YT YouTube2Text

YouTube2Text → Transcripts → Andrew Ng: The Biggest Opportunities in AI Aren't Where You Think

Andrew Ng: The Biggest Opportunities in AI Aren't Where You Think — Transcript

by Silicon Valley Girl · 7,388 words · 1,134 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00There's been a lot of misinformation
0:01about AI.
0:02>> This is Andrew. He co-founded Google
0:05Brain and Coursera. His machine learning
0:07course has reached millions of learners
0:09and he is one of the most influential
0:11voices in AI today.
0:13>> So, a handful of leading AI companies
0:14have been very loud voices fearmongering
0:17around AI to try to get regulations
0:19passed. This Trumpian fear-based
0:21messaging has skewed societal's [music]
0:24perception to be really negative on AI.
0:26>> People talk to me about data centers and
0:29job loss.
0:30>> Maybe AI could do 30-40% of many jobs.
0:33And what that means is well, that 60%
0:35that a human does has become even more
0:38valuable.
0:38>> What about loss of human control over
0:41AI?
0:42>> I think about something else that we
0:43can't control.
0:46>> I think you're one of the voices in AI
0:49who comes with a huge background in
0:51machine learning and teaching AI. And
0:53also you're a positive voice cuz this is
0:56something that I've been seeing
0:58especially this summer, how polarized
1:00the society has become especially on
1:02social media when I'm posting about AI
1:04and people talk to me about data centers
1:06and job loss. Why do you think this wave
1:10started recently? What do you think the
1:12causes are?
1:13>> There's been a lot of misinformation
1:15about AI.
1:16And the root cause of a lot of this is
1:18an unfortunate attempt that started 2-3
1:21years ago of I think PR and regulatory
1:24capture.
1:25It turns out that the most one of the
1:27most valuable things in AI right now is
1:29a giant AI models, giant large language
1:31models that some companies have trained.
1:33But if you spend billions of dollars
1:35training a model, it's really
1:36inconvenient if someone else trains a
1:38model and wants to give it to anyone in
1:40the world to use for free.
1:42So, a handful of leading AI companies, I
1:44think as you know, have been very loud
1:45voices fearmongering around AI to try to
1:49get regulations passed to create an
1:51unfair playing field that favors
1:54incumbents so that we all have to pay a
1:57high toll for use of AI
2:00while stymieing the other teams, be it
2:02researchers or other companies that want
2:04to just give away open way to open
2:06source models that anyone can use much
2:08cheaper. Unfortunately, fearmongering
2:09works
2:10when you go and say AI is like nuclear
2:13weapons, which is an analogy that has no
2:15basis in fact. What do they have to even
2:17do with each other?
2:18Or when you go around and cherry-pick
2:20cases of AI, you're making a misstep and
2:23making much bigger than it is.
2:25Or even spread misinformation about how
2:28AI uses data centers uses a lot more
2:31water than the actual reality. This
2:33drumbeat of fear-based messaging has
2:36skewed societal's perception to be
2:38really negative on AI,
2:40which is unfortunate because this is
2:42slowing down American adoption in AI.
2:44This is making America less competitive.
2:48And unless we get the truth about AI out
2:50there, which is that it's fantastic
2:52benefit with some problems, but not
2:54nearly the way it was to be blown up to
2:55be, it will hurt individuals.
2:58>> I'm going to read out some of the
2:59problems that people are highlighting.
3:01Job loss and inequality. What do you
3:03think?
3:04>> The job apocalypse or job apocalypse,
3:06this idea that AI will take over 50% of
3:09jobs, people will be out of work,
3:11rioting in the streets, that's just not
3:13going to happen. With every wave of
3:14technology, including AI,
3:16>> [gasps]
3:17>> the skills we need to do great work
3:19shifts. And so AI is changing job
3:22professions.
3:23But boy, I wish AI were AI just doesn't
3:26work well enough. I know that handful of
3:28businesses want to hype up AI to say
3:31we'll have superintelligence or we'll
3:32have artificial general intelligence or
3:34whatever and can do all the stuff that
3:35humans do. I wish AI worked better. It
3:38was just not good enough to make AI do
3:39everything a human does.
3:41And if you look at the analysis of jobs,
3:44economists, like my friend Erik
3:46Brynjolfsson at Stanford and Daron
3:48Acemoglu at MIT,
3:50economists have analyzed many people's
3:52jobs and by breaking down into
3:54individual tasks, and maybe AI could do,
3:57you know, 30-40% of many jobs. And what
4:00that means is, well, that 60% that a
4:03human does has become even more valuable
4:05because it's called an economic
4:07complement to the 30-40% that's now
4:09cheaper. And so, what will happen is
4:13people that use AI, maybe people that
4:16use AI will replace people that don't
4:18use AI. But AI is not in a position for
4:21the vast majority of jobs to replace
4:24people. Of all the different
4:25professions, the one that's most
4:27affected by AI now is software
4:28engineering because AI is actually
4:30fantastic at writing code.
4:32And [clears throat] what we see is that
4:33the number of job openings in software
4:35engineering is up, contrary to what, you
4:38know, the doom fear mongers would say,
4:41right? AI is not actually able to
4:43replace software engineers.
4:44And all the good software engineers I
4:46know are busier than ever. Now, the flip
4:48side of it is if someone still write
4:50code like it's 2022 before ChatGPT,
4:54they're in trouble. They need new
4:55skills. Like, don't don't do stuff that
4:57that 30-40% of AI can automate. You've
5:00got to stop doing that. Let AI do that.
5:02But then gain new skills to do the other
5:0560-70% that AI cannot do.
5:07>> What would your advice be to new
5:09graduates cuz I talked to Eric on this
5:11podcast and he was talking about that
5:14there is not really a lot of impact on
5:16the job market except for I think he
5:18mentioned people from 18 to 25 which is
5:20graduated. What would be your advice to
5:22those people who don't have the
5:25expertise maybe to strategize in their
5:27job yet?
5:28They can only do manual work that AI can
5:30do as well.
5:31>> So, one real challenge
5:34for fresh college grads is that the
5:36university system is slow to adapt and
5:40so, um
5:42you know, I love academia. I think we
5:43should all support academia and
5:44universities. And when AI comes and
5:48transforms the way software is written,
5:50universities often take like a year or
5:52two for the faculty to master skills,
5:54then create new courses, get curriculum
5:57committee approval, whatever, get the
5:59faculty Senate to vote. It just takes
6:01years. And that speed of change in
6:03academia is very poorly matched to speed
6:06of change in AI.
6:08So, sadly, many universities are still
6:10teaching students to be ready for the
6:12jobs of 2022,
6:14when we shouldn't even be teaching them
6:15for the jobs of 2026. We should be
6:17teaching them for the jobs of 2028 and
6:19beyond.
6:21And what this means is the job openings
6:23are there. Uh tons of employers I know
6:25just can't find enough skill, you know,
6:28uh people at any level of seniority.
6:31But, um it turns out in my office right
6:33now, we have a lot of interns. They're
6:35current college students, fresh college
6:36grad, we also have one high school
6:38intern. And they're amazing productive.
6:40But, the key is they're all very AI
6:41native. They all use AI tools to do the
6:44things that I could do, but then also,
6:46you know, lean in to doing the things
6:47that uh humans can do that AI can't for
6:50the for for a long time. So, there's
6:52plenty work for people to do. But, so my
6:54advice to fresh college grads or to
6:56people currently in college is, um
6:59by all means work hard in classes, you
7:01know, get good grades, learn from the
7:02instructors. But, to the extent that
7:05there are still additional skills the
7:06university has not yet adapted to
7:08teaching, then find other ways to learn
7:11online, be it from Coursera or
7:13deeplearning.ai or Udemy or other places
7:16we can gain the more cutting-edge
7:18skills, especially AI skills, that um
7:22universities have not yet worked in the
7:23curricula.
7:24>> Quick pause. Because what Andrew just
7:26said about workflows connects to
7:27something that HubSpot just put out for
7:29free. The thing is, the real opportunity
7:31right now is not just in the models
7:33themselves, but in what you build on top
7:36of them. How you turn AI from a chat
7:38window into workflows that actually do
7:41work for you. And the good news is you
7:43don't have to start by building a full
7:44agent. You can start much simpler with
7:47better prompts. HubSpot just released
7:49the advanced ChatGPT prompt engineering
7:51playbook and the idea is very simple. In
7:557 days it helps you move from getting
7:57generic AI answers to building prompts
7:59that give you much more consistent and
8:01useful options because we've all had
8:03this moment. You type in one vague
8:05sentence, get something mediocre back,
8:08and then spend the next 20 minutes
8:10fixing it yourself. This playbook is
8:12basically built for that exact problem
8:15and it walks you through a single
8:16progression. First, you learn how to
8:18structure prompts with things like role,
8:21context, format, and parameters. Then it
8:24gets more advanced, few-shot examples,
8:26chain-of-thought reasoning, more precise
8:28prompt structures, and ways to make
8:30outputs more reliable. What I like is
8:32that it doesn't stop at individual
8:34prompts. It also shows you how to build
8:35reusable AI personas, modular prompt
8:38components, and eventually your own
8:40signature prompt engineering system. So,
8:42instead of starting from scratch every
8:44time, you're building a repeatable way
8:46to work with AI, whether you're creating
8:48content, analyzing data, or making
8:51business decisions. The playbook is
8:52free, link's in the description. Thanks
8:55to HubSpot for sponsoring this video.
8:57And now back to our conversation with
8:59Andrew.
9:00>> I want to say one other thing. Um
9:03it turns out one if if look at the skill
9:06map changes,
9:07one of the most important changes is um
9:10it's so much easier to build with AI
9:12than before.
9:13When something becomes much easier, a
9:15lot more people should do it. And
9:18so now not only should professional
9:20software engineers build software with
9:21AI, it's becoming much easier for
9:23everyone to build with AI. And people
9:27that embrace that and do so will be more
9:29productive and will accomplish more and
9:31I think have more fun than the ones that
9:32don't.
9:33And AI lets you build really fast. So,
9:36for people that are not just software
9:37engineers, but you know, marketers,
9:40recruiters, HR professionals, operations
9:42specialists, I think if they learn to
9:44build with AI, uh they'll really just do
9:48much more whatever their job role is.
9:49>> How do you by the way measure uh the
9:52increase in productivity uh when you
9:54deploy AI? Do you have like a KPI in
9:57your company?
9:58>> I wish there was a simple answer. I find
10:00that the business outcome of AI is more
10:03function of the business than the
10:04function of the AI.
10:07For for some, it may be an increase, you
10:09know, uh customer growth and retention,
10:11or maybe a faster serve customers, or be
10:13increased accuracy in some tasks. So,
10:16the KPIs tend to be related to the
10:17business rather than the AI.
10:18>> Mm. So, you can't like directly measure
10:20just the AI. Cuz it's it's an
10:21interesting thing to do cuz we've been
10:23deploying AI actively in my company and
10:25I think for me as a media company, it's
10:26probably the amount of views, the
10:27output.
10:28>> [snorts]
10:28>> Uh it's just interesting how Yeah, it's
10:31just interesting how different people
10:32measure. Even revenue. Like if you're
10:34becoming more effective I see. Uh with
10:36um how you make money.
10:37>> Actually, say more how are you using AI
10:38in your business?
10:39>> Oh my god. Uh I have the So, first of
10:42all, we have Claude for all of us, and
10:44we have certain projects for every
10:46social media that we're on. So, for
10:47example, for this podcast, we have
10:49[snorts] a project that's called guests,
10:51and it knows all the analytics from
10:53previous guests, and it has certain
10:54criteria on which we rank every single
10:56person who comes to the podcast, whether
10:59he's he or she's cited, whether they
11:02have a certain opinion on AI, whether
11:04they've been active with AI in their
11:06company, or if they're a recent founder
11:07in AI. So, it gives them different
11:09weights, and it comes up with a grade
11:11based out of 40. 40 meaning tier one, 30
11:15meaning tier three, etc. And then we
11:18have another one that analyzes every
11:20single podcast and gives me tips on how
11:22to ask questions.
11:23>> Oh, wow.
11:23>> Same for Instagram, same for LinkedIn.
11:26It has my tone of voice, personal
11:28dossier, my business strategy. So, it
11:30whenever it writes something, it knows
11:32all the facts about me, how I sound.
11:34Every social media is run by a person.
11:36So a person makes a strategic call. And
11:38by the way, if you can give me feedback
11:40on this, if I can improve. So what I'm
11:42working on right now is closing the
11:44loop, because sometimes they send me a
11:45text. I'm like, "Oh, we need to change
11:47this, this, and that." But that happens
11:49in a chat in Telegram. And we have this
11:52feedback. We We have a bot that scans
11:54all of our chats.
11:56But I really want AI to be able to learn
11:58continuously from this feedback to just
12:00know my taste better.
12:02>> There's one thing I see a lot in AI,
12:03which is
12:04It turns out for AI data scientists or
12:06AI as brainstorming ponder, it often
12:08comes up with, you know, one or two good
12:10ideas, two or three mediocre ones, and
12:12like,
12:13you know, four atrocious ones. And
12:15sometimes you wonder, "How could my AI
12:17have thought, you know, like that could
12:19even be a plausible idea?" And to me,
12:22this relates to the jobocalypse point of
12:24view, which is that for a long time,
12:27humans, you, me, everyone watching this,
12:30will have a significant context
12:31advantage over AI, which is that you
12:34know something that's incredibly obvious
12:35to you that, you know, that was an awful
12:37idea, but the AI did not. And it turns
12:40out that one of the reasons why AI will
12:42not replace our jobs or whatever of that
12:45sort of thing anytime soon, is because
12:47humans have a massive context advantage
12:51compared to AI.
12:52We know so much that, you know, from our
12:54years of experience, that we talked to a
12:56customer, we saw the funny facial
12:57expression that told us, "Ah, they don't
12:59like this." Or we talked to business or,
13:01you know, our manager said, "Hey, blah
13:03blah really care about this." And so it
13:05turns out that almost all humans, what
13:07maybe all humans, just know a lot of
13:09stuff that the plumbing does not exist,
13:12and I don't think it exists for the
13:13foreseeable future for AI to get. I know
13:16sometimes people talk about the
13:18importance of human judgment or human
13:20taste. And something you want to All
13:22right, what is taste? It's this fuzzy
13:24thing. But to me, the technical thing
13:26that underlies why humans have better
13:28judgment and better taste than AI is
13:30this context advantage. And because this
13:33is a long-term advantage, like no one's
13:35going to solve this, you know, in a few
13:36years, this is why we just need a lot
13:39more humans with that judgment and taste
13:41to keep on complementing AI.
13:43>> And doesn't this make education even
13:45more important? Because education also
13:46gives us context. Cuz it's another
13:48another thing I'm hearing about AI, like
13:49you won't need education cuz all the
13:51information is at your fingertips. You
13:52just ask ChatGPT. But when you say
13:55context and taste, for me, that's years
13:58of acquiring knowledge and learning from
14:00the best and seeing how they perform
14:02versus just asking a chat.
14:04>> I'm going to say something that may be
14:06controversial. I don't have said this
14:07publicly, but I think it's true, which
14:09is frankly,
14:11AI models are terrible for learning. Um
14:14I know people think
14:17AI is
14:18AI is wonderful getting things done. Use
14:20it all the time. Love it. But all the
14:23data that's coming out is that when say
14:26college students use AI, we know this.
14:29It's just the studies now back up as
14:30well. So we also have numbers, but the
14:32data is very clear.
14:33Students score higher on homeworks when
14:36they use AI. Yay, higher homework
14:37scores.
14:39But retention,
14:40their long-term performance is much
14:42worse because they AI do the work for
14:43them. More and more studies are coming
14:44out to back this up now that I think
14:48people think oh, it turns out, you know,
14:51I think Wikipedia is a wonderful tool.
14:53It has tons of facts. Our web search is
14:54a wonderful tool. It has tons of facts.
14:57But it turns out that when you
14:59ask AI to do work for you, you are
15:02cognitive offloading to AI, which is
15:05great because that's how society moves
15:06forward and it gets work done. But human
15:09retention is much worse. It's just so
15:11clear that LLMs, as they are most
15:14commonly used, are terrible for
15:16learning. I'm not saying there's no way
15:17to use it in a way that is good for
15:19learning. I think there there ways to
15:20use that are good for learning. But even
15:21for myself, there's so many things I
15:23asked the AI model over the last, you
15:24know, 6 months or whatever, like, I
15:26don't know,
15:26>> [gasps]
15:27>> building some project, how does this
15:28front-end, back-end component work,
15:30whatever. Give me the answer, get the
15:32job done. It was fantastic. But 6 months
15:34later, I don't remember the answer when
15:36I need to redo that front-end, back-end
15:39component, I asked the AI again. So,
15:41data is really clear. We should stop
15:43thinking of AI as helpful for learning,
15:45at least the vast majority of ways that
15:48the vast majority of people are using AI
15:50models today is absolutely terrible for
15:52learning.
15:53>> But you're building a company helping
15:54solve that, right? Cuz the one
15:56one-to-one tutoring with AI is that
15:58where you just announced with a 100
16:00million investment from Coursera.
16:02>> Yes. So, I recently about leading a new
16:04organization called Learn Vector that is
16:07focused on um
16:08building new learning experiences that
16:11is much more one-to-one than
16:12one-to-many. So, you know, 15 years ago,
16:14I I was privileged to participate in the
16:17online courses movement that I think
16:20changed the way a lot of people learn.
16:22But that was and still remains largely a
16:24one-to-many experience where everyone,
16:26you know, kind of watches the same
16:28video, which is actually okay. It
16:29actually works well. But the technology
16:31now exists to create much more
16:33personalized, customized one-to-one
16:35experiences, and so our team is working
16:38hard on that. I think we'll have a a lot
16:39more to show by early next year. When
16:42think about human skill development, I I
16:44feel like because AI has so heavily
16:46impacted software engineering, um what
16:49we see happening in the job market for
16:51software engineering is a harbinger or
16:53the forerunner of what we'll see in
16:55other disciplines as well. And in
16:56software engineering, um people need to
16:59learn new skills, but when they do, they
17:01are thriving and creating more value and
17:04frankly getting raises and doing even
17:05more exciting projects. And what I've
17:07seen the early signs of in other
17:09disciplines as well. For example, in
17:10software engineering, you know, most
17:12developers, like front-end, back-end
17:14developers, have now become full-stack
17:16developers because of AI help, you could
17:18take on broader scope.
17:20I'm seeing early signs of this in other
17:22disciplines as well, where for example,
17:24someone that in marketing that did
17:26marketing coordination, uh coordinating
17:28marketing campaigns, with AI help, they
17:30can now become more of a full cycle
17:32marketer and take on a broader scope.
17:34And I'm seeing, you know, frankly,
17:35sources in recruiting become more full
17:37cycle, do end-to-end recruiting. So,
17:41now the good news and bad news is for
17:42people to step up to these broader
17:44roles, you do need to learn AI skills,
17:46but also it's not just learning AI, you
17:47also need to learn these other skills,
17:49uh like how do you do the other parts of
17:51marketing or recruiting or software
17:52engineering or AI engineering. So, I
17:54think this actually creates a heavy
17:56need, a big need for people to gain new
17:58skills, but when they do, which is both
18:00AI skills, but also disciplinary skills,
18:03then they can do much more, hopefully
18:05have more fun, work on more exciting
18:07projects, hopefully get paid more as
18:09well. And one reason I kind of worry
18:11about the fearmongering is um I got an
18:14email from someone that was about to
18:15enter college, and you know, he emailed
18:18me saying, "Hey Andrew, taking online
18:20courses, but I'm really struggling with
18:21what I should major in college because
18:24in 4 years, won't AI do all this and
18:26everything I learn will be obsolete?"
18:28And the answer is, "No, of course it
18:29won't all be obsolete." But when we keep
18:31on pushing these fear messages, uh
18:35we make people wonder if they will even
18:37be relevant, and it makes people not
18:39lean in to gain these skills that'll put
18:42them in much better position. So, I see
18:44very clearly that these fearmongering
18:47are distorting how many people,
18:50including, you know, high school
18:51students, college students, fresh grads,
18:53think about the economy, and frankly,
18:56making people give up is one of the
18:57worst things we could be doing. In this
18:59era, when people that lean in will
19:02thrive.
19:03>> Andrew has taught over 8 million people
19:05AI. He started teaching machine learning
19:08online back in 2011, years before the
19:11current AI boom. Now, one of the
19:12companies he's building is focused on AI
19:14agents. From the way that it sounds, it
19:17can still feel way too technical. So, I
19:19put together a step-by-step guide to
19:21building your first AI agent with no
19:23coding required. It walks you through
19:25what to automate, how to set it up, and
19:27how to make it actually useful. It is in
19:29my newsletter this week. The
19:31newsletter's called Future Proof. It's
19:33free. Link is in the description.
19:35>> What would you reply back to that email
19:38that somebody sent to you? What would
19:39you say
19:40>> is the best major to study now to thrive
19:42in AI era? Do you think it's like going
19:44deep into niche or just broader computer
19:46science so that you can acquire AI
19:48skills really fast?
19:49>> You know, I don't know what's the best
19:51major. There are awful lot of great
19:52majors. It's It's like
19:54I kind of feel like what's the best job
19:56in the world? It's like what's the best
19:57major in the world? I guess so.
19:58>> that you love, right?
19:59>> Yeah, my daughter wants to be an
20:00astronaut. I don't know if she can major
20:02in becoming an astronaut. I have to
20:04think about that. When she gets older,
20:05she may change her mind. I see so many
20:06opportunities across across so many job
20:10profiles. It all seems very exciting to
20:11me. But, do learn AI. Do learn to build
20:13with AI. The other thing that my team's
20:16been working on the AI engineering
20:17skills map to try to map out you know,
20:19the most important skills for AI
20:20engineering. One thing that I felt
20:22intuitively, but I was surprised to see
20:24it show up in the data, was that a lot
20:26more job descriptions seem to be saying
20:28they want people that demonstrate a very
20:30high sense of agency. Because it turns
20:32out with AI, there are a lot more
20:34opportunities for individuals to spot
20:36problems and go build something or do
20:40something to go solve it. So, I think
20:42we're really evolving Well, we've long
20:44been evolving, but we're accelerating
20:46past the era where people should sit
20:48around and wait for their boss to tell
20:49them what to do.
20:50>> This is what I've been feeling a lot,
20:52especially when we started doing remote
20:53work. I want people to be entrepreneurs
20:56within their niche. Like if you're
20:58helping me with LinkedIn, you're an
21:00entrepreneur there. You can hire more
21:01contractors. You can deploy different
21:03tools. You make the strategic decision
21:05whether this topic is good or not. Shall
21:07we proceed with it? I really think, and
21:09tell me if you agree with me, we're
21:11moving into that job market where
21:12everyone is kind of independent in their
21:15workplace.
21:15>> I think people will have much more
21:17autonomy and creativity. So, I agree
21:19with that. I'd even go one step further,
21:20which is I talked to a lot of people is
21:23the engineers and others in large
21:24companies that tell me that their
21:26manager tells them to stay in their swim
21:28lane. They'll say, "Oh, I have this
21:29creative idea, but the manager says,
21:31"No, I need you to focus on this one
21:33thing." Frankly, often because the
21:35manager's career depends on it. But, I
21:36feel like the number of opportunities
21:38for people to spot things outside their
21:40swim lane,
21:41um and then in a responsible way explore
21:44how to get it done. That feels very
21:46exciting to me, and I think that in the
21:48future, the businesses that set up a
21:50culture that encourage people to learn
21:53AI, build fast, responsibly, um
21:57talk to customers will drive a lot more
21:59value than the more hierarchical silo
22:02organizations.
22:03>> Yeah, it it starts with hiring the right
22:05people and then nurturing this in your
22:06organization. When you say learn how to
22:09use AI and become proficient with AI,
22:11can you give me some benchmarks? Like of
22:13a person who's like say marketer,
22:15knowledge worker, advanced with AI, what
22:17are you looking for when you're
22:19interviewing this person?
22:20>> I'm pretty sure my team's ahead of the
22:21curve. All of my marketers know how to
22:23code. So, as part of how I interview
22:25marketers, we ask them what they built,
22:27and uh if they have not built any
22:30software
22:31>> If it's a dashboard, is it good or bad?
22:32Like is it too basic or
22:34>> A dashboard's Okay, my team's pretty,
22:36you know, somewhat ahead of the curve.
22:38Whereas
22:39>> Well, it's good to hear like that.
22:40>> All of my marketers have built much more
22:41sophisticated things than dashboards.
22:43>> me?
22:43>> Oh, I I I feel like um
22:45I don't know. The other day, uh one of
22:47our someone on the marketing team was
22:48talking about the tools that he had
22:49built to uh when he's considering
22:52writing an article on something, it will
22:54um crawl the web, find related work, has
22:56a custom desktop app. He actually built
22:58a desktop app that runs on his Mac to um
23:00highlight related articles for him. then
23:02he can chat to the whole system,
23:04navigate, you know, the thing he's
23:06writing as well as the related work, and
23:08he had a large dashboard for trawling
23:10the internet to highlight to him
23:12exciting things that are popping up. Um,
23:15now even on my team, I think that
23:16marketers is ahead of the curve.
23:18>> But that's that's great to hear. Any
23:20other interesting use cases
23:22uh that will inspire people to build
23:24something similar?
23:26>> Let's see. Maybe a uh
23:28my finance team uses AI extensively. So,
23:32I think uh
23:33uh my um
23:35one of my CFOs uh realized that, you
23:37know, her team was spending hours every
23:39week clicking through documents, open
23:41this, copy paste this number here. And
23:43so, um she started building uh
23:45automation scripts that runs on a
23:47routine that um automatically opens
23:50files, checks what's in there, checks
23:52the consistency, highlights to her team
23:53if there's something uh if there's
23:55something they need to be paying
23:56attention to or if a new document has
23:58showed up. So, I find that rather than
24:00waiting around for an engineer to do the
24:02work for them, the team's ability to to
24:05kind of a not just build dashboards, but
24:07build kind of a data management
24:10infrastructures. They can just data,
24:12alert them if something's happening.
24:13>> Yeah.
24:14>> I think uh my finance and marketing
24:16teams are doing that. Oh, my recruiting
24:18team um well, we actually have
24:20recruiting engineers, which are really
24:22professional engineers that sit in a
24:24recruiting team that are building very
24:26sophisticated tools for recruiting. And
24:28this is actually the other trend. I
24:29think marketers, recruiters, HR
24:30professionals should all learn AI. But
24:33the other thing is when you take an
24:34engineer and embed them in these teams,
24:36then that further accelerates what you
24:38can do.
24:38>> We do the same. We we start with
24:40something basic, build it ourselves,
24:41then we hit the wall, an engineer comes
24:43in, we build it further.
24:45>> Frankly, when you look at not just
24:46software engineers, but recruiting
24:47engineers, marketing engineers, HR
24:48engineers, I think there's so much
24:50valuable engineering work that can now
24:52be done. I'm just, you know, not worried
24:55about running out of
24:56>> [laughter]
24:56>> Frankly, all my friends were so busy. We
24:58think, "Boy, how could we run out of
24:59engineering jobs?"
25:00>> Yeah, yeah. There are so many cool ideas
25:02you can experiment on. But, you touched
25:03up on something that is actually one of
25:05the fears when when we talk about like
25:06financial information, how much you're
25:08giving to AI. So, I gave my
25:11Perplexity permission to scan my
25:12Fidelity account so it can track my
25:15portfolio, tell me when to rebalance. It
25:17doesn't do anything on my behalf, but it
25:18has access.
25:20Do you think there is any problem with
25:23that?
25:23>> This is complicated. I think AI privacy
25:26is a complex area. Um
25:29and it depends a lot on the company that
25:31you are sharing your data with.
25:33So, for example, I trust all the
25:35hyperscalers to really 100%, you know,
25:39follow the terms of service and to do
25:41what they say. My personal opinion not
25:42not giving legal business advice, but
25:44I'd be shocked if, you know, the largest
25:46hyperscalers publish the terms of
25:47service or some privacy notice and if
25:49they breach that that because that'd be
25:51not the culture, it'd be so damaging to
25:53the long-term business model.
25:54Now, that's on the largest hyperscaler
25:56side.
25:57If you look at AI company side, there's
25:59been, you know, at least one company
26:01that I won't name that seems to
26:03occasionally change the terms of service
26:06and you're using it, you go to the
26:08website, say you pop up, "Hey, we
26:10changed the terms of service to retain
26:11your data or train your data." And
26:13you're not paying attention and click
26:14the wrong button, then they suddenly
26:17gave themselves permission to access
26:19your data in a way that I'm not that
26:20comfortable with. I feel like I handle,
26:22you know, some sensitive information.
26:24So, then I tend to be very careful with
26:26the businesses that I just don't feel
26:29that culture and their DNA and frankly
26:32the long-term business model is as tied
26:34to protecting individual user privacy
26:37than the hyperscalers.
26:39And um I see businesses, you know, get
26:42this as well. For example, one of my
26:44teams at Aspire, we work with very large
26:46corporations including banks with
26:49incredibly sensitive financial data. And
26:53as you can imagine
26:54AI spiring and our clients do not
26:57willingly share, you know, very
27:00sensitive and often material non-public
27:02information, right? NPI
27:04with Frontier with with Frontier Labs
27:07without really careful thinking about
27:09the guardrails and privacy. So I think
27:11it's complicated.
27:12>> So trusting hyperscale, but also
27:14another thing that you can do you can
27:15download an open-source model and just
27:17run it on your computer and then it just
27:19stays on your computer.
27:20>> Yes. I think yes.
27:22It turns out a lot of banks will
27:23actually run the things in
27:26a virtual private cloud on prem. So they
27:28so they never even leave their control.
27:30But I think for individuals it's for the
27:32really sensitive things
27:34I sometimes run a local model and it's
27:36been interesting with the open way
27:38models. Some of the latest open way
27:39models are approaching Frontier
27:41capability and that are, you know,
27:44actually small enough. They're actually
27:45really good models now that you can run
27:47on your
27:47>> Yeah, the one from Meta, right? The
27:48recent one.
27:49>> Oh, yes. Meta music is a good model and
27:50I'm thinking also the latest version of
27:52Quen is also very good. But I think
27:54frankly these models change every other
27:56week. So I think the best practice is to
27:58not get stuck on one, but to keep on
28:00trying new models.
28:01>> So basically when there's a situation
28:03that you don't trust anyone, you run a
28:05local model and this is how you keep
28:06your data safe.
28:08>> I do trust the hyperscalers, but
28:10sometimes for, you know, literally NPI,
28:13material non-public information, that I
28:15won't even send to that I just can't
28:17even send that to the cloud. So that uh
28:19I either do it manually without AI help
28:22or if I really need to use AI, then, you
28:24know, really carefully only use a local
28:27model.
28:27>> Interesting. Okay, this is this is this
28:29is an interesting one. Okay, what about
28:33loss of human control over AI? Cuz I've
28:36talked to I talked to Yoshua Bengio, who
28:38is very um negative when it comes to
28:42open free AI without any regulation and
28:46he painted me some very scary pictures
28:49of AI taking over control because we
28:51basically the the whole scenario is we
28:54can't control something that's smarter
28:56than us and if AI gets smarter and
28:58smarter, what where do we end up? What
29:00do you think about that?
29:01>> about something else that we can't
29:02control, which is airplanes. No one can
29:06build an airplane that you can fly
29:07perfectly. Winds will buffet to the
29:09realm.
29:10And then incidentally in the early days
29:11of developing airplanes, some airplanes
29:14crash and people died. Then it was
29:15tragic and awful.
29:17But through the early lessons learned,
29:19we then learned to control airplanes
29:21better and better so that today, you
29:23know, we can mostly get in an airplane
29:25and not fear too much for our lives. And
29:28it's really like that too with AI. No
29:29one can perfectly control AI because it
29:32generates tokens or outputs that are a
29:34little bit random so we don't really
29:35know what exactly it'll do.
29:37But as we run them and you know, there's
29:39been a small number of mishaps on which
29:41is unfortunate and some number of
29:43mishaps have done some road damage. But
29:46the way we engineer almost any system
29:49from an airplane to electric circuits to
29:52now AI is carefully
29:55grow their capabilities so that we can
29:57have a controlled environment in which
29:59to measure what's wrong and then to
30:02shape it to make sure we can control it
30:03well enough that it behaves responsibly
30:06and safely. And to this day, we can't
30:07perfectly control any airplane and we
30:10will never perfectly control AI either.
30:12But I think
30:14we are safely controlling them well
30:16enough that this loss of control doesn't
30:18feel like science. It feels like science
30:20fiction.
30:22>> What about deep fakes?
30:24>> Deep fakes are a problem. Well, one of
30:26the most disgusting things I've ever
30:28seen or heard of is non-consensual
30:30intimate deep fake imagery or
30:33I'm really glad that you know, US
30:35Congress has been moving to right let's
30:37pass laws to get rid of that. Penalties
30:39for that are just I think there's some
30:40really problematic uses of AI that we
30:43should outlaw, heavily penalize. Let's
30:45just get rid of that.
30:46>> What do you think about children and
30:48social connection when it comes to AI
30:50with kids using more of AI?
30:52Cuz we've seen social media, how
30:55you know, there are people who are dumb
30:56scrolling all day. And my daughter, who
30:58is 5 years old now, whenever I don't
31:00have an answer, she's like, "Ask
31:02ChatGPT." And like, "Who's that person?"
31:04I'm like, "I don't know." "Ask ChatGPT."
31:06And like, she thinks ChatGPT knows
31:09everything. What would you say about,
31:12you know, kids' future with AI?
31:15>> First, I think kids have a bright
31:16future. It's such an exciting time to be
31:18a child to grow up in this environment
31:19with tools that none of us ever had
31:21before.
31:21>> [gasps]
31:22>> At the same time, we're seeing that
31:23social media, um I think social media
31:26has probably been blamed a bit more than
31:28it deserves, but it does deserve blame.
31:30Uh it has kind of not been great for
31:32kids. I actually worry a lot about
31:35It's a wonderful tool, but AI damaging
31:39learning is something I worry a lot
31:40about. So, it turns out, um I have a
31:435-year-old and a 7-year-old.
31:45When I teach them math, they're still
31:46young enough that I can basically, you
31:48know, not let them use a calculator. Can
31:51say,
31:51"How do you multiply these numbers?" And
31:53I don't give them a calculator and
31:55practice that with them. But as they get
31:56a little bit older,
31:57I worry a lot about students using
32:00cognitive offloading to AI in a way that
32:03damages the long-term learning and
32:04retention. Um but then at the same time,
32:07oh, I actually built an app. I did not
32:08like any of the you know, free online
32:11learning type type type of things. So, I
32:13actually built my own to um have my
32:15daughter learn to type. And I'm hoping
32:18that she's actually getting pretty
32:19decent now for a 7-year-old.
32:20>> Oh, so she's typing
32:21>> Oh, yeah, she actually typed all the
32:23lowercase letters. She's still a little
32:24bit fit you know, not her shift
32:27uppercase letters is a little bit not
32:29quite there. But I think that this um
32:31unlocks, you know, responsible adult
32:34supervised use of online tools. And I
32:37think it it it's really tricky, you
32:39know.
32:40I think um
32:41adult supervised use of digital tools
32:43seems to be a great thing for kids, but
32:45too many adults don't have time to
32:46supervise the use of the tools and then
32:48the incentives of
32:49>> [gasps]
32:49>> say social media, right? To do funny
32:52things.
32:53>> Yeah, has to be the right incentive when
32:55it comes to AI.
32:56Okay, [snorts] you mentioned we we
32:58talked about the fears, we talked about
32:59how you can improve your work with AI.
33:03Can you name some of the biggest
33:04opportunities in AI in 2026 for people
33:07who want to build?
33:08>> A friend individual that wants to build,
33:10I don't think it's one-size-fits-all,
33:13but because the cost of building has
33:15plummeted,
33:17um
33:17>> [clears throat]
33:18>> I encourage people to learn AI,
33:21build fast, and talk to customers. I
33:25find myself building things, I don't
33:27know, every week, every weekend,
33:30because I or someone on our team with
33:32some problem and I have some idea for
33:35building some AI thing
33:37to automate it. Last weekend, I had
33:40really I was using a frontier model to
33:42analyze a lot of our key business
33:43metrics cuz I didn't have time to do it
33:45myself, but it was kind of
33:47measuring, you know, dividend yield and
33:49key business metrics and I didn't have
33:51time to go find a data scientist to go
33:53work with me on it, so I just did it a
33:55variety of frontier models, being really
33:57careful on their data retention
33:59policies, so I did not use models with
34:01data retention policies that I don't
34:02like
34:03uh in order to analyze data.
34:05But and then I find that um
34:08what's happened AI is the cost of
34:09building has plummeted and so the
34:11challenge is shifting to deciding what
34:13to build, which I was calling which I've
34:15been calling the product management
34:16bottleneck.
34:17And so people, you know, founders,
34:19engineers, product managers, they can
34:22talk to customers, get a sense for the
34:24taste of judgment on what to build, and
34:26then build with AI and iterate quickly.
34:28I think there's just a ton of the things
34:31to do.
34:31>> And you've been starting so many
34:32companies. You're like when I looked at
34:34your portfolio, do you think for
34:36beginners when you said you built
34:37something during the weekend, how do you
34:39decide what to focus on or you can
34:41pursue multiple ideas because of AI now
34:43and you can just be, you know, playing
34:45in different companies at the same time?
34:47>> It turns out building a company is still
34:48really, really hard. And so there's a
34:50lot to be said for a single threaded
34:52leadership or someone that's fully
34:54focused on just one thing. I find that,
34:56you know, over a weekend I can often
34:58build an own wrapper, build a simple
34:59application, but
35:01I wish it was that easy to build a large
35:05company. Um I find that building
35:07something meaningful often takes either
35:10real technical depth or and or deep
35:13customer insight and integration of
35:15customers. And yes, we can now, you
35:17know, use AI to code something in a few
35:19hours, but that's a small piece of the
35:21puzzle. So spend time
35:23understanding the technical complexity
35:25of building the really complex software
35:27that takes us like months, you know,
35:29maybe years. Or having that deep
35:31customer insight to decide what to
35:32build, that also just takes a lot
35:34talking to people, reading facial
35:35expressions, surveys, doing that over
35:37and over until we figure what to build.
35:39Um and so I think there's sometimes a
35:41lot of value to sampling widely, but
35:44then having that focus for an individual
35:46to go really deep in a couple sectors
35:48that that still seems important for
35:50building a business.
35:51>> My last question, I know it's we don't
35:52have much time, but I wanted to ask you
35:54about AGI just because people use this
35:57word so much and some people say I think
35:59Jensen Huang said we already reached
36:01AGI. You said it's decades away. What's
36:04the one criteria when you're going to
36:06say we reached AGI?
36:08>> So different people say we reached AGI
36:10at different times because of different
36:11definitions of AGI. The definition I'm
36:13most familiar with is AI that can do any
36:16intellectual task that a human can.
36:19But so the human brain can take, say,
36:21five years to study and do a PhD thesis.
36:24Uh or and so can AI write a PhD thesis
36:27or a human can learn to drive a truck
36:30through a dense rainforest with you tens
36:32of minutes of practice. So when can AI
36:35do that to drive new environment with
36:37tens of minutes of practice. It feels
36:38like there's a long list of these things
36:41that AI cannot do
36:44for what feels to me decades. I hope
36:46it's only decades. Maybe it turn out to
36:48be longer. So that's what I think for
36:50that definition of AI or AGI, AGI is
36:53still very far away.
36:55But um it turns out because of you know
36:57economic incentives, I think Open AI and
36:59Microsoft had an agreement. There's
37:00actually been renegotiated now, so
37:02that's gone away. But Open AI had
37:04economic incentive to try to declare
37:06reaching AGI earlier. Uh and so it turns
37:09out that if you come up with other
37:11definitions of AI,
37:13depending on how far you lower the bar,
37:15then you could totally have reached AGI,
37:17you know, already or even 30 years ago
37:19depending on how you want to define it.
37:21>> Yeah, true. Andrew, thank you so much
37:23for this positive conversation and very
37:25applicable. I like when
37:27you watch something and then you go and
37:29you measure yourself against what people
37:30are doing with AI, look at your process
37:33and maybe expand it. So thank you so
37:35much for showing what your team is doing
37:37and thank you for your insights.
37:40>> Yeah, I think given the huge benefits of
37:42AI to come, I hope whoever's watching
37:44this is motivated to really go learn AI,
37:47apply it
37:49and and even to go build some things.

About this transcript

This page contains the full transcript of Andrew Ng: The Biggest Opportunities in AI Aren't Where You Think by Silicon Valley Girl, generated from the public captions YouTube serves with the video. The transcript has 7,388 words across 1,134 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.