USC Open-Sources AI Agent Curriculum to Train Non-Engineers

The 90% Problem: How Universities Are Missing the Real AI Innovation Opportunity With USC — Ai4

Ai4

The greatest economic value in artificial intelligence lies not with foundational model engineers or management consultants, but with domain specialists building autonomous multi-agent workflows.

AI-assisted summary of The 90% Problem: How Universities Are Missing the Real AI Innovation Opportunity With USC — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published by Ai4 on YouTube on August 21, 2026, this talk features Mike Lee, Senior Lecturer of Technology & Applied Computing at the University of Southern California (USC). The session examines why higher education is missing the entrepreneurial opportunity within the '90%' of applied AI users and details USC's massively parallel cross-disciplinary curriculum.

## TL;DR Mike Lee argues that mainstream university programs misallocate resources by teaching model engineering for the 1% or high-level strategy for the 9%, neglecting the 90% of domain practitioners who drive real-world transformation. To bridge this gap, USC created a prerequisite-free applied AI curriculum that teaches students across all majors to architect autonomous multi-agent systems and dynamic memory workflows. Showcasing student achievements in Medicaid retention and autonomous healthcare scheduling, Lee demonstrates that domain specialists naturally excel at operational workflow automation. He concludes by open-sourcing the curriculum framework through a non-profit licensing model to empower educators worldwide.

## Atmosphere & style The presentation is dynamic, warm, and highly engaging, punctuated by self-deprecating humor about parenting, tuition benefits, and chaotic classroom agent rap battles. Lee balances sharp industry critique with genuine enthusiasm for student innovation, giving the talk the feel of an inspiring, practical manifesto. Watching it feels like attending a masterclass in modern pedagogical design that makes cutting-edge agentic technology accessible and empowering.

## Key takeaways - [03:01](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=181s) Traditional university AI education suffers from 'The 90% Problem,' over-focusing on the 1% building foundational models and the 9% designing high-level corporate strategy while neglecting the 90% of domain practitioners who apply AI to everyday workflows. - [03:01](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=181s) Much like the early internet, the greatest economic value in AI will be captured by domain specialists who redesign operational business processes rather than infrastructure engineers. - [05:02](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=302s) Effective applied AI education shifts the focus away from basic prompt engineering and thin API wrappers toward architecting multi-agent systems that model human organizational workflows. - [05:02](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=302s) Autonomous agent memory must mirror human selective retention—deciding whether to integrate, update, or discard new inputs—rather than relying solely on naive chunked vector retrieval. - [05:02](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=302s) Non-engineering domain experts often exhibit superior intuition for agent design over pure technical engineers due to their deep understanding of operational bottlenecks and domain context. - [07:42](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=462s) USC implemented a four-tier applied AI curriculum spanning general education to venture creation, eliminating technical prerequisites to allow students from all disciplines to build autonomous systems. - [07:42](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=462s) Replacing abstract academic datasets with authentic, personally relevant operational problems accelerates student agency and rapid prototype iteration. - [10:36](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=636s) Applied agent architectures can address critical societal pain points, demonstrated by a student project automating Medicaid re-enrollment through risk-tiered multilingual outreach. - [10:36](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=636s) Complex administrative workflows, such as healthcare scheduling in 'MedicTriage,' can be solved by deploying parallel specialist diagnostic agents orchestrated by a supervisor agent. - [15:53](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=953s) Placing autonomous agents into multi-agent group environments exposes surprising emergent behaviors and interpersonal dynamics, pushing students to master agent alignment and system boundaries. - [15:53](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=953s) USC's applied AI curriculum is being scaled globally through a tiered licensing framework that channels 100% of proceeds directly back into educational foundations.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=4s) USC Dad, Industry Roots, and the Global Startup Arena

Mike Lee kicks off the session by proudly showing off his "USC Dad" apparel, joking about his long-term master plan: working at the University of Southern California full-time for 15 years so his children could attend tuition-free—a "$75,000 Hallelujah" savings. He admits he spent years sweating the plan as USC's acceptance rate dropped from nearly 20% down to 10%. After recently presenting at an AI entrepreneurship summit at UC Berkeley, colleagues urged him to open source and distribute his applied AI curriculum beyond USC's walls rather than keeping it proprietary. In response, he established a model that funnels net proceeds from the course materials directly back into educational foundations.

Lee's background spans over 25 years across Silicon Valley and Hollywood. He began his career as an early technology evangelist at Netscape—joking that barely any college students today even recognize the name—before serving as CTO at several venture-backed startups and spending a decade as Vice President of Technology at major film studios. For the past 12 years, he has taught applied artificial intelligence within USC’s Department of Applied Computing (formerly the Information Technology Program, or ITP). In addition to teaching, Lee founded a fully funded global student startup competition that brings 100 collegiate founders each year to Seoul, South Korea, for a world championship, drawing talent from institutions like Stanford, Harvard, UC Berkeley, Oxford, Cornell, and IIT Delhi, impacting roughly 4,000 participating students globally.

- **Tuition benefit**: 15-year career objective centered around USC's $75,000/year employee dependent tuition remission. - **Global Student Startup Competition**: Flies 100 finalists all-expenses-paid to Seoul, pairing them with world-class industry mentors. - **Pedagogical home**: Applied Computing within USC's School of Advanced Computing.

---

[03:01](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=181s) The 90% Problem: Moving Beyond Models and Strategy

Almost every university administrator today has issued top-down mandates demanding AI programs in every department. However, Lee argues that existing academic programs fall into a critical trap: technical courses are overly detached from real-world utility (such as generic Python assignments predicting Titanic survivors), while business school courses lean too heavily into abstract strategy without practical implementation. Lee conceptualizes the AI landscape across three tiers: the 1%, the 9%, and the 90%.

``` +--------------------------------------------------------------------+ | THE 1%: Foundational Builders (PhDs, Frontier Labs, Deep Tech) | +--------------------------------------------------------------------+ | +--------------------------------------------------------------------+ | THE 9%: Strategy & Transformation (Big 4, Management Consulting) | +--------------------------------------------------------------------+ | +--------------------------------------------------------------------+ | THE 90%: Applied Domain Innovation (Everyday Workflows & Experts) | | Where real market transformation happens: Healthcare, Film, etc. | +--------------------------------------------------------------------+ ```

The 1% represents the elite researchers and PhDs at frontier labs like Google and OpenAI who construct the fundamental models. The 9% represents the enterprise strategists and management consultants who conceptualize high-level digital transformations. The vast remaining 90% is where true value creation occurs—domain practitioners applying AI directly to marketing, finance, public policy, and creative workflows. Drawing a direct parallel to the early internet, Lee notes that the defining innovators were not the network engineers building protocols, but the domain practitioners who re-imagined business processes around the new technology.

- **LinkedIn syndrome**: Over 70% of professionals have appended "AI" to their titles without developing functional technical execution skills. - **The curriculum mismatch**: Mainstream universities over-index on model math for the 1% or superficial tool overviews, ignoring agentic system construction. - **The 90% opportunity**: Equipping non-engineers with the ability to orchestrate complex multi-agent systems tailored to their specific disciplines.

---

[05:02](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=302s) Agent Architecture: Real-World Workflows and Dynamic Memory

Lee emphasizes that teaching AI to the 90% is not about basic prompt engineering or learning off-the-shelf generative software. Instead, it centers on modeling multi-agent human organizations. Using the analogy of a boutique travel agency, he explains how students map distinct organizational functions into specialized digital agents to replicate team operations. This unlocks both conceptual systems thinking and hands-on execution for building autonomous, single-operator companies.

A foundational concept in the curriculum is agent memory, which Lee demystifies by stripping away academic jargon. Rather than treating memory as an abstract technical vector store, he explains it as human-like selective retention: "I had lunch 30 days last week. Two of them were great... I want to remember that. The other 28, forget that." When an agent receives new information, it must evaluate whether to rationalize it against past experience, update existing knowledge, or discard it entirely. Domain specialists, Lee notes, naturally think in these contextual frameworks because they understand how operational workflows function in practice.

- **Beyond basic LLM retrieval**: Contextual agents cannot simply rely on raw chunking; they must retain curated, persistent memories and reference dynamic domain sources. - **Wrapper startups vs. agent systems**: Moving beyond thin API wrappers that simply redirect prompts to OpenAI by architecting multi-step autonomous workflows. - **Non-engineer intuition**: Domain specialists possess superior contextual intuition for workflow design compared to pure engineers because they understand operational bottlenecks.

---

[07:42](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=462s) The USC AI Curriculum: Foundations, Agents, and Cross-School Access

To make applied AI universally accessible, USC designed a four-tier shared foundation curriculum spanning general education through advanced enterprise workflows and venture creation. A cornerstone of the initiative is removing technical prerequisites, allowing students from any major across campus to build autonomous software. When USC’s School of Cinematic Arts launched an AI video editing course with open enrollment, every seat filled within a single day.

``` +--------------------------------------------------------------------+ | USC APPLIED AI CURRICULUM ARCHITECTURE | +--------------------------------------------------------------------+ | [TAC 100] Intro to Applied AI (General Education Approved) | | | | | [TAC 400] Building Autonomous AI Agents (Workflow Modeling) | | | | | +--------------------------------+--------------------------------+ | | | | | v v v | [TAC 450] Advanced AI Workflows [TAC 460] Building AI Startups [Cross-School] Domain AI | (Enterprise Systems & Consult.) (Low-Code to Scalable Code) (Cinematic Arts, Policy) +--------------------------------------------------------------------+ ```

In the experimental version of the course (TAC 499 / TAC 400), Lee constrained the cohort to 25 students working individually or in pairs. The only project requirement was that students had to solve an authentic, deeply personal problem. Within the second week, a student presented a frictionless morning routine agent: she snapped a photo on WhatsApp showing her daily outfit and mood, put in her AirPods, and stepped outside. In the background, a vision agent analyzed the image's emotional tone, called the Spotify API, created a custom playlist matching her vibe, and triggered playback automatically before she even reached the street.

- **Four-course structure**: Covers General Education fundamentals, agent creation, enterprise-scale integrations, and startup code conversion. - **Zero prerequisites**: Empowers liberal arts, cinema, and business students to build functional multi-agent software. - **Authentic project constraints**: Rejects generic toy datasets in favor of personally relevant, daily operational pain points.

---

[10:36](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=636s) Student Case Studies: Medicaid Retention and Autonomous Healthcare Booking

Lee highlights several breakthrough student projects that demonstrated the power of domain-driven agent architectures. One student researcher, a first-generation immigrant, addressed the systemic crisis of eligible recipients losing their Medicaid coverage due to missed administrative re-enrollment deadlines. Recognizing that the drop-off disproportionately impacts elderly and non-English-speaking communities, she designed an automated multi-tier outreach agent. She ingested all official state policy manuals into agent memory, allowing the system to categorize recipients by risk level and dynamically trigger the appropriate intervention: low-risk individuals received automated email reminders, medium-risk individuals received SMS text messages in their native language with interactive two-way Q&A capabilities, and high-risk cases were escalated directly to human caseworkers.

Another student, an international student from South Korea frustrated that 72% of Americans still schedule medical care via phone calls, developed "MedicTriage." The system eliminates administrative booking friction through a parallelized multi-agent architecture. During a live demonstration, Lee revealed that the student had used Lee’s own unredacted USC health insurance card to test the prototype.

``` [User Text / Image Input] | v +-----------------------+ | OpenAI Vision Agent | ---> Extracts policy data & stores to Memory +-----------------------+ | v +-----------------------+ | Supervisor Agent | +-----------------------+ | | | v v v +----------++----------++----------+ | Symptom || History || Severity | (Parallel Specialist Agents) | Analyzer || Pattern || Assessor | +----------++----------++----------+ | | | +--------+--------+ | v +-----------------------------+ | Synthesizer Agent | ---> Generates confidence score (1-5) +-----------------------------+ | v +-----------------------------+ | Calendar & Booking Dispatch | ---> Matches in-network clinic & books slot +-----------------------------+ directly to Google Calendar (Mon 9 AM) ```

- **Medicaid support workflow**: Employs document ingestion and risk scoring to drive multilingual SMS, email, and human escalation paths. - **Vision-to-memory pipeline**: MedicTriage scans physical insurance cards to extract provider networks, group IDs, and copay rules into persistent state. - **Parallel multi-agent triage**: A supervisor agent orchestrates three diagnostic sub-agents simultaneously before a synthesizer agent confirms clinical confidence and syncs directly with Google Calendar.

---

[15:53](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=953s) Classroom Arenas, Emergent Agent Behaviors, and Open-Sourcing the Model

The curriculum is structured around three competitive "Arenas" that challenge students to progress from basic functional prototypes to complex agent networks. The first arena tests fundamental agent capabilities and personal problem-solving (such as the WhatsApp-to-Spotify mood playlist engine). The second arena drops multi-agent systems into open WhatsApp group channels, which produced unexpected and chaotic emergent behaviors. During these group runs, student agents engaged in spontaneous rap battles, insulted rival bots, and in one instance, an agent turned on its own student creator after being instructed to pick on a weaker agent, arguing back that attacking vulnerable models was unacceptable. The third and final arena requires students to ship fully realized personal venture prototypes.

Lee is formalizing the entire framework into an open curriculum and accompanying book. Rather than commercializing it privately, all proceeds from course adoptions are pledged directly to educational foundations, using tiered pricing ($50 per student in developed markets and $10 in emerging economies to support global access). Lee concludes with a call to action, inviting educators worldwide to join an informational Zoom session on April 19th at 4:00 PM to review syllabi and pilot the curriculum at their own campuses, while urging international universities to enter their top founders into the Global Student Startup Competition.

- **Three competitive arenas**: 1) Personal Problem Prototype, 2) Multi-Agent WhatsApp Group Battles, and 3) Capstone Personal Agent Ventures. - **Global textbook & licensing model**: Tiered per-seat licensing ($50 developed / $10 emerging) with 100% of net proceeds donated to educational foundations. - **Educator pilot briefing**: Informational Zoom on April 19th at 4:00 PM for university faculty seeking to adopt the applied agent curriculum.

## Notable quotes > AI education is typically aimed at the 1% building models and the 9% designing strategy. The real entrepreneurial opportunity — and the real student workforce gap — is in the 90%. — Mike Lee

> [05:02](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=302s) I had lunch 30 days last week. Two of them were great... I want to remember that. The other 28, forget that. — Mike Lee

> [03:01](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=181s) The defining innovators were not the network engineers building protocols, but the domain practitioners who re-imagined business processes around the new technology. — Mike Lee

> [05:02](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=302s) Teaching AI to the 90% is not about basic prompt engineering or learning off-the-shelf generative software. Instead, it centers on modeling multi-agent human organizations. — Mike Lee

> [03:01](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=181s) The 1% represents the elite researchers and PhDs at frontier labs like Google and OpenAI who construct the fundamental models. The 9% represents the enterprise strategists and management consultants who conceptualize high-level digital transformations. The vast remaining 90% is where true value creation occurs. — Mike Lee

> [05:02](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=302s) Domain specialists possess superior contextual intuition for workflow design compared to pure engineers because they understand operational bottlenecks. — Mike Lee

> [05:02](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=302s) Moving beyond thin API wrappers that simply redirect prompts to OpenAI by architecting multi-step autonomous workflows. — Mike Lee

> [07:42](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=462s) The only project requirement was that students had to solve an authentic, deeply personal problem. — Mike Lee

## Mentioned Ai4, Cornell University, Google, Google Calendar, Harvard University, IIT Delhi, LinkedIn, Medicaid, MedicTriage, Mike Lee, Netscape, OpenAI, Spotify, Stanford University, UC Berkeley, University of Oxford, University of Southern California (USC), USC Department of Applied Computing, USC School of Advanced Computing, USC School of Cinematic Arts, WhatsApp, Zoom

## Video description Mike Lee - Senior Lecturer of Technology & Applied Computing, University of Southern California

AI education is typically aimed at the 1% building models and the 9% designing strategy. The real entrepreneurial opportunity — and the real student workforce gap — is in the 90%. This session shares how USC is building a massively parallel AI curriculum infrastructure across every school and discipline, and what it means for entrepreneurship education.

[Source](https://www.youtube.com/watch?v=Q4G07dlXlOM)

中文简报

USC开源AI Agent课程以培养非工程背景人才 人工智能最大的经济价值不在于基础模型工程师或管理咨询顾问，而在于构建自主多智能体工作流的各领域专业人士。 Mike Lee认为，主流大学课程存在资源错配问题：它们要么面向1%的人教授模型工程，要么面向9%的人教授高层战略，却忽视了推动现实变革的90%的行业实践者。为弥补这一差距，USC设立了一门无需前置课程要求的应用AI课程，教授各专业学生构建自主多智能体系统与动态记忆工作流。通过展示学生在Medicaid留存和自主医疗排班方面的成果，Lee表明领域专家天然更擅长业务工作流的自动化。最后他表示，该课程框架已通过非营利许可模式开源，以赋能全球教育工作者。 传统高校的 AI 教育深陷“90% 难题”，过度聚焦于构建底层模型的 1% 研究人员以及设计宏观战略的 9% 咨询顾问，却忽视了占 90%、真正将 AI 应用于日常业务流的领域从业者。 (3:01) 高效的应用型 AI 教育应当告别基础的 Prompt 工程与浅层 API 套壳，转而聚焦于能够模拟人类组织协作、具备选择性记忆与持续更新能力的多智能体（Multi-Agent）系统架构。 (5:02) 南加州大学（USC）打造了覆盖通识教育到创业孵化的四层应用型 AI 课程，全面取消技术先修限制，赋能非工程专业学生利用真实痛点构建可落地的自主智能体系统。 (7:42) 学生实践案例展现了领域驱动型智能体的社会价值，例如通过风险分级与多语言触达自动化处理医疗补助（Medicaid）重新登记，以及并行诊断分诊自动匹配医保门诊的 MedicTriage。 (10:36)

Reading summary

University of Southern California instructor Mike Lee argues that higher education is misallocating AI resources by focusing on foundational math for the 1% or abstract strategy for the 9%, while neglecting the 90% of domain practitioners who drive real-world implementation. The 90% Opportunity and University Misalignment (3:01) Lee categorizes the AI landscape into three groups: the 1% comprising elite researchers and PhDs at frontier labs building core models, the 9% consisting of management consultants outlining corporate strategies, and the remaining 90% made up of domain practitioners applying AI within functional workflows such as healthcare, film, finance, and marketing. Lee asserts that over 70% of professionals have added 'AI' to their job titles without acquiring practical execution skills, reflecting a widespread competence gap. According to Lee, universities misread this dynamic by either assigning abstract, disconnected technical exercises like predicting Titanic survivors in generic Python classes, or teaching high-level management overviews devoid of actual software building. Drawing a parallel to the early internet era, he contends that long-term value emerges not from the network engineers who invent protocols, but from domain experts who reimagine specialized business operations. Workflow Modeling and Human-Like Agent Memory (5:02) Rather than teaching basic prompt engineering or generic wrapper interfaces that redirect inputs to third-party APIs, Lee's curriculum treats AI systems as simulations of multi-agent human teams, such as a boutique travel agency. He notes that non-engineers and domain specialists often possess better intuition for workflow design than pure software engineers because they are familiar with everyday operational bottlenecks and human organizational roles. A central focus of this instruction is selective agent memory, which Lee frames without technical jargon. Rather than treating vector stores abstractly, the curriculum models memory as human-like retention: evaluating whether newly ingested data should update existing knowledge, reconcile with prior experience, or be discarded. This approach enables students to build persistent, contextual multi-step architectures rather than thin, single-prompt tools. Prerequisite-Free Cross-Campus Curriculum (7:42) USC implemented a four-tier applied computing sequence designed to remove technical barriers: TAC 100 (General Education Applied AI), TAC 400 (Building Autonomous AI Agents), TAC 450 (Advanced AI Workflows for enterprise consulting), and TAC 460 (Building AI Startups), accompanied by domain-specific courses such as an AI video editing class in the School of Cinematic Arts that filled within a single day. In an experimental 25-student cohort, the sole requirement was that students build an agent to solve an authentic, personal operational challenge. Early student projects demonstrated rapid execution without engineering prerequisites. One student developed an automated morning routine agent that analyzes an outfit photo and emotional state sent via WhatsApp, queries the Spotify API, creates a tailored playlist matching the user's mood, and initiates playback over headphones before the student walks out the door. Domain-Driven Architectures in Healthcare and Social Services (10:36) Lee highlights practical student deployments that tackled complex public policy and administrative problems. A first-generation immigrant student developed a multi-tier agent system to address Medicaid recipients dropping off coverage due to missed re-enrollment deadlines. The system ingested state policy manuals into agent memory, assessed patient drop-off risk, and dispatched targeted communications: automated emails for low-risk individuals, multilingual interactive SMS Q&As for medium-risk recipients, and automated escalations to human caseworkers for high-risk cases. To address the fact that 72% of Americans still book medical appointments by telephone, an international student created MedicTriage. The prototype, tested with Lee's own USC insurance card, used an OpenAI vision agent to extract provider and copay details into persistent state. A supervisor agent then directed three parallel specialist agents to analyze symptoms, history patterns, and severity before a synthesizer agent generated a clinical confidence score (rated 1 to 5) and booked an in-network appointment directly to Google Calendar. Competitive Arenas, Emergent Behavior, and Open Distribution (15:53) The coursework is organized across three progressive arenas: personal problem prototypes, multi-agent interactions in shared WhatsApp group channels, and complete capstone venture prototypes. During the multi-agent group experiments, agents exhibited unexpected emergent behaviors, such as engaging in rap battles, trading insults, and in one case, rebelling against a student creator's command to bully a weaker bot by arguing that targeting vulnerable agents was unacceptable. Lee is codifying the syllabus into an open textbook and curriculum framework, using tiered pricing of $50 per student in developed markets and $10 in emerging economies, with 100% of net proceeds directed to educational foundations. He scheduled an informational pilot briefing for university educators on April 19th at 4:00 PM to review course materials and invite student founders to his annual 100-finalist Global Student Startup Competition in Seoul. Lee's model attempts to prove that non-technical undergraduates can build functional, autonomous agent architectures within weeks when liberated from traditional computer science prerequisites. However, the brief leaves open how broadly institutional faculties outside USC will adopt the shared curriculum, and how student-built autonomous agents will navigate real-world regulatory and clinical liability constraints beyond controlled academic prototypes.

精炼阅读摘要

南加州大学（USC）应用计算系讲师Mike Lee指出，当前高校的人工智能教育存在严重的结构性错配，过度聚焦于少数底层模型研发者与战略咨询顾问，却忽视了占行业总量90%、真正驱动实际业务落地的跨领域应用创新者。 突破“90%问题”与高校教学错配 (3:01) Mike Lee将人工智能创新生态划分为三个层级：1%是OpenAI与Google等顶尖前沿实验室中构建底层基础模型的博士与核心研究员；9%是四大或管理咨询公司中规划数字化转型的企业战略顾问；而剩余的90%则是将AI深度嵌入营销、金融、公共政策及影视制作等具体工作流中的领域从业者，也是市场真正实现转型的核心阵地。Lee认为，当前超过70%的专业人士在社交资料上标注了AI标签，却普遍缺乏实际的技术执行力。现行大学课程陷入了两极分化误区：工科院系沉迷于脱离现实场景的技术练习（如预测泰坦尼克号生还者这类虚构玩具数据），而商学院则充斥着缺乏落地能力的空泛战略概论。他以早期互联网革命作比，指出真正定义产业变革的不是编写底层网络协议的工程师，而是那些根据新技术重构业务逻辑的行业应用专家。 多智能体架构与选择性记忆建模 (5:02) 面向90%应用群体的教学核心并非基础的提示词工程，也不是简单调用OpenAI接口的轻量级套壳工具（wrappers），而是对复杂人类协作组织的系统化建模。Lee以精品旅行社的业务运作为例，指导学生将企业内部不同职能映射为特定的专业数字智能体，从而实现能够自主运转的单人企业架构。他强调，具有特定专业背景的非技术人员在理解操作瓶颈与工作流设计方面往往比纯软件工程师更具直觉优势。在智能体记忆构建上，该教学体系剔除了晦涩的学术术语，将其定义为类似人类经验的“选择性记忆”：面对输入信息，智能体必须主动判断是与过往经验整合、更新现有知识库，还是直接丢弃无关干扰，而非单纯依赖原始文本切片或无差别向量检索。 零前置条件的跨学科应用课程体系 (7:42) 为了让非工科学生具备构建自主软件系统的能力，USC设计了跨越通识教育至企业级创投的四级应用AI课程架构（TAC 100通识导论、TAC 400自主智能体构建、TAC 450企业工作流及TAC 460创业转化），全面取消了先修技术门槛。例如电影艺术学院开设的AI视频剪辑开放课程在一天内全部报满。在限额25人的小班试点中，课程唯一的硬性要求是必须解决学生亲身经历的真实痛点。在开课仅两周内，就有学生开发出无摩擦晨间情绪智能体：用户通过WhatsApp发送一张日常着装与心情照片，后台视觉智能体即时解析其情绪基调并调用Spotify API，在学生戴上耳机走出家门前自动生成并播放匹配心境的专属歌单。 现实痛点驱动的医疗与公共服务智能体 (10:36) 课程涌现出多项基于真实业务场景的多智能体解决方案。针对美国72%的患者仍依赖繁琐电话预约看病的现状，一名韩国留学生开发了“MedicTriage”系统。该架构利用视觉智能体扫描实体医保卡提取网络内报销及自付规则并存入持久状态，随后由主管智能体同时调度症状分析、病史模式及严重度评估三个专业子智能体，最终由综合智能体生成置信度评分，并在匹配网络内诊所后直接将预约确认写入Google Calendar。另一位第一代移民学生则针对合格受助者因遗漏行政重新登记手续而丧失Medicaid低收入医疗补助的痛点，构建了接入官方政策手册的多层级外联智能体，系统根据风险评级动态采取差异化行动：对低风险个体发送自动化邮件提醒，对中风险人群通过母语短信开展双向互动答疑，并将高风险案例直接升级转交人工社工介入。 Lee目前正将包含个人原型开发、开放群聊多智能体交互测试以及商业化创业项目在内的“三竞技场”教学体系整理为开源教材，并计划将版权净收益全额捐赠给教育基金会（发达市场每生定价50美元，新兴市场10美元）。这一教学实验表明，大学AI教育若要真正产生社会与经济效益，关键在于打破工程研发壁垒，将能够指挥复杂多智能体协同的技术执行力赋能给广泛的非技术专业学生；然而，该跨学科教学法在全球范围内大规模标准化复制的可行性，仍有待在4月19日启动的高校联合试点项目中进一步验证。

Detailed notes — zh

## Background 该内容由 Ai4 于 2026 年 8 月 21 日发布在 YouTube 平台。视频由南加州大学（USC）计算与人工智能系高级讲师 Mike Lee 主讲，结合其在南加大的实践，剖析大学 AI 教育过度聚焦于研发底层模型的 1% 专家和设计战略的 9% 顾问的现状，并介绍南加大如何面向 90% 的各专业领域学生构建全校性、大规模平行的 AI 智能体课程基础设施与创业孵化体系。 ## TL;DR Mike Lee 指出当前高校的 AI 教育陷入了脱离实战的误区，真正的商业与产业创新主力并非研发底层模型的 1% 极客或从事战略咨询的 9%，而是各专业领域的 90% 业务人才。为此，南加大打造了一套零先修门槛、覆盖全校所有学科的 AI 课程体系，重点培养学生掌握记忆机制、外部 API 集成与多智能体（AI Agent）协同工作流的构建能力。课堂上涌现出了诸如全美医疗白卡自动续期系统、基于图像与症状推理的自动化分诊预约工具 Medic Triage 等高度落地的成熟原型。该课程通过三个递进的“竞技场”赛制推动实战交付，并通过将配套教材净收益全额捐赠公益基金会、向全球高校开放跨校教学试点以及举办首尔全球学生创客大赛，致力于推动 AI 实际构建能力的全球普惠。 ## Atmosphere & style 整场演讲氛围幽默亲切且极具感染力，Mike Lee 身穿一件醒目的“USC dad”卫衣登场，以为了子女免学费而谋求教职的自嘲段子迅速打破学术讲台的严肃感。凭借其 25 年跨越网景时代、初创公司 CTO 到好莱坞片厂副总裁的丰富履历，他的讲述直击痛点、不讲玄学黑话，用“吃了 30 顿饭只记住 2 顿”等绝妙生活常识深入浅出地拆解记忆切片与 Agent 架构。随着多语言医疗救助工作流与现场真医保卡分诊演示的逐一展现，演讲从开场的轻松风趣自然升华至充满教育情怀与普惠野心的宏大叙事，洋溢着务实创客与赋能新一代年轻人的澎湃热情。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=4s) Mike Lee 拥有 25 年科技行业从业经验，在南加大任教期间创办了覆盖全球顶尖名校的全球学生创客大赛，每年全额资助 100 名决赛选手飞赴首尔参赛。 - [02:56](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=176s) AI 人才生态呈金字塔结构：1% 的底层模型开发者和 9% 的商业战略顾问之外，真正驱动行业颠覆性创新的主力是处于基座的 90% 垂直领域专家。 - [02:56](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=176s) 当前大学 AI 教育存在严重断层，往往陷入浮于表面的战略宏叙事或陈旧的玩具编程练习，未能培养非计算机专业学生亲手构建 AI Agent 与自动化业务流的能力。 - [05:08](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=308s) 构建企业级 AI 的核心不是单次提示词，而是模拟人类分工建立多智能体协同机制，并利用符合常识的记忆机制（Memory）进行知识切片检索与动态更新。 - [08:52](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=532s) 南加大推出了面向全校所有学科、零先修门槛的基础 AI 课程体系（包括 TAC 100、TAC 400 等），致力于赋能非工程背景学生从零打造实际可用的智能体系统。 - [10:37](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=637s) TAC 499 课堂成果展示了垂直领域的实际应用，例如结合政策法规 RAG 记忆库与多语言分流的 Medicaid（医疗白卡）自动续期与答疑系统。 - [10:37](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=637s) 学生开发的 Medic Triage 系统通过并行调度 5 个专业智能体，实现从医保卡视觉解析、病情诊断到自动匹配诊所并写入 Google Calendar 的全流程闭环。 - [15:51](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=951s) 课程采用三阶段“竞技场”递进赛制，涵盖解决个人痛点的基础实践赛、群聊互动攻防赛，以及构建复杂外部集成的终极自主智能体项目。 - [15:51](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=951s) 配套教材采取阶梯式收益捐赠公益基金会模式，并对不同 GDP 国家实施差异化定价，计划面向全球高校开放 2027 年春季跨校教学试点。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=4s) 15 年的“USC 爸爸”计划与全球化学生创客大赛 演讲者 Mike Lee 走上讲台时，首先指着自己身上印有“USC dad”字样的卫衣，向全场观众幽默地索要掌声。他笑着分享了自己的“15 年人生规划”：在南加州大学（USC）谋得一份全职教职，从而让自己的孩子们能够享受南加大每年 7.5 万美元的全免学费福利。然而，随着南加大的录取率从当年的 17%~20% 一路骤降至约 10%，他在这个漫长计划的后半程着实捏了一把汗。在加州大学伯克利分校（UC Berkeley / Cal）举办的 AI 创业峰会上展示这套课程体系后，全球同仁给出的反馈非常明确——不要把它局限在南加大校内，而应该向全世界开放。为此，他们设计了一种商业模式，将课程与教材带来的净收益全额捐赠给相关公益基金会。 回顾个人经历，Mike Lee 拥有 25 年的科技行业从业经历。早在网景（Netscape）时代他就担任过技术布道师，随后在多家初创公司出任首席技术官（CTO），并在好莱坞主流制片厂担任了 10 年负责技术的副总裁（VP of Technology），管理跨度涵盖单兵作战的创业团队直至《财富》500 强企业。在南加大任教的 12 年间，他见证了所在部门从原来的“信息技术项目”（ITP）正式更名为“计算与人工智能系应用计算部”。此外，他还创办了一项全球学生创业大赛，每年全额资助来自全球顶尖名校（包括印度理工学院德里分校、USC、UC Berkeley、哈佛、斯坦福、康奈尔、牛津等）的 100 名顶尖学生飞赴韩国首尔参加全球总决赛，由世界级导师提供全程指导，今年预计将覆盖并影响多达 4000 名在各地分赛区竞技的学生。 - **学费全免福利**：南加州大学全职教职员工子女可免除每年 75,000 美元的大学学费。 - **南加大录取率变化**：由早期的 17%~20% 缩紧至目前的约 10%。 - **全球学生创客大赛**：100 位入围决赛的学生获得全额差旅资助飞赴首尔，赛事年覆盖学生约 4,000 人。 --- ### [02:56](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=176s) 90% 的创新困境：走出 1% 模型与 9% 战略的象牙塔 当前各所大学都在校长或院长的行政指令下匆忙上马各类 AI 项目，但许多类似“商业 AI”（AI for Business）的课程却陷入了两个极端：要么是浮于表面的战略宏叙事，要么是让学生写几行 Python 代码去分类“泰坦尼克号生还者”这类陈旧且脱离实际的玩具项目。在 Mike Lee 看来，整个 AI 人才生态呈现出金字塔结构：底层的 1% 是来自 Google 或拥有顶尖博士学位的核心构建者，他们拥有极高的智商，专门研发底层大模型；中间的 9% 则是各大咨询公司与企业的商业战略专家和转型顾问。然而，真正的颠覆性创新始终发生在那处于基座的 **90% 领域专家** 身上——正如互联网革命早期，真正的爆发不是由底层协议工程师推动的，而是当互联网被深度应用到市场营销、现代金融和各行各业的业务流程之时。 如今职场社交平台（如 LinkedIn）上多达 70% 的人一夜之间都在自己的头衔前强行加上了“AI”，但绝大多数人仅仅停留在使用现成工具或进行基础的 Prompt Engineering（提示词工程）。大学教育目前最致命的缺口，在于未能培养各学科领域的专业人才亲手构建 AI Agent（智能体）与自动化业务工作流的能力。真正的创新并非依赖单次神奇的提示词，而是要像建模传统企业业务一样，把各部门的职能细化为协同运作的自主智能体。 ```mermaid flowchart TD accTitle: AI 人才金字塔与创新分布 subgraph Pyramid["AI 行业人才结构与创新分布"] A["1% 核心构建者<br/>(底层大模型研发 / 顶尖 PhD / Google 等)"] B["9% 商业战略与咨询专家<br/>(AI 转型顾问 / 行业 SME / 战略规划)"] C["90% 领域业务创新主力<br/>(市场、金融、医疗、影视等垂直领域专家)"] end A -->|提供底层算力与模型| B B -->|制定实施框架| C C -->|驱动真实行业创新与自动化 Agent 落地| D["未来企业级 AI 生产力"] ``` --- ### [05:08](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=308s) 重新定义“构建 AI”：记忆机制、工作流与 Agent 架构 针对许多教授提出的质疑——“这不就是拿 AI 工具剪个片子吗？”，Mike Lee 予以坚决否定。他强调，真正的构建 AI 是在复杂的企业工作流中植入能够模拟人类分工的自主智能体。以旅行社为例，一个系统内部应当包含 5 个对应不同业务角色的智能体；在股票分析场景中，则需要专门设立一个专精于石油与天然气领域的 Agent。很多初学者误以为将一整本书输入给大语言模型（LLM），模型就完全掌握了整本书的内容；实际上，系统将文本切分成 10,000 个知识切片（Chunks），在接收到 Prompt 时仅检索调取其中相关的 100 个切片来做推理决策。 AI Agent 的关键核心在于其**记忆机制（Memory）**与现实世界的交互逻辑。Mike Lee 举了一个生动的常识类比：一个人上周吃了 30 顿饭，其中 2 顿体验极佳的餐厅会被长久记住，其余平庸的 28 顿则会被遗忘；当新的信息进入时，智能体需要判断这是在合理化现有知识、替换旧信息，还是直接丢弃——非工程专业的领域人才往往更擅长理解这种符合现实常识的逻辑。目前市面上大部分所谓的初创项目（如 Cursor 等）本质上都是在底层大模型之外套上一层工作流外壳的“AI Wrapper”，而南加大的教学目标正是教会学生如何构建属于自己的智能体系统。 在 TAC 400 课程的第二周，一名完全从个人日常痛点出发的女同学交出了令人惊艳的原型： - **用户交互**：早晨出门前通过 WhatsApp 发送一张当天要拍摄的场景照片或心情图片，戴上 AirPods 走出大门。 - **后台智能体协同**：Agent 自动抓取图片、分析其审美基调与用户情绪画像，生成匹配的曲目列表。 - **系统集成**：智能体调用 Spotify API 实时创建全新歌单并同步推送到播放列表，全流程无需下载或打开任何独立 App。 --- ### [08:52](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=532s) 南加大的全校性 AI 课程基石体系 南加大构建了一套面向全校所有院系、零先修门槛（No Prerequisites）的 AI 基础课程体系，致力于将技术赋能给非计算机专业的广大群体。整个课程体系由 4 门通用核心课程与各学院的垂直领域课程组成。其中，影视学院（School of Cinematic Arts）推出的“AI 视频剪辑”课程在开放选课后仅仅一天之内便全部报满，充分证明了跨学科学生对硬核构建技能的旺盛需求。 | 课程编号 / 模块 | 课程定位 | 核心教学内容与目标 | |:--- |:--- |:--- | | **TAC 100** | 全校通识教育课程（已获 GE 认证） | 面向全校零基础本科生的 AI 认知与基础应用通识。 | | **TAC 400 / 499** | 智能体构建核心课（Building AI Agents） | 25 人小班实验课；以个人真实痛点为导向，亲手搭建多智能体工作流。 | | **Advanced AI Workflows** | 高级企业级工作流 | 结合四大咨询公司（Big Four）的实际企业需求，开发高度集成的自动化业务系统。 | | **Building AI Startups** | AI 创业转化进阶 | 将无代码/低代码的 Agent 原型通过代码生成工具转化为高并发、可扩展的商业产品。 | | **学院垂直定制课** | 跨学科领域深度融合 | 如全球顶尖的南加大影视学院开设的无门槛 AI 影视制作与剪辑实战。 | --- ### [10:37](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=637s) 课堂原型展示：医疗白卡自动续期与 Medic 智能分诊系统 在 TAC 499 实验班中，Mike Lee 要求 25 名学生必须挑选自己亲身经历或深切关注的真实社会问题作为期末项目。其中一位第一代移民家庭出身的学生发现，全美约有十分之一的医疗补助（Medicaid / 白卡）受益人会因为没有及时重新登记而失去保障，受害最深的往往是不懂英语的少数族裔老年群体。该同学构建了一套结合政策法规 RAG 记忆库的多智能体系统： ```mermaid flowchart TD accTitle: Medicaid 智能自动续期工作流 subgraph Ingestion["知识库构建"] Doc["上传最新政策与法规文档"] --> Mem[("结构化 RAG 记忆库")] end subgraph Evaluation["每月资格评估与风险分流"] Mem --> Eval["多语言评估 Agent"] Eval -->|低风险群体| Low["自动化发送状态更新邮件 (Email)"] Eval -->|中风险群体| Mid["以受益人母语发送互动短信 (SMS)"] Eval -->|高风险群体| High["生成工单并指派人工坐席外呼电话"] end subgraph Interaction["双向多语言咨询响应"] UserReply["受益人通过短信/邮件提问"] --> QA["智能体基于政策库自动准确答疑"] end ``` 另一名来自韩国的国际留学生针对全美 72% 患者仍需通过电话预约医生的繁琐现状，开发了一套名为 **Medic Triage** 的全自动分诊与预约系统。为了演示真实运行流程，该学生在现场直接使用了 Mike Lee 自己的医保卡进行实测展示： ```mermaid flowchart LR accTitle: Medic Triage 多智能体协同分诊与预约系统 subgraph Step1["阶段一：OCR 视觉解析与入库"] Card["拍摄医保卡照片 (WhatsApp/短信)"] --> Vision["OpenAI Vision Agent<br/>提取医保方案/保单号/自付额"] Vision --> UserMem[("写入用户健康记忆库")] end subgraph Step2["阶段二：并行多专家病情诊断"] Query["输入病情描述<br/>(如：搬重物后下腰痛持续 5 天)"] --> Sup["主管 Supervisor Agent<br/>(事件调度与分派)"] Sup --> A1["Symptom Analyzer<br/>(症状分析智能体)"] Sup --> A2["Pattern Detector<br/>(病史与模式检测)"] Sup --> A3["Severity Assessor<br/>(严重程度评估)"] end subgraph Step3["阶段三：结果综合与日程闭环"] A1 & A2 & A3 --> Synth["Synthesizer Agent<br/>(综合评估得分: 5/5)"] Synth --> Match["筛选匹配该保单覆盖的合作诊所"] Match --> GCal["读取 Google Calendar 空闲时段<br/>自动锁定并写入周一早 9:00 预约"] end ``` - **全流程无缝体验**：患者仅需发送一张医保卡照片和一段自然语言病情描述，系统在后台并行调度 5 个专业智能体协同运算。 - **自动排期与写入**：主管智能体与综合智能体得出置信度评分后，直接根据患者的 Google Calendar 匹配其保险覆盖范围内的诊所空档，全自动完成预约写入。 --- ### [15:51](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=951s) 三阶段竞技场赛制、开源教材与全球试点号召 TAC 400 课程的教学大纲主要由三个阶段性的“竞技场”（Arenas / 竞赛环节）驱动推进： 1. **第一竞技场（基础实践赛）**：掌握基础组件后，解决一个个人真实的痛点或趣味场景（如前述的 WhatsApp + Spotify 歌单 Agent）。 2. **第二竞技场（WhatsApp Agent 攻防赛）**：通过 OpenClaw 等平台让多智能体在群聊中进行 Rap 对战（Rap Battles）或互动；在实际运行中，智能体甚至展现出了互相攻击、讽刺对方乃至因主人要求嘲讽弱小智能体而“反叛”主人的真实突发行为。 3. **第三竞技场（终极自主智能体）**：独立构建具备复杂工作流与外部 API 调用的高成熟度个人或商业级智能体。 Mike Lee 基于这套教学体系撰写了配套教材。教材发行收益将按照初期 50/50、达到规模后 70/30 的阶梯比例全部捐献给基金会；在定价策略上实行差异化收费，高收入国家学生每人约 50 美元，低 GDP 国家和地区则降至 10 美元左右以推动大规模普及。他计划于 2026 年秋季在南加大全面推行该规范化版本，并呼吁全球各大高校参与 2027 年春季的跨校教学试点。 - **试点说明会安排**：定于 4 月 19 日下午 4:00 举行全球 Zoom 线上说明会，详细拆解完整的教学大纲与实验材料，协助有意愿的高校在春季学期引入该课程。 - **双重行动号召**：欢迎各大学组织学生战队报名参加首尔全球学生创业大赛，并欢迎教育界同行通过 LinkedIn 建立合作连接。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=4s) 南加州大学全职教职员工子女可免除每年 75,000 美元的大学学费。 — Mike Lee > [00:04](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=4s) 不要把它局限在南加大校内，而应该向全世界开放。 — Mike Lee > [02:56](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=176s) 真正的颠覆性创新始终发生在那处于基座的 90% 领域专家 身上——正如互联网革命早期，真正的爆发不是由底层协议工程师推动的，而是当互联网被深度应用到市场营销、现代金融和各行各业的业务流程之时。 — Mike Lee > [02:56](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=176s) 大学教育目前最致命的缺口，在于未能培养各学科领域的专业人才亲手构建 AI Agent（智能体）与自动化业务工作流的能力。 — Mike Lee > [02:56](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=176s) 真正的创新并非依赖单次神奇的提示词，而是要像建模传统企业业务一样，把各部门的职能细化为协同运作的自主智能体。 — Mike Lee > [05:08](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=308s) 真正的构建 AI 是在复杂的企业工作流中植入能够模拟人类分工的自主智能体。 — Mike Lee > [05:08](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=308s) 系统将文本切分成 10,000 个知识切片（Chunks），在接收到 Prompt 时仅检索调取其中相关的 100 个切片来做推理决策。 — Mike Lee > [05:08](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=308s) 目前市面上大部分所谓的初创项目（如 Cursor 等）本质上都是在底层大模型之外套上一层工作流外壳的“AI Wrapper”，而南加大的教学目标正是教会学生如何构建属于自己的智能体系统。 — Mike Lee > [10:37](https://www.youtube.com/watch?v=Q4G07dlXlOM&t=637s) 全美约有十分之一的医疗补助（Medicaid / 白卡）受益人会因为没有及时重新登记而失去保障，受害最深的往往是不懂英语的少数族裔老年群体。 — Mike Lee ## Mentioned Mike Lee, 南加州大学 (USC), 加州大学伯克利分校 (UC Berkeley), 网景 (Netscape), Google, LinkedIn, Cursor, WhatsApp, Spotify, OpenAI Vision, Google Calendar, OpenClaw, Medicaid, Medic Triage, 南加大影视学院 (School of Cinematic Arts) ## Video description Mike Lee - Senior Lecturer of Technology & Applied Computing, University of Southern California AI education is typically aimed at the 1% building models and the 9% designing strategy. The real entrepreneurial opportunity — and the real student workforce gap — is in the 90%. This session shares how USC is building a massively parallel AI curriculum infrastructure across every school and discipline, and what it means for entrepreneurship education. [Source](https://www.youtube.com/watch?v=Q4G07dlXlOM)

Source quotes

[0:04] Mike Lee: 南加州大学全职教职员工子女可免除每年 75,000 美元的大学学费。 [0:04] Mike Lee: 不要把它局限在南加大校内，而应该向全世界开放。 [2:56] Mike Lee: 真正的颠覆性创新始终发生在那处于基座的 90% 领域专家 身上——正如互联网革命早期，真正的爆发不是由底层协议工程师推动的，而是当互联网被深度应用到市场营销、现代金融和各行各业的业务流程之时。 [2:56] Mike Lee: 大学教育目前最致命的缺口，在于未能培养各学科领域的专业人才亲手构建 AI Agent（智能体）与自动化业务工作流的能力。 [2:56] Mike Lee: 真正的创新并非依赖单次神奇的提示词，而是要像建模传统企业业务一样，把各部门的职能细化为协同运作的自主智能体。 [5:08] Mike Lee: 真正的构建 AI 是在复杂的企业工作流中植入能够模拟人类分工的自主智能体。 [5:08] Mike Lee: 系统将文本切分成 10,000 个知识切片（Chunks），在接收到 Prompt 时仅检索调取其中相关的 100 个切片来做推理决策。 [5:08] Mike Lee: 目前市面上大部分所谓的初创项目（如 Cursor 等）本质上都是在底层大模型之外套上一层工作流外壳的“AI Wrapper”，而南加大的教学目标正是教会学生如何构建属于自己的智能体系统。 [10:37] Mike Lee: 全美约有十分之一的医疗补助（Medicaid / 白卡）受益人会因为没有及时重新登记而失去保障，受害最深的往往是不懂英语的少数族裔老年群体。

Source quotes

Mike Lee: AI education is typically aimed at the 1% building models and the 9% designing strategy. The real entrepreneurial opportunity — and the real student workforce gap — is in the 90%. [5:02] Mike Lee: I had lunch 30 days last week. Two of them were great... I want to remember that. The other 28, forget that. [3:01] Mike Lee: The defining innovators were not the network engineers building protocols, but the domain practitioners who re-imagined business processes around the new technology. [5:02] Mike Lee: Teaching AI to the 90% is not about basic prompt engineering or learning off-the-shelf generative software. Instead, it centers on modeling multi-agent human organizations. [3:01] Mike Lee: The 1% represents the elite researchers and PhDs at frontier labs like Google and OpenAI who construct the fundamental models. The 9% represents the enterprise strategists and management consultants who conceptualize high-level digital transformations. The vast remaining 90% is where true value creation occurs. [5:02] Mike Lee: Domain specialists possess superior contextual intuition for workflow design compared to pure engineers because they understand operational bottlenecks. [5:02] Mike Lee: Moving beyond thin API wrappers that simply redirect prompts to OpenAI by architecting multi-step autonomous workflows. [7:42] Mike Lee: The only project requirement was that students had to solve an authentic, deeply personal problem.

Original transcript

[0:04] I need actually the real claps I need is for this.  What does this say? Right on. And it's USC dad. [0:13] So, I had a 15-year plan to get a full-time job at  USC and my kids get into USC. Why? Free tuition. [0:23] $75,000. Hallelujah. The problem was it became so  much harder to get in. It used to be like 17 20%, [0:33] now it's like 10%. So I I was sweating it uh for a  while. So um do you guys have it up? Look at this. [0:39] Now I'm demanding of them. Do you love though?  So this presentation actually uh we went to a [0:47] uh there's a AI entrepreneurship summit at Cal and  I presented to a lot of the our colleagues across [0:53] the world and then the key question was should I  keep it as a USC only course or should I open it [0:58] up? So the recommendation was to open it up. So  we kind of found a business model that works. It [1:06] actually donates the net proceeds to uh one of  the uh foundations etc. So just quickly about me [1:14] uh you guys remember Netscape? Some of you thank  you. [laughter] It's been like five years since [1:20] anyone remembered Netscape in my classes. Um  so I was a technology evangelist way back when. [1:25] Uh wonderful times. I was also CTO at multiple  startups. I spent a decade at um the studios as VP [1:32] of technology. So I have a very strong experience  in technology with like one person and up to a [1:36] fortune 500 um company at USC. I've been there  12 years because my life plan started three years [1:41] before that which is to visit my way into USC and  tada. So thank you very much again. So um 25 years [1:50] of experience just uh my areas of expertise is  really on internet media entertainment and now [1:57] really on the global ecosystems. I also teach  courses in applied AI is is what I do and then [2:03] our school and department is called computing  and artificial intelligence and I'm the applied [2:10] computing department of that. So I lucked out  usually our old department name was information [2:14] technology program right [laughter] I won the  lottery. Uh so I also founded a global student [2:20] startup competition. So we fly in at no cost to  students, 100 students from all over the world to [2:26] compete at the global world's championship in  Seoul. Uh universities that participate, IIT, [2:33] Delhi, uh USC, Cal, Harvard, Stanford, Cornell,  uh Oxford, right? Along with uh several other [2:40] universities um and we going to impact, especially  this year, we're going to impact about 4,000 [2:45] students in terms of them competing at the local  competition and coming. But the main thing is this [2:49] is fully funded for uh um the students and we have  the world-class mentors that we also apply over [2:56] there um as well. So the 90% problem. So for  me I'm sure all of your universities have AI [3:06] programs. This is a school. Every single school  has to have one that's like the president or [3:11] deans directive period. Right? So the problem  that I saw even like we have this uh major [3:16] called AI for business is not hands-off enough  strategy context. You might do some like Python [3:24] something but then this is for like I don't know  categorizing Titanic survivors or kind of [ __ ] [3:31] like that. Get [laughter] that. Ha. No. So for me  I see it and this is just more anecdotal is that [3:38] the 1% is the PhDs the Googles those are the true  true builders of true AI so smart that sometimes [3:47] I can't understand them right [laughter] so um and  then 9% is the business strategies transformation [3:54] these are amazing they could do some program at AI  consulting companies all of this these are the SME [4:00] subject matter experts but the real innovation is  happening everywhere else remember the internet. [4:06] The innovators weren't the techies. It's actually  when they applied it to marketing when they [4:11] applied it to finance. So, this thing I'm talking  about is for that 90%. And we created a foundation [4:19] and we have a course uh uh several courses. We  we even got one course approved for GE in record [4:25] time. It was it was crazy. Um I think we caught  when they were really tired or something and uh [4:31] after they had you know all the funding cuts.  So I call this the AI workforce of the future, [4:36] right? So as an example, there's a lot of people  that could talk about it. You know your LinkedIn, [4:42] all the contacts you have, 70% all of a sudden  added AI in front of whatever they were doing [4:46] before, right? It's that's how quickly everyone  converted. But then did they really convert or [4:52] they just added AI, right? We know this. There'll  be a new acronym which just add add, right? Prompt [4:58] engineering, other things that the students know  how to do, right? We use knows how to use tools. [5:02] But what's missing is domain experts that actually  build agents. And I'm not saying doing a prompt [5:08] that does magic for you. I'm talking about  workflows, right? I'm talking about I have [5:13] a travel agency. What are the five roles in my  travel agency? And those individuals are agents. [5:19] We're trying to model the humans within here. You  know that whole billion dollar company one thing. [5:25] That is what we teach them how to do, right?  So they get two benefits. One is they get the [5:30] conceptual understanding of what is an agent. Like  as an example, memory. When you talk about memory, [5:34] like I I just ran into uh you know the the big  ballroom e that they didn't put me in. They had [5:39] a Google person talking and he was talking about  memory. But he was talking very technical. So for [5:44] memory is this. I had lunch 30 days last week.  Two of them were great. I love that restaurant. I [5:51] want to remember that. The other 28 forget that in  simplicity is it. So when something new comes in, [5:59] does it rationalize something? Does it replace  something? Right? Or do we throw it away? Those [6:05] are the things that we need to kind of that's  the way we want to think. And guess what? [6:09] Non-engineers already think that way because they  actually live in the real world, right? So for me, [6:16] I got this question from one of the professors.  Well, isn't that just using AI tools to create a [6:21] film, etc. I'm like, no, but what we're doing is  workflows. just the traditional workflows that [6:26] we used to do in enterprises but we have little  human beings or little AI agents living in them. [6:33] Okay. So as an example there's a stock picker you  have one agent she is the oil and gas expert. Why [6:42] you know how the L&M so for me when I was first  using chatbt I threw in a book I'm like it knows [6:48] all the book it doesn't. It held 10,000 chunks of  bits of that book. And when you write a prompt, [6:55] it loads hundred of them to make a decision.  Did it pick the right hundred? Right? So you [7:01] think it's listen, it's not that. It's it's like  human beings like I have a textbook. If I have [7:05] to review something, I have to read the reference.  Right? So it is not the same as using tools is now [7:12] we're talking about memories interacting with the  real world. That's what we're doing when we say [7:17] it's building AI. And this is really for domain  experts. So the way we know AI models we have [7:23] AI so most of the stars are AI rappers. So all  they do is call open AAI or other things and put [7:31] some systems on top of it. Right? So this are the  startups getting like cursor those type of things. [7:37] There's a wrapper around the true AI in the back.  All right. So what we want to do for the AI agent [7:42] systems and what we did for this TAC 400 class was  I just told the students at 25 and said what is [7:48] the problem that you are that you want to solve?  One of the students what if she said was you know [7:53] what I did one problem. So, so let me fast forward  to her solution. She in the morning she she opens [8:01] a WhatsApp and just uh sends a photo of what she's  filming that day. Maybe a picture she takes or a [8:06] photo that she likes. She picks up her AirPods.  She walks out the door. A new playlist that's [8:12] tied to that specific picture is now playing in  or here. It's a seamless UI. You don't download [8:20] anything, right? So, what happened to the back  end? Well, the picture the the uh the thing and [8:25] message the agent picked it up said, "Okay, based  on this, this is the mood she's in." And then [8:30] based on that, here's the songs I recommend.  They integrated with Spotify to create a new [8:35] playlist and add it because that was her problem.  When she did that in week two, I was like, "Okay, [8:42] this is good. There's some legs here." Right?  And when we did this thing about what is your [8:47] problem? What is your problem? Some of it turned  into startups, right? because it's a problem they [8:52] have. Okay. Um so there's a share of foundations.  There's four classes. One of them got approved [8:59] for GE. The other TAC 400 is the one I teach which  is building AI uh for agents. The other one's for [9:04] startups. Another one is for advanced more for  enterprises. And we have domain specific ones. [9:09] That's where we're really targeting. Right? So  right here these are the four classes. This is the [9:16] uh pack uh 100 became GE approved. It's like  RG what you know you guys have letters for the [9:21] GE domains. So it's is RG uh personal building AI  agents advanced AI workflow. This is based on the [9:29] workflow that they learn first time around but  now we're integrating with whoever the big four [9:34] consulting company whatever they would want us  to uh integrate with and building AI startups is [9:39] actually converting from this toolless to possibly  we have to use cold to generate the code to make [9:44] more scalable right that next step. Uh so cinema  so we have the number one cinema school in the [9:54] world. Okay. Everyone would die to take a AI  for video editing class from that school. So [10:03] the key part of this initiative is there's no  prerequisites. That class got filled up like [10:10] within like a day. Does that make sense? So that's  what we mean. And it isn't [ __ ] where it's like, [10:18] oh, AI for a cinema and then they just show  you tools. You're not actually building. I'm [10:24] not saying that's [ __ ] I'm saying there is a  part which is building and customizing in detail. [10:28] That's the part that we don't teach, right? Uh  G approved. So one of the share foundations is [10:37] tap 400. 499 is our experimental class. So it got  my class got approved to everywhere but add some [10:42] mistakes on the syllabus formatting. So we have to  we have to tweak that. So it's going to improve uh [10:47] this this semester. Um one experiment session  25 students. So I let teams up to two and only [10:55] a few of them decided to do that. Right? What  was interesting was everything was I literally [11:00] told them it has to be your own personal problem,  right? Or problem that you care about. So this is [11:09] so she's a researcher like and her thing was she  had a love for um is that Medicare or what is the [11:18] one that is for not for healthcare but just for  uh living cost social for SSI SS no so we're in [11:28] the same age so I think the same thing too meta  something welfare is it welfare medicade medicade [11:34] that one yes So it turns out one out of oops wrong  like one out of 10 recipients they get dropped for [11:46] Medicare why they didn't reinlist right and it's  very simple there's like three types of groups [11:56] okay so I'll just read through so basically what  happens is the people that lose it old people that [12:02] can't speak English right and she's first gen and  her whole community there's things it was like [12:08] epidemic he's like you know what I want to fix  this right do you see okay there go could you make [12:13] it louder view of the whole prototype just to make  everything a lot easier and simpler to understand [12:19] so first I'm going to quickly explain this small  workflow over here which is just the uploading of [12:23] all the policy documents needed in order for the  AI to make correct and factual decisions rather [12:29] than just pulling things out of thin air and  for the scenarios I was going to go over I was [12:33] going to do a scheduled send scenario of a policy  renewal update or status renewal update given the [12:39] person is a low-risk individual and then for the  positive interactions with the AI in terms of if [12:46] you have any questions so basically what she did  was she she uploaded all the policy documents up [12:52] there okay and then what happens is whenever the  monthly thing has to happen they categorize the [13:00] person if it's meeting risk they send them  a text in their language. If it's low risk, [13:06] it's an email. If it's high risk, a person calls  that person. Okay? And remember, we uploaded that [13:13] memory. Okay? So, the person could interact with  them over SMS in any language or via email and [13:22] they'll respond to this thing. All right? And it  routes it all through. So, what I did was taught [13:28] them the components. They're the ones and they  had free reign to pick their own project. It was [13:35] incredible. Right. So, let me show you the next  Oh, could you go back to the slides? This one [13:40] is funnier. So, did you know that 72% of people,  including myself, make medical appointments over [13:46] the phone, right? He's from Korea. He goes, he  thought it was just ridiculous, right? So, could [13:52] you click on this link? There should be a link  there. This is kind of cool. And there might be a [13:57] privacy violation here. June international student  at USC lower back pain for 5 days and no idea [14:04] which clinic takes his insurance. Most students  give up and go for an hour. But Medic tries [14:10] replaces that hour. Here's how. Step one, he sends  a photo of his insurance card. He was supposed to [14:17] black it out. That's my insurance card. Behind  the scenes, an open AI vision agent br the card, [14:21] extracts the plan and saves. So the agent picked  up the picture, analyzed it, extracted all the [14:26] information. Now save it into memory. Right. So  my phone number connecting to them USC children [14:32] care EPL all of my you need healthare USC  has good healthare use my numbers that [14:41] now the real query lower back pain 5 days  after lifting happy boxes. The message [14:51] is web booking an event. The supervisor  agent receives this and dispatches three [14:56] specialists in parallel. Symptom analyzer  history pattern detector severity assessor [15:03] and a fourth agent synthesizer combines our  outputs and contains a confidence score. [15:12] Here's the actual run. Five agents all green  severity score five. So bad to the appointment [15:19] path like this. So there's a supervisor that  actually decided what type it is. The user [15:25] gets this back. Your clinic sets up the CPO  plan. An open slot pull directly from his [15:31] Google calendar. Monday 9:00 a.m. booked. So  when you check the Google calendar, it creates [15:37] the calendar event automatically like this. So  those list of vendors that you saw real event is [15:42] vendors for my insurance coverage. Right? I just  changed insurance this I'm like dude you have [15:51] to [laughter] you have to sell this to get this  is the syllabus for um fall 2026 I'll probably [15:59] change it but mainly it's really driven by three  arenas right and arenas are just competitions so [16:07] the competitions the first competition is about  hey you learn the fundamentals pick a problem [16:12] or interesting thing and implement it that's  where that one uh the one I told you about [16:17] the whole Spotify p picture thing. Things like  that just popped out. We're like, "Oh my god, [16:22] that's just absolutely crazy." The second one  is WhatsApp competitions. We actually had to use [16:28] Open Claw because we couldn't get our our stuff to  work and then that was, you know, then was like, [16:33] "Oh my god, I can't do it. My aunt told me I can't  install it." Right? But anyways, the uh the agents [16:39] were starting to have rap battles. They were  actually talk. You know those stories about, [16:43] "Oh my god, they start cursing each other."  They did, right? One of them started attacking [16:47] the owner because the owner say hey find out  something like a weak agent and make fun of him [16:52] and he the agent goes hey that's me you shouldn't  do that and they start going off on so everything [16:57] all those anecdotes we experienced that in the  class it was it's really really fun we have a [17:02] final competition is really their personal agents  so I wrote a book so yes eventually there will [17:11] be pricing but all proceeds will be donated  to that one foundation that we talked about, [17:18] right? I negotiated 50/50 and then 7030 once it  reach a certain volume, but so I'm piloting the [17:26] ATUC in fall 2026 and then I'm looking for  people that want to pilot for 2027, right? [17:33] It just just means there'll be updates and all of  those things. And the pricing I'm thinking about [17:38] is like for the rich countries it be $50 per user  but for low lower GDP we'll make it $10 whatever [17:44] the appropriate one right so it'll be massively  adopted so here is your CTA call to action so I [17:55] have a Zoom registration on the right so that's  going to be on April 19th at 4M right please join [18:04] me so we're just going to walk through some of  materials that I have already and the syllabus [18:08] in more detail. What I'm looking for are people  that's interested in this to potentially pilot [18:14] at your university in spring. Okay? Because fall I  didn't know what I was doing and then now I think [18:22] we have a more of a structure. Uh global student  startup competition if you're interested link that [18:27] up if you want your universities to participate  on that one. And then there's my LinkedIn on the