NY Academy Chief Dirks Sees Niche AI Startups Winning Over Giant Labs

Economic Implications of AI Policy: Innovation, Competition, and Market Access — Ai4

Ai4

While frontier labs capture the bulk of compute and capital, specialized application startups will secure the ultimate market advantage if the industry maintains public trust.

AI-assisted summary of Economic Implications of AI Policy: Innovation, Competition, and Market Access — original source. Check the original for full context.

Source published: 2025-08-28

Detailed notes

## Background Published by the Ai4 channel on YouTube on August 28, 2025, this session features a dialogue between Sanjay Puri, Chairman and Founder of RegulatingAI, and Nicholas Dirks, President and CEO of The New York Academy of Sciences. According to the page description, the conversation examines how AI policy shapes global economic ecosystems, addressing innovation pipelines, competitive dynamics, and market access.

## TL;DR Nicholas Dirks analyzes the AI revolution through a historical lens, noting that its compressed timeline demands faster institutional adaptation than any prior industrial transformation. The discussion reviews AI's uneven progress across the sciences, from biomedical breakthroughs to lagging energy frontiers, while warning that automated coding tools require higher education to prioritize critical thinking over narrow technical syntax. Evaluating geopolitical competition, Dirks assesses regulatory approaches across the United States, Europe, and China, observing how capital concentration among a few frontier labs challenges traditional universities. Ultimately, Dirks expresses qualified optimism for smaller, specialized application startups, identifying societal public trust as the ultimate constraint on sustainable innovation.

## Atmosphere & style The session has an intellectual, reflective atmosphere driven by Nicholas Dirks's perspective as both a historian and academic executive, countering typical tech hype with historical grounding. Sanjay Puri maintains an engaged, conversational rhythm, steering the exchange through clear thematic pivots aimed at technical practitioners and founders. Moments of collegial warmth and shared empathy—such as joking about the relentless grind of fundraising—punctuate an otherwise serious assessment of geopolitical and labor challenges. As off-stage timekeepers wrap up the panel, the brief closing imparts a focused, cautionary verdict on the indispensable role of public confidence in technological adoption.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=u25ZDHryTKI&t=4s) Unlike prior historical inflection points that unfolded across generations, the AI revolution is compressing sweeping economic and cultural reorganization into an unprecedentedly narrow window. - [00:04](https://www.youtube.com/watch?v=u25ZDHryTKI&t=4s) The exact timeline for AGI matters far less than the immediate structural remake AI is driving across work, education, healthcare, and daily human existence. - [03:03](https://www.youtube.com/watch?v=u25ZDHryTKI&t=183s) AI is advancing biomedical research and extending health spans, but other crucial physical frontiers—such as fusion energy, quantum computing, and green hydrogen—remain critical bottlenecks. - [07:31](https://www.youtube.com/watch?v=u25ZDHryTKI&t=451s) Automated coding tools have diminished the market advantage of entry-level programming degrees, requiring modern workers to pivot toward data literacy and critical oversight. - [07:31](https://www.youtube.com/watch?v=u25ZDHryTKI&t=451s) The integration of AI necessitates a shift away from rigid four-year educational tracks toward lifelong learning frameworks focused on critical thinking, verification, and catching algorithmic hallucinations. - [12:10](https://www.youtube.com/watch?v=u25ZDHryTKI&t=730s) Targeted regulatory frameworks can drive multi-year productivity gains, whereas erratic policy whiplash and instability represent the greatest risks to technical and capital deployment. - [12:10](https://www.youtube.com/watch?v=u25ZDHryTKI&t=730s) Western policymakers risk falling behind if they embrace insular AI nationalism, as China's state-backed coordination and elite institutions are rapidly matching premier Western research centers. - [16:03](https://www.youtube.com/watch?v=u25ZDHryTKI&t=963s) The massive capital and GPU requirements for training foundational models have created a tight oligopoly of frontier labs, squeezing out both small businesses and elite universities from building proprietary base models. - [18:19](https://www.youtube.com/watch?v=u25ZDHryTKI&t=1099s) Startups will thrive by serving as the indispensable implementation layer, building targeted, domain-specific vertical applications on top of frontier foundation models. - [18:19](https://www.youtube.com/watch?v=u25ZDHryTKI&t=1099s) Sustaining the momentum of the AI economy depends fundamentally on instilling public trust, which serves as the ultimate bottleneck for widespread technological adoption.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=u25ZDHryTKI&t=4s) Historical Inflection Points and the Uniqueness of the AI Revolution

Sanjay Puri opens the session by welcoming the audience and introducing Nicholas Dirks, underscoring Dirks’ distinguished academic leadership at the University of California, Berkeley and his stewardship of The New York Academy of Sciences. Acknowledging that every technological shift is routinely hailed as unprecedented, Puri prompts Dirks to draw on his multi-sector vantage point across corporate boards, academic institutions, and policy circles to evaluate the true character of this current moment. Rather than evaluating artificial intelligence solely through contemporary Silicon Valley hype, Dirks approaches the dilemma through his original discipline as a historian. He emphasizes that while modern AI literature reflexively begins with the agricultural revolution, the four distinct iterations of the industrial revolution, or the early digital revolution, every single one of those historic transformations unfolded over multi-generational spans, giving human societies extended runways to adapt their economic structures and cultural institutions.

In contrast, the velocity of the artificial intelligence transformation compresses profound societal reorganization into an extraordinarily narrow window. Whether one forecasts that artificial general intelligence (AGI) or superintelligence is slated to emerge in three, five, ten, or twenty years, Dirks argues the exact horizon matters far less than the sheer comprehensiveness of the disruption underway. Echoing remarks made earlier by congressional leaders at the conference, he stresses that artificial intelligence is poised to fundamentally rewrite the future of work, higher education, healthcare delivery, leisure, social relationships, and the baseline human condition itself. We are confronting a structural remake of identity, daily existence, and societal expectations that departs radically from all prior inflection points along the historical timeline.

[03:03](https://www.youtube.com/watch?v=u25ZDHryTKI&t=183s) AI Across the Sciences: Breakthroughs, Lagging Sectors, and Risk

Recalling his tenure at UC Berkeley in 2012 and 2013, Dirks reflects on watching modern artificial intelligence take root in one of the nation's premier computer science departments. Even then, researchers recognized that the grand ambitions first articulated by John McCarthy when he coined the term "artificial intelligence" at Dartmouth in 1955—alongside early experimentation with neural networks—were on the precipice of realization. To prepare the institution, Berkeley established a pioneering data science program explicitly designed to bridge computer science and statistics with broader academic disciplines across campus, anticipating that foundational algorithmic tools would quickly transform public health, computational biology, and the social sciences. The recent explosion of generative artificial intelligence and large language models has fully vindicated that strategy. While prominent figures across his circle like Reid Hoffman and Gary Marcus hold sharply divergent perspectives on developments such as the release of frontier models like ChatGPT-5, Dirks notes that there is universal agreement regarding the historic weight of these breakthroughs.

```mermaid flowchart TD accTitle: AI's Dual Impact Across Scientific Disciplines subgraph Accelerating["Rapidly Advancing Horizons"] A1["Biomedical & Immunology"] --> A2["Targeted Cancer Therapeutics"] A2 --> A3["Extended Human Health Span"] end subgraph Bottlenecks["Critical Lagging Frontiers"] B1["Clean Fusion Energy"] B2["Room-Temp Quantum Computing"] B3["Green Hydrogen Scaling"] end subgraph Risks["Systemic Systemic Hazards"] R1["Autonomous Lethal Weapons"] R2["Training Data Biases & Limits"] end ```

Yet the pace of scientific discovery remains uneven. Invoking economic commentators like Tyler Cowen, Dirks points out that while certain biomedical and immunological disciplines are advancing at breakneck speeds—yielding revolutionary cancer therapeutics and expanding the realistic human health span—other critical frontiers are lagging behind. To truly catalyze the new innovation economy, artificial intelligence must simultaneously be deployed to unlock complex roadblocks in fusion energy, room-temperature quantum computing, and green hydrogen. Alongside this vast promise lies genuine peril: AI systems are already redefining modern warfare through autonomous lethal weapon architectures, and their cognitive utility remains strictly bounded by the integrity, provenance, and accuracy of the data sets upon which they are trained.

- John McCarthy originally coined the term "artificial intelligence" at the Dartmouth Summer Research Project in 1955. - Key technical frontiers requiring immediate acceleration alongside AI include fusion energy, room-temperature quantum computing, and green hydrogen. - Biomedical innovations are extending human health span, challenging historical assumptions about chronic illness and longevity.

[07:31](https://www.youtube.com/watch?v=u25ZDHryTKI&t=451s) The Labor Shift: Computer Science Degrees, Coding Tools, and Life-Long Learning

Addressing an audience largely composed of AI developers and engineers, Puri points to automated coding platforms like Cursor and Windsurf, asking how these productivity multipliers will alter the labor landscape when even baseline STEM and computer science credentials no longer guarantee immediate employment for university graduates. Dirks observes that university campuses have spent the last two decades responding to intense student demand by funneling enrollments out of the humanities—such as English, history, and political science—and into business, economics, and predominantly computer science. He recalls that when he arrived at Berkeley in 2012, legendary computer science chair David Patterson asked for an additional one hundred full-time equivalent (FTE) faculty lines solely to satisfy the crush of undergraduates demanding coding instruction. Today, however, possessing entry-level coding skills is no longer enough to secure a career in tech.

Instead of narrow technical specialization, the modern workforce demands a return to the foundational virtues of the liberal arts: critical thinking, data literacy, the ability to independently evaluate data provenance, and the judgment required to catch algorithmic hallucinations. Dirks scrutinizes the popular aphorism that "it's not that AI is going to displace jobs, but people who use AI will take over from the people who don't," arguing that fulfilling this premise requires dismantling legacy educational pipelines. Unlike the early days of massive open online courses (MOOCs) fifteen years ago—when 170,000 students enrolled in Stanford’s online AI course only for the vast majority to drop out—AI tools are now deeply integrated into the fabric of scholarly research and course design. Dirks notes his own wife, a professor teaching at Berkeley, routinely incorporates various large language models into her lecture preparations. Just as academia once adjusted from the slide rule to the pocket calculator and then to personal computers, it must now reconsider whether rigid four-year undergraduate and master's degree tracks are adequate for longer working lives that will require professionals to continuously "learn to learn."

- In 2012, UC Berkeley computer science chair David Patterson requested 100 additional full-time equivalent (FTE) faculty lines to accommodate enrollment spikes. - Early digital learning initiatives, such as Stanford's massive open online courses (MOOCs) developed around 2011, saw enrollments reach 170,000 students but suffered steep attrition rates. - Generative AI shifts educational priorities away from basic syntax writing toward data verification, source analysis, and critical oversight.

[12:10](https://www.youtube.com/watch?v=u25ZDHryTKI&t=730s) Global Competition: US Innovation vs. EU Regulation and China’s State Strategy

Turning to the global geopolitical arena, Puri contrasts three competing governance approaches: the European Union's prescriptive framework under the EU AI Act, China's centralized state-led focus, and the United States' historic "innovation first, regulation later" mindset. Dirks contextualizes the European approach by recalling his attendance at an AI summit in Paris, where French President Emmanuel Macron actively pitched France as an open destination for AI entrepreneurship and showcased domestic startups, even while operating under stringent European regulatory overhangs. Drawing on insights from Stanford economist Erik Brynjolfsson, Dirks asserts that well-targeted regulation, though costly upfront, actually drives greater long-term productivity within a three- to five-year window. The paramount danger is not the presence of rules, but the lack of regulatory stability; unpredictable policy whiplash paralyzes capital deployment and technical experimentation.

```mermaid flowchart TD accTitle: Divergent Global Approaches to AI Policy and Innovation subgraph US["United States Model"] direction TB U1["Innovation-First Posture"] --> U2["Venture-Backed Frontier R&D"] U2 --> U3["Risk: Regulatory Volatility"] end subgraph EU["European Union Model"] direction TB E1["EU AI Act Framework"] --> E2["Comprehensive Pre-Deployment Rules"] E2 --> E3["Risk: Compliance Choke on Startups"] end subgraph CN["China Model"] direction TB C1["State-Enterprise Integration"] --> C2["Targeted Strategic Funding"] C2 --> C3["Elite University Output (Tsinghua)"] end ```

Dirks contrasts the Western landscape with China's unique ecosystem, which combines strict internal social controls with aggressive state-sponsored coordination and immense capital investments in fundamental research. Based on his firsthand engagements at Tsinghua University in Beijing, Dirks warns that China's top tier of engineering institutions will very soon match and compete directly with world-leading centers like Stanford, MIT, UC Berkeley, and Carnegie Mellon. Disputing the idea that Chinese AI is limited to viral headlines surrounding DeepSeek, Dirks warns Western policymakers against sliding into an insular, "faux kind of AI nationalism." Taking comfort in European regulatory sluggishness is a strategic mistake when China is building world-class capabilities, while nations like India represent an emergent competitive frontier.

[16:03](https://www.youtube.com/watch?v=u25ZDHryTKI&t=963s) Frontier Lab Hegemony and the Squeeze on Universities and Small Business

Puri pivots the competitive discussion to domestic market concentration, highlighting how the contemporary AI landscape is dominated by a tight oligopoly of three to five frontier companies, leaving small business owners, startups, and open-source advocates uncertain of their position. Dirks acknowledges that out of the staggering $250 billion deployed in AI capital investments in 2024, the lion's share was absorbed directly by a handful of frontier laboratories. When high-profile executives like Sam Altman and Dario Amodei dominate headlines, or when celebrated Nobel laureates like Google DeepMind’s Demis Hassabis herald that AI will fundamentally rewrite the boundaries of scientific discovery, smaller players often wonder whether independent technological development remains viable.

Dirks expresses deep solidarity with small business entrepreneurs, noting that major research universities find themselves in an identical quandary. Despite their academic prestige and prominent media presence, universities possess nowhere near the financial reserves, server clusters, or specialized compute infrastructure required to train and deploy proprietary large language models from scratch. As hyper-scaled commercial entities consolidate the computational bedrock of artificial intelligence, higher education and small enterprises alike are forced to navigate an ecosystem where they must innovate on top of private platforms rather than building independent foundational models.

[18:19](https://www.youtube.com/watch?v=u25ZDHryTKI&t=1099s) Startup Opportunities in Specialized AI and the Crucial Role of Public Trust

Nicholas Dirks points out that while acquiring the massive volume of GPUs required for heavy-duty compute is a daunting hurdle, smaller companies can still thrive through strategic partnerships. These smaller, niche-focused startups—concentrated on targeted product development—will be indispensable for the real-world implementation and commercial success of the large foundational models coming out of frontier labs. Dirks emphasizes that these specialized players "are actually going to win in the end," provided they have the grit to persevere through the relentless demands of startup life. Acknowledging the grinding cycle of continuous work and pitching for capital, Dirks draws an empathetic parallel to his own role leading The New York Academy of Sciences, joking, "I run a nonprofit. I spend my time doing fundraising too. I know, I feel your pain."

As the panel draws to a close, Sanjay Puri remarks that they could easily continue the conversation for hours and asks Dirks for a quick closing verdict: is he ultimately positive or negative on the future of AI? Mindful of the timekeeper signaling them from offstage, Dirks delivers a condensed, single-sentence condition for his optimism. He explains that his outlook is positive provided society can sustain the momentum of the innovation economy while simultaneously instilling a deep sense of public trust in AI deployment, warning that a lack of trust could become the "ultimate constraint" on technological progress. Puri sums up the response as a "qualified thumbs up" and brings the session to a close.

- **Niche Innovation as the Implementation Layer:** Smaller startups focused on vertical domain applications are essential partners that will drive the practical adoption and value realization of frontier lab models. - **The Public Trust Bottleneck:** While compute infrastructure and funding are major hurdles, societal trust in AI systems is highlighted as the ultimate limiting factor for long-term technological and economic advancement.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=u25ZDHryTKI&t=4s) Whether one forecasts that artificial general intelligence (AGI) or superintelligence is slated to emerge in three, five, ten, or twenty years, Dirks argues the exact horizon matters far less than the sheer comprehensiveness of the disruption underway. — Nicholas Dirks

> [00:04](https://www.youtube.com/watch?v=u25ZDHryTKI&t=4s) artificial intelligence is poised to fundamentally rewrite the future of work, higher education, healthcare delivery, leisure, social relationships, and the baseline human condition itself. — Nicholas Dirks

> [07:31](https://www.youtube.com/watch?v=u25ZDHryTKI&t=451s) it's not that AI is going to displace jobs, but people who use AI will take over from the people who don't — Nicholas Dirks

> [07:31](https://www.youtube.com/watch?v=u25ZDHryTKI&t=451s) possessing entry-level coding skills is no longer enough to secure a career in tech. — Nicholas Dirks

> [07:31](https://www.youtube.com/watch?v=u25ZDHryTKI&t=451s) learn to learn — Nicholas Dirks

> [12:10](https://www.youtube.com/watch?v=u25ZDHryTKI&t=730s) innovation first, regulation later — Sanjay Puri

> [12:10](https://www.youtube.com/watch?v=u25ZDHryTKI&t=730s) The paramount danger is not the presence of rules, but the lack of regulatory stability; unpredictable policy whiplash paralyzes capital deployment and technical experimentation. — Nicholas Dirks

> [12:10](https://www.youtube.com/watch?v=u25ZDHryTKI&t=730s) faux kind of AI nationalism — Nicholas Dirks

> [18:19](https://www.youtube.com/watch?v=u25ZDHryTKI&t=1099s) are actually going to win in the end — Nicholas Dirks

> [18:19](https://www.youtube.com/watch?v=u25ZDHryTKI&t=1099s) I run a nonprofit. I spend my time doing fundraising too. I know, I feel your pain. — Nicholas Dirks

> [18:19](https://www.youtube.com/watch?v=u25ZDHryTKI&t=1099s) ultimate constraint — Nicholas Dirks

> [18:19](https://www.youtube.com/watch?v=u25ZDHryTKI&t=1099s) qualified thumbs up — Sanjay Puri

## Mentioned Sanjay Puri, Nicholas Dirks, RegulatingAI, The New York Academy of Sciences, University of California, Berkeley, John McCarthy, Dartmouth Summer Research Project, Reid Hoffman, Gary Marcus, ChatGPT-5, Tyler Cowen, Cursor, Windsurf, David Patterson, Stanford University, Emmanuel Macron, Erik Brynjolfsson, EU AI Act, Tsinghua University, MIT, Carnegie Mellon University, DeepSeek, Sam Altman, Dario Amodei, Demis Hassabis, Google DeepMind

## Video description Sanjay Puri, Chairman/Founder, RegulatingAI Nicholas Dirks, President and CEO, The New York Academy of Sciences

Abstracts

In this conversation, we’ll explore how AI policy is shaping global economic ecosystems—from innovation pipelines and competitive dynamics to equitable access to emerging markets.

[Source](https://www.youtube.com/watch?v=u25ZDHryTKI)

中文简报

纽约科学院院长Dirks：细分领域AI初创企业将胜过巨头实验室尽管前沿实验室占据了绝大部分算力与资本，但如果该行业能维持公众信任，专注于特定应用的初创企业将锁定最终的市场优势。 Nicholas Dirks从历史视角审视了人工智能革命，指出其大幅压缩的时间跨度要求制度适应的速度超过以往任何一次工业转型。讨论梳理了AI在各科学领域的不均衡进展——从生物医药突破到相对滞后的能源前沿探索，同时警示自动化编程工具正促使高等教育必须将批判性思维置于狭隘的技术语法之上。在评估地缘政治竞争时，Dirks分析了美国、欧洲和中国的监管路径，并指出少数前沿实验室的资本集中正对传统大学构成挑战。最终，Dirks对规模更小、更专注的应用型初创企业表达了审慎乐观，并认为社会公众信任是决定可持续创新的最终制约因素。与以往历经数代演进的历史转折点不同，AI 革命正将深刻的经济与社会重组压缩在极窄的时间窗口内，无论 AGI 何时到来，它都已在实质性重塑工作、教育与医疗体系。 (0:04) AI 虽然显著加速了生物医学与癌症疗法的研发并有望延长健康寿命，但在核聚变、室温量子计算及绿氢等关键物理科学前沿仍面临发展瓶颈。 (3:03) 自动化编程工具的普及削弱了初级计算机学位的市场壁垒，促使高等教育与职场需求从基础语法编写转向数据素养、批判性思维以及对算法幻觉的鉴别能力。 (7:31) 精准的监管框架有助于释放长期生产力，而政策的不确定性与摇摆才是资本和技术部署的最大风险；西方若陷入封闭的“AI 民族主义”，将面临被快速崛起的竞争者超越的危险。 (12:10) 基座模型对巨额资金与算力的需求造就了少数前沿实验室的寡头垄断，使高校与小企业难以自建底座，但深耕垂直细分领域的初创企业与建立社会信任将成为推动落地的核心。 (16:03)

Reading summary

Nicholas Dirks, President of The New York Academy of Sciences, argues that artificial intelligence compresses societal adaptation into an unprecedentedly narrow window, requiring educational systems to prioritize data literacy over basic coding, while smaller enterprises must focus on domain-specific applications amidst heavy frontier compute concentration. Compressed Historical Timelines and Uneven Scientific Discovery (0:04) As a historian, Dirks contrasts the artificial intelligence revolution with earlier inflection points—including the agricultural, industrial, and early digital transitions—which unfolded over multi-generational spans that allowed human institutions runway to adapt. AI, by contrast, compresses structural disruptions to labor, higher education, healthcare, and human identity into an exceptionally brief timeline, rendering the exact arrival date of artificial general intelligence less critical than the immediate breadth of societal reorganization. Simultaneously, Dirks highlights an uneven rate of discovery across scientific domains. While AI has catalyzed major breakthroughs in biomedical research, immunology, and targeted cancer therapeutics to extend the human health span, critical technical areas remain bottlenecks. Dirks emphasizes that AI must be systematically applied to accelerate lagging frontiers such as clean fusion energy, room-temperature quantum computing, and green hydrogen scaling, all while managing severe systemic hazards like autonomous lethal weapons and training data biases. The Labor Shift Toward Critical Data Literacy (7:31) The rapid emergence of automated coding platforms such as Cursor and Windsurf has undermined the assumption that baseline computer science credentials guarantee tech employment. This upends a two-decade trend where universities redirected resources away from the humanities to meet student demand for programming, illustrated by UC Berkeley computer science chair David Patterson requesting 100 additional faculty lines in 2012 to handle enrollment spikes. Dirks contends that entry-level syntax generation is increasingly commoditized, elevating the value of traditional liberal arts competencies. Workers now require critical thinking, the capacity to independently assess data provenance, and the judgment to detect algorithmic hallucinations. Rather than relying solely on rigid four-year degree programs, Dirks argues that longer working lives will require educational pipelines to transition toward continuous, lifelong learning frameworks. Geopolitical Models and the Risk of Policy Whiplash (12:10) Examining international AI governance, Dirks contrasts the United States' historic market-driven model with the European Union's precautionary EU AI Act and China's state-directed coordination. Citing Stanford economist Erik Brynjolfsson, Dirks notes that well-crafted regulation can enhance productivity within three to five years; the primary hazard to capital deployment and technical experimentation is not the presence of rules, but unpredictable regulatory whiplash. Dirks warns Western leaders against adopting an insular 'faux AI nationalism' or assuming European regulatory delays ensure Western dominance. Drawing on his firsthand interactions at Tsinghua University, Dirks observes that China's combination of centralized funding and elite engineering pipelines will soon compete directly with leading Western institutions like Stanford, MIT, and UC Berkeley, while nations such as India represent an expanding competitive frontier. Frontier Lab Hegemony and Niche Enterprise Opportunities (16:03) Market consolidation poses severe resource constraints for both independent startups and higher education. Dirks notes that a handful of frontier laboratories absorbed the vast majority of the approximately $250 billion in AI capital deployed in 2024. Because major research universities and small businesses lack the financial capital, server infrastructure, and GPU volume needed to train foundation models from scratch, they are forced to build atop proprietary corporate platforms. Despite this infrastructure squeeze, Dirks maintains that specialized startups focusing on vertical domain applications hold a decisive strategic advantage. These smaller, niche players represent the essential implementation layer required to operationalize raw foundation models into practical commercial products, positioning them to capture substantial long-term value if they can navigate early capital-raising pressures. Dirks offers a qualified positive verdict on the trajectory of AI, cautioning that long-term advancement ultimately hinges on maintaining public trust, which he identifies as the ultimate operational constraint on technological adoption.

精炼阅读摘要

纽约科学院负责人兼加州大学伯克利分校前校长Nicholas Dirks指出，人工智能转型的推进速度远超以往的技术变革，正在重塑劳动力结构与科研模式；与此同时，算力高度集中和地缘政策差异对市场准入构成挑战，社会信任则成为决定技术全面落地的核心约束。历史级转型速度与科研突破的不均衡性 (0:04) 从历史学家视角出发，Nicholas Dirks对比了农业革命、四次工业革命与早期数字革命，强调以往的重大变革均历经数代人演进，为社会和经济结构的适应留出了充足周期。相比之下，AI转型的压缩周期极短，无论通用人工智能（AGI）在数年还是数十年内到来，其引发的涵盖就业、教育、医疗乃至人类生存状态的全面重构都无法避免。自1955年John McCarthy在达特茅斯确立AI概念以来，基础算法在过去十余年间快速向交叉学科渗透，并在大语言模型浪潮中得以验证。然而，科学领域的落地进展并不均衡。生物医药与免疫学受益于AI加速，催生出靶向肿瘤疗法并拓展了健康寿命；但在清洁核聚变、室温量子计算以及绿氢规模化等硬科技领域，技术突破依然滞后。Dirks提醒，技术推进伴随着自主致命武器等新型风险，且AI系统的可靠性严格受制于训练数据的完整性、溯源与准确性。编程工具普及重构高校教育与劳动力要求 (7:31) 随着Cursor和Windsurf等自动化编码工具的兴起，基础STEM与计算机科学学历不再是获得就业的铁律。Dirks回顾了加州大学伯克利分校在2012年的招生热潮，当时计算机系主任David Patterson曾申请增设100个全职教职以应对爆炸性选课需求，但如今单纯的初级编程技能已被自动化大幅稀释。现代劳动力市场对纯技术专长的依赖逐步转向对批判性思维、数据素养、数据溯源评估以及识别算法幻觉之能力的追求。这也促使高等教育模式发生转变。尽管早期的慕课（如2011年斯坦福大学吸引17万名注册者的人工智能在线课程）遭遇了严重的辍学率，但当前生成式AI已深入渗透到大学日常教学与教案设计中。Dirks主张，高校必须重新评估传统的四年制本科和硕士阶段设置，为从业者在更长的职业周期中建立持续“学习如何学习”的终身教育体系。全球监管路线分野与国际竞争力格局 (12:10) 针对全球政策生态，讨论梳理了三种代表性模式：欧盟通过《人工智能法案》（EU AI Act）推行前置性规范，中国依托集中体制主导战略研发，美国则沿袭“先创新、后监管”的自由市场路径。Dirks援引斯坦福大学经济学家Erik Brynjolfsson的观点指出，设计得当的监管虽然会带来短期合规成本，但有望在三至五年内提升长期生产力；对于企业而言，政策反复无常造成的波动比合规要求本身更具破坏性。在全球竞争层面，Dirks提醒西方决策者警惕狭隘的“AI民族主义”，并强调中国的AI发展绝不仅限于DeepSeek等个别热门事件。基于对清华大学等学术机构的实地交流，他评估中国顶尖工程学府的研发输出在不久后将与斯坦福大学、麻省理工学院、加州大学伯克利分校及卡内基梅隆大学直接竞争。此外，在欧洲受制于监管负担的同时，印度等新兴市场也正在形成不可忽视的技术研发力量。算力垄断与垂直领域初创企业的市场空间 (16:03) 2024年全球AI领域投入的高达2500亿美元资本开支绝大多数被少数几家前沿实验室吸纳，形成了算力与模型研发的高度集中。在Sam Altman、Dario Amodei以及Google DeepMind的Demis Hassabis等领跑者主导行业议程的背景下，包括顶尖大学在内的非巨头机构由于缺乏同等规模的算力集群和资金储备，无法从零训练大型专有基础模型，只能被迫在商业公司的基础平台上展开二次创新。尽管获取GPU集群的门槛极高，Dirks认为专注于垂直领域场景的初创企业依然具备独特的生存空间。前沿实验室的基础模型需要依赖这些细分领域公司进行产品化集成和商业落地。只要具备应对融资与开发周期的韧性，深耕垂直应用层的创新企业终将成为AI技术落地的主要承载者。在总结展望时，Nicholas Dirks对AI前景给出了有条件的积极评价。他强调，技术红利的可持续释放取决于在维持创新动力的同时，能否建立起深厚的公众信任；一旦缺失社会信任，技术推广将遭遇根本性瓶颈。

Detailed notes — zh

## Background 本视频由 Ai4 频道于 2025 年 8 月 28 日发布在 YouTube 平台。内容为 RegulatingAI 创始人兼主席桑杰·普里（Sanjay Puri）与纽约科学院院长、加州大学伯克利分校前校长尼古拉斯·德克斯（Nicholas Dirks）的高端对谈，探讨 AI 政策如何塑造涵盖创新管线、竞争格局以及新兴市场准入的全球经济生态。 ## TL;DR 对谈深入剖析了人工智能相较于历史工业革命更迅猛的颠覆性，以及它对生物医药等前沿科学创新与能源物理瓶颈带来的不对称影响。针对行业变革，德克斯指出代码自动生成正在打破纯技术工种优势，迫使高等教育重塑为终身学习与批判性思维导向。在全球竞合方面，德克斯对比了美、欧、中三地的监管与科研路径，警示过度规制与狭隘民族主义，并断言面对少数前沿实验室的算力与资本垄断，中小企业突围的关键在于深耕垂直利基场景与实际应用。 ## Atmosphere & style 整场对话洋溢着学术思辨的厚重感与行业一线的务实观察，节奏从容而视野宏大。普里的提问直击产业从业者的焦虑痛点，而德克斯则凭借历史学家与顶尖高校掌舵者的复合背景，以沉稳幽默且富有同理心的语调拆解复杂议题。整体氛围呈现出一种审慎、理性且充满建设性的“有条件乐观”，既不盲目鼓吹神话，也深刻洞察着技术落地的现实重力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=u25ZDHryTKI&t=4s) AI 变革与以往农业或工业革命本质不同，其颠覆速度不在同一量级，将以极短周期重塑工作、教育、医疗乃至重新定义人类本质。 - [03:09](https://www.youtube.com/watch?v=u25ZDHryTKI&t=189s) 当前科学创新呈现不均衡的双轨态势：生物医药与健康寿命领域因 AI 狂飙突进，但核聚变、室温量子计算与绿氢等底层物理瓶颈仍亟待突破。 - [03:09](https://www.youtube.com/watch?v=u25ZDHryTKI&t=189s) AI 作为新型创新经济的核心引擎面临潜在安全威胁（如自主致命武器系统），且模型表现始终受限于训练数据的质量。 - [07:31](https://www.youtube.com/watch?v=u25ZDHryTKI&t=451s) 代码自动化生成工具的普及打破了计算机科学专业的就业“黄金定律”，仅具备基础编程技能的初级程序员正面临淘汰危机。 - [07:31](https://www.youtube.com/watch?v=u25ZDHryTKI&t=451s) AI 浪潮倒逼教育体系重回通识博雅教育，批判性思维、甄别高质量数据以及校验大模型幻觉的能力成为新的核心竞争力。 - [11:00](https://www.youtube.com/watch?v=u25ZDHryTKI&t=660s) 传统的线性学制难以适应知识更新周期，未来教育必须转向构建终身学习体系，重点培养学生“学会如何学习”的元能力。 - [12:10](https://www.youtube.com/watch?v=u25ZDHryTKI&t=730s) 全球 AI 竞争呈现分化格局：欧洲深陷过度监管，美国坚持创新优先，中国则依托集中规划与强劲的基础科研快速追赶。 - [12:10](https://www.youtube.com/watch?v=u25ZDHryTKI&t=730s) 稳定且可预测的监管框架虽在短期带来合规成本，但能在未来 3 到 5 年内释放生产力红利；科研资助体制应加大对高风险前沿探索的容忍度。 - [12:10](https://www.youtube.com/watch?v=u25ZDHryTKI&t=730s) 以清华大学为代表的中国顶尖高校在 AI 研发实力上正逼近美国名校，西方应警惕封闭的“虚假 AI 民族主义”，积极寻求全球市场共识。 - [16:06](https://www.youtube.com/watch?v=u25ZDHryTKI&t=966s) 2024 年全球 2500 亿美元 AI 投资及媒体关注高度集中于少数前沿模型实验室，即便是顶尖大学也无力承担底层通用大模型的算力军备竞赛。 - [16:06](https://www.youtube.com/watch?v=u25ZDHryTKI&t=966s) 中小企业和初创团队的商业破局点在于避开底层基座模型，深耕垂直利基市场（niche）并打造与具体应用场景结合的产品落地方案。 - [19:04](https://www.youtube.com/watch?v=u25ZDHryTKI&t=1144s) 对 AI 未来的有条件乐观依赖于两大支柱：一方面保持创新经济飞轮运转，另一方面必须在全社会建立起对 AI 技术的深层公众信任。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=u25ZDHryTKI&t=4s) 历史转折点：AI 为何不同于以往的一切工业革命对话在拉斯维加斯 Ai4 大会的现场拉开帷幕。主持人、RegulatingAI 主席兼创始人桑杰·普里（Sanjay Puri）在开场中向观众隆重介绍了纽约科学院（The New York Academy of Sciences）院长兼首席执行官、加利福尼亚大学伯克利分校（UC Berkeley）前校长尼古拉斯·德克斯（Nicholas Dirks）。普里风趣地表示，德克斯的资历极其深厚，如果要把他在学术界、科研机构和公共政策领域的领导履历一一列尽，整场研讨会的时间恐怕都不够用。面对当下的技术浪潮，普里抛出了核心议题：历史上人类经历过多次自称“独一无二”的技术拐点，而身处连接企业界、学术圈与政策制定核心枢纽的德克斯，究竟如何看待当前人工智能所带来的冲击？作为一名历史学家出身的学者，德克斯给出了自己的深度观察。他坦言，如今探讨 AI 的文章无一例外都会从农业革命、工业革命（从第一次到第四次工业革命）以及早期的数字革命谈起。然而，以往那些深刻重塑人类劳动分工乃至整个社会生活肌理的重大历史转折，其演变过程往往横跨数十年甚至几个世纪，有着非常漫长的适应期。而今天摆在全人类面前的这场变革，不论各方所争论的通用人工智能（AGI）或超级智能（Superintelligence）究竟会在 3 年、5 年、10 年还是 20 年后降临，其颠覆速度都完全不在同一个量级上。正如前一位发言的国会议员所言，AI 正在改变一切。它不仅将重构工作的未来、教育的未来、医疗保健的未来，还将彻底颠覆休闲娱乐、人际社交关系，乃至重新定义“何以为人”（what it means to be human）的本质。 ```mermaid flowchart TD accTitle: 历史技术革命演进与 AI 变革模式对比 subgraph PastRevolutions["历史上的技术革命（长周期演进）"] A["农业革命 / 工业革命 1.0-4.0 / 数字革命早期"] B["生产力与工作方式重组"] C["历经数十年至数百年扩散，社会肌理缓慢适应"] A --> B --> C end subgraph AIRevolution["AI 与 AGI 变革（瞬时全面渗透）"] D["现代 AI 爆发：大语言模型 / 自主系统"] E["生产工具与认知边界全面重构"] F["极短周期内彻底重塑未来：工作、医疗、教育乃至人类本质定义"] D --> E --> F end ``` ### [03:09](https://www.youtube.com/watch?v=u25ZDHryTKI&t=189s) 从伯克利到前沿科学：大模型爆发与全学科创新瓶颈当被问及科研界与知名学术机构如何应对这一变革时，德克斯回忆起了自己主政 UC Berkeley 时的往事。伯克利拥有全美排名前三的计算机科学专业，汇聚了众多人工智能领域的学术奠基人。早在 2012 至 2013 年间，学术前线就已经明确感知到了技术范式转换的暗流——自 1955 年约翰·麦卡锡（John McCarthy）在达特茅斯会议上确立“人工智能”这一概念并开启人工神经网络探索以来，那个宏大梦想终于开始显现突破的迹象。为此，德克斯在伯克利力推设立了跨学科的数据科学项目，旨在将计算机科学、统计学与全校所有学科相连接。因为校方当时已预见到，数据与计算将横扫公共卫生、计算生物学乃至社会科学的每一个角落。随着过去几年大语言模型（LLM）的全面爆发，技术的聚光灯集中落在了特定赛道上。德克斯提到，自己身边的各路好友——从科技投资人里德·霍夫曼（Reid Hoffman）到认知科学家加里·马库斯（Gary Marcus），尽管两人在技术路线上观点截然不同（甚至对最新发布的各类大模型走向争论不休），但他们都一致公认：人类正在经历极其重大的科学发现。然而，德克斯援引经济学家泰勒·考恩（Tyler Cowen）等人的观察指出，当前科学创新的步伐其实呈现出极度不均衡的双轨态势： - **极速狂飙的领域**：在免疫学研究、生物医药及抗癌疗法开发中，AI 进驻实验室工作台带来了惊人的突破，为攻克人类数百年来不得不忍受并致死的顽疾、甚至大幅延长健康寿命（health span）带来了曙光。 - **亟待突破的深水区**：人类仍未能彻底解决核聚变（Fusion）、室温量子计算（Room-temperature quantum computing）、绿氢（Green hydrogen）等关乎能源与算力底座的重大物理瓶颈。德克斯强调，AI 将成为这种全新“创新经济”的绝对核心引擎。这既令人振奋，同时也令人担忧。除了社会对“末日论”的极端恐慌之外，AI 已经展现出潜在的致命风险，例如正在从根本上改变现代战争形态的自主致命武器系统（autonomous lethal weapon systems）。此外，一个最基本的底层现实始终无法回避：“AI 的优劣完全取决于其训练所依赖的数据质量”（AI is only as good as the data it trains on）。 ### [07:31](https://www.youtube.com/watch?v=u25ZDHryTKI&t=451s) 程序员面临退场危机？教育转向与文科批判性思维的复兴普里将话题切回现场的行业从业者与现实痛点：Ai4 大会的参会者绝大多数身处 AI 产业一线，其中大量工程师依赖编程与代码为生。然而当下 Cursor、Windsurf 等新一代 AI 辅助与自动化编码工具层出不穷，代码编写正在被大模型成批量代劳。过去社会上“只要学 STEM、拿到 CS 学位就能高枕无忧”的黄金定律正在迅速瓦解，应届毕业生甚至名校计算机毕业生的招聘形势变得愈发严峻。德克斯对此深有感触。他梳理了高等教育过去几十年间学生专业选择的演进历程： - 最早的传统大学阶段，学生普遍青睐英语文学、历史学和政治学等文科专业； - 其后，经济学与工商管理逐渐成为热门； - 在过去 15 至 20 年间，计算机科学（CS）彻底统治了校园。2012 年德克斯刚到伯克利任职时，计算机系主任戴维·帕特森（David Patterson，计算机体系结构宗师、图灵奖得主）就曾直接找上门要求：“我还需要额外 100 个全职教师（FTE）编制，才能勉强应付那些挤破头要选修我们课程的学生。” 然而时过境迁，德克斯直言，如果今天的大学生走出校门时“仅仅是一个平庸的初级代码敲砖工”（a bare coder），已经很难找到一份体面的工作，更遑论在人机协同的全新范式下立足。令人意想不到的是，这种技术拐点正在逼迫教育体系重新回归更广泛意义上的“通识博雅教育”（liberal arts）。德克斯指出，当下的核心能力不再是死记语法敲代码，而是深刻的批判性思维（critical thinking）：如何寻找高质量数据？如何验证数据的准确性？如何亲手校验大模型在展现神奇能力时所混杂的幻觉？谈及当下广泛流传的行业金句——“淘汰你的不是 AI，而是掌握了 AI 的人”——德克斯指出，这虽然准确概括了 AI 的泛在渗透性，但高等教育必须重新审视如何培养下一代人才。他回顾了 15 年前第一波慕课（MOOC）兴起时的尴尬教训：当年斯坦福大学的人工智能在线公开课瞬间吸引了 17 万人注册报名，但随后几乎全员退课，并未引发实质性的教育变革。而今天，AI 已经不可逆转地嵌入到教学实践的每一处肌理，成为未来学生求知与探索世界最底层的必备工具。 ### [11:00](https://www.youtube.com/watch?v=u25ZDHryTKI&t=660s) AI 重构教育范式与“终身学习”体系在探讨 AI 对学术研究与教学的冲击时，Nicholas Dirks 分享了身边的切实例子：他的妻子在加州大学伯克利分校（UC Berkeley）任教，在准备秋季学期的新课程时，已经开始广泛使用各种大语言模型（LLM）进行资料检索与教学设计。面对学生可能提交出“完美得令人难以置信”（too good to be true）的 AI 代写论文这一现实困境，Dirks 表现得相对乐观。他认为这属于教育技术演进中的适应问题——正如历史上学术界成功从计算尺过渡到计算器、再从计算器演进到计算机一样，教育界终会找到甄别与融合 AI 的技术与评估方案。然而，更深层次且尚未被系统思考的挑战在于：AI 将如何根本性地重塑“人才培养与毕业”的定义。在人类寿命不断延长、知识更新周期急剧缩短的时代，传统的“高中—四年制大学—硕士研究生”线性学制是否还能满足未来需求被打上了问号。Dirks 强调，未来的教育必须转向培养学生在整个生命周期中持续学习的能力，即真正掌握“学会如何学习”（learn to learn）的元能力，而不是单纯为了获取一纸文凭去通过几场阶段性考试。 ### [12:10](https://www.youtube.com/watch?v=u25ZDHryTKI&t=730s) 全球 AI 竞争格局：美欧监管路径与中国的科研崛起 Sanjay Puri 引导话题进入全球地缘竞争维度，指出当今世界 AI 格局呈现出鲜明的三极分化：欧盟出台了具有强约束力的《人工智能法案》（EU AI Act）；中国依托举国体制与集中规划；而美国目前则奉行“创新优先、事后监管”的自由发展路径。针对这一态势，Dirks 分享了他在巴黎参加 AI 峰会的见闻。尽管法国总统埃马纽埃尔·马克龙（Emmanuel Macron）极力向外界释放“法国对 AI 商业敞开大门”的信号并大力扶持本土初创公司，但整体而言，欧洲依然深陷“过度监管”（overregulated）的泥潭。 ```mermaid flowchart TD accTitle: 全球主要经济体AI发展与监管路径对比 subgraph EU["欧洲联盟 (EU)"] E1["强监管优先"] --> E2["《欧盟人工智能法案》"] E2 --> E3["合规成本高 / 创新受限"] end subgraph US["美国 (US)"] U1["创新与市场优先"] --> U2["试错与高风险研发"] U2 --> U3["寻求长期稳定的监管共识"] end subgraph CN["中国 (China)"] C1["集中规划与公私协同"] --> C2["清华等顶尖高校高强度基础科研"] C2 --> C3["全产业链创新与全球角逐"] end ``` 谈及监管哲学，Dirks 表示赞同斯坦福大学经济学家埃里克·布莱恩约弗森（Erik Brynjolfsson）的观点：一定程度的监管短期内虽会增加合规成本，但只要建立起“稳定且可预测的监管框架”，在未来 3 到 5 年的中期时间窗口内反而会促进生产力的大幅跃升。过早过度的行政干预会扼杀技术试错，美国过去二三十年在科学资助体系上对“高风险探索”的支持显得有些保守，合理的路径应当是加大对前沿冒险科研的资金支持，并在技术成熟后进行追溯性规制。在谈到中国时，Dirks 指出中国虽然在公私协同与国有企业联动层面有着高度的规制，但在基础研究领域的投入与创新爆发力极强。根据他在北京清华大学的深入交流经验，中国顶尖工科院校的 AI 研发实力正在迅速逼近斯坦福（Stanford）、麻省理工（MIT）、加州大学伯克利分校（UC Berkeley）和卡内基梅隆大学（Carnegie Mellon）。他警示西方不能因 DeepSeek 等引发轰动的单一事件而退缩至封闭的“虚假 AI 民族主义”（faux AI nationalism），而应建立全球化的市场理解与监管共识，同时印度等新兴市场的崛起同样不容忽视。 - **监管稳定性假说**：短期合规成本换取 3 至 5 年内的生产力红利，前提是政策不能朝令夕改。 - **科研资助体制反思**：呼吁改革现有科研拨款机制，鼓励承担更高风险，而非因噎废食地实施事前严管。 - **中国顶尖科研追赶**：以清华大学为代表的中国顶级学术机构在 AI 领域正与美国顶尖名校形成并驾齐驱的竞争态势。 ### [16:06](https://www.youtube.com/watch?v=u25ZDHryTKI&t=966s) 巨头垄断困局下的破局点：垂直利基与算力突围针对市场上少数几家前沿实验室（Frontier Labs）垄断话语权与资源的现状，Sanjay 提出了中小企业与开源生态应如何生存的疑虑。Dirks 坦言，仅 2024 年全球 AI 领域的投资就高达 2500 亿美元，而这笔巨资的绝大部分市场份额与媒体头条都被三至四家超级机构所瓜分。无论是 OpenAI 的山姆·阿尔特曼（Sam Altman）、Anthropic 的达里奥·阿莫代（Dario Amodei），还是谷歌 DeepMind 兼诺贝尔奖得主德米斯·哈萨比斯（Demis Hassabis），这些行业领袖对 AI 颠覆科学发现的宏大预言虽然振奋人心，但也让许多中小参与者产生了无从下手的焦虑。 Dirks 从自身背景出发表达了对中小企业和研究者的声援。即使是声誉卓著的一流大学，也根本无力承担训练万亿级基础模型所需的庞大 GPU 算力集群。但他坚定地认为，未来的商业胜负手将发生在垂直应用层： “那些深耕特定利基领域（niche）、专注于具体应用场景与产品开发的小型企业，最终必将赢得胜利。” 因为超级模型必须依靠这些精准贴合实际业务场景的落地方案，才能真正兑现商业价值。面对初创企业在无休止的研发与融资中所承受的巨大压力，身为非营利机构负责人的 Dirks 幽默地表示自己常年奔波于筹资，因此对创业者面临的艰辛“感同身受”（"I know, I feel your pain"）。 - **资金高度集中**：2024 年 2500 亿美元 AI 投资主要流向极少数几家底层模型研发实验室。 - **算力鸿沟现实**：传统大学科研机构缺乏足够的 GPU 算力硬件来训练自主基础大模型。 - **中小企业破局路径**：避开底层通用模型的军备竞赛，聚焦垂直行业应用（Niche-focused product development）与生态落地。 ### [19:04](https://www.youtube.com/watch?v=u25ZDHryTKI&t=1144s) 结语：有条件的乐观与公众信任的基石在对谈的尾声，Sanjay 询问 Dirks 对 AI 的未来发展究竟持积极还是谨慎态度。由于现场时间所限，Dirks 给出了一个“有保留的肯定”（qualified thumbs up）。他总结道，保持乐观的前提在于我们必须同时做到两点：一方面维系创新经济的飞轮持续运转，另一方面必须在公众群体中建立起对 AI 技术落地的深层信任（public trust）。如果无法消除社会的疑虑与恐慌，公众信任的缺失将成为锁死 AI 发展潜力的最大瓶颈。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=u25ZDHryTKI&t=4s) AI 正在改变一切。它不仅将重构工作的未来、教育的未来、医疗保健的未来，还将彻底颠覆休闲娱乐、人际社交关系，乃至重新定义“何以为人”（what it means to be human）的本质。 — Nicholas Dirks > [03:09](https://www.youtube.com/watch?v=u25ZDHryTKI&t=189s) AI 的优劣完全取决于其训练所依赖的数据质量（AI is only as good as the data it trains on）。 — Nicholas Dirks > [07:31](https://www.youtube.com/watch?v=u25ZDHryTKI&t=451s) 我还需要额外 100 个全职教师（FTE）编制，才能勉强应付那些挤破头要选修我们课程的学生。 — David Patterson > [07:31](https://www.youtube.com/watch?v=u25ZDHryTKI&t=451s) 仅仅是一个平庸的初级代码敲砖工（a bare coder） — Nicholas Dirks > [07:31](https://www.youtube.com/watch?v=u25ZDHryTKI&t=451s) 淘汰你的不是 AI，而是掌握了 AI 的人 — Nicholas Dirks > [11:00](https://www.youtube.com/watch?v=u25ZDHryTKI&t=660s) 学会如何学习（learn to learn） — Nicholas Dirks > [12:10](https://www.youtube.com/watch?v=u25ZDHryTKI&t=730s) 虚假 AI 民族主义（faux AI nationalism） — Nicholas Dirks > [16:06](https://www.youtube.com/watch?v=u25ZDHryTKI&t=966s) 那些深耕特定利基领域（niche）、专注于具体应用场景与产品开发的小型企业，最终必将赢得胜利。 — Nicholas Dirks > [16:06](https://www.youtube.com/watch?v=u25ZDHryTKI&t=966s) I know, I feel your pain — Nicholas Dirks > [19:04](https://www.youtube.com/watch?v=u25ZDHryTKI&t=1144s) 保持乐观的前提在于我们必须同时做到两点：一方面维系创新经济的飞轮持续运转，另一方面必须在公众群体中建立起对 AI 技术落地的深层信任（public trust）。 — Nicholas Dirks ## Mentioned Ai4, Sanjay Puri, RegulatingAI, Nicholas Dirks, The New York Academy of Sciences, UC Berkeley, John McCarthy, Reid Hoffman, Gary Marcus, Tyler Cowen, Cursor, Windsurf, David Patterson, Emmanuel Macron, EU AI Act, Erik Brynjolfsson, 清华大学, Stanford, MIT, Carnegie Mellon, DeepSeek, OpenAI, Sam Altman, Anthropic, Dario Amodei, Google DeepMind, Demis Hassabis ## Video description Sanjay Puri, Chairman/Founder, RegulatingAI Nicholas Dirks, President and CEO, The New York Academy of Sciences Abstracts In this conversation, we’ll explore how AI policy is shaping global economic ecosystems—from innovation pipelines and competitive dynamics to equitable access to emerging markets. [Source](https://www.youtube.com/watch?v=u25ZDHryTKI)

Source quotes

[0:04] Nicholas Dirks: AI 正在改变一切。它不仅将重构工作的未来、教育的未来、医疗保健的未来，还将彻底颠覆休闲娱乐、人际社交关系，乃至重新定义“何以为人”（what it means to be human）的本质。 [3:09] Nicholas Dirks: AI 的优劣完全取决于其训练所依赖的数据质量（AI is only as good as the data it trains on）。 [7:31] David Patterson: 我还需要额外 100 个全职教师（FTE）编制，才能勉强应付那些挤破头要选修我们课程的学生。 [7:31] Nicholas Dirks: 仅仅是一个平庸的初级代码敲砖工（a bare coder） [7:31] Nicholas Dirks: 淘汰你的不是 AI，而是掌握了 AI 的人 [11:00] Nicholas Dirks: 学会如何学习（learn to learn） [12:10] Nicholas Dirks: 虚假 AI 民族主义（faux AI nationalism） [16:06] Nicholas Dirks: 那些深耕特定利基领域（niche）、专注于具体应用场景与产品开发的小型企业，最终必将赢得胜利。 [16:06] Nicholas Dirks: I know, I feel your pain [19:04] Nicholas Dirks: 保持乐观的前提在于我们必须同时做到两点：一方面维系创新经济的飞轮持续运转，另一方面必须在公众群体中建立起对 AI 技术落地的深层信任（public trust）。

Source quotes

[0:04] Nicholas Dirks: Whether one forecasts that artificial general intelligence (AGI) or superintelligence is slated to emerge in three, five, ten, or twenty years, Dirks argues the exact horizon matters far less than the sheer comprehensiveness of the disruption underway. [0:04] Nicholas Dirks: artificial intelligence is poised to fundamentally rewrite the future of work, higher education, healthcare delivery, leisure, social relationships, and the baseline human condition itself. [7:31] Nicholas Dirks: it's not that AI is going to displace jobs, but people who use AI will take over from the people who don't [7:31] Nicholas Dirks: possessing entry-level coding skills is no longer enough to secure a career in tech. [7:31] Nicholas Dirks: learn to learn [12:10] Sanjay Puri: innovation first, regulation later [12:10] Nicholas Dirks: The paramount danger is not the presence of rules, but the lack of regulatory stability; unpredictable policy whiplash paralyzes capital deployment and technical experimentation. [12:10] Nicholas Dirks: faux kind of AI nationalism [18:19] Nicholas Dirks: are actually going to win in the end [18:19] Nicholas Dirks: I run a nonprofit. I spend my time doing fundraising too. I know, I feel your pain. [18:19] Nicholas Dirks: ultimate constraint [18:19] Sanjay Puri: qualified thumbs up

Original transcript

[0:04] truly an honor to have you here and [0:06] thank you all for coming which I think [0:08] is an important session because Nick [0:11] truly brings a really incredible [0:14] perspective. Uh she talked about his [0:17] background. We won't spend a lot of time [0:19] because we could do the whole session [0:20] about you uh in terms of the leadership [0:23] you have at the National Academy of [0:25] Sciences, your leadership at the UC [0:28] University of California, but we are [0:30] living in very interesting times uh Nick [0:34] and not trying to age you in any way but [0:38] when you look back at many different you [0:42] know inflections that we've had people [0:45] say you know this is unique this is [0:48] special when we talk about all those [0:50] different inflections when you look at [0:52] artificial intelligence especially this [0:54] session is about innovation it's about [0:57] competition [0:59] what is your perspective given you know [1:02] you deal with companies a lot you deal [1:04] with academics a lot you deal with [1:07] policy leaders a lot so you have a [1:08] unique perspective so our audience would [1:11] be interested to understand this moment [1:13] how does it feel [1:15] >> well thank you Sanjay and it's great to [1:17] be here with you all. Uh I am actually [1:19] by uh going way back a long time ago a [1:22] historian and so I've been thinking [1:24] about the question of you know is this [1:26] revolution fundamentally different or is [1:29] it just another revolution along the [1:31] historical timeline and I think it is [1:34] and it is for several different reasons. [1:36] If you go back and you know everybody [1:37] who writes about AI these days usually [1:39] begins with the agricultural revolution, [1:41] the industrial revolution, the first [1:43] phase of the digital revolution or you [1:46] know different modes of industrial [1:48] revolution first, second, third, fourth [1:51] but all of those revolutions which were [1:54] major in terms of the reorg [1:56] reorganization not just of work but of [1:58] our entire societies of our the fabric [2:01] of life took place over very long [2:03] periods of time. And what is happening [2:06] today whether you believe that AGI or [2:10] something like super intelligence or [2:12] whatever is coming in 3 years, 5 years, [2:14] 10 years or 20 doesn't really matter. [2:18] The point of course that we're all here [2:20] at in Las Vegas to to to to try to [2:23] understand more about is that AI is [2:26] going to and we heard this from the [2:28] congressman just now change everything. [2:30] And in changing everything, whether it's [2:32] our uh the future of work, whether it's [2:35] the future of education, whether it's [2:37] the future of health care, whether it's [2:38] the future of leisure, whether it's the [2:40] future of social relationships, whether [2:42] it's the future even of how we [2:43] understand what it means to be human. [2:46] Mhm. [2:48] We're confronting something that could [2:51] involve a total remake of who we are, [2:55] how we live, how we uh imagine the world [2:58] and how we uh ultimately uh will live in [3:01] this world. [3:03] >> So it is a a completely unique as you [3:06] said how we are, how we live, how we [3:08] work [3:09] from a scientific perspective. Is your [3:12] organization making any changes? tell us [3:16] a little bit about how is the National [3:18] Academy of Sciences looking at this. Uh [3:21] so first of all uh as a historian I [3:23] didn't ne necessarily encounter AI just [3:26] you know years and years and years ago [3:28] although when I was at Berkeley [3:30] obviously Berkeley is a great university [3:32] computer science one of the top three [3:34] programs in the country a place where [3:36] some of the leading uh uh real pioneers [3:39] in artificial intelligence have worked [3:42] over the years and it was already clear [3:45] back in 2012 2013 that things were going [3:47] to change that there was going to be a [3:50] revolution in terms of the you know [3:52] ultimate uh dream of people like John [3:55] McCarthy who invented the term [3:57] artificial intelligence back at [3:58] Dartmouth in 1955 [4:00] and of course that was when neural [4:02] networks were beginning to be used and [4:04] explored and so on. Uh and I watched [4:08] this this revolution begin and saw it [4:11] unfold. What we did at Berkeley was to [4:14] establish a data science program which [4:16] was actually meant to bridge computer [4:18] science and statistics with everything [4:20] else in the university because we [4:22] already understood that it was going to [4:24] impact everything whether it was public [4:25] health whether it was computational [4:27] biology whether it was even you know [4:29] questions in social science and so on. [4:32] So we uh uh we established this and of [4:35] course you know that that sense of the [4:38] future uh has has been vindicated and [4:41] over the last few years we've all [4:43] witnessed together the explosion in [4:45] terms of large language models but you [4:47] know when we're actually looking at the [4:49] future of AI right now we're looking at [4:51] a at a at a at a technology we're [4:53] looking at a particular domain within AI [4:56] and there are you know people who [4:58] friends of mine ranging from Reed [5:00] Hoffman to Gary Marcus who have [5:02] different views about uh you know even [5:05] the release the other day of chat GBT5 [5:09] but they nevertheless I think both agree [5:12] about the extent to which this is a [5:14] momentous [5:16] >> uh uh set of discoveries but it's going [5:18] to impact a lot of other things in [5:19] science and this will go to the question [5:20] of innovation [5:22] >> in some ways science is moving faster [5:25] than ever in other ways and various [5:27] commentators have talked about this [5:28] Tyler Cowan among others uh it's it's [5:31] it's not moving so fast. Healthcare, [5:34] some areas in biomedical science, [5:36] amazing breakthroughs. Some of them [5:38] related to immunological work, some of [5:41] them related to the power that AI has [5:44] brought to uh the the the lab bench. But [5:48] other areas really we need to [5:50] simultaneously as it were uh explode uh [5:54] all you know uh all over the place [5:56] because we need to figure out [5:59] fusion. We need to figure out room [6:02] temperature quantum computing. We need [6:04] to figure out green hydrogen. We need to [6:06] be able to figure out a whole variety of [6:09] things. And then going back to the world [6:11] of biomedical research, we really need [6:14] to make good on the kinds of [6:16] developments that are taking place [6:17] around cancer therapeutics, which are [6:20] really extraordinary, around a whole set [6:22] of other diseases that we have over the [6:24] centuries learned to live with and die [6:27] from, but which all of a sudden we now [6:31] have the capacity to do major things. uh [6:34] to address if not to live forever as [6:37] some people uh uh close to my [6:40] neighborhood in Berkeley believe but uh [6:42] for a much longer health span perhaps [6:45] than we have ever experienced in in [6:47] life. But the point is that AI is going [6:50] to be central to this new innovation [6:52] economy and that's both extraordinary. [6:57] It's also frightening. [6:59] AI is both incredibly powerful and again [7:04] uh and there are naysayers and [7:06] doomsayers and others who will take this [7:08] to a very very extreme point uh but [7:11] potentially dangerous and we already [7:12] know the extent to which it has uh [7:14] played a role in terms of the [7:16] development of autonomous lethal weapon [7:17] systems that are going to change [7:20] completely the nature of warfare. But we [7:22] also know that AI is only as good as the [7:27] data it trains on. [7:31] >> You are at the AI4 conference which you [7:34] never thought you would be but thank you [7:35] for coming. Most of the people here are [7:38] in the business of artificial [7:40] intelligence. A lot of them happen to be [7:43] doing AI programming and coding for a [7:46] living. [7:47] Now the beauty of AI is as you said it [7:50] does great things but it can also help [7:52] you code and it is impacting a lot of [7:57] those kinds of things when you look at [7:58] technology you know tools like cursor [8:01] winds surf and many of those things and [8:04] you see that implication in the hiring [8:07] especially for fresh graduates it is [8:10] getting harder and harder for people who [8:12] are coming out even with people used to [8:14] say get a STEM degree you'll be okay get [8:16] a CS degree you'll be okay. Um from your [8:21] perspective we are seeing some major [8:24] shifts happen in the economy. Some of [8:26] these people work in those right now. [8:28] Obviously hopefully they're all doing [8:31] good because they're here. But how do [8:34] you see this thing playing out Nick? [8:36] >> Yes question and it's one that we ask in [8:39] the world of higher education a great [8:40] deal over the years and I've had a few [8:43] in the world of universities. uh we saw [8:46] you know a a a change in terms of [8:50] student preferences. Uh for a long time [8:52] students would major in English, major [8:54] in history, major in political science. [8:56] Increasingly they began to major in [8:58] economics, business administration and [9:00] then of course it's been computer [9:03] science and it's been computer science [9:04] for 15 20 years. is when I first got to [9:07] Berkeley in 2012, I was asked by the [9:09] chair of the department, uh, David [9:11] Patterson, some of you know the name, I [9:13] need another hundred FTE, full-time [9:16] equivalent faculty to teach the students [9:19] who want to take our courses. But now, [9:22] as you say, uh if you graduate from [9:24] college and you're just kind of a a be [9:28] uh coder, it's not clear you're going to [9:30] get either a job or really be able to uh [9:33] figure out how how to work in uh in and [9:37] with AI. Uh increasingly, again, there's [9:40] a way in which we're going back to a [9:42] broader sense of liberal arts because [9:44] you need, as the congressman also said [9:46] just now, you need to develop critical [9:48] thinking. You have to know something [9:49] about data. How do you find data? How do [9:51] you test the accuracy of data? How do [9:53] you actually uh yourself uh check the [9:57] different kinds of ways in which LLMs [10:00] can sometimes do extraordinary things [10:02] and sometimes hallucinate. But the [10:05] question of how and and and what it will [10:08] take to train the next generation of [10:11] graduates is one in which I think we [10:12] have to really parse what it means when [10:14] people say as a kind of uh a kind of [10:17] slogan. They say it's not that AI is [10:19] going to dis displace jobs but people [10:21] who use AI will take over from the [10:24] people who don't which is again an [10:27] extrapolation of the general point which [10:29] is that it's going to affect virtually [10:31] everything but what will that take and [10:33] is it possible now to even think about [10:35] AI as a uh as a tool in education that [10:38] will do much more than we predicted 15 [10:41] years ago when the first muks the [10:43] massive online open courses were [10:45] developed at Stanford and 170,000 and [10:47] people signed up to take a course in [10:48] artificial intelligence and then dropped [10:51] out. Uh uh now it's clear that AI is [10:56] going to be embedded. It's going to be [10:58] how students uh uh often do their [11:00] research. I know I'm uh I I'm I'm a [11:03] professor, but my wife is preparing for [11:05] her she's teaching at Berkeley and she's [11:07] preparing for her classes in September [11:10] and she's using a variety of different [11:12] uh LLM models to do research and it's uh [11:16] it's of course a question what is going [11:18] to happen when you get a paper that [11:20] looks too good to be true. But that is [11:23] something we can figure out. uh we [11:25] figured out how to, you know, use the uh [11:27] the calculator over the slide rule and [11:29] the computer over the over over the [11:31] calculator. We'll figure this out. But [11:33] what we haven't really begun to figure [11:35] out is how is this going to change what [11:37] it means to prepare students to [11:39] graduate? And are even the fundamental [11:42] structures of high school, the four-year [11:44] college, uh you know, maybe the master's [11:46] degree and so on and so forth, is that [11:48] even close to adequate for a world in [11:50] which we'll live longer? We'll need to [11:52] learn more across the full uh span of [11:55] our lives. We'll need to prepare [11:57] students to do a lot of different kinds [11:59] of things and to have the capacity to [12:01] actually learn to learn over the course [12:05] of their lives and not just pass some [12:07] tests on the way to getting a degree. [12:10] Um since competition is a part of this [12:13] panel, I have two questions on that. [12:15] Nick, one is you've had exposure in [12:18] South Asia. Right now when you look at [12:21] the world of AI, there are basically [12:23] three spheres or you could argue maybe [12:26] two spheres. It's you know the European [12:28] Union has the EU AI act. We have China [12:31] which has a more centralized focus and [12:34] in the US now our focus is it's [12:37] innovation first regulation later. [12:40] How do you look at it from a global [12:42] competitive perspective? Then I'll move [12:44] to the US competitive environment. [12:46] >> Yeah. [12:47] >> So, you know, the obvious uh uh [12:50] distinctions between the US and the EU [12:53] are that the EU is overregulated. I was [12:56] in Paris uh you know for the February [12:58] summit. uh President Mcron was very [13:01] interested in both announcing that [13:03] France was open for AI business and he [13:05] was in fact promoting of course a number [13:06] of startups that uh were based in in [13:09] France uh and that in the US were much [13:12] more innovative and so on although [13:14] there's obviously a great deal of [13:16] conversation and debate about the role [13:18] of regulation right now I tend to agree [13:21] with Eric Bernolson from Stanford who [13:24] has said that uh a certain level of of [13:26] regulation uh is actually going to be [13:29] even though it's costly in the short [13:31] term uh uh something that will conduce [13:34] to greater productivity in the long term [13:36] and and we're not talking about 20 years [13:38] down the line he's talking about 3 to 5 [13:40] years uh uh but the real key there is [13:43] going to be uh the stability of the [13:46] regulatory framework if it comes and [13:48] goes it's like every other political [13:52] constraint or externality that [13:55] effectively will uh make for real [13:57] difficulty if you can't predict what's [13:59] going to come. So I think it's important [14:00] that we establish a certain level of [14:02] consensus around regulation in the US [14:04] and we abide by it. Understanding that [14:08] too much regulation early on can be very [14:10] uh uh uh problematic for for innovation. [14:13] You have to be able to try a lot of [14:15] different things out. And I think we've [14:17] seen over the last 20 30 years and read [14:20] this conversation about science and [14:21] innovation that sometimes we haven't [14:23] taken enough risks in science. And I [14:26] think uh we have a system and a [14:28] structure for funding science which is [14:31] another conversation but nevertheless [14:34] that needs to be revised not with less [14:36] support but with more in part for taking [14:39] risks and then understanding that you [14:41] may have to come in and regulate after [14:43] the fact as opposed to before. But China [14:46] is a different matter altogether. [14:49] China's much more highly regulated in [14:51] some domains obviously with respect to [14:54] you know the collaboration between [14:56] public and private stateowned enterprise [14:58] and so on. Uh but it's also of course uh [15:01] incredibly innovative right now and that [15:03] also speaks to investments in uh in [15:06] fundamental research. I spent a lot of [15:09] time in Chinua University in Beijing and [15:11] I can assure you that AI uh at the top [15:14] engineering university in China is very [15:16] soon going to compete very very closely [15:20] with Stanford, MIT, Berkeley, Carnegie [15:23] Melon and so on. Uh and I think that's [15:26] going to be reflected in uh in the [15:28] economy overall. So there are real [15:31] issues I think beyond you know sort of [15:33] the headlines of deepseek uh as to how [15:36] we think about a global uh a global [15:39] marketplace but also a global uh [15:42] regulatory set of understandings around [15:45] AI uh that will allow us uh not to [15:49] retreat into a kind of faux kind of AI [15:52] nationalism that leaves us [15:54] uncompetitive. We have taken a lot of [15:56] comfort from the fact that Europe has [15:58] not China is India is another matter and [16:01] that is something we could perhaps talk [16:03] about later. [16:03] >> Yeah. Kind of getting a small hook but I [16:07] can't resist asking you since we've [16:08] talked about competition globally. AI is [16:11] if you look around is dominated by the [16:14] same three four five companies here. Is [16:17] that good bad? What does it mean for [16:20] small business? a lot of small business [16:22] entrepreneurs who rely on uh openness, [16:25] open source, etc. Message for them. What [16:28] do you think? So, we all live in this [16:31] world in which yeah, four or five [16:33] frontier labs steal all the headlines uh [16:36] and certainly get uh uh the massive [16:39] market share of the $250 billion that [16:42] went into AI uh investment just in 2024 [16:46] 2024. Uh and of course you know just [16:49] this also speaks a little parenthesis [16:51] here but this also speaks to the [16:53] question of innovation because when you [16:56] have three or four people who are saying [16:59] uh you know this is going to happen in [17:01] the space of AI uh with usually the [17:04] release of a new uh of a new product and [17:08] you know good people I don't mean just [17:10] Sam Alman and uh Dario Amade I mean Deis [17:14] Hassabus who does the same even though [17:17] He's in the research side of Google and [17:19] he doesn't uh uh himself talk about [17:22] Gemini or notebook or any of the other [17:24] products that are coming out of Google. [17:26] But you know the kinds of statements [17:27] he's making about and he's a Nobel [17:29] laureate. The kinds of statements he's [17:30] making about the future uh impact of AI [17:33] on on scientific research and and [17:35] discovery makes you think oh okay why [17:37] are we even bothering with this now you [17:40] know just as a form of solidarity for [17:43] those of you in smaller businesses. [17:45] First of all, coming from universities [17:47] which are large but don't have anything [17:51] like the resources to uh to to to create [17:55] our own uh our own systems of uh our own [17:59] platforms, our own large language [18:01] models. uh you know we've we've been [18:03] watching you know this sort of growth of [18:06] the frontier labs with a certain level [18:08] of you know wouldn't we and of course [18:11] again university uh research is in the [18:13] news every day but we have nothing like [18:16] the capacity uh to to to have the the [18:19] GPUs needed to to to to really do uh the [18:22] kind of work that needs to be done but I [18:24] think there are partnerships and ways in [18:26] which the smaller uh the smaller [18:29] companies which are going to be more [18:30] niche [18:31] more focused on particular kinds of [18:34] product development that will be [18:36] absolutely critical for the ultimate [18:38] implementation and success of the larger [18:41] models that come out of the frontier [18:42] labs are actually going to win in the [18:45] end. But you have to stick in there and [18:47] obviously when you're in a startup that [18:49] sometimes means [18:52] you know just endless work and and [18:54] endless uh efforts at fundraising which [18:57] uh uh which by the way I run a [18:59] nonprofit. I spend my time doing [19:00] fundraising too. I know I feel your [19:03] pain. [19:04] >> Awesome. Uh we can continue this [19:07] discussion uh for hours on and end. [19:09] Nick, thank you so much. Uh this was [19:12] fantastic. Really love it. Are you [19:14] positive on AI or [19:17] >> if we can figure out ways to in short [19:20] answer because she's looking at me [19:21] >> but if we can figure there's one [19:23] sentence I know it's a long one [19:24] >> but uh if we can figure out a way to [19:27] ensure not only that we keep the [19:29] innovation economy going but that we can [19:31] instill in the public a real sense of [19:33] trust in what we're doing with AI [19:36] because that I think could be the [19:38] ultimate constraint. [19:40] Thumbs up. Okay. qualified thumbs up. [19:43] Thank you, Nick.