Andrew Ng Rejects AI Job Apocalypse, Cites Ongoing Tech Shortages

The Economist

Artificial intelligence will not cause mass unemployment because it automates discrete tasks rather than full jobs, while physical infrastructure bottlenecks slow progress and obsolete university curricula leave junior graduates unprepared.

AI-assisted summary of Will AI put millions out of work? | The Economist — original source. Check the original for full context.

Source published: 2026-10-14

Key ideas

Andrew Ng dismisses fears of an AI-induced jobs apocalypse, arguing that artificial intelligence automates discrete subtasks rather than entire roles, leaving overall demand for human labor intact. (0:00)
Predictions of rapid exponential AI advancement via recursive self-improvement remain largely hype, bounded by physical constraints such as power plant construction, data center deployment, and chip manufacturing lead times. (2:02)
Friction for entry-level workers stems from outdated university curricula rather than an outright disappearance of junior jobs, with computer science programs graduating students who have never used an AI coding agent or API. (4:54)

Detailed notes

## Background Published on YouTube by The Economist on October 14, 2026, this video features an interview between The Economist’s AI writer Alex Hern and Andrew Ng, CEO and founder of LearnVector. According to the description, Ng argues that fears of an AI jobs apocalypse are overblown because AI automates only partial tasks, progress faces real-world physical bottlenecks like data center construction, and the primary hurdle for young workers is that universities are too slow to teach the modern skills employers demand.

## TL;DR Andrew Ng dismisses catastrophic predictions of an AI employment apocalypse, arguing that artificial intelligence automates discrete subtasks rather than whole professions, which actually drives hiring demand for broader, AI-fluent roles. He clarifies that recursive self-improvement and overnight leaps to human-level versatility are overhyped, as real-world development is fundamentally constrained by slow physical realities like power grids, chip fabrication, and data center buildouts. Addressing concerns about incoming junior workers, Ng notes that entry-level demand remains strong for AI-native talent, pointing instead to outdated university computer science curricula as the real obstacle holding young graduates back.

## Atmosphere & style The exchange carries a calm, analytical, and grounded tone that cuts through Silicon Valley alarmism with sober engineering pragmatism. Alex Hern serves as a sharp, probing interlocutor who presses on pervasive industry fears regarding compounding intelligence and junior worker displacement, while Andrew Ng responds with relaxed confidence and practical clarity. The conversation feels reassuring without dismissing economic friction, pivoting the emotional register away from existential dread and toward actionable institutional challenges in power infrastructure and higher education.

## Bulletin - [00:00](https://www.youtube.com/watch?v=J-sMqnzHwI0&t=0s) Andrew Ng dismisses fears of an AI-induced jobs apocalypse, arguing that artificial intelligence automates discrete subtasks rather than entire roles, leaving overall demand for human labor intact. - [02:02](https://www.youtube.com/watch?v=J-sMqnzHwI0&t=122s) Predictions of rapid exponential AI advancement via recursive self-improvement remain largely hype, bounded by physical constraints such as power plant construction, data center deployment, and chip manufacturing lead times. - [04:54](https://www.youtube.com/watch?v=J-sMqnzHwI0&t=294s) Friction for entry-level workers stems from outdated university curricula rather than an outright disappearance of junior jobs, with computer science programs graduating students who have never used an AI coding agent or API.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=J-sMqnzHwI0&t=0s) Early empirical data, even from institutions previously predicting up to 50% unemployment, shows AI is not causing catastrophic net job losses. - [00:25](https://www.youtube.com/watch?v=J-sMqnzHwI0&t=25s) Because AI automates specific subtasks rather than whole occupations, human workers remain essential for executing the remaining components of their jobs. - [00:45](https://www.youtube.com/watch?v=J-sMqnzHwI0&t=45s) The labor market is currently experiencing acute hiring shortages for software engineers as well as AI-fluent talent in sales, marketing, and operations. - [01:03](https://www.youtube.com/watch?v=J-sMqnzHwI0&t=63s) Software development serves as an indicator for the broader knowledge economy, demonstrating that worker value expands as developers transition into broader full-stack roles augmented by AI. - [01:25](https://www.youtube.com/watch?v=J-sMqnzHwI0&t=85s) Front-end engineers whose routine interface coding is automated are successfully adapting by moving into back-end architecture, mobile clients, and system security. - [02:30](https://www.youtube.com/watch?v=J-sMqnzHwI0&t=150s) Achieving human-level versatility across 100% of tasks requires multiple fundamental breakthroughs that will likely take decades or longer rather than arriving by 2030. - [02:58](https://www.youtube.com/watch?v=J-sMqnzHwI0&t=178s) Recursive self-improvement is frequently overhyped; computers have assisted in designing new computer chips for decades without triggering runaway, instant leaps. - [03:30](https://www.youtube.com/watch?v=J-sMqnzHwI0&t=210s) The pace of AI scaling is physically bound by real-world infrastructure constraints, including electrical grid capacity, power generation, data center buildouts, and empirical testing cycles. - [04:54](https://www.youtube.com/watch?v=J-sMqnzHwI0&t=294s) While Erik Brynjolfsson's early research suggested AI exposure disproportionately harmed entry-level hiring prospects, subsequent data reflects a more optimistic outlook. - [05:32](https://www.youtube.com/watch?v=J-sMqnzHwI0&t=332s) Universities are lagging industry needs by awarding computer science degrees to students who have never utilized frontier model APIs or AI coding agents. - [06:15](https://www.youtube.com/watch?v=J-sMqnzHwI0&t=375s) Employers actively recruit junior talent, high schoolers, and college interns whenever candidates possess proven proficiency with AI-native tools and workflows.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=J-sMqnzHwI0&t=0s) Why There Will Be No AI Jobs Apocalypse

Andrew Ng rejects the catastrophic narratives surrounding artificial intelligence and the labor market. Addressing the fear of mass unemployment—often framed as half the population losing their livelihoods and descending into civil unrest—he insists there is simply no plausible path to such an outcome. Early empirical findings, even from institutions that previously sounded the alarm on 50% unemployment, are beginning to demonstrate that artificial intelligence is not devastating the workforce as anticipated. The fundamental reason is task composition: AI automates specific subtasks rather than entire roles, meaning human workers remain indispensable for executing the remaining responsibilities.

```mermaid flowchart LR accTitle: AI Automation Scope and Human Labor Demand accDescr: Diagram showing how AI automates specific subcomponents of a job while humans remain necessary for the remaining tasks, shifting market demand toward AI-native specialists.

subgraph JobRole["Traditional Job Role"] AutoTasks["Subtasks Automated by AI"] HumanTasks["Tasks Requiring Human Judgment & Execution"] end

JobRole --> Expansion["Upskilling & Broader Scope"] Expansion --> Market["Demand for AI-Native Talent<br>(Engineers, Sales, Marketing, Ops)"] ```

Looking at software development as a bellwether for the broader knowledge economy, Ng sees an optimistic pattern unfolding. Rather than facing widespread job cuts, employers are facing acute hiring shortages across multiple domains. Companies are struggling to secure enough software developers, and the same shortage is already emerging for AI-fluent professionals in sales, marketing, and operations. As Andrew Ng puts it:

> “And if what happens in software engineering is a predictor of what will happen in other job sectors as well, I think it'll be fantastic.”

The primary labor friction is therefore not a disappearance of roles, but the necessity of rapid upskilling. Workers must broaden their functional capabilities alongside learning AI tooling. In software, developers who previously operated strictly as front-end engineers—crafting user interfaces—find that AI handles interface coding effectively. In response, those engineers are expanding into full-stack engineering, managing back-end architecture, mobile clients, and system security while orchestrating AI systems. Far from becoming obsolete, workers who expand their scope become extraordinarily sought-after talent.

- Early data from organizations that initially warned of 50% unemployment are already showing that AI lacks the feared negative impact on net employment. - High industry demand persists for software developers, alongside a growing deficit of AI-native marketers, sales professionals, and operations staff. - Front-end developers are transitioning into full-stack engineers by combining broader responsibilities (back-end infrastructure, security, mobile) with direct AI workflow integration.

---

[02:02](https://www.youtube.com/watch?v=J-sMqnzHwI0&t=122s) Physical Bottlenecks and the Limits of AI Progress

*The Economist*’s Alex Hern presses Ng on whether his dismissal of an employment catastrophe rests on an assumption that AI capabilities will plateau. Hern notes that anxiety in the tech sector focuses not on models existing in 2026, but on where the technology might stand by 2030 or even six months out if progress compounds exponentially. Hern questions whether physical limits, architectural roadblocks, or breakdown in scaling laws will prevent models from ever mastering certain complex human tasks.

Ng clarifies that while he does not envision a hard brick wall for AI, progress will remain incremental and stretched across long horizons. While he jokes that he would happily take an extended vacation if artificial intelligence could replicate every human intellectual capability, achieving that level of general competence requires multiple fundamental breakthroughs that could take decades or longer. He firmly discounts claims that human-level versatility across 100% of tasks will occur within six months or by 2030. Addressing the widespread industry rhetoric around recursive self-improvement, Andrew Ng states:

> “And you use the term compounding progress. Um, one of the terms of hype in Silicon Valley has been um compounding progress or what recursive improvements and so on. And at least right now that's much more hype than reality.”

Ng explains that computers aiding in the creation of better computers is standard practice: prior-generation microchips have engineered newer microchips for decades, just as frontier artificial intelligence labs currently rely heavily on AI to write their own system code. However, the pacing of intelligence is governed by physical reality rather than pure algorithmic software loops. Advancing frontier capabilities demands physical infrastructure: deploying GPU server clusters, constructing dedicated electrical power plants, and conducting physical and empirical experiments. Superintelligent systems cannot escape real-world constraints, and those physical lead times represent hard, slow-moving bottlenecks.

- Achieving full parity with human intellectual tasks requires multiple novel scientific breakthroughs that may take decades or more. - Recursive self-improvement is seen largely as Silicon Valley hype; semiconductor manufacturing has utilized recursive tooling for decades without causing runaway instant leaps. - Major physical bottlenecks constraining AI advancement include electrical power grid capacity, data center construction lead times, server hardware fabrication, and empirical testing cycles.

---

[04:54](https://www.youtube.com/watch?v=J-sMqnzHwI0&t=294s) Why Universities Are Failing Young Workers

Hern shifts the conversation from macroeconomic collapse to targeted harm, specifically referencing research by Stanford economist Erik Brynjolfsson. Brynjolfsson's work suggests that higher exposure to AI within an industry correlates with worse employment prospects for incoming junior talent. With routine junior tasks like basic coding being handed off to software agents, entry-level candidates without an established track record risk being locked out of the classic apprenticeship path.

Ng notes that Brynjolfsson has since gathered newer data pointing to a more favorable outlook, while identifying the true problem: traditional higher education is failing to evolve at the speed of modern industry. Universities continue to graduate computer science majors who have never integrated a frontier model API or used an AI coding agent, leaving them unprepared for modern engineering positions. Describing the hiring standards of his own organizations, Andrew Ng makes clear:

> “And today I'm not going to hire a software engineer that's never done that. Um it's still possible to get a bachelor's degree in computer science in some universities without using a AI coding agent. And frankly none of my teams will hire a software engineer that doesn't know how to use a AI coding agent.”

Ng emphasizes that junior talent remains valuable when properly prepared. During the recent summer, his teams hired cohorts of high schoolers, university students, and recent graduates who completed significant work. The differentiator was that every intern hired was already AI-native, utilizing state-of-the-art tools and coding agents to execute tasks. The root challenge facing entry-level workers is not a structural elimination of junior positions, but an outdated university curriculum that fails to equip students with contemporary developer workflows.

- Erik Brynjolfsson’s research on AI-exposed sectors raised alarms about junior job prospects, though more recent data indicates a more positive trajectory. - Major universities still grant computer science degrees to students who have never executed an API call to a frontier model or used AI coding agents. - Organizations are actively hiring early-career workers, high schoolers, and college interns, provided they are proficient in AI-native tools and methods.

## Notable quotes > [04:54](https://www.youtube.com/watch?v=J-sMqnzHwI0&t=294s) And frankly none of my teams will hire a software engineer that doesn't know how to use a AI coding agent. — Andrew Ng

## Mentioned Alex Hern, Andrew Ng, Erik Brynjolfsson, LearnVector, Silicon Valley, Stanford University, The Economist

## Video description Will AI put millions of people out of work? Andrew Ng, CEO and founder of LearnVector, argues the fears are overblown. He tells Alex Hern, The Economist’s AI writer, that AI automates only part of most jobs, that its progress will be slowed by physical bottlenecks like building data centres and that the real problem for young workers is universities are too slow to teach the skills employers want now.

#ai #jobs #datacentres #artificialintelligence #technology

00:00 - Why there will be no AI jobs apocalypse 01:03 - How AI is reshaping software engineering 02:02 - Is there a limit to AI progress? 04:54 - Are universities failing young workers?

Watch the full show: http://bit.ly/4ety1EO Sign up to the Insider newsletter: https://econ.st/4nOyzIb

Subscribe to The Economist: https://www.economist.com/subscribe Download our app: https://www.economist.com/get-the-app

Follow The Economist on Instagram: https://www.instagram.com/theeconomist/ Follow The Economist on Facebook: https://www.facebook.com/TheEconomist Follow The Economist on X: https://x.com/TheEconomist

[Source](https://www.youtube.com/watch?v=J-sMqnzHwI0)

中文简报

吴恩达驳斥AI就业末日论，称技术人才持续短缺人工智能不会导致大规模失业，因为它自动化的是独立任务而非完整岗位，同时物理基础设施瓶颈减缓了进展速度，而过时的大学课程使应届毕业生缺乏足够准备。吴恩达驳斥了关于AI将引发就业末日的灾难性预测，认为人工智能自动化的是独立的子任务，而非整个职业，这实际上推动了对更广泛、熟悉AI的人才的招聘需求。他澄清道，递归自我改进以及一夜之间跃升至人类水平的通用能力被过度炒作了，因为现实世界的发展从根本上受到电网、芯片制造和数据中心建设等缓慢物理现实的制约。针对初级工人的担忧，吴恩达指出，对AI原生人才的初级岗位需求依然强劲，并指出大学过时的计算机科学课程才是阻碍年轻毕业生发展的真正障碍。吴恩达驳斥了对AI引发就业末日的担忧，认为人工智能自动化的是独立子任务而非整个岗位，对人类劳动力的总体需求依然完好。 (0:00) 通过递归自我改进实现AI快速指数级进步的预测在很大程度上仍是炒作，受制于电厂建设、数据中心部署和芯片制造周期等物理约束。 (2:02) 初级工作者的阻力源于过时的大学课程，而非初级岗位的彻底消失，计算机科学专业培养出的毕业生甚至从未用过AI编程智能体或API。 (4:54)

Reading summary

Computer scientist Andrew Ng dismisses warnings of an AI-induced labor apocalypse, arguing that artificial intelligence automates discrete subtasks rather than entire roles while physical infrastructure constraints prevent runaway technological leaps. Subtask Automation and Broadening Labor Demand (0:00) Ng rejects catastrophic scenarios of mass joblessness, noting that early empirical data from institutions that initially warned of 50% unemployment show no devastating impact on aggregate employment. Because AI automates specific subtasks instead of eliminating full positions, human judgment and execution remain necessary. In fields like software development, which Ng views as a bellwether for the wider knowledge economy, companies continue to experience hiring shortages rather than widespread layoffs. Similar talent deficits are emerging for AI-fluent workers across sales, marketing, and operations. The real labor challenge centers on worker upskilling. Rather than facing obsolescence, workers who integrate AI tools are expanding their functional responsibilities. For example, front-end developers whose basic interface coding is handled by AI are transitioning into full-stack engineers by managing back-end architecture, mobile clients, and system security while directing AI workflows. Physical Constraints and Gradual Model Progress (2:02) Addressing concerns raised by Alex Hern about potential compounding, exponential improvements by 2030 or within months, Ng characterizes rhetoric around recursive self-improvement as Silicon Valley hype rather than current reality. While computers assisting in the design of next-generation hardware has been routine in semiconductor manufacturing for decades, advancing frontier AI systems remains tethered to slow-moving physical dependencies. Achieving general human-level competence across all intellectual tasks will require multiple fundamental breakthroughs spanning decades or longer, according to Ng. The pacing of frontier capabilities is bound to material bottlenecks, including server hardware fabrication, empirical testing cycles, data center buildouts, and electrical grid capacity, precluding sudden runaway leaps. Curriculum Gaps and Junior Worker Prospects (4:54) Addressing research by Stanford economist Erik Brynjolfsson regarding junior workers facing diminished entry-level prospects in AI-exposed sectors, Ng notes that more recent data points to a more positive outlook. The primary obstacle confronting new entrants is not the structural destruction of entry-level jobs, but university programs that fail to update their curricula to match industry practices. Ng points out that many institutions continue to graduate computer science majors who have never called a frontier model API or worked with an AI coding agent, disqualifying them from teams that require these proficiencies. He stresses that organizations continue to actively hire junior talent, including interns and recent graduates, provided they have mastered AI-native workflows and contemporary developer tools. Ng maintains that artificial intelligence will not cause systemic unemployment, but the transition hinges on universities modernizing curricula to teach AI workflows and workers actively broadening their technical capabilities to match industry demands.

精炼阅读摘要

计算机科学家吴恩达（Andrew Ng）驳斥了人工智能将引发大规模失业的灾难性论调，强调AI仅自动化具体子任务而非完整职业角色；劳动力市场的核心问题不在于工作岗位消亡，而在于物理基础设施瓶颈对技术演进速度的制约，以及大学教育脱节未能培养出适应现代工作流的AI原生人才。任务级自动化与新型劳动力需求 (0:00) 吴恩达指出，有关AI导致半数人口失业并引发社会动荡的观点缺乏现实依据，早期实证数据表明AI并未对整体就业造成灾难性冲击。由于AI技术取代的是职业内部的特定子任务（subtasks），人类员工在执行综合判断与剩余职责上依然不可或缺。以软件工程领域为例，当前行业不仅未出现裁员潮，反而在软件工程以及熟练运用AI的销售、营销和运营等职能上普遍面临人才紧缺。物理瓶颈打破指数级演化神话 (2:02) 面对有关技术将在数月内或2030年前产生全面替代人类能力的担忧，吴恩达反驳称，实现覆盖100%人类任务的智力水平需要多项基础性科学突破，这一过程可能历时数十年。他认为硅谷宣扬的“递归自我改进”或指数级复合增长目前存在大量夸大成分；半导体行业数十年间始终使用上一代芯片设计新一代芯片，前沿AI实验室使用模型编写代码亦属常规辅助，并未引发失控的爆发式跳跃。吴恩达强调，AI的能力提升受限于现实物理世界。GPU集群的部署、电网与专用发电厂的建设、硬件制造交付周期以及实体经验实验的迭代，构成了纯算法循环无法绕过的硬性物理瓶颈，决定了AI技术的演进必然是渐进且耗时的长周期过程。高校培养滞后危及初级就业通道 (4:54) 针对斯坦福大学学者埃里克·布林约尔松（Erik Brynjolfsson）关于AI暴露度较高的行业对初级人才产生挤压的研究，吴恩达援引后续更新数据指出前景正趋于乐观，而年轻劳动者面临的真正障碍在于高等教育课程的严重滞后。目前许多大学计算机科学专业仍允许学生在未调用前沿模型API或从未使用AI编程代理的情况下毕业，导致其无法满足现代工程团队的实际要求。吴恩达明确表示，其团队绝不招募没有AI编程代理使用经验的软件工程师，但这并不意味着初级工作机会已被结构性消除。其机构近期依然雇佣了能够熟练运用现代AI工具的高中生、大学生及应届毕业生实习生。初级岗位的危机本质上并非职位消失，而是高校未能为学生提供与现代开发工作流匹配的AI原生技能训练。吴恩达认为劳动力市场所面临的转型并非不可逆的职业消亡，而是对技能广度与工具整合能力的重塑。就业生态的未来平衡不仅受制于数据中心和电网等现实物理瓶颈的发展节奏，更取决于教育体系能否迅速转向，将AI原生工具全面融入初级人才培养中。

Source quotes

[4:54] Andrew Ng: And frankly none of my teams will hire a software engineer that doesn't know how to use a AI coding agent.

Original transcript

[0:00] there will be no AI job apocalypse. This [0:03] idea of 50% being out of work, 50% of [0:06] people being out of work, maybe rioting [0:08] in the streets, I don't know, is is I I [0:10] I don't see any realistic path to that. [0:12] And in fact, early data even out of some [0:15] of the organizations that were hyping up [0:17] the 50% unemployment job scenarios are [0:21] showing that um AI is not having the [0:24] negative impact on jobs that you know [0:26] people were fearing. um because AI [0:31] automates only part of someone's job, it [0:34] turns out we really need people to do [0:36] the rest of their job still and in I've [0:40] been very and if what happens in [0:42] software engineering is a predictor of [0:45] what will happen in other job sectors as [0:47] well, I think it'll be fantastic. And if [0:50] anything, we can't find enough software [0:52] engineers right now. And I'm already [0:54] seeing we can't find enough AI native [0:56] marketers or enough AI native sales [0:58] people, enough AI native operations [1:00] people. Um, [1:03] but the challenge is upskilling because [1:06] uh for all of these job roles, we have [1:09] to help people learn new skills and I [1:12] don't just mean learn AI skills, but it [1:14] turns out that maybe I'll take one [1:16] example of software engineering. A lot [1:18] of people used to be maybe still are [1:20] front-end engineers, meaning they build [1:22] a website or an app to make it look [1:24] good, but AI is pretty good at front-end [1:26] engineering. So, a lot of front-end [1:28] developers are now full stack [1:30] developers, meaning they don't just [1:32] build the part that users look at, but [1:34] they build what runs in the background [1:36] as well. They build a mobile app, they [1:37] build security things. And so, a lot of [1:39] developers are now taking on broader [1:41] scope in their work. And what that means [1:44] is instead of just focusing on what it [1:46] looks like, they have to learn to build [1:48] all these other pieces as well. And they [1:50] have to learn AI skills. When they do [1:53] that, they're amazingly valuable in the [1:55] job market. We just can't find enough of [1:56] these people. I'm seeing early signs of [1:59] a similar transformation happening as [2:01] well in other job categories. [2:03] >> I'm curious though, you know, you say [2:05] there's not going to be a jobs [2:07] apocalypse. I'm curious what that [2:08] implies about your sense of the eventual [2:12] progress of AI because the fear about a [2:14] jobs apocalypse isn't the AI we have in [2:16] 2026, it's the AI we have in 2030 or if [2:20] your timelines are short, it's the AI we [2:22] have in 6 months, right? Few people are [2:24] arguing that the AI today can do every [2:28] task in a job like software engineering. [2:30] But the fear is that if we get this [2:33] compounding rate of progress for another [2:35] year or two, it's hard to see where it [2:38] stops. Are you implicitly arguing that [2:40] you think it will stop that we will hit [2:43] a hard limit at some sort of of your [2:46] scaling laws or or some architectural [2:48] change will be needed and there will be [2:51] some tasks that the AI systems of [2:53] tomorrow simply will not be able to do [2:54] and probably never will be able to do. [2:56] Well, I don't think we'll hit a hard [2:58] limit, but progress will be um [3:01] incremental for a long time to come. So, [3:04] I would love if AI could do any [3:06] intellectual task a human can. Uh maybe [3:08] I'll go on holiday or something. Um but [3:12] realistically, that would take multiple [3:15] novel breakthroughs and I'd love to see [3:18] it within our lifetime, but whether [3:20] it'll take decades or even longer than [3:22] decades to get there, I don't know. Um, [3:25] and until AI could do 100% of what [3:28] people can do, which seems incredibly [3:30] far away, I don't think we'll get there [3:31] anytime, certainly not in six months, [3:33] almost a very unlikely by 2030. There's [3:36] still plenty of work for people to do. [3:39] And you use the term compounding [3:41] progress. Um, one of the terms of hype [3:44] in Silicon Valley has been um [3:46] compounding progress or what recursive [3:49] improvements and so on. And at least [3:52] right now that's much more hype than [3:54] reality. For decades computers have made [3:57] computers better. I mean we use the last [4:00] generation of semiconductor microchips [4:02] to build the current generation of [4:03] semiconductor microchips. And yes at the [4:06] frontier labs AI writes all the code or [4:09] almost all the code. So of course AI is [4:10] helping us make better AI. But it turns [4:13] out that if we look at the bottlenecks [4:15] AI progress, a lot of the bottlenecks [4:18] are um we have to build new GPU servers, [4:21] we have to get power plants built, um we [4:24] have to run a lot of experiments to see [4:26] what works. So unfortunately even super [4:29] intelligent AI, if we can get there, [4:31] needs to obey the laws of physics and [4:34] the bottlenecks are the time to build [4:37] data centers and run experiments to [4:38] figure out what to do rather than just, [4:40] you know, thinking in the box. um then [4:42] those are fundamental bottlenecks [4:44] because the bottleneck to more [4:45] intelligence is not just to be smarter [4:48] about things that these other [4:50] bottlenecks as well that seem very [4:52] difficult to overcome. [4:54] >> If you're not concerned then about a [4:56] jobs apocalypse are you still concerned [4:58] about some narrower impacts? I'm [5:00] thinking firstly of the impact on early [5:02] stage careers. You brought up your [5:04] Stanford colleague uh Bren Jolson who [5:06] did some fantastic research last year [5:09] showing that it seems like the more [5:12] exposed a sector is to AI, the worse job [5:15] prospects are for people entering that [5:17] sector. That research was pretty [5:20] convincing to me and it seems to be [5:21] borne out anecdotally that getting a job [5:24] as an entry-level uh software engineer [5:27] with no skills, no uh previous work [5:31] history is getting quite hard because [5:33] you don't have much to offer in a way [5:35] that you may have if there were roles of [5:38] you know just being a code monkey while [5:40] you learn for a couple of years. [5:41] >> Yeah. So Eric's um Canary paper was very [5:44] thoughtful. I think he actually has new [5:46] data since then that paints a more [5:49] optimistic picture. Um but to take take [5:53] software engine as example, I think the [5:56] challenge is not the lack of jobs. The [5:58] challenge is the slowness of our [6:01] educational system to adapt to change to [6:04] train graduates to do the jobs that are [6:06] now in demand. So I think it's really [6:08] unfortunate that today in many [6:10] universities it's still possible to get [6:13] a bachelor's degree in computer science [6:15] without having made a single API call or [6:18] network call to a frontier AI model. And [6:22] today I'm not going to hire a software [6:23] engineer that's never done that. Um it's [6:26] still possible to get a bachelor's [6:28] degree in computer science in some [6:29] universities without using a AI coding [6:31] agent. And frankly none of my teams will [6:33] hire a software engineer that doesn't [6:34] know how to use a AI coding agent. So to [6:38] me the challenge is not this job [6:40] apocalypse. We don't need junior people. [6:42] Oh, over this last summer my team had a [6:45] lot of interns uh including high school [6:47] students, current college students, [6:49] fresh grads, and they were great. They [6:51] did a ton of work. I'm really glad I had [6:53] all these interns. But the trick is I [6:55] hire only AI native interns. So the jobs [6:58] are there, but we've got to fix the [7:00] educational system.