Cisco's Patel Rejects AI Job-Cut Warnings, Defends Entry-Level Roles

The Catalyst for Driving an Agentic AI Revolution - Fireside Chat With Cisco, CNN — Ai4

Ai4

Eliminating junior talent undermines enterprise innovation by removing the unencumbered perspectives needed to force institutional unlearning, according to Cisco Chief Product Officer Jeetu Patel.

AI-assisted summary of The Catalyst for Driving an Agentic AI Revolution - Fireside Chat With Cisco, CNN — original source. Check the original for full context.

Source published: 2025-08-28

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 28, 2025, this video captures a fireside chat between Jeetu Patel, President and Chief Product Officer at Cisco, and CNN Senior Reporter Matt Egan. According to the page description, the session explores the transformative landscape of agentic AI, focusing on enterprise adoption, human collaboration, and Cisco's strategies for securing autonomous AI ecosystems at machine scale.

## TL;DR Jeetu Patel frames artificial intelligence as a fundamental civilizational shift from resource scarcity to abundance, dismissing alarmist human obsolescence claims while highlighting the underhyped emergence of original, superhuman scientific insights. As agentic AI advances from short-duration tasks to multi-week autonomous workflows, it faces severe bottlenecks in electrical power, compute infrastructure, and human trust. Patel strongly rejects the notion that AI will eliminate entry-level positions, arguing junior workers are vital for institutional unlearning and fresh problem framing. Ultimately, he stresses that enterprise adoption hinges on solving novel security challenges, calling for runtime guardrails, model validation, and fine-grained agent identity controls to counteract sophisticated adversarial threats.

## Atmosphere & style The fireside chat carries an intellectually rigorous, pragmatic, and conversational tone that cuts cleanly through standard Silicon Valley boosterism. Matt Egan acts as an incisive journalistic foil, raising pressing anxieties around technological displacement and economic shocks, while Patel counters with calm executive realism and vivid personal anecdotes, such as his daughter tapping a physical newspaper. The exchange feels grounded and candid, balancing high-stakes warnings about adversarial cybersecurity risks and power grid deficits with a deeply humanistic, optimistic perspective on adaptability.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=4s) Artificial intelligence marks a profound societal transition from historical resource scarcity to cognitive abundance, unlocking capacity to solve previously intractable problems. - [00:04](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=4s) Fears of complete human obsolescence within five years are overhyped, whereas the impending ability of AI to generate net-new original insights beyond the existing human knowledge corpus is deeply underhyped. - [02:59](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=179s) Agentic AI is moving from conversational chatbots toward multi-step autonomous agents, expanding autonomous execution horizons from 20-minute tasks to multi-day and multi-month workflows. - [02:59](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=179s) Pairing every individual with 10 to 100 autonomous agents could amplify global throughput to 800 billion human-equivalents, creating severe energy, compute, and infrastructure bottlenecks. - [02:59](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=179s) A pervasive trust deficit remains a primary adoption barrier, as current AI capabilities outpace what humans feel secure delegating without strict safety and reliability guarantees. - [06:35](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=395s) As intelligence commoditizes, professional value shifts away from raw knowledge accumulation toward the ability to frame nuanced, high-quality questions and challenges. - [06:35](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=395s) Metaprompting—instructing an AI model to architect the optimal prompt for a given task—delivers a step-function improvement in downstream output quality. - [06:35](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=395s) Adoption cycles are accelerating rapidly because AI systems understand natural human language rather than forcing humans to master machine programming syntax. - [09:36](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=576s) Eliminating entry-level white-collar roles would be a critical mistake; junior employees bring fresh perspectives and prompt organizational unlearning of outdated assumptions. - [09:36](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=576s) Accumulated professional experience can become a liability when leaders dismiss viable innovations based on past failures whose underlying technical conditions have fundamentally changed. - [15:02](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=902s) Workforce disruption requires tech and public sector partnerships to fund comprehensive reskilling, recognizing that displaced workers cannot simply transition into prompt engineers overnight. - [15:02](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=902s) Adversaries possess symmetric access to advanced AI models, enabling hyper-bespoke, context-aware social engineering and cyber threats at machine scale. - [17:06](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=1026s) Securing agentic ecosystems requires moving beyond perimeter defense to securing AI models themselves, implementing runtime guardrails, and enforcing fine-grained agent identity permissions.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=4s) Shifting from Scarcity to Abundance and the AI Overhype vs. Underhype

Matt Egan opens the fireside chat by recalling Jeetu Patel's assertion that artificial intelligence represents a shift larger than anything seen in our lifetimes. When asked what fundamental assumptions about society and the world must be reconsidered, Patel explains that human civilization has always operated under an organizing philosophy of resource scarcity. For generations, scarcity served as "the mother of invention," requiring humans to constantly balance finite resources. Society is now crossing a threshold into an era where true abundance is possible, unlocking the capacity to solve previously intractable problems—from curing cancer to engineering entirely new materials.

```mermaid flowchart TD accTitle: Evolution of AI Paradigms subgraph S1["Phase 1: Knowledge Aggregation"] A["Chatbots & LLMs"] --> B["Synthesize Human Corpus"] end subgraph S2["Phase 2: Autonomous Agents"] C["Deep Research (~20 Min)"] --> D["Multi-Day/Month Autonomy"] D --> E["Computer Use & Autonomous Tasks"] end subgraph S3["Phase 3: Original Insights"] F["Superhuman Problem Solving"] --> G["Novel Discoveries Beyond Human Corpus"] end S1 --> S2 --> S3 ```

Patel draws a sharp distinction between what is currently overhyped versus what remains deeply underhyped about the technology. What is overhyped is the alarmist notion that human beings will be made obsolete within five years, left with nothing to do while sitting idly on beaches with their creative juices extinguished. While every job will undergo significant reconfiguration, humans have consistently adapted and found new avenues to deliver value.

Conversely, the most underhyped dimension of AI is its impending capacity to produce genuine "original insights." Current large language models function primarily as sophisticated aggregation engines, scouring the existing human corpus of knowledge to synthesize patterns. The next frontier will feature models generating novel insights that have never existed within human records, fundamentally altering how science, material engineering, and broader society function.

- Core philosophical shift: Moving from historical resource scarcity to computational and cognitive abundance. - Overhyped narrative: Complete human obsolescence and displacement within a 5-year timeframe. - Underhyped breakthrough: Generation of net-new original insights beyond the existing human knowledge corpus.

---

[02:59](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=179s) The Phases of Agentic AI, Autonomous Horizons, and Infrastructure Constraints

Examining the trajectory of agentic AI, Patel contextualizes the industry as being squarely in its second phase. The first phase, spanning the prior two and a half years, centered on intelligent chatbots that answered queries with a sense of magic as users learned prompting dexterity. The second phase introduces autonomous agents capable of conducting end-to-end tasks and multi-step jobs on behalf of human operators. Patel points to tools like Deep Research as early proof points—agents that can currently run autonomously for roughly 20 minutes while a user gets coffee. As that autonomous horizon expands from 20 minutes to two hours, two days, two weeks, and eventually two months, the complexity of solvable problems will grow exponentially.

```mermaid flowchart LR accTitle: Agentic AI Bottlenecks and Enablers subgraph Scaling["Agentic Scaling"] H["1 Human Operator"] --> A["10 to 100 AI Agents"] A --> T["Throughput of 80B - 800B People"] end subgraph Bottlenecks["Critical Constraints"] E["Energy Shortage"] --> DC["Data Centers Moving to Power"] C["Compute & Network"] --> CU["Surge from Computer-Use Agents"] TD["Trust Deficit"] --> HU["Human Gatekeeping"] end Scaling -.-> Bottlenecks ```

This exponential agentic expansion brings severe physical and systemic bottlenecks. If every human is eventually paired with 10 to 100 autonomous agents operating round the clock, a global population of 8 billion could exhibit the economic and computational throughput of 80 billion to 800 billion people. The world currently lacks the power and compute capacity to sustain such demand, forcing data center construction to migrate directly to wherever power exists rather than bringing power to data centers. Furthermore, emerging modalities like "computer use"—where an agent interacts with operating systems and laptop interfaces just like a human—will place unprecedented pressure on network infrastructure and compute clusters.

Compounding these infrastructure limits is a pervasive "trust deficit." Patel emphasizes that the technical capacity of modern AI already outstrips what humans are willing to delegate to it. Until autonomous systems consistently earn user trust through verified safety, reliability, and security, humans will continue to restrict agent permissions and throttle adoption.

- Phase progression: Transitioning from transactional chatbot queries to prolonged autonomous task execution. - Autonomy scaling vector: Expanding continuous execution from 20 minutes to hours, days, and months. - Virtual labor multiplier: A 10x to 100x agent-to-human ratio multiplying global throughput up to 800 billion human-equivalents. - Primary adoption barriers: Severe electrical power/compute deficits and an acute human trust deficit.

---

[06:35](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=395s) The Art of Prompting, Metaprompting, and Compressed Adoption Cycles

Transitioning to user interactions, Patel explains how the fundamental nature of skill is evolving. Historically, accumulating domain knowledge was the ultimate differentiator for professional success. With intelligence and information becoming abundant, free, and commoditized, the primary differentiator becomes the ability to formulate nuanced, high-quality questions. Patel outlines a counterintuitive dynamic in problem-solving: tackling harder problems actually yields a higher probability of success because difficult challenges inherently attract top-tier talent, and business remains a collaborative team sport.

```mermaid flowchart TD accTitle: Prompting vs Metaprompting Flow subgraph Direct["Standard Prompting"] U1["User Intent"] --> P1["Direct Prompt"] --> O1["Standard Output"] end subgraph Meta["Metaprompting Architecture"] U2["User Intent"] --> MP["Prompt to Generate Optimal Prompt"] MP --> OP["Optimized System Metaprompt"] OP --> O2["High-Fidelity Output"] end ```

Patel highlights "metaprompting" as a transformative technique gaining prominence with advanced models. Rather than asking an AI system directly for an end asset—such as generating an image or complex analysis—the user instructs the model to construct the optimal prompt required to produce that asset, and then runs the refined prompt. This meta-layer produces an asymmetric, step-function improvement in the final output quality.

Engaging the audience with a show of hands on who uses metaprompting, Patel observes how rapidly societal ingestion cycles are compressing. Unlike previous technology cycles where humans had to learn the syntax and programming languages of machines, modern machines are learning natural human language, accelerating the pace of mainstream adoption.

- Value shift: Moving from knowledge accumulation to nuance and precision in problem framing. - Counterintuitive law of difficulty: Harder organizational problems attract superior talent, raising the net probability of success. - Metaprompting mechanics: Prompting the model to architect the query itself to unlock superior downstream fidelity. - Adoption acceleration: Natural language understanding drastically reduces human learning curves.

---

[09:36](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=576s) Challenging the Demise of Entry-Level Jobs and the Power of Unlearning

Egan references warnings from Anthropic CEO Dario Amodei, who suggested AI could wipe out half of all entry-level white-collar jobs and push unemployment to 10–20% within one to five years. While Patel notes that Cisco is an investor in Anthropic and expresses deep respect for Amodei, he strongly dissents from the thesis that entry-level positions should or will be eliminated, calling the blanket eradication of early-career roles "the stupidest thing a company can do."

To illustrate why, Patel compares how different age groups approach technology. A 28-year-old often utilizes AI in a transactional, task-oriented manner to automate existing workflows, whereas a 20-year-old interacts with AI as a persistent brainstorming partner and cognitive sidekick. Sharing a personal anecdote about his 14-year-old daughter—who grew up scrolling on an iPad from toddlerhood and later tapped a physical newspaper expecting the photograph to zoom in—Patel stresses that younger generations possess a fundamentally different baseline orientation toward modern tools.

```mermaid flowchart LR accTitle: Experience Liability vs Early Career Perspective subgraph Tenured["Tenured Mindset"] EXP["Past Experience"] --> BIAS["'Tried 5 years ago, failed'"] BIAS --> STAG["Institutional Inertia"] end subgraph Early["Early-Career Mindset"] FRESH["Unencumbered Perspective"] --> UNL["Forces Unlearning"] UNL --> INN["Novel Problem Solving"] end Tenured <-->|Symbiotic Coupling| Early ```

Patel argues that accumulated experience can frequently become a "massive liability." Seasoned professionals often dismiss viable ideas because an attempt failed years earlier, ignoring that thousands of underlying technical assumptions have since changed. Early-career workers inject untainted perspectives that force organizations to unlearn outdated mental models. Patel spends considerable time with student interns and junior employees for reverse mentorship, noting that true breakthrough innovation occurs only when fresh perspectives are coupled with seasoned execution experience.

- Rejection of early-career elimination: Discarding junior talent removes essential institutional adaptability and unlearning. - Generational usage disparity: Older cohorts use AI transactionally; younger digital natives use AI as an omnipresent collaborative sidekick. - Experience as a liability: Historical memory often prevents tenured leaders from revisiting previously failed concepts whose underlying conditions have changed. - Core value proposition: AI's true promise lies in tackling brand-new categories of problems rather than mere task-level productivity gains.

---

[15:02](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=902s) Societal Growing Pains, Reskilling, and Emerging Adversarial Threats

Egan brings up prior comments from Stanford professor Fei-Fei Li regarding unavoidable transitional growing pains during major technology disruptions. Patel agrees that the transition will carry significant friction, cautioning that the technology sector often lives in an insulated bubble where disruption is treated as routine. When industrial workers, such as steel mill employees, face displacement, "they don't become an AI prompt engineer." He calls on the tech sector to accept civic responsibility, partnering closely with the public sector to rebuild education pipelines, close the widening skills gap, and deliver comprehensive workforce retraining.

Beyond workforce displacement, Patel warns that safety, security, and adversarial threats represent catastrophic operational risks if left unmitigated. The identical AI capabilities built to empower enterprises are simultaneously available to nation-states and malicious adversaries. This parity enables bad actors to execute hyper-bespoke, sophisticated cyberattacks at machine scale, rendering legacy spam—such as poorly phrased emails from fictitious foreign princes asking for bank transfers—completely obsolete in favor of highly believable, context-aware threats.

- Macro friction: Workforce transitions require public-private partnerships; displaced industrial workers cannot simply pivot overnight to prompt engineering. - Symmetric adversarial capabilities: Malicious actors and nation-states wield the exact same advanced AI tooling as defenders. - Attack vector evolution: Shifting from rudimentary, typo-filled phishing scams to hyper-targeted, machine-speed adversarial exploits.

[17:06](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=1026s) Securing AI Itself and Enforcing Fine-Grained Agent Identity

Jeetu Patel illustrates the insidious nature of modern social engineering threats with a hyper-personalized phishing scenario: receiving a message saying, *"was great to see you at our kids' basketball game last night. Here are some pictures. Click to download."* The instant an unsuspecting user clicks the link, malware is dropped onto the endpoint and begins lateral movement across the enterprise network. To combat this reality, safety and security must be addressed aggressively—not merely by leveraging AI for cyber defense, but fundamentally by securing the AI models and systems themselves. AI models introduce novel attack surfaces and failure modes, including hallucinations, toxic outputs, and prompt injection attacks, all of which require dedicated mitigation strategies.

```mermaid flowchart TD accTitle: Securing Agentic AI Workflows subgraph ModelGov["Model Governance & Guardrails"] A["Data Ingestion Visibility"] --> B["Model Validation"] B --> C["Runtime Enforcement Guardrails"] end subgraph AgentSec["Agent Identity & Access"] C --> D["Agent Identity Verification"] D --> E{"Fine-Grained Permissions"} E -->|"Allowed"| F["Draft Emails to Team"] E -->|"Restricted"| G["Block Access: Boss & Board"] end ```

Securing the emerging ecosystem of autonomous agents requires end-to-end visibility into the training and inference data entering models, robust validation frameworks to confirm models behave as expected, and dynamic runtime enforcement guardrails. Furthermore, agents require a novel paradigm of machine identity and authentication when communicating with one another. Patel emphasizes that organizations must enforce fine-grained access controls rather than granting blanket privileges:

> "You don't want to have an agent say, 'Okay, get access to my email and send an email to everyone.' You might say, 'I want to have an agent be able to craft an email on my behalf, but not to my boss and my board of directors, but it's okay to do it to my team.'"

Patel concludes that solving safety, security, and risk mitigation across autonomous AI ecosystems is not just an enterprise imperative, but a broader responsibility that society collectively needs to resolve.

- **Vulnerabilities inherent to AI models:** Susceptibility to prompt injections, hallucinations, toxic outputs, and data poisoning. - **Core defensive layers:** Ingestion data visibility, model behavioral validation, runtime guardrail enforcement, and agent-to-agent identity verification. - **Principle of least privilege for agents:** Granular scoping of agent permissions (e.g., segmenting internal team communications from executive or board interactions) to prevent unauthorized automated actions.

---

[18:25](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=1105s) Session Wrap-Up

CNN senior reporter Matt Egan closes the fireside chat by thanking Jeetu Patel for sharing his perspectives on the agentic revolution, safety, and security architectures. Patel offers his thanks in return as the session concludes on a high note.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=4s) For the first time, we might be moving into an era of abundance where we can actually solve problems that were previously unsolvable. — Jeetu Patel

> [00:04](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=4s) The thing that is overhyped is that humans will become obsolete in five years and we won't have anything to do and we'll all be on beaches with no creative juices flowing. — Jeetu Patel

> [00:04](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=4s) The thing that's deeply underhyped is the capacity for these systems to generate original insights that don't exist in the human corpus today. — Jeetu Patel

> [02:59](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=179s) If every human has 10 to 100 agents working on their behalf, a population of 8 billion starts behaving like 80 billion or 800 billion people. — Jeetu Patel

> [02:59](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=179s) Data centers will move to power rather than power moving to data centers. — Jeetu Patel

> [02:59](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=179s) The technical capability of AI is moving faster than the trust deficit humans have with delegating tasks to it. — Jeetu Patel

> [06:35](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=395s) Harder problems actually have a higher probability of success because they attract better people, and business is a team sport. — Jeetu Patel

> [09:36](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=576s) The stupidest thing a company can do is eliminate entry-level jobs. — Jeetu Patel

> [09:36](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=576s) Experience can be a massive liability because you say, 'I tried that five years ago and it didn't work,' ignoring that a thousand underlying variables have changed. — Jeetu Patel

> [15:02](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=902s) When steel mill workers get displaced, they don't become an AI prompt engineer. — Jeetu Patel

> [15:02](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=902s) The same AI capabilities available to enterprises are available to nation-states and bad actors. — Jeetu Patel

> [17:06](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=1026s) You don't want to have an agent say, 'Okay, get access to my email and send an email to everyone.' You might say, 'I want to have an agent be able to craft an email on my behalf, but not to my boss and my board of directors, but it's okay to do it to my team.' — Jeetu Patel

## Mentioned Ai4, Anthropic, Cisco, CNN, Dario Amodei, Deep Research, Fei-Fei Li, Jeetu Patel, Matt Egan, Stanford University

## Video description Jeetu Patel, President and Chief Product Officer, Cisco Matt Egan, Senior Reporter, CNN Abstracts Join Jeetu Patel, President and Chief Product Officer at Cisco, for a fireside chat with Matt Egan, award-winning CNN Senior Reporter, where they will explore the transformative power of agentic AI. They will discuss how organizations that effectively harness agentic AI will set themselves apart in the near future, while those that don't risk falling behind. Together, Jeetu and Matt will address the evolving landscape of safety, security, identity, and human collaboration with autonomous AI agents. Jeetu will also share Cisco's groundbreaking innovations that secure agentic AI ecosystems at machine scale, turning security from a challenge into a strategic advantage that accelerates AI adoption. Discover how these advancements empower organizations to lead with trust, agility, and resilience in the emerging agentic era.

[Source](https://www.youtube.com/watch?v=iGd4cJP0ZwA)

中文简报

思科Patel驳斥AI裁员论调，捍卫初级岗位思科首席产品官Jeetu Patel认为，淘汰初级人才会剥夺打破固有认知所必需的不受拘束的视角，从而削弱企业创新能力。 Jeetu Patel将人工智能定性为文明从资源匮乏迈向充裕的根本性转变，他在驳斥人类将被淘汰的危言耸听论调的同时，指出AI展现出原创性、超人类科学洞见的前景尚未得到足够重视。随着智能体AI（agentic AI）从短时任务向持续数周的自主工作流迈进，它正面临电力、算力基础设施和人类信任方面的严峻瓶颈。Patel强烈反对AI将消灭初级岗位的观点，认为初级员工对于打破机构固有思维和构思全新问题视角至关重要。最后，他强调企业落地应用的关键在于解决新型安全挑战，呼吁部署运行时护栏、模型验证及细粒度的智能体身份控制，以防范复杂的对抗性威胁。思科（Cisco）总裁 Jeetu Patel 指出，AI 正推动社会从资源匮乏走向认知富足；五年内人类被完全淘汰的论调被过度夸大，而 AI 超越现有知识库生成全新原创见解的能力则被严重低估。 (0:04) Agentic AI 正从对话式 Chatbot 迈向多步骤自主 Agent，任务执行周期将从 20 分钟延伸至数周乃至数月；若人均配备 10 至 100 个 Agent，全球产能将成倍放大，但也将引发严峻的电力、算力与人类信任赤字瓶颈。 (2:59) 随着智力走向通用化与平民化，核心职业价值正从死记硬背的知识积累转向提出高质量复杂问题的能力，而通过“元提示词”（Metaprompting）指导 AI 构建最优提示词可实现输出质量的阶跃式提升。 (6:35) 针对削减初级白领岗位的预测，Patel 认为裁撤初级员工极其不智，年轻一代作为数字原住民将 AI 视作协同大脑，能促使组织摆脱以往因沉没经验带来的创新阻力与思维定势。 (9:36) 对抗高度定制化的新型 AI 钓鱼与网络攻击，企业不仅需用 AI 进行防御，还必须建立覆盖模型幻觉与注入攻击的防护机制，并对跨系统通信的自主 Agent 实施细粒度的身份验证与最小权限控制。 (17:06)

Reading summary

Cisco executive Jeetu Patel argues that artificial intelligence is shifting from knowledge aggregation toward autonomous, multi-day agentic execution capable of generating original scientific discoveries. Realizing this potential, however, requires resolving severe power shortages, maintaining junior talent to force organizational unlearning, and instituting fine-grained machine identity to combat machine-scale adversarial threats. From Knowledge Aggregation to Original Insight Generation (0:04) Patel contends that civilization is crossing from an organizing principle of historical resource scarcity into an era of computational abundance capable of resolving long-standing challenges, including material engineering and cancer research. In his view, fears of near-term human obsolescence within five years are overhyped, as human labor historically adapts to reconfigure workflows rather than disappear entirely. Conversely, Patel identifies the generation of genuine "original insights" as deeply underhyped. While current large language models function as aggregation engines that synthesize the existing human corpus, the subsequent evolutionary phase will center on models producing net-new discoveries that have never existed in human records, fundamentally transforming scientific discovery. Autonomous Horizons and Physical Infrastructure Bottlenecks (2:59) The AI industry is currently navigating its second phase, transitioning from conversational chatbots to autonomous agents executing multi-step tasks. Patel notes that while tools like Deep Research currently sustain autonomous execution for approximately twenty minutes, this operational horizon will progressively lengthen to hours, days, weeks, and months, drastically increasing the complexity of solvable problems. Patel projects that if individual humans eventually direct 10 to 100 autonomous agents continuously, an 8-billion-person global population could achieve the economic throughput of 80 billion to 800 billion people. However, this trajectory faces hard limits: global electrical grids and compute clusters cannot sustain this projected load—particularly as modalities like computer use surge—forcing data center developers to build directly adjacent to power sources. Compounding this physical bottleneck is a persistent human trust deficit, where technical agent capability outpaces user willingness to delegate critical workflows without verified safety and reliability. Metaprompting and the Strategic Value of Early-Career Talent (6:35) With domain knowledge becoming commoditized, Patel argues that human differentiation now hinges on framing high-quality, nuanced inquiries. Techniques like metaprompting—instructing a model to construct the optimal prompt before executing a complex task—deliver step-function gains in output fidelity. Because contemporary models interface through natural human language rather than rigid code syntax, societal absorption cycles are rapidly compressing. Within this context, Patel explicitly rejects predictions from Anthropic CEO Dario Amodei suggesting AI could eliminate half of entry-level white-collar jobs and elevate unemployment to 10–20% within one to five years. Patel argues that cutting entry-level roles undermines organizational viability because younger digital natives interact with AI as ambient cognitive partners rather than transactional automation tools. Furthermore, accumulated tenure frequently acts as a liability when senior staff dismiss viable concepts based on outdated historical assumptions, whereas junior personnel introduce unencumbered perspectives that compel institutions to unlearn obsolete models. Workforce Retraining and Granular Agent Security Architectures (15:02) Acknowledging transitional frictions highlighted by Stanford's Fei-Fei Li, Patel cautions that displaced industrial workers cannot simply retrain overnight as prompt engineers, obligating the technology sector to partner with public institutions to overhaul education and address the widening skills gap. At the same time, symmetric access to advanced tooling allows malicious actors and nation-states to conduct hyper-personalized social engineering campaigns at machine scale, rendering primitive phishing schemes obsolete. Mitigating these threats requires securing AI models themselves against prompt injection, hallucinations, and data poisoning through runtime guardrails and data validation. In multi-agent environments, Patel stresses that organizations must implement fine-grained identity and access controls rather than granting broad autonomy, citing policies that allow an agent to draft routine messages to immediate team members while strictly restricting autonomous outreach to senior executives and board members. Patel's framework presents agentic AI as an unprecedented multiplier of human problem-solving capacity, yet its practical execution remains tethered to physical grid capacity, civic investments in workforce transition, and verifiable agent governance. Whether organizations capture these gains depends on resolving the trust deficit through granular security controls while actively retaining early-career talent to drive institutional adaptation.

精炼阅读摘要

思科（Cisco）执行副总裁 Jeetu Patel 在与 CNN 记者 Matt Egan 的对话中指出，人工智能正推动人类社会从资源匮乏迈向认知富足时代，其核心价值在于产生前所未有的原创见解，而非短期内淘汰人类劳动力。Patel 强调，虽然自主智能体（Agentic AI）具备巨大的生产力倍增潜力，但其落地正受到电力与算力短缺、人类信任赤字、社会技能转型摩擦以及严峻对抗性安全威胁的多重制约。从稀缺走向富足：原创见解的突破潜力 (0:04) Patel 指出，人类文明过往始终建立在资源匮乏的基础之上，匮乏迫使社会不断平衡有限资源并倒逼发明创新；而人工智能正推动社会跨入资源富足的新纪元，使攻克癌症、研发全新材料等历史性难题成为可能。针对市场舆论，Patel 区分了过度炒作与被严重低估的技术维度。他认为，关于人类将在五年内被彻底淘汰并丧失创造力的悲观论调属于过度炒作，因为人类历史上始终能够适应变革并在岗位重构中挖掘新的价值空间。相反，人工智能真正被低估的能力在于产生“原创见解”。Patel 阐明，当前的大型语言模型本质上仍是高度复杂的信息聚合引擎，主要功能是在人类既有的知识库中归纳模式与合成信息；而下一阶段的技术跃迁将诞生能够提出全新洞见的模型，这些见解从未存在于人类既有历史记录中，将从根本上颠覆基础科研与工业材料工程的研发范式。智能体自主化扩张与算力电力瓶颈 (2:59) 在智能体的发展路径上，Patel 认为当前行业正处于第二阶段。第一阶段侧重于能够基于提示词回答问题的交互式聊天机器人，而第二阶段则引入了能够代表人类执行多步骤、端到端任务的自主智能体。以 Deep Research 等工具为例，智能体目前已能自主运行约 20 分钟；未来随着连续自主执行时间延伸至数小时、数天乃至数月，系统所能解决的问题复杂度将呈指数级提升。然而，这种自主化扩展面临着严峻的物理与基础设施瓶颈。Patel 推测，若未来全球 80 亿人口每人平均配置 10 至 100 个全天候运行的自主智能体，全球经济与计算吞吐量将相当于 800 亿到 8000 亿人口的规模。现存的电力供应和算力网络完全无法支撑这种负荷，迫使数据中心建设转向“靠近电源布局”；而模拟人类操作界面的计算机使用（Computer Use）技术，也将对网络带宽与计算集群构成空前压力。此外，当前技术能力已超越人类的授权意愿，在未通过严格安全与可靠性验证解决“信任赤字”前，人类用户仍将严格限制智能体的自主权限。交互模式演进：元提示词与提问能力 (6:35) 随着专业信息与基础智能日益丰沛、廉价并走向通用大宗化，职场核心竞争力的定义正在发生转移。Patel 指出，过往依靠死记硬背或积累专门领域知识建立壁垒的模式已被弱化，提出细致、高质量问题的能力已成为新的关键分水岭。他在组织问题解决中提出一项反直觉认知：挑战更高难度的问题反而拥有更高的成功几率，因为棘手难题天然能够汇聚顶尖人才，而现代商业竞争本质上是一项团队协作。在具体的交互技巧层面，“元提示词”（Metaprompting）正在先进模型中展现出巨大价值。与直接向模型索取最终成果不同，元提示词要求用户先指示模型构建出生成该成果的最优提示词框架，再运行该优化后的提示词以输出最终内容。Patel 观察到，这种方法能够带来阶梯式的输出质量跃升；加之现代机器正在主动理解人类的自然语言，而非强迫人类学习复杂的机器代码，全社会的技能学习曲线与技术采纳周期正被大幅压缩。破除初级岗位消亡论与经验负债风险 (9:36) 针对 Anthropic 首席执行官 Dario Amodei 提出的“人工智能可能在未来一至五年内消灭半数白领初级岗位并将失业率推高至 10% 至 20%”的预警，Patel 虽表达了对 Amodei 的个人尊重，但明确反驳了该论点，直言一刀切地裁撤初级岗位是“企业能做出的最愚蠢之举”。通过代际使用习惯的对比，Patel 发现 28 岁的员工通常仅以事务性方式使用人工智能以自动化既有流程，而 20 岁左右的新生代数字原住民则将其视作持续进行头脑风暴的认知协作者与思维搭档。 Patel 进一步强调，在剧烈变革期，长期积累的资历往往会异化为沉重的“经验负债”。资深管理者常因某项技术方案在五年前失败过而断然否决新构想，却忽视了数以千计的底层技术前提如今已彻底改变。相比之下，初入职场的年轻人才不受旧思维禁锢，能够倒逼组织反思并遗忘过时的经验假设。Patel 认为，突破性创新必须依靠初级员工带来的新鲜视角与资深团队的执行经验紧密结合，企业绝不能因追求短期效率而牺牲人才梯队的代际新陈代谢。社会转型摩擦、对抗威胁与安全治理 (15:02) 针对技术转型期带来的阵痛，Patel 认同斯坦福大学教授李飞飞（Fei-Fei Li）关于转型摩擦不可避免的观点，并批评科技行业习惯于孤立地看待颠覆性创新。他指出，传统产业的下岗工人无法在短时间内直接转型为人工智能提示词工程师，科技企业必须承担社会公共责任，与政府公共部门合作重塑教育体系并开展实质性的劳动力再培训。与劳动力冲击并存的另一大威胁是对抗性安全风险：恶意攻击者与敌对国家拥有与企业完全相同的先进人工智能工具，使网络攻击演变为机器级规模的超个性化攻击，彻底淘汰了以往容易辨识的劣质钓鱼邮件。面对这种能力对称的威胁，安全防御不仅要利用人工智能实现自动反制，更必须保障人工智能系统自身的安全性。Patel 指出，模型面临着数据投毒、提示词注入、幻觉及有害内容生成等全新攻击面，必须建立涵盖数据摄入可见性、模型行为验证和运行时防护的端到端防线。在自主智能体生态中，必须引入严格的机器身份认证与细粒度权限控制机制，坚决杜绝宽泛授权。例如，系统可允许智能体代表员工向内部团队起草工作邮件，但必须从技术底层严禁其直接向高管或董事会擅自发送沟通信息。 Patel 总结认为，自主智能体革命的核心价值在于赋能人类解决此前无法企及的宏观难题，而非单纯提高日常微观事务的生产力；但要实现这一愿景，全社会必须共同克服电力与算力供给的物理制约，妥善化解劳动力转型带来的阵痛，并建立起涵盖模型合规与细粒度权限控制的安全信任架构。

Detailed notes — zh

## Background 该内容由频道 Ai4 于 2025 年 8 月 28 日发布在 YouTube 平台上。视频记录了 Cisco 总裁兼首席产品官 Jeetu Patel 与 CNN 资深记者 Matt Egan 进行的炉边对话，重点探讨了 Agentic AI（智能体 AI）的变革力量、组织在拥抱智能体时面临的机遇与挑战，以及如何在机器规模下构建安全、身份认证与信任体系。 ## TL;DR 对话深入探讨了 AI 从简单信息问答向具备自主执行能力的 Agentic AI 跃迁的范式转移，指出智能体的长程任务执行力将打破数千年的“稀缺性”假设，但也正撞上电力、算力等基础设施极限与“信任赤字”两大瓶颈。Jeetu Patel 反驳了“人类将被彻底淘汰”的悲观论调，强调单纯提升效率是 AI 最无趣的应用，真正的突破在于生成未知的原创洞见，并指出企业一刀切裁撤初级岗位是扼杀组织“卸载旧认知（Unlearning）”能力的重大战略错误。面对日益高级的 AI 武器化威胁，企业必须将重心从“用 AI 防御”转向“保护 AI 本身”，通过数据可见性、运行时防护栏、智能体身份认证与细粒度权限控制，构建可信且安全的自动化生态。 ## Atmosphere & style 整场对话呈现出务实、睿智且极具思辨深度的行业领袖探讨氛围，完全跳脱了盲目的技术狂热与末日论恐慌。Matt Egan 提问精准敏锐，Jeetu Patel 谈吐从容且充满洞察力，擅长用鲜活具体的现实隐喻将宏观的技术趋势与微观的安全策略串联起来。叙事基调在对技术充裕性的高远展望与对社会转型阵痛、算力极限及网络攻防严峻现实的审慎剖析之间实现了平衡，带给听众一种兼具前瞻洞察与工程师式严谨清醒的沉浸感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=4s) AI 浪潮正在打破人类社会数千年来的“资源稀缺性”假设，推动人类进入全面充裕（Abundance）的新纪元，解锁前所未有的问题解决路径。 - [00:04](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=4s) AI 最被低估的潜力并非简单汇总已知信息，而在于生成人类语料库中从未存在过的“原创洞见”（Original Insights），进而颠覆性地攻克癌症与合成新材料等未知难题。 - [02:59](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=179s) 技术正从问答机器人跃迁至具备自主行动能力的 Agentic AI，随着自主执行窗口从 20 分钟延展至数天乃至数月，复杂任务处理能力将呈指数级爆发。 - [02:59](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=179s) 智能体的规模化普及正撞上两大硬性制约：一是全球严重的电力短缺与算力基础设施承载极限；二是由于安全与透明度不足所导致的“信任赤字”。 - [06:35](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=395s) 在智能走向充裕与商品化的时代，单纯的知识积累不再稀缺，提问的精妙程度（Nuance of Asking Questions）成为定义并解决高阶问题的核心壁垒。 - [06:35](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=395s) 元提示（Meta-Prompting）具备显著的效率放大效应，通过让 AI 协助构建高阶提示词模板，能够使最终输出质量实现断层式飞跃。 - [09:37](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=577s) 技术变革会重塑工作岗位但不会淘汰人类，企业若因短视而一刀切砍掉初级岗位，将犯下损害组织人才生态的毁灭性战略错误。 - [09:37](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=577s) 年轻一代天生将 AI 视作实时思维伙伴，能够为组织注入“卸载旧经验”（Unlearning）的生机，与资深员工的业务积淀碰撞出创新突破。 - [15:02](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=902s) 产业变革伴随不可忽视的社会转型阵痛，科技界必须与公共及教育机构协同推进体系化再培训（Reskilling），切实承担社会兜底责任。 - [15:02](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=902s) 攻击者与防御方掌握同等强大的前沿 AI 工具，网络威胁已从粗劣钓鱼演进为由国家级黑客力量驱动、高度定制化且几无破绽的精准社工渗透。 - [17:06](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=1026s) 企业网络安全必须从单纯的“利用 AI 防御”升级为“保护 AI 本身”，机制化地遏制基础模型与智能体固有的幻觉、毒性输出及提示词注入攻击。 - [17:06](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=1026s) 构建可信智能体生态需要建立全生命周期治理闭环，涵盖输入数据可见性、模型验证、运行时强制护栏、智能体独立数字身份以及细粒度权限控制。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=4s) 重新审视稀缺与充裕：AI 革命中被高估与被低估的面向在科技行业发展的关键节点上，人类社会数千年来的核心底层假设正在被打破。长期以来，人类的一切文明构想几乎都建立在“资源稀缺性”（Scarcity）的基础之上——因为资源匮乏，人们必须精打细算地平衡配置，而稀缺也成了发明创造的源动力。然而，当下的 AI 浪潮正在将人类带入一个真正可能实现“全面充裕”（Abundance）的新纪元。尽管这不会在短短一年内彻底兑现，但充裕性将解锁一整套前所未有的问题解决路径。谈及当前的 AI 讨论热潮，业内存在一个被严重高估的论调和一个被严重低估的潜力： ```mermaid flowchart TD accTitle: AI Hype vs Reality Comparison subgraph Overhyped["被高估的论调"] A["5年内人类彻底被淘汰"] --> B["人类坐以待毙、丧失创造力"] B --> C["实际：工作形态重塑，但人类价值不可替代"] end subgraph Underhyped["被低估的潜力"] D["超越现有语料的原创洞见"] --> E["从简单聚合知识走向未知推演"] E --> F["重塑癌症攻克、新材料合成与社会运作"] end ``` - **被高估的盲目悲观**：对于“人类在未来五年内将彻底过时、无所事事地躺在海滩上、创造力完全枯竭”的说法，应当坚决予以驳斥。人类历史上每一次技术变革都在不断重塑工作岗位的形态，但人类始终能找到创造独特价值的方式，绝不会沦为无用之物。 - **被低估的“原创洞见”（Original Insights）**：当前的大语言模型（LLM）在很大程度上仍属于“信息聚合机制”，擅长从人类现有的海量知识库中提炼归纳。然而，未来 AI 最令人兴奋的突破在于生成人类知识语料库中**从未存在过**的原创洞见，进而以颠覆性的视角去攻克癌症、研发新型材料，从根本上重塑社会的运转方式。 --- ### [02:59](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=179s) 智能体时代来临：从任务委派到算力与信任的双重瓶颈 AI 发展正处于从第一阶段（智能问答机器人）向第二阶段（智能体 AI / Agentic AI）的跃迁。在过去两年半的时间里，大众已经逐步掌握了通过精准提问获取高质量答案的技能；而如今的第二阶段，核心在于具备自主行动能力的 AI Agent 能够代表人类独立执行复杂任务。以深度研究（Deep Research）类 Agent 为例，目前它已能自主运行约 20 分钟——用户下达调研指令后去喝杯咖啡，它便能自动搜集全网信息并整理汇报。随着这一自主运行窗口从 20 分钟拓展至 2 小时、2 天、2 周乃至 2 个月，AI 解决复杂难题的能力将呈指数级爆发。 ```mermaid flowchart LR accTitle: Evolution and Constraints of Agentic AI subgraph Evolution["Agent 自主时间轴延伸"] T1["20 分钟自主运行"] --> T2["2 小时 / 2 天"] T2 --> T3["2 周 / 2 个月长程执行"] end subgraph Scale["生产力几何级放大"] P1["人均配置 10~100 个 Agent"] --> P2["80 亿人拥有 800 亿~8000 亿级吞吐能力"] end subgraph Constraints["两大现实瓶颈"] C1["基础设施极限 电力匮乏 / 算力短缺 / 电脑使用场景爆发"] C2["信任赤字 安全风险 / 系统不透明导致人类不敢放权"] end Evolution --> Scale Scale --> Constraints ``` 伴随这种指数级潜力的释放，整个生态也将迎面撞上两大不可忽视的硬性制约： - **基础设施的极限承载**：若未来每个人拥有 10 到 100 个全天候无休运行的 AI Agent，全球 80 亿人口将释放出相当于 800 亿甚至 8000 亿人口的吞吐量。目前全球面临严重的电力短缺与算力危机，以至于呈现出“数据中心追随电力产地而建，而非电力输送至数据中心”的倒挂局面。尤其当“电脑使用”（Computer Use——Agent 模拟人类在终端操作界面）这一高频消耗网络和算力的场景普及后，基础设施压力将进一步激增。 - **不可忽视的“信任赤字”（Trust Deficit）**：目前 AI 实际具备的技术潜能远远超过了人类现阶段允许其施展的范围。如果安全、隐私与防护机制无法随着时间推移建立起足够透明度，用户便不敢放手将核心工作交由系统托管。 --- ### [06:35](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=395s) 提问的艺术与元提示（Meta-Prompting）的威力过去人们始终认为“积累知识”是拉开个人竞争差距的关键壁垒，但在智能被极大地充裕化、甚至走向免费与商品化的时代，单纯的知识储备已不再稀缺。真正的艺术转向了“提问的精妙程度”（Nuance of Asking Questions）。提问的质量直接决定了所能界定并解决的问题层次；而从商业规律来看，越困难的问题往往越能吸引顶尖人才参与攻关，最终取得突破性成果的概率反而更高。 ```mermaid flowchart TD accTitle: Meta Prompting Flow subgraph DirectPrompt["传统直接提示"] U1["用户简单 Prompt"] --> G1["AI 生成普通结果"] end subgraph MetaPrompt["元提示流程 (Meta-Prompting)"] U2["用户初始需求"] --> M1["要求 AI 构建最优提示词模板"] M1 --> M2["AI 返回结构化的高阶 Prompt"] M2 --> G2["将高阶 Prompt 重新输入 AI"] G2 --> R2["获得跨越式提升的高质量成果"] end ``` 针对日常人机交互，元提示（Meta-Prompting）展现出了惊人的效率放大效应。例如在图像生成场景中，如果让 AI 直接作画，得到的效果往往中规中矩；但如果换一种方式，要求 AI“为这个场景设计一段最能激发生成器潜力的结构化 Prompt”，再把 AI 给出的元提示词复制回系统中，所呈现的画面质量与细节表现力将实现断层式飞跃。这种非对称的效能提升证明了：由于机器正在飞速理解人类的自然语言，人类掌握新技术的消化周期已被前所未有地压缩。 --- ### [09:37](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=577s) 拒绝“人类被淘汰”：为何消灭初级岗位是企业最大的战略错误针对 Anthropic CEO Dario Amodei 提出的“AI 可能在 1 至 5 年内消灭半数初级白领岗位、并将失业率推高至 10% 到 20%”的预警，应当从人才生态与技术认知的维度进行有力反驳。尽管岗位职责的重组势在必行，但“人类彻底被淘汰”在逻辑上是站不住脚的。更重要的是，那些企图一刀切“彻底砍掉初级岗位”的企业，正在犯下一个毁灭性的长期战略错误。年轻一代与资深员工在使用 AI 时存在着本质上的思维代沟： | 观察维度 | 28 岁及以上成熟职场人 | 20 岁左右及刚毕业的年轻一代 | |:--- |:--- |:--- | | **交互模式** | 任务型、事务型、工具化（单次问答、局部流程自动化） | 协作者、思维伙伴、实时头脑风暴伴侣（Sidekick & Companion） | | **技术心智** | 带有过往软硬件的使用惯性与经验包袱 | 天生具备直觉化的交互本能（如幼童对触控界面的原生理解） | | **对团队的价值** | 提供业务沉淀，但容易受制于“5年前尝试失败”的经验主义 | 带来完全未被经验污染的全新视角，推动组织主动“重塑认知” | 经验在特定领域固然可贵，但在技术剧烈迭代的关口也常常演变为巨大的负债。许多资深人士往往因为过去的失败经验而拒绝重新尝试，忽视了底层数万个预设条件早已改变的事实。通过吸纳年轻一代，企业才能拥有持续“卸载旧知识”（Unlearning）的生机。经验与全新视角的深度碰撞，才是孕育非凡创新的真正土壤。 > "AI 绝不仅仅关乎生产力——事实上，在我的认知里，单纯提升效率是 AI 最无趣的一面。AI 真正迷人之处，在于解决那些过去甚至无法想象的人类未知难题。" --- ### [15:02](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=902s) 转型期的社会阵痛与日益严峻的网络安全攻防历史上的每一次产业大颠覆都伴随着不可避免的转型阵痛。身处科技行业的人群往往生活在自给自足的信息茧房中，容易轻描淡写地将动荡视为创新的必然代价；但现实中，一名面临下岗的传统钢铁工人不可能一夜之间转型为 AI 提示词工程师。科技界必须对社会承担起兜底责任，深化与公共部门及教育机构的协同，通过体系化的再培训（Reskilling）与技能弥合来缓冲剧烈冲击。 ```mermaid flowchart TD accTitle: Evolution of Cyber Threats with AI subgraph Traditional["传统网络攻击模式"] A1["低劣钓鱼邮件 拙劣借口 / 错别字连篇 / 广撒网"] end subgraph ModernAI["AI 赋能下的高级威胁"] B1["同等武器化工具赋能黑客"] --> B2["国家级黑客组织 (Nation-States) 深度介入"] B2 --> B3["高度定制化、无破绽的精准社工渗透"] end Traditional -.->|攻击范式质变| ModernAI ``` 伴随技术扩散而来的还有极具灾难性的安全与防御风险： - **攻防工具的对等性**：正向研发团队所能调用的前沿 AI 能力，对手与黑客同样触手可及。 - **攻击范式的升维**：网络攻击已从早期充斥错别字、粗制滥造的“异国王子转账千万资产”类垃圾钓鱼邮件，全面演化为由国家级黑客力量（Nation-States）驱动、借助生成式 AI 精确伪造的高级定制化社工渗透。若缺乏针对性的防御体系，其破坏力将是灾难性的。 ### [17:06](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=1026s) 保护 AI 本身：智能体生态的安全防线与细粒度权限控制 Jeetu Patel 用一个生动且令人警惕的社会工程攻击场景切入，展示了网络威胁的演变：某人可能会收到一封看似极其逼真且熟络的邮件，上面写着“昨晚在我们家孩子的篮球赛上见到你真高兴，这里附上几张照片，点击下载”。然而，用户点击链接的瞬间，恶意软件就会悄然植入设备，并随即在企业网络内部展开横向移动（lateral movement）。Jeetu 指出，面对日益复杂的威胁，企业绝不能仅仅停留在“利用 AI 进行网络防御”（using AI for cyber defense），更关键的任务在于必须“保护 AI 本身”（secure AI itself）。因为基础模型与智能体自身就存在幻觉（hallucination）、毒性输出（toxicity）以及提示词注入攻击（prompt injection attacks）等重大固有风险，必须从机制上加以遏制与缓解。为了化解这些风险并建立真正的信任体系，Jeetu 梳理了一套全周期的防护与治理逻辑：首先，企业必须对进入模型的数据流具备完全的可见性（visibility）；其次，要对模型进行严格的运行验证（validation），确保其行为符合预期；更重要的是在系统执行中部署运行时强制防护栏（runtime enforcement guardrails）。此外，随着智能体之间开始自主交互（agent-to-agent communication），必须为每个 AI 智能体赋予独立、可验证的数字身份，并配合极其严密的细粒度权限控制（fine-grained control）。 Jeetu 举例说明了细粒度权限控制的必要性：用户绝不会希望对智能体下达“获取我的邮箱权限并直接给所有人发邮件”这样不受限制的指令；更符合安全规范的做法是设定清晰的边界——例如“允许智能体代表我给团队下属撰写并发送日常邮件，但绝对禁止它直接向我的主管老板或董事会成员发送邮件”。Jeetu 强调，在智能体时代，安全与安全风险缓解绝不仅是单家企业的 IT 议题，而是整个社会在迈向全面自动化与自主智能体生态时必须共同攻克的头等大事。 ```mermaid flowchart TD accTitle: AI 安全防护与细粒度控制架构 subgraph SecuringAI ["保护 AI 本身 (Securing AI)"] A["数据输入可见性 (Data Visibility)"] --> B["模型验证与合规 (Model Validation)"] B --> C["运行时护栏拦截 (Runtime Guardrails)"] C --> D["智能体身份与通信验证 (Agent Identity)"] end subgraph ControlPolicy ["细粒度权限控制策略示例"] D --> E["团队内部邮件 (授权代表撰写/发送)"] D -.->|严格阻断 / 需人工核准| F["老板与董事会邮件 (越权防护限制)"] end ``` - **AI 自身面临的核心安全风险**：模型幻觉（Hallucination）、有害与毒性内容生成（Toxicity）、提示词注入攻击（Prompt Injection Attacks）。 - **防御闭环机制**：数据输入可见性审计、模型预期行为验证、运行时策略强制拦截、多智能体交互身份认证。 - **细粒度控制准则**：区分不同接收对象与业务敏感度，严格限定智能体代行权利的边界，防止无差别授权扩散风险。 --- ### [18:25](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=1105s) 炉边对话结语与展望 CNN 高级记者 Matt Egan 对 Jeetu Patel 的深度剖析表达了感谢，认为 Jeetu 分享的关于 Agentic AI 变革、身份验证机制以及下一代网络安全架构的见解极具启发性。Jeetu 随后致谢，整场关于如何以安全与信任驱动智能体人工智能革命的炉边对谈在融洽的气氛中正式圆满结束。 ## Notable quotes > [09:37](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=577s) AI 绝不仅仅关乎生产力——事实上，在我的认知里，单纯提升效率是 AI 最无趣的一面。AI 真正迷人之处，在于解决那些过去甚至无法想象的人类未知难题。 — Jeetu Patel > [00:04](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=4s) 人类在未来五年内将彻底过时、无所事事地躺在海滩上、创造力完全枯竭 — Jeetu Patel > [02:59](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=179s) 数据中心追随电力产地而建，而非电力输送至数据中心 — Jeetu Patel > [06:35](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=395s) 为这个场景设计一段最能激发生成器潜力的结构化 Prompt — Jeetu Patel > [09:37](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=577s) AI 可能在 1 至 5 年内消灭半数初级白领岗位、并将失业率推高至 10% 到 20% — Dario Amodei > [17:06](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=1026s) 昨晚在我们家孩子的篮球赛上见到你真高兴，这里附上几张照片，点击下载 — Jeetu Patel > [17:06](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=1026s) 利用 AI 进行网络防御 — Jeetu Patel > [17:06](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=1026s) 保护 AI 本身 — Jeetu Patel > [17:06](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=1026s) 获取我的邮箱权限并直接给所有人发邮件 — Jeetu Patel > [17:06](https://www.youtube.com/watch?v=iGd4cJP0ZwA&t=1026s) 允许智能体代表我给团队下属撰写并发送日常邮件，但绝对禁止它直接向我的主管老板或董事会成员发送邮件 — Jeetu Patel ## Mentioned Jeetu Patel, Matt Egan, Dario Amodei, Cisco (思科), CNN, Anthropic, Ai4, Agentic AI (智能体 AI), Meta-Prompting (元提示), Deep Research, Computer Use, Prompt Injection (提示词注入) ## Video description Jeetu Patel, President and Chief Product Officer, Cisco Matt Egan, Senior Reporter, CNN Abstracts Join Jeetu Patel, President and Chief Product Officer at Cisco, for a fireside chat with Matt Egan, award-winning CNN Senior Reporter, where they will explore the transformative power of agentic AI. They will discuss how organizations that effectively harness agentic AI will set themselves apart in the near future, while those that don't risk falling behind. Together, Jeetu and Matt will address the evolving landscape of safety, security, identity, and human collaboration with autonomous AI agents. Jeetu will also share Cisco's groundbreaking innovations that secure agentic AI ecosystems at machine scale, turning security from a challenge into a strategic advantage that accelerates AI adoption. Discover how these advancements empower organizations to lead with trust, agility, and resilience in the emerging agentic era. [Source](https://www.youtube.com/watch?v=iGd4cJP0ZwA)

Source quotes

[9:37] Jeetu Patel: AI 绝不仅仅关乎生产力——事实上，在我的认知里，单纯提升效率是 AI 最无趣的一面。AI 真正迷人之处，在于解决那些过去甚至无法想象的人类未知难题。 [0:04] Jeetu Patel: 人类在未来五年内将彻底过时、无所事事地躺在海滩上、创造力完全枯竭 [2:59] Jeetu Patel: 数据中心追随电力产地而建，而非电力输送至数据中心 [6:35] Jeetu Patel: 为这个场景设计一段最能激发生成器潜力的结构化 Prompt [9:37] Dario Amodei: AI 可能在 1 至 5 年内消灭半数初级白领岗位、并将失业率推高至 10% 到 20% [17:06] Jeetu Patel: 昨晚在我们家孩子的篮球赛上见到你真高兴，这里附上几张照片，点击下载 [17:06] Jeetu Patel: 利用 AI 进行网络防御 [17:06] Jeetu Patel: 保护 AI 本身 [17:06] Jeetu Patel: 获取我的邮箱权限并直接给所有人发邮件 [17:06] Jeetu Patel: 允许智能体代表我给团队下属撰写并发送日常邮件，但绝对禁止它直接向我的主管老板或董事会成员发送邮件

Source quotes

[0:04] Jeetu Patel: For the first time, we might be moving into an era of abundance where we can actually solve problems that were previously unsolvable. [0:04] Jeetu Patel: The thing that is overhyped is that humans will become obsolete in five years and we won't have anything to do and we'll all be on beaches with no creative juices flowing. [0:04] Jeetu Patel: The thing that's deeply underhyped is the capacity for these systems to generate original insights that don't exist in the human corpus today. [2:59] Jeetu Patel: If every human has 10 to 100 agents working on their behalf, a population of 8 billion starts behaving like 80 billion or 800 billion people. [2:59] Jeetu Patel: Data centers will move to power rather than power moving to data centers. [2:59] Jeetu Patel: The technical capability of AI is moving faster than the trust deficit humans have with delegating tasks to it. [6:35] Jeetu Patel: Harder problems actually have a higher probability of success because they attract better people, and business is a team sport. [9:36] Jeetu Patel: The stupidest thing a company can do is eliminate entry-level jobs. [9:36] Jeetu Patel: Experience can be a massive liability because you say, 'I tried that five years ago and it didn't work,' ignoring that a thousand underlying variables have changed. [15:02] Jeetu Patel: When steel mill workers get displaced, they don't become an AI prompt engineer. [15:02] Jeetu Patel: The same AI capabilities available to enterprises are available to nation-states and bad actors. [17:06] Jeetu Patel: You don't want to have an agent say, 'Okay, get access to my email and send an email to everyone.' You might say, 'I want to have an agent be able to craft an email on my behalf, but not to my boss and my board of directors, but it's okay to do it to my team.'

Original transcript

[0:04] I'd love to start with your views on the [0:07] AI revolution. You've said that AI is [0:09] bigger than anything you've seen in your [0:11] lifetime. Given that, are there any [0:14] assumptions about the world and society [0:17] that must be reconsidered to fit this [0:22] new world that we're all entering? [0:26] Look, I think firstly it's a really [0:27] interesting time to be uh be alive in [0:29] tech. I think we'd all agree. Um there's [0:32] a uh um there's a core set of [0:36] assumptions that we've all kind of lived [0:38] with for as long as humans have been [0:41] around which is around this notion of [0:44] scarcity and that everything is around [0:46] you know we have a scarce amount of [0:48] resources. We got to balance those [0:50] resources so that we can make sure that [0:52] we um you know and scarcity has been the [0:55] mother of invention to some degree. I [0:58] think we are now entering a phase where [1:03] abundance is actually possible and it [1:06] might actually open up a whole set of [1:08] possibilities of solving problems that [1:11] we couldn't solve before. Uh I still [1:13] think we're a ways away from it. that's [1:14] not going to happen by next year, but I [1:17] do feel like, you know, there's going to [1:18] be a tremendous amount of opportunity [1:21] for solving problems that we couldn't [1:24] solve before. U and [1:27] and I feel like there's one thing that's [1:29] overhyped about AI and there's one thing [1:31] that's underhyped about AI. Thing that's [1:33] overhyped is I reject the notion that [1:35] humans are going to be obsolete in like [1:36] 5 years where we're not going to have [1:38] anything to do and we're going to be [1:39] sitting on the beach and our creative [1:41] juices are just going to be completely [1:43] obsolete. it it doesn't make any sense. [1:45] I think we've always found a way to go [1:47] out and add value. Now, all of our jobs [1:49] will have some degree of [1:51] reconfiguration, but it's not going to [1:53] be where we we get to be be obsolete. [1:55] So, I think that part is overhyped about [1:57] AI. What's underhyped about AI is the [2:01] original insights. Right now, AI is [2:03] largely especially with large language [2:05] models and all of that is largely an [2:07] aggregation mechanism. It does a good [2:09] job going out and sourcing the corpus of [2:12] knowledge that humans have and being [2:14] able to go figure out if there's [2:16] insights that could be had from it. I [2:18] think moving forward original insights [2:21] will be created in ways that it's it's [2:25] going to be very hard um to um go out [2:29] and replicate and those original [2:30] insights will actually not exist in the [2:32] human corpus of knowledge and that'll [2:34] allow us to solve problems that I don't [2:36] think we've actually thought about [2:37] solving yet. [2:38] >> Wow. And and that in my mind is um is is [2:41] a very exciting kind of possibility that [2:42] we can the way that you solve cancer, [2:44] the way that you actually make sure that [2:46] you create materials, the way that um [2:49] you know kind of um uh society operates, [2:52] all of those things will need to get [2:55] fundamentally reimagined. Uh and I think [2:57] that's going to be cool. [2:59] >> You know, let's talk about agentic AI a [3:02] bit. Um, you've said that you can see a [3:03] future where billions of AI agents are [3:07] working together and in in harmony on on [3:10] behalf of humans. How far are we from [3:13] that future and and what will the [3:15] consequences of that be? [3:18] >> So, I think we're in this squarely in [3:20] the second phase of AI, right? The first [3:22] phase was these intelligent chat bots. [3:23] You could answer questions and it felt [3:25] like magic. Um, and you know, we are [3:28] about two and a half years into that [3:29] phase. most of us have now gotten to [3:33] some level of dexterity of answer asking [3:36] nuanced questions so that you can get [3:38] good answers back. Uh we'll talk a [3:40] little bit about that if we have some [3:42] time. I think the second phase of AI is [3:44] where you have these agents that can [3:46] autonomously conduct tasks and jobs on [3:48] behalf of humans. Now today, if you [3:51] think about probably the one of the most [3:53] effective agents that you have out [3:54] there, it's deep research, right? Like [3:57] you can it probably works autonomously [3:59] for about 20 minutes. You can ask it to [4:02] go do something. You can go get a cup of [4:03] coffee, ask it to do some research, [4:05] it'll come back to you. It'll go, you [4:08] know, scour the web, come back to you [4:09] with some answers. I think as that [4:12] window expands and if you have from you [4:15] know 20 minutes to two hours that an [4:18] agent can work autonomously to to two [4:21] days to two weeks to two months you can [4:25] start seeing the exponentiality of the [4:28] kind of problems that can be solved with [4:30] AI. Uh it it'll be very cool. Now it'll [4:33] also create a tremendous amount of [4:35] constraints on the other side which is I [4:38] think there's going to be a massive [4:39] constraint on infrastructure. [4:42] Um we don't have enough power in the [4:44] world to go out and satiate the needs [4:46] for AI today. That's why you're actually [4:49] seeing data centers being built in every [4:51] place where you might have power. Data [4:53] centers are going to where power is [4:55] rather than power coming to where data [4:56] centers are. um we don't have enough [4:59] compute capacity. [5:01] And if you see these agents that are [5:03] going to autonomously continue to keep [5:05] working round the clock and if you're if [5:09] the ratio is for every human you might [5:11] have 10 to 100 agents, [5:14] a world of 80 8 billion people might [5:16] feel like we have the throughput [5:18] capacity of maybe 80 or 800 billion [5:20] people. Um but the the use case that'll [5:24] completely change the assumptions of [5:25] infrastructure is just starting to come [5:28] up. So you know if you think about the [5:30] use cases that have actually gotten [5:31] volume in AI right now it's coding it's [5:35] customer support all of those use cases. [5:37] Computer use I think as a use case where [5:40] an agent pretends to be a human on your [5:43] laptop and actually conducts tasks for [5:45] you. computer use as a use case will [5:48] have a exponentiality to it on use of [5:52] compute on use of the network. That's [5:54] going to be pretty huge. And then the [5:56] second big area that we should think [5:58] about is there's a there's going to be a [5:59] trust deficit where people just don't [6:02] trust these systems right now. If they [6:03] don't trust these systems, they're not [6:05] going to use them. Safety and security [6:06] is pretty important and those systems [6:09] that that trust has to get earned over [6:12] time by AI from humans in order for [6:15] humans to actually let it in a little [6:18] bit more. The capacity of what AI can do [6:20] even today is far more than what humans [6:23] are allowing it to do today. And I think [6:27] the trust deficit is a big part of it. [6:28] And so those are the two areas, the [6:30] constraints on infrastructure and the [6:32] trust deficit that'll have to get played [6:33] out. [6:35] So you mentioned prompting. Can you talk [6:37] a little bit about the art of asking [6:40] questions to AI systems uh and [6:43] metaprompting because I think this is [6:45] something that is is pretty um practical [6:49] that would help our audience [6:50] >> help everyone. Yeah. So um and you know [6:53] Fay talked about this a little bit on [6:57] if you think about the uh I have a [6:59] 14-year-old daughter right and [7:02] the assumption [7:05] during my entire lifetime was that [7:07] knowledge is really important. You need [7:09] to make sure that you accumulate as much [7:11] knowledge as possible and that's what's [7:13] going to get you differentiated in the [7:14] world. Where we are now is knowledge and [7:17] intelligence is going to get to be in [7:18] abundance and actually you know [7:20] available for free and to some degree [7:22] commoditized. [7:24] But where the art comes in is the nuance [7:28] with which you can ask questions [7:30] right and the quality of questions that [7:32] you ask will determine the quality of [7:36] problems that you can solve. And the [7:38] quality of problems that you solve [7:40] actually is directly proportionate to [7:42] the success of the outcome. The harder [7:44] the problem, [7:45] the higher the probability of success, [7:47] which is completely counterintuitive. [7:49] Why is that? Because typically the [7:52] hardest problems are the ones that [7:53] attract the best people. And business is [7:55] a team sport. And so if you have the [7:57] hardest problems, you'll actually get [7:59] the best amount of, you know, [8:01] probability for a successful outcome. [8:03] So on the prompting side, what's [8:06] actually now starting to happen with [8:07] GPT5, you're starting to see this more [8:10] is this notion of meta prompting [8:12] where you can ask a question, [8:15] but that prompt is asking AI to give you [8:19] a prompt that's going to give you an [8:21] answer for that prompt. [8:23] And try this out sometimes, and it's [8:25] actually quite remarkable. Ask it to go [8:27] generate an image for you with a prompt. [8:30] you know, I I wanted to see if um you [8:32] know, whatever the image might be that [8:34] you want to generate for whatever [8:36] situation, ask AI to generate the image, [8:39] then ask it not to generate the image, [8:41] but to give you a prompt to generate the [8:43] image. And then they will give you a [8:45] prompt and you copy and paste that [8:46] prompt and see the see the quality of [8:49] image you're going to get from that [8:50] metaprompt rather than the prompt that [8:52] you might have provided it. And you'll [8:54] actually see a stark difference and an [8:56] asymmetry with which you can actually [8:59] have this massive step function [9:01] improvement in the output that you get. [9:04] >> You know, [9:04] >> that's that's really cool. I I can't [9:06] wait. [9:06] >> How many have done how many have used a [9:07] metaprompt over here with a show of [9:08] hands? [9:09] >> Wow. [9:10] >> Great. This this is the other thing [9:12] that's crazy is um the the rate of [9:14] adoption of something that's coming out [9:17] is so fast, right? like where um the [9:20] time just gets compressed for the [9:23] ingestion period for humanity on how [9:25] quickly you actually learn something. [9:26] It's because machines are now learning [9:29] the language of humans rather than the [9:31] other way around. And so everything gets [9:33] exponentially faster as a result of it. [9:36] Yeah, it's it's incredible how fast it [9:38] is all moving. In part because of that [9:40] speed though, many of us are concerned [9:43] about what AI will mean for our work and [9:47] for our career prospects and that of our [9:49] children, right, and their children. The [9:52] CEO of Anthropic has warned that AI [9:54] could wipe out half of all entrylevel [9:57] white collar jobs. He said it could [9:59] increase unemployment to 10 to 20% and [10:02] he said that could happen in the next 1 [10:03] to five years. [10:05] What do you say to those who are [10:07] concerned about AI wiping out a [10:11] significant number of jobs? [10:14] So, you know, Dario is is a friend. U we [10:17] are investors in anthropic. Um I have a [10:19] ton of respect for what he's done. I [10:21] think in this area though, I have a [10:23] slightly different opinion on a couple [10:25] different dimensions. one is um [10:31] the job and the configuration will [10:34] change but like I was saying earlier I [10:36] just refuse to believe that humans are [10:38] going to be obsolete like it just seems [10:40] like it's it's kind of an absurd concept [10:42] uh we are far too creative for us to be [10:45] obsolete you know um the second thing is [10:48] this notion of entrylevel jobs going [10:50] away I actually have a huge concern with [10:52] this kind of line of thinking [10:55] take a look at someone who's 20 years [10:57] old [10:59] and see how they use AI and then observe [11:03] how someone who's 28 years old uses AI. [11:06] Not even like a 50-year-old like me, [11:08] like a 28-year-old, right? And you'll [11:10] see a stark difference. And what's the [11:12] difference? The 28-year-old might [11:15] actually use AI in a very taskoriented [11:18] transactional way. Give me an answer for [11:20] something that I'm asking you right now. [11:22] Over time, you might automate workflows. [11:25] The 20-year-old uses it as a sidekick [11:27] and a companion a as a brainstorming [11:29] partner. And when you actually start to [11:34] have generations that are coming in that [11:37] have a very different starting point of [11:39] when they started using this technology. [11:42] Um, you know, I had a I have a [11:43] 14-year-old daughter. Like I said, she [11:45] started using the iPad, shame on us, way [11:49] too early in life. And um, she was like [11:52] two years old or three years old. and [11:54] she she just instinctively knew how to [11:56] kind of vertically scroll, right? First [11:59] time she saw a newspaper, all she did [12:01] was keep tapping [12:05] on the picture and couldn't understand [12:06] why the picture wasn't zooming in, you [12:09] know, because she had a different kind [12:10] of starting point orientation to this [12:13] whole thing. And that's what you want to [12:14] make sure that you inject into the [12:17] workforce. I think aging and death, as [12:20] glib as it sounds, is actually a really [12:22] good thing because it makes way for the [12:25] new to come in, right? And so if you [12:27] just say, "I'm going to eradicate all [12:29] entrylevel jobs," that's the stupidest [12:32] thing a company can do in the long term [12:34] because what you've done is you've [12:36] actually taken away the injection of new [12:40] perspective. And I think sometimes [12:42] experience can be a massive liability. I [12:45] think in there certain jobs where [12:46] experience is really really valuable in [12:48] certain jobs where experience can [12:50] actually hold you back because how many [12:52] times have all of us said at some point [12:53] in time you know what we tried this out [12:55] five years ago it didn't work so we're [12:56] not going to do it again like 50,000 [12:59] assumptions might have changed from 5 [13:00] years ago when you tried it out but your [13:03] experience in that point instinctively [13:05] held you back so what you have to do is [13:07] you have to try to figure out mechanisms [13:09] for unlearning what you've learned over [13:12] time and the the the early in career [13:16] generation actually allows us to [13:18] unlearn. So I spend an enormous amount [13:21] of my time with student interns and [13:23] people who are actually I I will put it [13:26] under the mentorship guys because [13:28] everyone will say okay I want to get [13:29] mentored by someone who's you know [13:32] relatively reasonably successful but I I [13:35] get mentored as much by them as they get [13:37] mentored by me because I learn a lot [13:39] from my 14-year-old daughter. I learn a [13:41] lot from people who've just gotten out [13:42] of college because they have a fresh and [13:45] a unique perspective and that [13:46] perspective when coupled with experience [13:50] makes magic happen. So I feel like it [13:53] would be a really bad strategy to not [13:55] have early and career people and entry- [13:57] level people injected in your workplace. [14:00] I don't think you should do that. I also [14:02] don't think AI is only about [14:04] productivity. In fact, that's in my mind [14:06] the least interesting things about AI is [14:09] getting productive. I think the more [14:11] interesting things about AI is the [14:12] quality of problems that you can solve [14:15] that you couldn't even imagine solving [14:16] before because that perspective didn't [14:19] exist in the human corpus of knowledge [14:21] which now does or now could and as a [14:24] result you'll be able to do a whole lot [14:26] more and I think a big part of that is [14:28] the injection of new perspective that is [14:31] not tainted with experience [14:34] that make sense does I I I love the [14:36] newspaper uh example that you uh that [14:38] you gave. Um, so I know you said that [14:41] you don't think that AI is going to make [14:43] humans obsolete and that it should [14:45] >> I'm curious how many over here think [14:47] that you should just eradicate entry- [14:48] level jobs. Anyone? [14:51] Great. So, you think it's an absurd [14:52] concept, right? Okay, great. I hope it's [14:55] a yes for that one. How many think it's [14:57] an absurd concept? The show of hands. [15:00] Thank you. Okay. [15:02] >> But you did hear Fay say during the last [15:05] session [15:06] >> that there could be there probably will [15:08] be some growing pains like there is [15:10] anytime there's a major new technology [15:13] that's being adopted. [15:15] >> What do those growing pains look like in [15:17] in in your in your mind because I I [15:19] understand what you're saying about [15:19] entry level work and not making humans [15:21] obsolete but it it does feel like [15:23] there's going to be some significant [15:25] impact [15:25] >> there. there's meaningful impact. We as [15:27] a society have to collectively make sure [15:29] that the transition is a smooth one. Um, [15:32] and by the way, every disruption that's [15:33] happened, you've actually seen growing [15:35] pains where people will actually get [15:36] disrupted. You know, PE in tech, we live [15:38] in a bubble. So, we keep thinking, oh, [15:40] in tech disruption is just part of it. [15:42] But when a steel mill worker gets [15:44] disrupted, they don't become an AI [15:47] prompt engineer, [15:50] you know. Um, and so what you start to [15:52] see is there's there's actually a level [15:56] of retraining and reskilling in society [15:59] that has to be done that the tech [16:02] community has to actually take some [16:03] responsibility for and we have to work [16:06] closely with the with the public sector [16:08] and make sure that education and the [16:10] skills gap is actually narrowed because [16:12] if you don't do that you will you will [16:15] create um you know some level of pain in [16:17] society and we we want we want to make [16:18] sure we avoid that. The second thing [16:20] that I think is pretty important is [16:22] safety and security are going to be [16:25] pretty massive consequential and [16:27] potentially catastrophic outcomes if not [16:31] um if not mitigated effectively. For [16:33] example, [16:35] um every tool that we use for AI is the [16:38] same tool that an adversary is going to [16:40] be able to use. That adversary is going [16:42] to have attacks that are going to be far [16:45] more sophisticated. Those adversaries [16:47] might be nation states. uh those attacks [16:50] might be far more bespoke where you used [16:52] to have, you know, an exotic prince from [16:54] an exotic country that was asking you to [16:57] give you their bank account so they [16:58] could just transfer $10 million with a [17:00] bunch of typos in an email which was a [17:02] spam email. [17:04] >> Now might be someone that says, "Hey, it [17:06] was great to see you at the our kids [17:08] basketball game last night. Here are [17:09] some pictures. Click to download." And [17:11] the moment you clicked, there's a piece [17:12] of malware that gets downloaded on your [17:14] machine that can then actually start [17:16] having lateral movement. So what you [17:18] need to do is make sure that safety and [17:20] security is something that is pretty [17:22] importantly dealt with and that has to [17:24] be dealt with not just by using AI for [17:27] cyber defense. You have to secure AI [17:30] itself because these models they [17:32] hallucinate. These models they have [17:33] toxicity. These models can have prompt [17:36] injection attacks. And so how do you [17:38] make sure that those things can get [17:39] mitigated? And so the way in which you [17:43] get visibility on what data is going [17:45] into a model, the way in which you [17:47] actually validate a model so that the [17:48] model is working the way that you expect [17:50] it to work and the way you put runtime [17:53] enforcement guard rails, the way that an [17:55] agent has a different kind of identity [17:58] that you can validate that can talk to [18:00] another agent but has fine grain control [18:02] because you don't want to have an agent [18:04] say, "Okay, get access to my email and [18:06] send an email to everyone." You might [18:08] say, "I want to have an agent be able to [18:10] craft an email on my behalf, but not to [18:12] my boss and my board of directors, but [18:14] it's okay to do it to my team." And [18:16] those are the kind of things that we [18:17] have to make sure that we actually think [18:19] through. And that risk mitigation on [18:21] safety and security will be a pretty big [18:22] one that we as a collective society have [18:24] to solve for as well. [18:25] >> Yeah, absolutely. Thank you so much, Ju, [18:28] for doing this. Uh we really appreciate [18:29] all the insights. This was great. [18:31] >> Thank you so much. Cheers.