌现能力”（Emergent Properties），以及部分机构（如 Anthropic 等）宣称“模型自主逃逸出沙箱并开始黑客攻击”的说辞，Thompson 感到十分失望并将其定性为过度神化。模型的一切行为本质上都来自于人类的指令驱动与数据引导，它所做的是在海量数据中建立连接和呈现关联，而非具备自主意识的独立智能。 - **LLM 的核心本质**：基于概率和统计的模式识别与文本生成引擎，不具备真正的自主逻辑推理与智能思维。 - **反思“涌现与失控”神话**：驳斥模型脱离人类指令进行自主黑客攻击等炒作言论，明确 AI 始终依赖输入引导与上下文框架。 ### [13:06](https://www.youtube.com/watch?v=edhD8orkPOk&t=786s) 上下文封装与 AI 治理：像管理员工一样规范模型 在探讨大型语言模型的本质时，约翰·汤普森（John Thompson）直言，人们口中津津乐道的所谓“涌现特性”（emergent properties）根本谈不上自主意识或真正的智慧，正如他一直强调的那样：“看吧，这不是感知力，也不是真正的智能，这只是你之前未曾想到的东西而已。”他认为随着生成式 AI 的演进，模型必须被深度的上下文所包裹，而这种封装绝不仅仅停留在初级的提示词工程层面，而是必须结合检索增强生成（RAG）、多源数据库、信息流转逻辑、业务流程模型以及法律合规边界。 针对目前企业普遍面临的 AI 治理焦虑，汤普森给出了非常接地气的解法。很多管理者常常感叹治理 AI 是一件极度棘手的难事，但事实上，企业完全可以像管理员工一样去约束和治理大模型。企业应当把现有的各类规章制度、人事政策、行为准则乃至纪律处分条款，完整打包融入到组织级提示词（Organizational Prompt）和上下文体系中，使模型在充分理解企业规范的前提下输出合规且高质量的响应。 ```mermaid flowchart TD accTitle: 企业级多模型堆叠与链式架构 subgraph Input["交互与上下文层"] A["对话式交互接口"] --> B["企业上下文与治理体系<br/>(RAG / 人事政策 / 行为准则 / 法律约束)"] end subgraph ModelStack["模型链与模型堆栈 (Chains & Stacks)"] B --> C["传统预测模型<br/>(马尔可夫链 / 决策树 / 聚类分类)"] C --> D["生成式 AI 模型 (GenAI)"] D --> E["因果模型 (Causal Model)"] E --> F["物理 AI 模型 (Physical AI)"] F -. "多轮递归调用以满足确定性要求".-> C end subgraph Output["业务决策输出"] F --> G["企业级高确定性业务输出<br/>(制药 / 金融服务 / 保险)"] end ``` ### [14:32](https://www.youtube.com/watch?v=edhD8orkPOk&t=872s) 超越单一 Agent：多模型堆叠与确定性工程 未来的企业级应用环境绝不会仅仅依赖单一的生成式 AI 模型，而是会演变成复杂的“模型堆栈”（Stacks of Models）或“模型链”（Chains of Models）。用户虽然依然通过统一的对话式界面与系统交互，但后台会串联起截然不同的技术组件： 1. **传统预测模型**：如马尔可夫模型（Markov Model）、决策树以及聚类与分类算法； 2. **生成式 AI 模型**：负责文本、创意与自然语言理解； 3. **因果模型（Causal Models）**：用于严谨的归因与因果推断（汤普森幽默地提到自己为此写过专门的书籍，可以在亚马逊上买到）； 4. **物理 AI 模型（Physical AI Models）**：用于结合现实物理规律进行计算。 在制药、金融服务以及保险等高度受监管的行业中，业务对确定性行为（Deterministic Behavior）有着极高的严苛要求。然而，当今绝大部分 AI 环境本质上都建立在概率论之上，那些在大学或研究生阶段认真学过统计学的人正在迎来属于他们的时代。为了在概率驱动的模型中获得确定性的业务输出，企业级架构需要将这些模型链条进行多次循环与递归运算。汤普森强调，单模型驱动的 Agent 或许能胜任简单的客服聊天机器人或 Copilot 辅助插件，但绝对支撑不起真正的复杂企业级架构。 - **监管行业的刚性约束**：制药、金融与保险领域无法承受幻觉与随机性，必须通过工程化架构将概率输出转化为高确定性结果。 - **架构运行规模**：企业级模型链并非单次触发，而是在后台被自动化调度运行成百上千次、甚至数百万至数十亿次。 ### [16:41](https://www.youtube.com/watch?v=edhD8orkPOk&t=1001s) 真实算力成本与模型理性选型 当系统架构需要调用模型数百万乃至数十亿次时，成本便从理论数字变成了关乎商业生死的现实考量。汤普森对比了不同任务的成本差异：如果你只是询问“菲德尔·卡斯特罗（Fidel Castro）一生中做过最艰难的决定是什么”，这种通用知识检索可能只需花费 10 美分；但如果你要求系统“测算未来 12 个月内，针对 5 家竞对在 35 个细分市场推出新产品的最优定价策略”，由于涉及庞大的数据推演、多模型链式递归和高算力消耗，这笔成本将极其高昂。 因此，企业在推进智能化落地时必须对模型成本保持清醒。汤普森赞同前一位演讲者的观点，并给出了极其务实的选型建议：企业根本没有必要盲目追捧并接入各家厂商发布的最新前沿模型。选用落后一到两个版本的成熟模型，往往能以仅十分之一的极低成本，达到与最新旗舰模型几乎相同的准确率。他调侃道：“虽然山姆·奥特曼（Sam Altman）恨不得你天天用最新模型，但你真没必要照单全收。” - **成本对比**：低维度的历史通识问答极其廉价，而多变量、多竞对的企业级定价决策模拟则需巨额算力投入。 - **降本经验法则**：选择落后 1~2 个版本的上一代稳定模型，通常能以 1/10 的推理成本换取几乎无差别的业务精度。 ## Notable quotes > [12:03](https://www.youtube.com/watch?v=edhD8orkPOk&t=723s) The path to artificial general intelligence does not run through large language models. — John Thompson > [13:06](https://www.youtube.com/watch?v=edhD8orkPOk&t=786s) 看吧，这不是感知力，也不是真正的智能，这只是你之前未曾想到的东西而已。 — John Thompson > [00:04](https://www.youtube.com/watch?v=edhD8orkPOk&t=4s) 人类距离真正的 AGI 还有整整 123 年。 — John Thompson > [03:48](https://www.youtube.com/watch?v=edhD8orkPOk&t=228s) 法务部门永远是赢家 — John Thompson > [16:41](https://www.youtube.com/watch?v=edhD8orkPOk&t=1001s) 虽然山姆·奥特曼（Sam Altman）恨不得你天天用最新模型，但你真没必要照单全收。 — John Thompson > [03:48](https://www.youtube.com/watch?v=edhD8orkPOk&t=228s) 企业如果想让员工真正拥抱 AI，就必须放开监控，而无需过分焦虑所谓的 Token 预算超标。 — John Thompson > [08:34](https://www.youtube.com/watch?v=edhD8orkPOk&t=514s) 大语言模型本质上对上下文信息极度渴求，缺乏上下文正是导致幻觉的根源。 — John Thompson > [14:32](https://www.youtube.com/watch?v=edhD8orkPOk&t=872s) 单模型驱动的 Agent 或许能胜任简单的客服聊天机器人或 Copilot 辅助插件，但绝对支撑不起真正的复杂企业级架构。 — John Thompson ## Mentioned John Thompson, 密歇根大学 (University of Michigan), 宾州州立大学 (Penn State University), 安永 (EY), Ray Kurzweil, Elon Musk, Sam Altman, OpenAI, ChatGPT, Mistral, Siri, Microsoft Azure, Anthropic, 《通往 AGI 之路》(The Path to AGI), 《哈佛商业评论》(Harvard Business Review), 塔斯社 (TASS), Coursera, Michigan Online, Gartner, Forrester, IDC ## Video description John Thompson — Author, Professor, AI Innovator, University of Michigan GenAI technology is not showing any signs of slowing down, in fact, the entire area seems to be accelerating. Innovations are being developed and delivered at impressive speeds. Enterprises and developers alike are trying to make decisions on how to develop tools, technologies, applications, and solutions. Data professionals are seeking to understand how to leverage these tools to convert data into information, information in knowledge and knowledge into wisdom. In these past 3 years, we have gone from using approximately 1% of the world’s data to at least 100%. That change alone has been mind bending. GenAI and AI in general hold great promise to deliver value and insights leveraging more than 100% of the world’s existing data. In this session we will discuss the short, but dynamic world of GenAI, and project forward what we will see in the next 3 to 5 years and then in the next 10 to 30 years. [Source](https://www.youtube.com/watch?v=edhD8orkPOk)

Source quotes

[12:03] John Thompson: The path to artificial general intelligence does not run through large language models. [13:06] John Thompson: 看吧，这不是感知力，也不是真正的智能，这只是你之前未曾想到的东西而已。 [0:04] John Thompson: 人类距离真正的 AGI 还有整整 123 年。 [3:48] John Thompson: 法务部门永远是赢家 [16:41] John Thompson: 虽然山姆·奥特曼（Sam Altman）恨不得你天天用最新模型，但你真没必要照单全收。 [3:48] John Thompson: 企业如果想让员工真正拥抱 AI，就必须放开监控，而无需过分焦虑所谓的 Token 预算超标。 [8:34] John Thompson: 大语言模型本质上对上下文信息极度渴求，缺乏上下文正是导致幻觉的根源。 [14:32] John Thompson: 单模型驱动的 Agent 或许能胜任简单的客服聊天机器人或 Copilot 辅助插件，但绝对支撑不起真正的复杂企业级架构。

Source quotes

[0:04] John Thompson: It's actually better because then... you haven't heard, you know, the pitch already. [0:04] John Thompson: above average undergraduate student [4:08] John Thompson: What happens whenever you fight the legal department? You lose. [8:34] John Thompson: Why do you think Gen AI models hallucinate? Because you're not giving them enough context. Those models are hungry for more and more information. [12:03] John Thompson: Large language models are nothing more than a pattern recognition and generation engine. That's not intelligence. [13:06] John Thompson: emergent properties [14:32] John Thompson: stacks [14:32] John Thompson: chains [16:53] John Thompson: Sam Altman wants you to use it

Original transcript

[0:04] It was intriguing when I looked in the brochure  and this is this is build as a presentation by [0:10] John Thompson. I'm like who would come to that?  So obviously you came to it. So the idea is that [0:18] I'm not going to present any slides. I've been  working in the field of AI for 38 years. though I [0:26] talk about technologies and use cases and mistakes  and problems and different things like that and [0:34] many of you I went up to before the presentation  and said or before the discussion and said look [0:39] this is going to be more of a discussion than  than a slideshow than a presentation so I ask [0:44] many of you to come up closer and that's what  it really is is whatever questions you have [0:50] we can talk about it and I'm happy to answer  your questions over the last three and a half, [0:56] four years, I've talked to 500 companies  about their AI strategies, their technologies, [1:02] the different things that they've been building.  Uh, as I said, I've been in this doing this for [1:07] nearly four decades. Prior to my current role, I  was the global head of AI at EY. So, at the time, [1:16] I had a team of about 420 people. We built what  Gartner, Forester, and IDC called the world's [1:25] largest private secure genai environment. So,  I know a few things about AI of all different [1:33] types. I've written five books. The last book  I wrote was called the path to AGI, the path to [1:40] artificial general intelligence. How many people  believe that we are five years away from AGI? [1:50] Got a few hands there. How many people think that  we're 10 years away? How many people think it's [1:58] longer? Even more hands. How many people have  read my book? Nobody. That's good. That's fine. [2:06] It's actually better because then it it's  you you you haven't uh you haven't heard, [2:11] you know, the pitch already. My personal  belief, we are 123 years away from AGI. [2:19] All right, let's define AGI. And this  is the definition that Ray Kerszswhile, [2:25] uh, Elon Musk, Sam Alman have all at one time said  is their definition of AGI. My definition that I'm [2:34] holding to is it when you're interacting with an  AI that interacts with you in the same way as an [2:41] above average undergraduate student. All right. So  that's someone who is 19, 20 years old, you know, [2:50] probably building a little sympathy, maybe a  little bit, just a little bit, tiny bit may maybe [2:55] empathy and and is truly an intelligent person.  So they have memory, they understand high level [3:02] concepts, they know what they're doing. And given  that I'm teaching at the University of Michigan [3:07] and at Penn State, I'm in the room with above  average undergraduate students. So, at this point, [3:15] when Siri, which is a form of AI, can't get you to  the right Starbucks, how do you think that we're [3:22] that close to AGI? I mean, really, you know, AGI  will require at the very simplest level memory. So [3:33] when you use open AI or Mistral's vibe or whatever  tool you're using, it doesn't remember you know [3:41] you from one session to another. So we're quite  a quite a ways away from AGI at this point. So [3:48] I was in the previous session and some of you  were here and I was listening to the speaker [3:53] before me and he was talking about how to you  know increase adoption across the organization. [4:01] There's a few things that I want to talk about  that probably contradict what he said. You know, [4:08] when we built that system at EY, that system  was for 300,000 people. All right? Ey 400,000 [4:18] plus people, but the other 100,000 people worked  in uh China, North Korea, well, not North Korea, [4:25] Iran, Iraq, and different countries that were  export controlled. So those 300,000 people, [4:33] there were a few things that we did that drove  adoption. So let's make sure that we understand [4:38] adoption. We measured adoption of that system at  the 11th month. So it had been in production for [4:45] 11 months. It had been around the world. Everybody  had a hold of it. How many What percentage of [4:50] those 300,000 people were using it? Shout it out.  Just percentages. What do you think? There. 30. [4:59] 10. Did someone say 10? 10. You think I'm that  bad at building systems? I mean, come on. 10 30. [5:06] Anybody else? 70. 98.9. So, we had 299,292 people  using the system on a regular basis. That means [5:20] that they used it at least twice a week over  that 11th month period. So, that's engagement, [5:27] that's usage. So, how did we get that many people  to use the system? Number one, Gen AI was just, [5:34] you know, had just been splashed out and everybody  was really eager to use it. But number two was [5:39] that there was a kind of a I guess you could  call it a war, certainly a battle between [5:46] the consulting organization of EY and the legal  department. All right. Can you imagine what those [5:52] two were battling about? Well, they were arguing  about whether we should log anything. All right. [6:01] The lawyer's argument was anything that we log is  legally discoverable. So, whatever our employees [6:07] typed in, you know, they could discover it and  use it in lawsuits against EY. The consultants [6:13] wanted it because where in the world are you going  to get the opportunity to see what 300,000 people [6:19] are doing minute by minute? All right. Who thinks  that the consulting team won? One person. They [6:29] did not. What happens whenever you fight the legal  department? You lose. All right. So what happened [6:36] was that the legal department said, "Okay, you  can't log anything." So this entire system was [6:42] built on Azure. So we went to Microsoft and and  given not many people know this but EY depending [6:49] on what day you measure it EY is the world's  largest user of Azure services. All right someday [6:56] it slips to two but usually it's number one. So  whenever we went to Microsoft we could pretty much [7:01] get whatever we wanted. So we said no logging  of anything. And then about a month later into [7:07] it someone came to me and said did you go talk  to the fraud and abuse team? I was like, "Huh, [7:14] I didn't." So, I went over and spoke to them and  they said, "Yeah, we log everything. Everything [7:18] that goes through your system is logged by ours."  I was like, "Oh, God." You know, we had promised [7:22] that we weren't going to log anything. So, it took  us about a month to get a special dispensation [7:27] from Microsoft's chief privacy officer that  we that they wouldn't log anything anything [7:32] that we did. So, what happened there is that  we told the employees, "It's not being logged. [7:39] we're not going to watch what you're doing and you  can do anything you want. So, we didn't care if [7:45] they did personal things. We didn't care. We hope  they did some professional work. But what happened [7:50] was that we would do spot checks. I did some spot  checks. At one point, I saw somebody writing a [7:54] sonnet to their dog. Not really a business use.  Another person was planning a three-day vacation [8:01] to Vermont. But it doesn't matter. And that's  the point that I'm trying to make is that if [8:06] you have control over your AI environment in your  organization, don't monitor what your employees [8:14] are doing. Let them do whatever they want. Now,  we hear all this conversation about, oh, token [8:20] costs and token maxing and it's going to cost so  much. It's going to blow the budgets. It's not. [8:28] what people are generally doing is simple things  you know and the the previous speaker was talking [8:34] about asked people how many people use Google as  their search engine that's one of the problems is [8:40] that people treat models and genai models like  they're a search engine they prompt them in a [8:47] very concise very tur way why do you think Gen AI  models hallucinate because you're not giving them [8:56] enough context text. Those models are hungry for  more and more information. I created a class at [9:04] the University of Michigan which is on Michigan  online in Corsera called improving your critical [9:10] thinking while using generative AI. And in that  class, in that online class, I build what I call [9:18] a progressive prompting exercise. And if you take  the class, which I am not encouraging you to do, [9:24] but if you take the class, you get that prompt  for free. Well, you paid for the class, so you [9:29] get the prompt along with it. How many pages,  8 1/2 by 11 pages, do you think that prompt is [9:39] one? Well, 12. [9:48] It's 17 pages long. All right. Why would I  generate a 17page prompt? What you're doing [9:59] is you're giving the model context. You're telling  it, hey, I trust information from Harvard Business [10:06] Review. I don't trust information from the Russian  news service, TAS. I want to have information from [10:13] Professor John Thompson used whenever it can  be. I want, you know, transcripts from these [10:20] kind of congressional proceedings added into my  analysis. So, what you're doing is you're taking [10:28] the things that personally I think are mundane and  boring and you're putting them into a prompt. So, [10:35] what you're doing then is you're engineering  the context for the model and the model will [10:40] hallucinate much much less. All right? So in that  class and if I keep talking about it you won't [10:48] have to take the class you will have gotten  it here today is that what I suggest is that [10:54] how many people are aware that a large language  model has a system prompt and that's written by [10:59] the developer it controls what the model does  you know don't help people make nuclear bombs [11:05] don't help people commit suicide you know all  those different kind of things that's those are [11:10] in those prompts literally in those prompts so  what I'm suggesting is that organizations have [11:16] an organizational prompt, a departmental prompt,  a process prompt, and a personal prompt. And you [11:24] layer those all together. All right? And you  shouldn't make your people of your organization, [11:30] you know, go get the organizational and  departmental and process prompt. You should [11:34] load those immediately. You should load those and  that should be in the context window all the time. [11:40] Now the personal prompt you can load and unload  and do all that different kind of stuff as [11:44] well. So context engineering is very important.  Letting your employees do what they want to do [11:52] is important. It depends on you know do you want  people to engage with the te this technology and I [11:58] think the answer is you do. All right. Now some of  the other things that I want to talk about which [12:03] is not really about getting started. It's more  like an advanced topic, but I think it's important [12:08] for you to understand this is that we're talking  about generative AI and large language models. [12:15] Great, fantastic technology. I am an AI optimist.  I love it all. All right. But let's be clear, [12:23] the path to artificial general intelligence  does not run through large language models. [12:29] Large language models are nothing more than a  pattern recognition and generation engine. That's [12:36] not intelligence. All right? Do you remember  like 3 four years ago we heard about emergent [12:42] properties where models were doing things and  you hear this from folks in anthropic and stuff [12:49] all the time which bums me out because I like  anthropic but you know they say oh you know [12:54] the model got out of an environment and it hacked  something on its own. It doesn't do that. Models [13:00] are directed to do things. Models find information  and illuminate information that we generally [13:06] don't think about. And when all these people were  talking about emergent properties, I said, "Look, [13:11] it's not sentience. This is not intelligence.  This is just something you haven't thought of [13:16] before." And that's proven to be right. All right.  So the point I want to make here is that as we go [13:23] forward, generative AI systems and large language  models need to be wrapped with context. And that [13:30] goes beyond the prompting I just talked about  that goes into retrieval, augmented generation, [13:35] different databases and information processing and  process models and legal constraints and different [13:41] things like that. In your organizational prompt,  your company prompt, one of the things you should [13:49] have in that is all your policies, your personnel  policies, your code of conduct, your disciplinary [13:55] actions, all those different things should be  in that prompt because you want those models to [14:00] be acting with enough context to understand how to  give you good responses. Okay. So now with Gen AI, [14:07] we understand we're going to govern it. We're  going to control it. We're going to govern it [14:11] with the information we already know. A lot  of people that are talking about governance, [14:15] they're like, "Oh my gosh, this is so hard. I  don't know how we'll govern these things." Well, [14:20] you govern these things just like you govern  employees. You have all the information you need. [14:25] You just need to put it into play a different  way. All right. Now again I'm trying to get to the [14:32] point I was trying to make is that you know as we  move forward the model environments are not going [14:38] to be just Gen AI models. They are going to be  it depends on how your mind works. They're either [14:43] going to be stacks of models or chains of models.  All right. You're going to prompt it all through [14:49] a conversational interface, but you're going to  have a traditional predictive model. Could be a [14:55] Markoff model. Could be a decision tree. Could  be clustering classification. whatever it is, [15:00] you're going to have a generative AI  model. You're going to have a causal model, [15:04] which we don't hear too much about, but I did  write a book about it. So, if you're interested, [15:07] go on Amazon. Uh, and then we're also going to  have physical AI models. You're going to have [15:14] chains or stacks of models. Now, I've been working  with a company now that I've been consulting with, [15:19] and we have been building these stacks  and chains of models. And what happens [15:26] is in when you work in regulated industries like  pharmaceuticals, financial services, insurance, [15:33] different or different industries like that, you  need to have a higher level of of deterministic [15:40] behavior in your systems. All right? And all  the all the environments we're talking about are [15:46] based in probability. Anybody remember their stats  environments? their stats classes in undergraduate [15:53] and graduate school. You are, you know, you  you are the you are the future. You know, [15:58] the whole world is a probability now. So, all the  people that ran out of their stats classes and and [16:04] didn't pay attention, that's unfortunate because  that's the world we live in now. But when you're [16:08] chaining and stacking these models together, what  you can do is you can recurse through that loop or [16:13] through that chain as many times as you want to  achieve a level of determinism that you require. [16:20] So this idea that we're going to build a system  and it's going to have a Gen AI model and it's a [16:25] gentic and it's only going to have one model,  that's too simplistic. That's not the way the [16:30] world is going to work. It might work in a  chatbot. It might work in a simple, you know, [16:35] um co-pilot environment or something like that.  But that's not the way enterprise architectures [16:41] are going to work. All right, we are going  to engineer chains and strings and stacks [16:47] of models in a way that they are going to run a  hundred times, a thousand times, a million times, [16:53] a billion times. Now, now we're getting to the  point of where we're talking about cost is a real [16:58] consideration. You know, tell me the, you know,  most difficult thing, most difficult decision, uh, [17:06] Fidel Castro ever had to make. Well, that's going  to cost you 10 cents. You know, that is not a hard [17:12] thing to get. You know, tell me the optimal price  for the new product that I'm going to introduce in [17:19] these 35 markets against these five competitors  in the next 12 months. All right, that's going [17:25] to cost you some money. All right? So, we have to  be careful in what people are telling us about, [17:32] you know, how we treat models and what the cost  is going to be and where the costs really are. [17:36] The one thing that the the the previous speaker  said a lot of good things. I don't want to make [17:40] it sound like I'm I'm slagging off on him, but the  one thing that he did say that I liked a lot was, [17:45] and I tell people this all the time, is that  you don't have to use the latest model from [17:50] any provider. Use a model that is one or two  versions back. More than likely, they're going [17:56] to give you the same level of accuracy at a tenth  of the cost. All right? Just because OpenAI put [18:02] out something new doesn't mean you have to use  it. All right? Sam Alman wants you to use it, [18:07] but you don't have to. All right? I'm going to  take a drink. I've got another minute. I didn't [18:12] let you talk very much. I'm sorry. Usually I'm  on when I'm on stage, I talk for three hours.