0:02Marina Mogilko： 这位是 Andrew。他是 Google Brain 和 Coursera 的联合创始人，他的机器学习课覆盖了几百万学员，是今天 AI 圈最有影响力的声音之一。你在机器学习和 AI 教学上底子很厚，而且你是个正面的声音。尤其是今年夏天，我明显感觉到大家变得很负面，特别是在社交媒体上：我一发 AI 相关的东西，就有人跟我说数据中心、说失业。你觉得这波情绪是怎么起来的？原因是什么？
1:13Andrew Ng： 关于 AI 的错误信息很多，其中很大一部分的根源，是两三年前开始的一场很不幸的操作，我认为是公关加监管俘获。现在 AI 里最值钱的东西之一，是有些公司训出来的超大语言模型。可如果你花了几十亿美元训一个模型，别人也训一个、还想免费给全世界用，那对你来说就非常不方便。
1:42Andrew Ng： 所以少数几家领先的 AI 公司一直在大声地贩卖对 AI 的恐惧，想推动监管立法，造出一个偏袒在位者的不公平赛场，让我们用 AI 都得交高额的过路费，同时打压其他团队，不管是研究者还是别的公司，只要他们想把开放权重、开源的模型免费放出来，让所有人用得便宜得多。
2:08Andrew Ng： 遗憾的是，贩卖恐惧是管用的。你说 AI 像核武器，这个类比根本没有事实依据，两者到底有什么关系？或者你专挑 AI 出错的个案，把它放大很多倍，甚至散布错误信息，说数据中心的用水量比实际多得多。这种持续不断的恐惧宣传，把社会对 AI 的看法扭得非常负面。这很可惜，因为它正在拖慢美国对 AI 的采用，让美国变得更没有竞争力。除非我们把真相讲出去，也就是 AI 带来巨大的好处、也有一些问题，但远没有被渲染得那么严重，否则受伤的会是每一个普通人。
3:05Andrew Ng： 工作末日，就是说 AI 会拿走 50% 的工作、人们失业、上街暴动，这根本不会发生。每一波技术，包括 AI，都会改变做好工作所需要的技能，所以 AI 确实在改变各个职业。但说实话，我真希望 AI 能更好用一点。有几家公司想把 AI 炒作成我们已经有了超级智能、通用人工智能，人能做的它都能做。可我们就是还没本事让 AI 做完人能做的所有事。
3:47Andrew Ng： 像我的朋友、斯坦福的 Erik Brynjolfsson，还有 MIT 的 Andy McAfee，这些经济学家把很多人的工作拆成一个个具体任务来分析。AI 大概能做很多工作里 30% 到 40% 的部分，这意味着人做的那 60% 反而更值钱了，因为它和那 30% 到 40% 是经济上的互补品，那部分变便宜了。所以也许会用 AI 的人会取代不会用的人，但对绝大多数工作来说，AI 还没到能取代人的地步。
4:24Andrew Ng： 所有职业里，现在受 AI 影响最大的是软件工程，因为 AI 写代码非常厉害。可软件工程的招聘岗位数量是在上升的，跟那些末日派的恐惧宣传正好相反。AI 其实取代不了软件工程师，我认识的好工程师现在都比以前更忙。反过来说，如果有人还在用 2022 年、ChatGPT 出现之前的方式写代码，那他就有麻烦了。他们需要新技能：别再做 AI 能自动化的那 30% 到 40%，交给 AI 去做，然后把能力练到能做 AI 做不了的另外 60% 到 70%。
05:07 · 给应届生的建议
5:07Marina Mogilko： 你会给应届毕业生什么建议？我在这个节目里跟 Erik 聊过，他说 AI 对就业市场的影响其实不大，除了 18 到 25 岁、刚毕业的那批人。他们可能还没有专业积累，能做的活 AI 也能做。
5:31Andrew Ng： 应届生面临的一个真实难题，是大学体系调整得太慢。我热爱学术界，也觉得大家都该支持大学。但当 AI 改变了写软件的方式，大学往往要花一两年：教师先学会新技能，再开新课，再过课程委员会审批，再让教授会投票。这个变化速度和 AI 的变化速度严重错配。很遗憾，很多大学还在按 2022 年的岗位培养学生，而我们连按 2026 年的岗位培养都已经晚了，应该按 2028 年及以后的岗位来培养。
6:22Andrew Ng： 岗位是有的。我认识的很多雇主，不管招哪个级别，都找不到足够多有本事的人。我办公室现在有很多实习生：在校大学生、应届生，甚至还有一个高中生，他们都非常出色、非常能干。关键在于他们都是 AI 原生的：AI 能做的事就交给 AI 工具，然后全力去做人能做、AI 做不了、而且很长时间内都做不了的事。所以我给应届生和在校生的建议是：该认真上课就认真上课，拿好成绩，向老师学习。但大学还没来得及教的技能，尤其是 AI 技能，就去网上找别的途径学，Coursera、DeepLearning.AI、Udemy 都行。
9:05Andrew Ng： 关于技能版图的变化，我再补一句。最重要的变化之一，是用 AI 做东西比以前容易太多了；一件事变得容易很多，就应该有更多人去做。不只是职业软件工程师该用 AI 写软件，现在每个人用 AI 做东西都容易得多，拥抱它的人会更有效率、做成更多事，我觉得也会更有乐趣。市场、招聘、HR、运营的人，只要学会用 AI 做东西，不管什么岗位，都能做到多得多。
09:35 · AI 带来多少，怎么衡量
9:49Marina Mogilko： 部署 AI 之后，生产力提高了多少，你怎么衡量？公司里有 KPI 吗？
9:58Andrew Ng： 我也希望有个简单的答案。我发现 AI 带来的业务结果，更多取决于业务本身，而不是 AI。有的公司看客户增长和留存，有的看服务客户更快，有的看某些任务更准确。所以 KPI 往往跟业务挂钩，而不是跟 AI 挂钩。
10:19Marina Mogilko： 所以没法直接衡量 AI 本身。我们公司一直在积极用 AI，作为一家媒体公司，对我来说大概就是播放量、产出。挺有意思的，大家衡量的方式都不一样，连收入怎么算都不一样。
10:41Marina Mogilko： 首先，我们所有人都有 Claude，我们做的每个社交平台都有一个专门的项目。比如这个播客，有个叫「嘉宾」的项目，它掌握以往每位嘉宾的数据，并按一套标准给每位来上节目的人打分：有没有被引用、对 AI 有没有鲜明观点、在自己公司里是不是积极用 AI、是不是新近的 AI 创业者。它给这些项加权，打出一个满分 40 的分数，40 是第一档，30 是第三档，以此类推。另一个项目会分析每一期节目，告诉我怎么提问。Instagram、LinkedIn 也一样：它知道我的语气、我的个人档案、我的商业策略，所以它写什么东西都知道关于我的事实、知道我说话是什么样。每个平台都有一个人负责，由这个人做战略判断。
11:42Marina Mogilko： 我现在在做的，你也可以给我点建议，是把反馈闭环接上。有时候团队发来一段文字，我说这里、那里要改，可这些都发生在 Telegram 的聊天里。我们有个机器人会扫所有聊天记录，但我真正想要的是 AI 能从这些反馈里持续学习，更懂我的品味。
03
人的优势是语境，AI 却会伤害学习
人从经验里知道大量 AI 拿不到的东西，这就是判断力和品味的来源；但让 AI 替人做事，人就记不住。
11:52 · 语境：人的优势
12:01Andrew Ng： 我经常看到一种情况。你把 AI 当数据科学家或头脑风暴伙伴用，它常常给出一两个好主意、两三个平庸的，再加四个糟糕透顶的，有时你会纳闷，AI 怎么会觉得这种主意说得通。在我看来，这和工作末日那个问题有关。在很长一段时间里，人类，你、我、每一个在看的人，相对 AI 都会有显著的语境优势：有件事你觉得再明显不过，你知道那是个馊主意，AI 却不知道。
12:40Andrew Ng： 这就是 AI 短期内不会取代我们工作的原因之一。我们从多年经验里知道太多东西：我们跟客户聊过，看到对方一个古怪的表情，就知道他们不喜欢这个；或者经理说过「这件事我真的很在意」。几乎所有人，也许是所有人，都知道大量这样的东西，而 AI 根本没有获取它们的管道，我认为在可预见的将来也不会有。大家常说人的判断力、人的品味，也有人觉得品味是个模糊的东西。在我看来，人的判断力和品味之所以比 AI 好，背后的技术原因就是这种语境优势。它是一种长期优势，几年内没人能解决，所以我们需要更多有判断力、有品味的人，持续和 AI 互补。
14:04Andrew Ng： 我要说一句可能有争议的话，不确定以前有没有公开讲过，但我认为是真的：说实话，AI 模型对学习很糟糕。AI 办事特别好用，我天天用，也很喜欢。但出来的数据非常清楚：大学生用了 AI，作业分数更高，好耶，分数高了，可他们的留存、长期表现要差得多，因为活是 AI 替他们干的。越来越多的研究在证实这一点。
14:51Andrew Ng： 维基百科是很好的工具，里面有大量事实；网页搜索也是很好的工具，里面也有大量事实。但你让 AI 替你干活，就是在把认知外包给 AI。这是好事，社会就是这样前进、把事情做成的，但人的留存会差很多。非常清楚：按大多数人最常见的用法，大语言模型对学习很糟糕。我不是说没有对学习有益的用法。可就连我自己，过去半年有太多事情，比如某个项目里前端、后端组件怎么配合：给我答案，把活干完，太好了。半年后我已经不记得答案，要重做那个组件时又去问 AI。所以我们不该再把 AI 当成对学习有帮助的东西，至少按今天绝大多数人使用 AI 模型的方式是这样。
15:52Marina Mogilko： 可你正在做一家公司来解决这个问题，对吧？用 AI 做一对一辅导，你刚宣布拿到了 Coursera 的 1 亿美元投资。
16:42Andrew Ng： 说到人的技能发展：因为 AI 对软件工程冲击最大，软件工程就业市场上发生的事，就是其他领域将要发生的事的先兆。在软件工程里，人们需要学新技能，可一旦学会了，他们就发展得很好，创造更多价值，说实话还涨了薪，做的项目也更有意思。大多数前端、后端开发者都变成了全栈开发者，因为有了 AI 帮忙，他们能承担更宽的范围。我在别的领域也看到了早期迹象：做市场活动协调的人，有了 AI 可以变成全流程的市场人员；招聘里做寻访的人，开始做端到端的招聘。好消息和坏消息是同一个：要胜任这些更宽的角色，你需要 AI 技能，也需要本专业的其他技能，市场、招聘、软件工程或 AI 工程的其余部分。这就产生了很大的学习新技能的需求，学会了的人能做得多得多，希望也更开心、薪水更高。
19:59Andrew Ng： 对。我女儿想当宇航员，我不知道能不能有个专业叫当宇航员。等她长大，也许会改主意。我看到各种岗位上都有很多机会，在我看来都很让人兴奋。但一定要学 AI，一定要学会用 AI 做东西。
20:16Andrew Ng： 我们团队还在做一张 AI 工程技能图谱，梳理 AI 工程最重要的技能。有件事我凭直觉早有感觉，但在数据里看到还是有点意外：越来越多的招聘描述写着，希望找主动性很强的人。因为有了 AI，个人发现一个问题、然后自己动手做点东西去解决它的机会多了很多。我们早就在离开那个大家坐等老板派活的时代，现在只是在加速。
22:34Andrew Ng： 数据看板……再说一遍，我的团队大概走得比较前。前几天，市场团队有个人讲他做的一个工具：他打算写一篇文章时，它会去爬网页、找相关内容。他还真做了一个跑在自己 Mac 上的桌面应用，把相关文章标出来，你还能跟整个系统对话，一边看他在写的东西，一边看相关内容。他还有一个很大的看板，在网上持续搜罗，把冒出来的有意思的东西推给他。
25:58Andrew Ng： 再看 AI 公司这边，至少有一家公司，我不点名，似乎时不时就改服务条款。你打开网站，弹出来一句「我们修改了服务条款，将保留你的数据」或者「将用你的数据训练」，你要是没留神点错了按钮，它就突然给了自己访问你数据的权限，而且是以一种让我不太舒服的方式。我经手一些敏感信息，所以对那些我觉得其文化、基因、说实话还有长期商业模式，跟保护用户隐私的绑定不如超大云厂商那么紧的公司，我会非常小心。企业也明白这一点。我有一个团队叫 AI Aspire，跟包括银行在内的超大企业合作，经手极其敏感的财务数据，无论是 AI Aspire 还是我们的客户，都不会随随便便把重大非公开信息交给前沿实验室，一定会非常仔细地考虑防护和隐私。
34:47Andrew Ng： 创办一家公司仍然非常非常难，所以单线程领导、有一个人全心专注于一件事，是很有道理的。一个周末我常常能做出一个大模型套壳应用、一个简单的应用，但我真希望建一家大公司也有那么容易。做出有意义的东西，往往需要真正的技术深度，或者深刻的客户洞察、和客户深度结合，或者两者都要。没错，现在用 AI 几个小时就能写出点东西，但那只是拼图里的一小块。搞懂技术复杂度、做出真正复杂的软件，可能要几个月甚至几年；靠深刻的客户洞察来决定做什么，也同样要花大量时间：跟人聊、看表情、做调研，一遍又一遍，直到想清楚做什么。所以广泛尝试是有价值的，但一个人还是要有专注，在一两个领域里真正钻深。对建立一门生意来说，这仍然很重要。
From the most mainstream seat in AI, Andrew Ng makes the full case against the doom story: fear-selling is regulatory capture, and the human edge is context.
Part 1 of 6 · 0:02
Fear-selling is regulatory capture
A few leading companies amplify AI's risks to push rules that favour incumbents and make it hard to give open models away.
Job ads increasingly ask for agency; his marketing, finance and recruiting teams write their own tools, with embedded engineers to go deeper. Read this part →
Sensitive data stays on local models; loss of control is like early aviation, manageable enough; his real worry is children offloading their thinking to AI. Read this part →
The bottleneck moves to what to build; AGI is decades off
Building has become cheap, so the hard part is deciding what to build; by the "any intellectual task" definition, AGI is decades away. Read this part →
Indigo's conclusion
Among those pushing back on AI fear, this is the weightiest statement yet: it turns regulatory capture into a debatable economic question, gives taste a technical explanation, and brings data on AI hurting learning. But his position sits on the acceleration and open side.
How to read this His stake is plain: he runs AI education (Coursera, DeepLearning.AI) and invests in AI start-ups (AI Fund), so more adoption and less fear is good for his business. Among the people pushing back on AI fear, though, he is the most evidence-based, leaning on economic analysis and data. Trust his data; discount the "don't worry, go learn, go build" call a little.
What to remember
Fear-selling is regulatory capture: a few incumbents play up risk to push rules that favour them and hold back open source.
No job apocalypse: AI can do 30 to 40% of a job and the human part gains value, as long as AI stays not quite good enough.
The human edge is context: what you know and AI can't reach is where judgment and taste come from.
AI hurts learning: homework scores go up, retention goes down; he is building Learn Vector for one-to-one learning because of it.
What would change my mind
AI capability really crosses a threshold and the 30 to 40% ceiling breaks.
How to read this
His stake is plain: he runs AI education (Coursera, DeepLearning.AI) and invests in AI start-ups (AI Fund), so more adoption and less fear is good for his business. Among the people pushing back on AI fear, though, he is the most evidence-based, leaning on economic analysis and data. Trust his data; discount the "don't worry, go learn, go build" call a little.
A few leading companies amplify AI's risks to push rules that favour incumbents and make it hard to give open models away.
00:45 · Where the fear comes from
0:02Marina Mogilko: This is Andrew. He co-founded Google Brain and Coursera, his machine learning course has reached millions of learners, and he is one of the most influential voices in AI today. You come with a huge background in machine learning and teaching AI, and you're a positive voice. This summer especially I've been seeing how negative people have become, especially on social media: when I post about AI, people talk to me about data centers and job loss. Why do you think this wave started? What are the causes?
1:13Andrew Ng: There's been a lot of misinformation about AI, and the root cause of a lot of it is an unfortunate attempt that started two or three years ago at, I think, PR and regulatory capture. One of the most valuable things in AI right now is the giant large language models that some have trained. But if you spend billions of dollars training a model, it's really inconvenient if someone else trains a model and wants to give it to anyone in the world to use for free.
1:42Andrew Ng: So a handful of leading AI companies have been very loud voices, fear-mongering around AI to try to get regulations passed, to create an unfair playing field that favors incumbents, so that we all have to pay a high toll to use AI, while stifling the other teams, be it researchers or other companies, that want to give away open-weight, open-source models that anyone could use much more cheaply.
2:08Andrew Ng: Unfortunately, fear-mongering works. When you say AI is like nuclear weapons, an analogy that has no basis in fact (what do they even have to do with each other?), or when you cherry-pick cases of AI making a misstep and make them much bigger than they are, or spread misinformation that data centers use a lot more water than they actually do, this drumbeat of fear-based messaging skews societal perception to be really negative on AI. That's unfortunate, because it is slowing down American adoption of AI and making America less competitive. Unless we get the truth out there, which is that AI is a fantastic benefit with some problems, but nowhere near the problems they're blown up to be, it will hurt individuals.
02
AI does a third of a job; the rest gains value
Break jobs into tasks: AI can do 30 to 40%, which makes the human part worth more; what people lack is the skill to build with AI.
02:57 · The "job apocalypse"
2:58Marina Mogilko: I'm going to read out some of the problems people are highlighting. Job loss and inequality. What do you think?
3:05Andrew Ng: The job apocalypse, this idea that AI will take over 50% of jobs and people will be out of work, rioting in the streets, is just not going to happen. With every wave of technology, including AI, the skills we need to do great work shift, so AI is changing professions. But boy, I wish AI worked better. A handful of businesses want to hype AI up and say we have superintelligence or artificial general intelligence and it can do all the things humans do. We're just not good enough to make AI do everything a human does.
3:47Andrew Ng: Economists like my friends Erik Brynjolfsson at Stanford and Andy McAfee at MIT have analyzed many people's jobs by breaking them down into individual tasks. Maybe AI can do 30 to 40% of many jobs, and what that means is that the 60% the human does becomes even more valuable, because it's an economic complement to the 30 to 40% that's now cheaper. So maybe people who use AI will replace people who don't, but for the vast majority of jobs AI is not in a position to replace people.
4:24Andrew Ng: Of all professions, the one most affected by AI now is software engineering, because AI is fantastic at writing code. And the number of job openings in software engineering is up, contrary to what the doom fear-mongerers would say. AI is not actually able to replace software engineers, and all the good software engineers I know are busier than ever. The flip side is, if someone still writes code like it's 2022, before ChatGPT, they're in trouble. They need new skills: stop doing the 30 to 40% that AI can automate, let AI do that, and build your skills to do the other 60 to 70% that AI cannot.
05:07 · Advice for new graduates
5:07Marina Mogilko: What would your advice be to new graduates? When I talked to Erik on this podcast, he said there isn't much impact on the job market except for people from 18 to 25 who just graduated. They may not have the expertise yet; they can only do work that AI can do as well.
5:31Andrew Ng: One real challenge for fresh college grads is that the university system is slow to adapt. I love academia and I think we should all support universities. But when AI transforms the way software is written, it often takes universities a year or two for the faculty to master the skills, create new courses, get curriculum committee approval and get the faculty senate to vote. That speed of change is very poorly matched to the speed of change in AI. Sadly, many universities are still teaching students for the jobs of 2022, when we shouldn't even be teaching them for the jobs of 2026. We should be teaching them for the jobs of 2028 and beyond.
6:22Andrew Ng: The job openings are there. Tons of employers I know just can't find enough skilled people at any level of seniority. In my office right now we have a lot of interns: current college students, fresh grads, even one high school intern, and they're amazing and productive. The key is they're all very AI-native. They use AI tools to do the things AI can do, and then lean into the things humans can do that AI can't, and won't be able to for a long time. So my advice to fresh grads and current students: by all means work hard in your classes, get good grades, learn from your instructors. But where there are skills the university hasn't yet adapted to teaching, especially AI skills, find other ways to learn online, be it Coursera, DeepLearning.AI, Udemy or elsewhere.
9:05Andrew Ng: One more thing about how the skill map is changing. One of the most important changes is that it's so much easier to build with AI than before, and when something becomes much easier, a lot more people should do it. Not only should professional software engineers build software with AI; it's becoming much easier for everyone, and the people who embrace it will be more productive, accomplish more and, I think, have more fun. Marketers, recruiters, HR professionals, operations specialists: if they learn to build with AI, they'll do much more, whatever their role.
09:35 · How do you measure what AI adds?
9:49Marina Mogilko: How do you measure the increase in productivity when you deploy AI? Do you have a KPI in your company?
9:58Andrew Ng: I wish there were a simple answer. I find the business outcome of AI is more a function of the business than of the AI. For some it may be customer growth and retention, for some serving customers faster, or higher accuracy on some tasks. So the KPIs tend to be about the business rather than the AI.
10:19Marina Mogilko: So you can't measure AI directly. We've been deploying AI actively in my company, and as a media company, for me it's probably views, the output. It's interesting how differently people measure it, even revenue.
10:37Andrew Ng: Actually, same question back: how are you using AI in your business?
10:41Marina Mogilko: First of all, we have Claude for all of us, and a project for every social media platform we're on. For this podcast we have a project called "guests" that knows the analytics from previous guests and ranks every person who comes on against certain criteria: whether they're cited, whether they have a certain opinion on AI, whether they've been active with AI in their company, whether they're a recent AI founder. It weights those and gives a grade out of 40, with 40 meaning tier one, 30 tier three, and so on. Another one analyzes every episode and gives me tips on how to ask questions. Same for Instagram, same for LinkedIn: it has my tone of voice, a personal dossier and my business strategy, so whenever it writes something it knows the facts about me and how I sound. Every channel is run by a person, who makes the strategic calls.
11:42Marina Mogilko: What I'm working on now, and tell me if I can improve it, is closing the loop. Sometimes my team sends me a text and I say we need to change this and that, but that happens in a Telegram chat. We have a bot that scans all our chats, but I really want AI to learn continuously from that feedback, so it knows my taste better.
03
Context is the human edge; AI hurts learning
People carry context AI can't reach, and that is where judgment and taste come from; but when AI does the work, you don't retain it.
11:52 · Context: the human advantage
12:01Andrew Ng: There's one thing I see a lot. When you use AI as a data scientist or brainstorming partner, it often comes up with one or two good ideas, two or three mediocre ones and maybe four atrocious ones, and sometimes you wonder how your AI could have thought that was even plausible. To me this relates to the job apocalypse question. For a long time, humans, you, me, everyone watching, will have a significant context advantage over AI: something is incredibly obvious to you, you know it's an awful idea, but the AI doesn't.
12:40Andrew Ng: That's one of the reasons AI won't replace our jobs any time soon. We know so much from years of experience: we talked to a customer and saw the funny facial expression that told us they don't like this, or our manager said "I really care about this". Almost all humans, maybe all humans, know a lot of things that the plumbing does not exist for AI to get, and I don't think it will for the foreseeable future. People talk about human judgment or human taste, and some wonder whether taste is a fuzzy thing. To me, the technical thing underneath why humans have better judgment and taste than AI is this context advantage. Because it's a long-term advantage that no one will solve in a few years, we need a lot more humans with that judgment and taste to keep complementing AI.
13:37 · "AI models are terrible for learning"
13:43Marina Mogilko: Doesn't that make education even more important, because education gives us context? Another thing I keep hearing is that you won't need education because all the information is at your fingertips; just ask ChatGPT. But when you say context and taste, to me that's years of acquiring knowledge, learning from the best and seeing how they perform, not just asking a chatbot.
14:04Andrew Ng: I'm going to say something that may be controversial; I don't know if I've said it publicly, but I think it's true. Frankly, AI models are terrible for learning. AI is wonderful at getting things done, I use it all the time and love it. But the data coming out is very clear: when college students use AI, they score higher on homework (yay, higher homework scores), but their retention, their long-term performance, is much worse, because the AI did the work for them. More and more studies are backing this up.
14:51Andrew Ng: Wikipedia is a wonderful tool with tons of facts; web search is a wonderful tool with tons of facts. But when you ask AI to do work for you, you're cognitively offloading to AI. That's great, because that's how society moves forward and gets work done, but human retention is much worse. It's very clear that LLMs, as they're most commonly used, are terrible for learning. I'm not saying there's no way to use them that's good for learning. But even for myself, there were so many things over the last six months, like how a front-end and back-end component works on some project: give me the answer, get the job done, fantastic. Six months later I don't remember the answer, and when I need to redo that component I ask AI again. So we should stop thinking of AI as helpful for learning, at least in the way the vast majority of people use AI models today.
15:52Marina Mogilko: But you're building a company to help solve that, right? One-to-one tutoring with AI, which you just announced with a $100 million investment from Coursera.
16:03Andrew Ng: Yes. I'm excited to be leading a new organization called Learn Vector, focused on building new learning experiences that are much more one-to-one than one-to-many. Fifteen years ago I was privileged to take part in the online courses movement, which I think changed the way a lot of people learn, but that was and still largely is a one-to-many experience where everyone watches the same video. That actually works well, but the technology now exists to create much more personalized, one-to-one experiences. Our team is working hard on it, and I think we'll have a lot more to show by early next year.
16:42Andrew Ng: On human skill development: because AI has hit software engineering so hard, what we see in that job market is a harbinger of what we'll see in other disciplines. In software engineering people need to learn new skills, but when they do, they're thriving, creating more value, frankly getting raises and doing more exciting projects. Most front-end and back-end developers have become full-stack developers because, with AI's help, they can take on broader scope. I'm seeing early signs of this elsewhere: someone in marketing who coordinated campaigns can, with AI, become a full-cycle marketer; sourcers in recruiting are doing end-to-end recruiting. The good news and the bad news is that to step up to these broader roles you need AI skills, but also the other skills of the discipline, the rest of marketing, recruiting, software or AI engineering. That creates a big need to learn new skills, and people who do can do much more, hopefully have more fun, and hopefully get paid more too.
18:10Andrew Ng: One reason I worry about the fear-mongering: I got an email from someone about to enter college. He was taking online courses but struggling with what to major in, because in four years won't AI do all this and won't everything he learns be obsolete? The answer is no, of course it won't all be obsolete. But when we keep pushing these fear messages, people wonder whether they'll even be relevant, and then they don't lean in to gain the skills that would put them in a much better position. These messages are distorting how many people, including high school students, college students and fresh grads, think about the economy. Frankly, making people give up is one of the worst things we could be doing in an era when people who lean in will thrive.
04
Agency wins; everyone should build tools
Job ads increasingly ask for agency; his marketing, finance and recruiting teams write their own tools, with embedded engineers to go deeper.
19:30 · What to major in, and agency
19:34Marina Mogilko: What would you reply to that email? What's the best major to thrive in the AI era: going deep into a niche, or broader computer science so you can pick up AI skills fast?
19:50Andrew Ng: I don't know what the best major is; there are an awful lot of great majors. Asking for the best major is like asking what the best job in the world is.
19:58Marina Mogilko: Something that you love, right?
19:59Andrew Ng: Yeah. My daughter wants to be an astronaut; I don't know if she can major in becoming an astronaut. When she gets older she may change her mind. I see so many opportunities across so many roles, and it all seems very exciting to me. But do learn AI, and do learn to build with AI.
20:16Andrew Ng: My team has also been working on an AI engineering skills map, to chart the most important skills for AI engineering. One thing I felt intuitively but was surprised to see show up in the data: a lot more job descriptions say they want people who demonstrate a very high sense of agency. With AI there are many more opportunities for individuals to spot a problem and go build something to solve it. We've long been moving away from an era where people sit around and wait for their boss to tell them what to do, and that's accelerating.
20:50Marina Mogilko: That's what I've been feeling, especially since we went remote. I want people to be entrepreneurs within their niche. If you're helping me with LinkedIn, you're an entrepreneur there: you can hire more contractors, deploy different tools, make the strategic call on whether a topic is good and whether we proceed. Tell me if you agree: we're moving into a job market where everyone is kind of independent in their workplace.
21:15Andrew Ng: I think people will have much more autonomy and creativity, so I agree. I'd go one step further. A lot of engineers and others in large companies tell me their manager tells them to stay in their swim lane: "I have this creative idea," and the manager says, "No, I need you to focus on this one thing," frankly often because the manager's career depends on it. But the number of opportunities for people to spot things outside their swim lane and responsibly explore how to get them done feels very exciting to me. In the future, businesses that build a culture encouraging people to learn AI, build fast responsibly and talk to customers will drive a lot more value than hierarchical, siloed organizations.
21:55 · How AI-native teams work
22:03Marina Mogilko: It starts with hiring the right people and then nurturing this in the organization. When you say become proficient with AI, can you give some benchmarks? For a marketer or knowledge worker who's advanced with AI, what are you looking for in the interview?
22:19Andrew Ng: I'm pretty sure my team is ahead of the curve: all of my marketers know how to code. When I interview marketers, we ask them what they've built, and whether they've built any software.
22:30Marina Mogilko: If it's a dashboard, is that good or bad? Too basic?
22:34Andrew Ng: A dashboard... again, my team is probably ahead of the curve. The other day someone on the marketing team was talking about a tool he'd built: when he's considering writing an article, it crawls the web and finds related work. He actually built a desktop app that runs on his Mac to highlight related articles, and you can chat with the whole system to navigate what he's writing alongside the related work. He also had a large dashboard trawling the internet to flag exciting things as they pop up.
23:28Andrew Ng: My finance team uses AI extensively too. One of my CFOs realized her team was spending hours every week clicking through documents: open this, copy and paste that number here. So she started building automation scripts that run on a routine, automatically open files, check what's in them and check for consistency, and alert the team if there's something to pay attention to or a new document has shown up. Rather than waiting for an engineer to do it for them, the team builds not just dashboards but a kind of data management infrastructure that ingests data and alerts them when something happens. And my recruiting team actually has recruiting engineers, real professional engineers sitting in the recruiting team and building very sophisticated recruiting tools. That's the other trend: marketers, recruiters, HR professionals and ops people should all learn AI, but when you also embed an engineer in these teams, it accelerates what you can do even further.
24:38Marina Mogilko: We do the same: we start with something basic, build it ourselves, hit the wall, then an engineer comes in and we build it further.
24:45Andrew Ng: When you look at not just software engineers but recruiting engineers, marketing engineers, HR engineers, there's so much valuable engineering work that can now be done. I'm not worried about running out of it. Frankly, all my friends are so busy we think: how could we run out of engineering jobs?
05
Privacy, control and kids
Sensitive data stays on local models; loss of control is like early aviation, manageable enough; his real worry is children offloading their thinking to AI.
24:53 · Privacy: what to hand to AI
25:04Marina Mogilko: You touched on one of the fears: how much financial information we give to AI. I gave Perplexity permission to scan my Fidelity account so it can track my portfolio and tell me when to rebalance. It doesn't do anything on my behalf, but it has access. Is there any problem with that?
25:23Andrew Ng: This is complicated. AI and privacy is a complex area, and it depends a lot on the company you're sharing your data with. I trust all the hyperscalers to follow their terms of service 100% and do what they say. This is my personal opinion, not legal or business advice, but I'd be shocked if one of the largest hyperscalers published terms of service with a privacy notice and then breached it, because that would be so damaging to their culture and long-term business model.
25:58Andrew Ng: On the AI company side, there's been at least one company, which I won't name, that seems to change its terms of service occasionally. You go to the website, a pop-up says "we've changed the terms of service to retain your data" or "to train on your data", and if you aren't paying attention and click the wrong button, they've suddenly given themselves permission to access your data in a way I'm not comfortable with. I handle some sensitive information, so I'm very careful with businesses whose culture, DNA and, frankly, long-term business model I don't feel are as tied to protecting user privacy as the hyperscalers'. Businesses get this too. One of my teams, AI Aspire, works with very large corporations, including banks with incredibly sensitive financial data, and neither AI Aspire nor our clients share material non-public information with frontier labs willy-nilly, without very careful thinking about guardrails and privacy.
27:15Marina Mogilko: Another thing you can do is download an open-source model and run it on your computer, and then the data just stays on your computer, right?
27:22Andrew Ng: Yes. A lot of banks run things in a virtual private cloud or on-prem so the data never leaves their control. For individuals, for the really sensitive things I sometimes run a local model. It's been interesting: some of the latest open-weight models are approaching frontier capability and are small enough to run locally. They're really good models now.
27:47Marina Mogilko: The recent one from Meta, right?
27:49Andrew Ng: Oh yes, Meta's Muse is a good model, and the latest version of Qwen is also very good. But frankly these models change every other week, so the best practice is not to get stuck on one but to keep trying new models.
28:03Marina Mogilko: So when you don't trust anyone, you run a local model, and that's how you keep your data safe.
28:08Andrew Ng: I do trust the hyperscalers. But for literally material non-public information, there are things I won't send to the cloud. For those I either do it manually without AI, or if I really need AI, I very carefully use only a local model.
28:15 · Loss of control, and deepfakes
28:31Marina Mogilko: What about loss of human control over AI? I talked to Yoshua Bengio, who is very negative about open, free AI without regulation, and he painted some very scary pictures of AI taking over. The whole scenario is that we can't control something smarter than us, so if AI gets smarter and smarter, where do we end up?
29:00Andrew Ng: I think about something else we can't control: airplanes. No one can build an airplane you can fly perfectly; winds buffet it around. And candidly, in the early days some airplanes crashed and people died, which was tragic and awful. But through those early lessons we learned to control airplanes better and better, so that today we can mostly get on a plane without fearing for our lives. It's really like that with AI. No one can perfectly control AI, because it generates tokens or outputs that are a little bit random, so we don't know exactly what it will do. There have been a small number of mishaps, which is unfortunate, and some did real damage. But the way we engineer almost any system, from airplanes to electric circuits to AI, is to grow its capabilities carefully, in a controlled environment where we can measure what's wrong and shape it so we can control it well enough that it behaves responsibly and safely. To this day we can't perfectly control any airplane, and we will never perfectly control AI either. But we're certainly controlling it well enough that loss of control doesn't feel like science fiction.
30:24Andrew Ng: Deepfakes are a problem. One of the most disgusting things I've ever seen or heard of is non-consensual intimate deepfake imagery. I'm really glad the US Congress has been moving. Let's pass laws, get rid of it, with penalties. There are some really problematic uses of AI we should outlaw and heavily penalize.
30:39 · Kids and AI
30:46Marina Mogilko: What about children and social connection, with kids using more AI? We've seen with social media people doom-scrolling all day. My daughter is five, and whenever I don't have an answer she says, "Ask ChatGPT." She thinks ChatGPT knows everything. What would you say about kids' future with AI?
31:15Andrew Ng: First, kids have a bright future. It's such an exciting time to grow up, with tools none of us ever had. At the same time, I think social media has probably been blamed a bit more than it deserves, but it does deserve some blame; it hasn't been great for kids. AI is a wonderful tool, but AI damaging learning is something I worry about a lot. I have a five-year-old and a seven-year-old. When I teach them math they're young enough that I can just not let them use a calculator: how do you multiply these numbers? We practice that. But as they get older, I worry a lot about students cognitively offloading to AI in ways that damage long-term learning.
32:06Andrew Ng: At the same time, I actually built an app. I didn't like any of the free online learn-to-type tools, so I built my own for my daughter to learn to type, and she's getting pretty decent for a seven-year-old. She can type all the lowercase letters; the shift key for uppercase isn't quite there yet. I think this unlocks responsible, adult-supervised use of online tools, and it's really tricky. Adult-supervised use of digital tools seems great for kids, but too many adults don't have time to supervise, and then there are the incentives of, say, social media.
32:54Marina Mogilko: The incentives have to be right when it comes to AI.
06
The bottleneck moves to what to build; AGI is decades off
Building has become cheap, so the hard part is deciding what to build; by the "any intellectual task" definition, AGI is decades away.
32:57 · The biggest opportunities in 2026
32:58Marina Mogilko: We've talked about the fears and about improving your work with AI. What are some of the biggest opportunities in AI in 2026 for people who want to build?
33:08Andrew Ng: For an individual who wants to build, I don't think it's one-size-fits-all, but because the cost of building has plummeted, I encourage people to learn AI, build fast and talk to customers. I find myself building things every week, every weekend, because I or someone on our team has a problem and I have an idea for an AI thing to automate it. Last weekend I used frontier models to analyze a lot of our key business metrics, because I didn't have time to do it myself or to find a data scientist to work on it with me. I used a variety of frontier models and was really careful about their data retention policies; I didn't use models whose data retention policies I don't like.
34:09Andrew Ng: What's happening with AI is that the cost of building has plummeted, so the challenge is shifting to deciding what to build, which I've been calling the product management bottleneck. Founders, engineers and product managers who can talk to customers, get a feel for the taste and judgment of what to build, then build with AI and iterate quickly: there are a ton of exciting things to do.
34:31Marina Mogilko: You've started so many companies. For beginners who build something over a weekend, how do you decide what to focus on? Or can you now pursue multiple ideas because of AI, playing in different companies at the same time?
34:47Andrew Ng: Building a company is still really, really hard, so there's a lot to be said for single-threaded leadership, someone fully focused on one thing. Over a weekend I can often build an LLM wrapper or a simple application, but I wish it were that easy to build a large company. Building something meaningful often takes real technical depth, or deep customer insight and integration with customers, or both. Yes, we can use AI to code something in a few hours, but that's a small piece of the puzzle. Understanding the technical complexity and building really complex software can take months, maybe years, and so does the deep customer insight to decide what to build: talking to people, reading facial expressions, doing surveys, over and over until we figure it out. So there's value in sampling widely, but then an individual needs the focus to go really deep in a couple of sectors. That still seems important for building a business.
35:48 · How far is AGI?
35:50Marina Mogilko: My last question: AGI, because people use the word so much. Some say, I think Jensen Huang said, we've already reached AGI. You've said it's decades away. What's the one criterion that would make you say we've reached AGI?
36:08Andrew Ng: Different people say we've reached AGI at different times because they have different definitions. The one I'm most familiar with is AI that can do any intellectual task a human can. A human brain can spend, say, five years studying and write a PhD thesis; can AI write a PhD thesis? A human can learn to drive a truck through a dense rainforest with tens of minutes of practice; when can AI drive in a new environment with tens of minutes of practice? There's a long list of things AI cannot do, for what feels to me like decades. I hope it's only decades; it may turn out to be longer. So by that definition, AGI is still very far away.
36:59Andrew Ng: Because of economic incentives, OpenAI and Microsoft had an agreement, which has now been renegotiated so that's gone away, that gave OpenAI an incentive to try to declare reaching AGI earlier. If you come up with other definitions, depending on how far you lower the bar, you could totally have reached AGI already, or even 30 years ago.
37:22Marina Mogilko: Andrew, thank you so much for this positive, very applicable conversation. I like watching something and then measuring myself against what people are doing with AI, looking at my process and maybe expanding it. Thank you for showing what your team is doing, and for your insights.
37:40Andrew Ng: Given the huge benefits of AI to come, I hope whoever is watching is motivated to really go learn AI, apply it, and even go build something.
Where Indigo landsFurther
Indigo's conclusion
Among those pushing back on AI fear, this is the weightiest statement yet: it turns regulatory capture into a debatable economic question, gives taste a technical explanation, and brings data on AI hurting learning. But his position sits on the acceleration and open side.
What to remember
Fear-selling is regulatory capture: a few incumbents play up risk to push rules that favour them and hold back open source.
No job apocalypse: AI can do 30 to 40% of a job and the human part gains value, as long as AI stays not quite good enough.
The human edge is context: what you know and AI can't reach is where judgment and taste come from.
AI hurts learning: homework scores go up, retention goes down; he is building Learn Vector for one-to-one learning because of it.
Claims you can check later
Claim
Who
When we will know
How firm
No job apocalypse (AI replacing more than half of jobs); AI does 30 to 40%, and the human part is worth more
Ng (citing Brynjolfsson, McAfee)
Near to medium term
First-hand plus economic analysis; he has a stake
Software engineering openings are still growing, and good engineers are busier
Ng
Now
First-hand observation
Context is the technical basis of judgment and taste, and no one will close the gap for years
Ng
Long term
First-hand judgment
AI hurts learning: homework scores up, retention down
Ng
Already observed
Cites studies plus his own example; checkable in direction
AGI (any human intellectual task) is still decades away
Ng
Decades
First-hand judgment
Loss of control isn't science fiction; it can be managed step by step, like aviation
Ng
Trend
First-hand judgment, against Bengio
Back on the long-running theses
confirms
What you own is not the model: value moves up to what cannot be rented A new answer to where taste lives: in a person's context, not in the model.
confirms
The AI jobs apocalypse: from two-pole prophecy to task-level replacement AI replaces tasks, not whole jobs: it does 30 to 40%, the human part gains value, and software openings are still growing. The most mainstream first-hand testimony for that view.
confirms
The safety politics of open weights: gatekeeping or competition Among those against gating AI on safety grounds, Ng is the most credible voice.
Jack Dorsey: the open frontier Both against gating on safety grounds: Dorsey argues from freedom and burden of proof, Ng from economics and adoption.
confirms
Seven principles for learning faster, and Tara Seshan: writing is thinking Three sources agree: AI can get the work done for you, but not understand or remember it for you.
What would change my mind
AI capability really crosses a threshold and the 30 to 40% ceiling breaks.
Finished. Indigo's take on this piece is in two places:
