back to index

Andrew Ng on AI's Potential Effect on the Labor Force | WSJ


Whisper Transcript | Transcript Only Page

00:00:00.000 | To sum it up to start, what would you say are going to be the biggest positive and
00:00:04.880 | negative impacts on the workforce from AI over the next, say, five years?
00:00:09.940 | >> I think there will be massive productivity boost for existing job roles and
00:00:14.600 | it will create many new job roles.
00:00:16.740 | And I don't want to pretend that there will be no job loss.
00:00:19.020 | There will be some job loss, but
00:00:20.880 | I think it may not be as bad as people are worried right now.
00:00:24.900 | I know that we're having an important societal discussion about AI's impact on
00:00:29.960 | jobs, and from a business perspective,
00:00:32.640 | I actually find it even more useful to not think about AI automating jobs, but
00:00:37.220 | instead AI is automating tasks.
00:00:39.600 | So it turns out that most jobs, you can think of as a bundle of tasks.
00:00:44.640 | And when I work with large companies, well often, many CEOs would come and
00:00:48.680 | say, hey, Andrew, I have 50,000 or 100,000 employees.
00:00:51.680 | What are all my people actually doing, right?
00:00:53.960 | Turns out none of us really know in detail what our workforces are doing.
00:00:57.640 | But I found that if you look at the jobs and break them down into tasks,
00:01:00.880 | then analyzing individual tasks for potential for AI automation or
00:01:05.280 | augmentation often leads to interesting opportunities to use AI.
00:01:10.960 | And maybe one concrete example, radiologists.
00:01:14.200 | We've talked about AI maybe automating some parts of radiology.
00:01:17.320 | But it turns out that radiologists do many tasks.
00:01:19.480 | They read x-rays, but they also do patient intake, gather patient histories.
00:01:24.440 | They consult with patients, mentor younger doctors.
00:01:26.400 | They operate the machines, maintain the machines.
00:01:28.440 | So they actually do many different tasks.
00:01:30.240 | And we found that when we go into businesses and do this task-based analysis,
00:01:34.880 | it often surfaces interesting opportunities.
00:01:38.320 | And regarding the job question, it turns out that for many jobs, if AI
00:01:42.740 | automates 20, 30% of the tasks in a job, then the job maybe is actually decently safe.
00:01:49.520 | But what will happen is not that AI will replace people, but
00:01:54.160 | I think people that use AI will replace other people that don't.
00:01:57.120 | >> What are the types of tasks that you're seeing?
00:02:01.880 | And if you can say, what professions do you think are most,
00:02:06.960 | where you have the highest concentration of those types of tasks?
00:02:10.440 | >> So some job roles really disrupted right now, call centers,
00:02:14.440 | call center operations, the customer support is one.
00:02:16.680 | It feels like tons of companies are using AI to nearly automate that or
00:02:22.160 | automate a large fraction of that, I think sales operations, sales back office,
00:02:26.480 | those routine tasks are being automated, and I think a bunch of others.
00:02:30.680 | I feel like we see different teams trying to automate some of the lower legal work,
00:02:37.080 | some of the lower level marketing work, a bunch of others.
00:02:39.480 | But I would say the two biggest I'm seeing are customer service and
00:02:43.440 | then maybe some sort of sales operations.
00:02:46.040 | But I think there's a lot of opportunities out there.
00:02:49.320 | How do you think it's going to change the role of CIOs, the folks in this room?
00:02:53.000 | >> I think it's an exciting time to be a CIO.
00:02:58.080 | One thing that my team AI fund does is we often work with
00:03:02.120 | large corporations to identify and then execute on AI projects.
00:03:07.600 | So over the last week, I think I spent almost half day sessions with two
00:03:12.160 | Fortune 500 companies where I got to hear from their technology leadership
00:03:17.360 | about the use cases they are pursuing.
00:03:19.560 | And some patterns I've seen every time we spend a little bit of time brainstorming
00:03:23.200 | AI projects, they're always way more promising ideas than anyone has
00:03:28.080 | the resources to execute.
00:03:29.720 | And so it becomes a fascinating kind of a prioritization exercise to
00:03:34.200 | just decide what to do.
00:03:36.360 | And then the other decisions I'm seeing is after you do a task-based analysis,
00:03:40.560 | identify tasks and jobs for ideas, or after your team learns about AI and
00:03:46.120 | brainstorms ideas, after you prioritize there's a usual kind of buy versus
00:03:51.320 | build decision.
00:03:54.040 | And it turns out that we seem to be in an opportunity-rich environment where
00:04:01.200 | actually what AI Fund winds up often doing is often the company will say,
00:04:06.720 | "These projects only keep close to my heart.
00:04:08.640 | I will pay for 100% of this."
00:04:10.800 | But there's so many other ideas that you just don't want to pay completely by
00:04:14.800 | yourself for the development of and then kind of we then help our corporate
00:04:19.440 | partners build it outside so you can still get the capability without needing
00:04:23.920 | having to pay for it entirely by yourself.
00:04:26.320 | But I find that at AI Fund we see so many startup and project ideas that we
00:04:33.200 | wind up having to use task management software to just keep track of all these
00:04:37.480 | ideas because no one on our team can keep straight of these kind of hundreds of
00:04:41.000 | ideas that we see and have to prioritize among.
00:04:43.600 | So you asked the generative AI to keep track of all the tasks that generative
00:04:50.120 | AI can do?
00:04:50.920 | Oh, that'd be interesting, but we actually use Asana to keep track of all the
00:04:54.480 | different ideas, so generally I summarize it.
00:04:58.600 | You've talked for years about the importance of lifelong learning, the
00:05:04.880 | enhanced importance of that in the AI world.
00:05:09.760 | Is it realistic to think that people will be able to re-skill, to educate
00:05:15.120 | themselves at a pace that keeps up with the development of this technology?
00:05:18.640 | Not necessarily the folks in this room, but the people whose tasks are being
00:05:24.120 | automated.
00:05:25.680 | How big of an issue do you think that's going to be, the displacement?
00:05:28.120 | Because when we talk about technology and jobs, we always talk about in the
00:05:32.040 | long run, "Look, we used to all be farmers, now we're not."
00:05:35.400 | In the long run it'll be fine, but in the meantime there's a lot of
00:05:39.200 | dislocation.
00:05:40.520 | Yeah, so this really can't answer.
00:05:42.880 | Honestly, I think it's realistic, but I'm a little bit nervous about it.
00:05:47.040 | But I think it is up to all of us collectively in leadership roles to make
00:05:51.000 | sure that we do manage this well.
00:05:54.040 | One thing I'm seeing, so the last wave of tech innovation, when deep learning,
00:05:58.560 | predictive AI, labeling technology, whether you call it started to work really
00:06:02.080 | well 10 years ago, it tended to be more of the routine, repetitive tasks like
00:06:07.040 | factory automation that we could automate.
00:06:09.680 | With generative AI, it seems to be more of the knowledge workers' work that AI
00:06:15.320 | can now automate or augment.
00:06:17.880 | And to the reskilling point, I think that almost all knowledge workers today can
00:06:22.760 | get a productivity boost by using generative AI right away, pretty much
00:06:27.600 | right now.
00:06:28.680 | But the challenge is there is reskilling needed.
00:06:31.680 | We've all seen the stories about a lawyer generating hallucinated
00:06:36.960 | court citations, and then getting in trouble with the judge.
00:06:39.800 | So I feel like people need just a little bit of training to use AI responsibly
00:06:45.000 | and safely.
00:06:46.080 | But with just a little bit of training, I think almost all knowledge workers,
00:06:49.440 | including all the way up to the C-suite, can get a productivity boost right away.
00:06:54.280 | But I think it is exciting, but also, frankly, daunting challenge to think about
00:06:59.960 | how do we help all of these knowledge workers gain those skills.
00:07:03.960 | Is that problem you alluded to, the hallucination problem, the accuracy,
00:07:09.600 | concern, is that fixable with AI, or is it more that we just have to learn to use it
00:07:16.520 | the right way and assume an error rate?
00:07:19.760 | Yeah, so I don't see a path.
00:07:22.400 | I myself do not see a path to solving hallucinations and making AI never
00:07:28.240 | hallucinate in the same way that I don't see a path to solving the problem that
00:07:32.440 | humans sometimes make mistakes.
00:07:35.080 | But we've figured out how to work with humans and for humans and so on.
00:07:39.320 | It seems to go OK most of the time.
00:07:41.560 | And I think because gen 2 AI bursts onto the scene so suddenly, a lot of people
00:07:46.080 | have not yet gotten used to the workflow and processes of how to work with them
00:07:50.240 | safely and responsibly.
00:07:51.960 | So I know that when an AI makes a mistake, sometimes it goes viral on social
00:07:58.200 | media or it draws a lot of attention, but I think that it's probably not as bad as
00:08:04.880 | the widespread perception.
00:08:06.160 | Yes, AI makes mistakes, but plenty of businesses are figuring out, despite some
00:08:11.240 | baseline error rate, how to deploy it safely and responsibly.
00:08:15.520 | And I'm not saying that it's never a blocker to get anything deployed, but I'm
00:08:19.320 | seeing tons of stuff deployed in very useful ways.
00:08:23.440 | Just don't use gen 2 AI to render medical diagnosis and output directly what drug
00:08:29.760 | to tell a patient to take.
00:08:30.760 | That would be really irresponsible.
00:08:32.560 | But there are lots of other use cases where it seems very responsible and safe
00:08:39.640 | to use gen 2 AI.
00:08:40.880 | Do you think improvements on error rates will increase the use cases?
00:08:46.760 | I mean, right now, maybe we'll never get to a point or not in the foreseeable
00:08:50.800 | future where you want the AI doctor to directly prescribe, but are there other
00:08:56.160 | cases that are not optimal now because we're still figuring out error rates that
00:09:01.960 | will become more usable over time?
00:09:05.320 | Yeah, it's been exciting to see how AI technology improves month over month.
00:09:11.240 | So I think today we have much better tools for guarding against hallucinations
00:09:15.840 | compared to, say, six months ago.
00:09:18.240 | But just one example, if you ask the AI to use retrieve augmented generation, so
00:09:23.760 | don't just generate text, but ground it in a specific trusted article and give a
00:09:28.400 | citation that reduces hallucinations.
00:09:31.120 | And then further, if AI generates something, you really want it to be right.
00:09:34.840 | It turns out you can ask the AI to check his own work.
00:09:37.440 | Dear AI, look at this thing you just wrote.
00:09:39.560 | Look at this trusted source.
00:09:40.920 | Read both carefully and tell me if everything is justified based on the
00:09:44.600 | trusted source.
00:09:45.680 | And this won't squash hallucinations completely to zero, but it won't massively
00:09:49.920 | squash it compared to if you ask AI to just say what it had on his mind.
00:09:55.600 | So I think hallucinations is-- it is an issue, but I think it's not as bad an issue
00:10:01.440 | as people fear it to be right now.
00:10:05.360 | You've been involved in AI for decades.
00:10:10.080 | And the technology has been through lots of multiple hype cycles and declines and
00:10:16.480 | winters, AI winters.
00:10:18.680 | What do you think is different about this moment, the last 15 months or so,
00:10:23.160 | since the-- of the boom of generative AI?
00:10:27.080 | Is this more lasting?
00:10:29.760 | So I think-- so I feel like compared to 10, 15 years ago, we've not really had
00:10:38.760 | another AI winter.
00:10:40.640 | I think it's been growing in value.
00:10:42.840 | So today, years back, I used to lead the Google Brain team, which seemed to help
00:10:46.920 | Google adopt deep learning.
00:10:49.400 | And the economics-- fundamental economics are very strong.
00:10:53.960 | I'm using deep learning to drive online advertising.
00:10:57.560 | Maybe not the most inspiring thing I've worked on, but the economic fundamentals
00:11:01.720 | have been really strong for 10-ish plus years now.
00:11:07.080 | And I feel like the economic-- the fundamentals for generative AI also feel
00:11:12.920 | quite strong in the sense that we can ultimately augment a lot of tasks and
00:11:18.240 | drive a lot of very fundamental business efficiency.
00:11:21.920 | Now, there is one question.
00:11:23.440 | I think Sequoia posted an interesting article asking, over the last year or last
00:11:29.000 | year, we collectively invested, I don't know, maybe something like $50 billion in
00:11:33.400 | capital infrastructure. We are buying GPUs and data centers.
00:11:37.000 | And I think we better figure out the applications to make sure that pays off.
00:11:41.120 | So I don't think we overinvested, but to me, whenever there's a new wave of
00:11:49.240 | technology, almost all of the attention is on the technology layer.
00:11:52.680 | So we all want to talk about what Google and OpenAI and Microsoft and Amazon and
00:11:57.800 | so on are doing, because it's fun to talk about the technology.
00:12:00.040 | There's nothing wrong with that.
00:12:01.480 | But it turns out that for every wave of technology, for the two builders, like
00:12:06.440 | these companies, to be successful, there's another layer that had better be even more
00:12:10.680 | successful, which is the applications you build on top of these tools.
00:12:14.640 | Because the applications that better generate even more revenue so that they
00:12:18.440 | can afford to pay the two builders.
00:12:20.960 | And for whatever reason, society of interest or whatever, the applications tend
00:12:25.120 | to get less interest than the two builders.
00:12:30.320 | But I think for many of you, in organizations where you are not trying to be
00:12:35.240 | the next large language model, foundation model provider, I think that as we look
00:12:41.520 | into the many years in the future, there would be more revenue generated, at least
00:12:45.760 | there better be, in the applications that you might build than just in the two
00:12:50.480 | builders.
00:12:51.640 | That gets to a question that I find fascinating about this.
00:12:55.000 | What is the effect on the power dynamics in the tech industry and the economy more
00:12:59.720 | broadly?
00:13:00.720 | And to what degree is this a technology that is a disruptive technology that is
00:13:06.680 | ushering in a new wave of companies that 10 years from now will be big, and even
00:13:12.360 | though we hadn't heard of them two years ago, to what degree is it just going to
00:13:15.440 | make Microsoft and Amazon and Google, et cetera, more powerful than they've ever
00:13:21.520 | been before?
00:13:23.080 | So I think the cloud businesses are decently positioned, because it turns out
00:13:31.440 | that AWS is your GCP.
00:13:34.560 | Those are beautiful businesses.
00:13:36.400 | They generate so much efficiency that even though I may have a huge bill, I need
00:13:41.080 | to pay them, I don't mind paying it because it's much better than the
00:13:44.280 | alternative most of the time.
00:13:45.840 | But they also are very profitable businesses.
00:13:49.480 | And it turns out that if you look at some of the generative AI startups today, the
00:13:53.720 | switching costs of my using, you know, one startup's API versus switching into AWS
00:14:00.680 | or zero or GCP, Google Cloud, the switching costs are actually still quite low.
00:14:06.440 | So the moat of a lot of the generative AI startups, I'm not quite sure how strong
00:14:12.120 | their moat is.
00:14:13.360 | But in terms of the cloud businesses have very high surface area.
00:14:16.840 | I mean, you know, frankly, once you build a deep tech stack on one of the clouds, it's
00:14:20.600 | really painful to move off that if you didn't, you know, design for multi-cloud
00:14:24.480 | from day one or whatever, which is part of what makes the cloud business such a
00:14:28.440 | beautiful business model.
00:14:30.720 | So I think a lot of the cloud businesses will do OK selling API calls and
00:14:35.920 | integrating this with the rest of their existing cloud offerings.
00:14:39.720 | And the market dynamics are very interesting, right?
00:14:41.760 | So Meta has been a really interesting kind of spoiler for some other businesses by
00:14:46.640 | releasing open source software, open-source generative AI software.
00:14:50.520 | And I think Meta, you know, it was my former team, Google Brain, that released
00:14:56.240 | TensorFlow.
00:14:57.360 | And I think that it would make logical sense for Meta.
00:15:01.720 | So Meta was really, you know, hurt by having to build on Android and iOS
00:15:08.040 | platforms, right?
00:15:08.880 | When Apple changed their privacy policies, that really damaged Meta's business.
00:15:13.200 | So, you know, kind of makes logical sense that Meta would be worried if Google
00:15:19.520 | Brain, my old team, released the dominant AI-developed platform, TensorFlow, and
00:15:25.400 | everyone had to build on TensorFlow, what are the implications on Meta's business?
00:15:28.360 | So frankly, Meta played its hand beautifully with open-source PyTorch as an
00:15:33.360 | alternative.
00:15:34.480 | I think again today, Genzware is very valuable for online advertising and also,
00:15:39.320 | you know, user engagement.
00:15:40.680 | And so it actually makes a very strong logical sense that Meta would be quite
00:15:45.800 | happy to have an open-source platform to build on, to make sure it's not locked
00:15:49.560 | into like an iOS-like platform in the Gen. AI era.
00:15:53.200 | Fortunately, the good news is for almost all of us, Meta's work and many other
00:15:57.880 | parties' work on open-sourcing AI gives all of us, you know, free tools to build
00:16:04.120 | on top of.
00:16:04.720 | It gives us those building blocks that lets us innovate cheaply and build
00:16:08.400 | maybe more exciting applications on.
00:16:10.640 | Sorry, not sure if there was two inside the baseball on, you know, tech comfy
00:16:14.960 | market.
00:16:15.320 | No, I think, I mean, I wouldn't answer for everybody out there, but I thought it was
00:16:18.840 | fascinating.
00:16:19.400 | And I want to come back to open-source in a minute, but from the point of view of
00:16:25.480 | CIOs and other corporate leaders across the economy, I think there are lots of
00:16:33.320 | options coming at you right now.
00:16:35.080 | You know, lots of people trying to sell products, lots of people saying this
00:16:38.800 | service will change your business, and part of the job is, you know, figuring out
00:16:43.720 | what's wheat, what's chaff.
00:16:44.920 | How do you -- do you have any advice on how to tell in a moment where the technology
00:16:49.960 | is fairly nascent and fast developing how to tell apart the sort of people who have
00:16:54.760 | real solutions from the snake oil salesman?
00:16:57.520 | You know, I'll tell you the thing that I think is tough.
00:17:00.000 | Even our VC friends right here on San Jo-- some of them on San Jo Road, the one thing
00:17:04.920 | that's still quite tough is the technical judgment, because AI is evolving so quickly.
00:17:10.360 | So I've seen, you know, really good investors here on San Jo Road.
00:17:14.160 | They'll be pitched on some startup, and sometimes, you know, someone, let's say,
00:17:18.680 | open AI or someone just released a new API, and the startup built something really
00:17:22.920 | cool over a weekend on top of a new API that someone just released.
00:17:26.360 | But unless you know, you know, about that new capability and what the startup really
00:17:30.960 | did, I've seen VCs come to me and say, "Wow, Andrew, this is so exciting.
00:17:35.320 | Look, these three, you know, college students built this thing.
00:17:38.480 | This is amazing.
00:17:39.480 | I want to fund it."
00:17:40.480 | And I'll go, "No, I just saw 50 other startups doing the exact same thing."
00:17:44.840 | And so I think that technical judgment, because the tech is evolving so quickly, that's the
00:17:48.760 | one thing that I find difficult.
00:17:52.120 | And then maybe I should say, for any of the work of corporate startups, we tend to work
00:17:56.240 | with corporates to go through a systematic brainstorming process, but I'll just mention
00:17:59.600 | one other thing that I think could probably in many CIOs' interests, which is we've all
00:18:04.000 | seen, when we buy a new solution, you know, often we end up creating another data silo
00:18:09.760 | within our organization.
00:18:12.120 | And I feel like if we're able to work with vendors that, you know, let us continue to
00:18:19.600 | have full access to our data in a reasonably interchangeable format, that significantly
00:18:26.000 | reduces vendor lock-in, so that if one vendor -- you decide to swap out for a different
00:18:30.840 | vendor in a month or two.
00:18:32.520 | So that's one thing I tend to pay heavy attention to myself, is if I buy it from a vendor, don't
00:18:39.080 | do stuff with my data, because I want them to, that's what I'm paying them for.
00:18:42.620 | But is there that transparency and interoperability to make sure that I control my own data and
00:18:47.680 | the ability to swap -- to my own team take a look at it or swap out for a different vendor?
00:18:52.160 | This does run counter, you know, to the interests of all the vendors that want right lock-in
00:18:56.640 | candidly, but this is one thing I tend to rate higher than, you know, some of my colleagues
00:19:02.960 | in my vendor selection and buying process maybe.
00:19:06.640 | >> It sounds like you see a world where the folks in this room are implementing AI, generative
00:19:13.360 | AI, through a multiple -- multiplicity of different providers.
00:19:19.400 | It's not going to be like, yeah, we're with Microsoft.
00:19:22.960 | It's going to be, yeah, we use Microsoft for this, we use this company over here for that.
00:19:26.920 | Is that right?
00:19:27.920 | >> Yeah.
00:19:28.920 | I would say it seems like the Microsoft sales reps -- oh, I'm actually -- well, should we
00:19:34.880 | do a poll?
00:19:35.880 | How many of you had that Microsoft sales reps push co-pilots really hard to you?
00:19:40.000 | Yeah, I thought so.
00:19:41.000 | I forgot everyone.
00:19:42.000 | Right?
00:19:43.000 | So Microsoft is great.
00:19:44.000 | You know, love the team, really capable, co-pilots can give a part of the reviews, but there's
00:19:47.840 | so much stuff out in the market.
00:19:49.480 | I think it's worthwhile to, you know, take a look at multiple options and buy the right
00:19:55.160 | two for the right job.
00:19:57.960 | >> You touched on the hardware costs, the amount that's been invested so far.
00:20:03.360 | How concerned are you about the hardware bottleneck and the lack of GPUs, TPUs, you know, whatever
00:20:09.680 | one wants to call it, and, you know, Nvidia's relative strength over the last year or two?
00:20:17.560 | And what do you think of what we reported as Sam Altman's plans to raise potentially trillions
00:20:24.480 | of dollars to solve this?
00:20:26.200 | >> Yeah.
00:20:27.200 | It would be -- Sam was my student at Stanford way back, so I've known him for years.
00:20:33.240 | He's a smart guy.
00:20:34.240 | I can't argue results.
00:20:35.400 | I don't know where we find trillions -- $7 trillion in the Washington County.
00:20:41.080 | >> Up to you.
00:20:42.080 | >> That lets you buy Apple twice, right, more than twice at this point, so it's an interesting
00:20:46.200 | figure to try to raise.
00:20:48.840 | I think that over the next year, I think in a year or so, the semiconductor shortage,
00:20:54.920 | I think will feel much better, and I want to give, you know, AMD credit, AMD and Intel
00:21:01.960 | maybe.
00:21:02.960 | So Nvidia's been -- Nvidia's -- one of Nvidia's mode has been the CUDA programming language,
00:21:10.040 | but AMD's open-source alternative called ROCM has been making really good progress, and so
00:21:15.840 | some of my teams, you know, would build stuff on AMD hardware, and sometimes -- I don't
00:21:21.600 | think it's not at parity, but it's also so much better than a year ago.
00:21:26.400 | So I think AMD is worth a careful look at, and Intel Gaudi is also, you know -- so we'll
00:21:33.360 | see how the market evolves.
00:21:37.320 | >> You mentioned open-source several times.
00:21:38.720 | I know you're a champion of that.
00:21:42.120 | It comes up in the regulatory discussion, and I think one argument that resonates is,
00:21:47.760 | well, if we have, you know, at least if we have these proprietary models, there's a handful
00:21:52.360 | of companies with these big powerful LLMs that we can focus on, make sure they're doing
00:21:58.360 | the right thing to prevent this technology from being misused.
00:22:01.800 | Open-source proliferates, and you're talking about not five or 10, but 500 or 1,000 or
00:22:07.560 | even larger numbers of people who have these tools, and, you know, how do you know what
00:22:11.240 | they're going to do with it, and how do you control it?
00:22:13.880 | What's your answer to the people who have that concern about open-source?
00:22:16.600 | >> Yeah.
00:22:17.600 | So I think over the last year or so, there's been intense lobbying efforts by a number
00:22:24.360 | of -- usually the bigger players in generative AI that would rather not have to compete with
00:22:30.560 | open-source.
00:22:31.560 | You know, they've invested hundreds of millions of dollars, right, to build a proprietary
00:22:34.880 | AI model.
00:22:35.880 | Boy, isn't it annoying if someone open-sources something similar for free?
00:22:40.520 | Just it's not good, so the level of intensity and lobbying in DC, it really took me by surprise.
00:22:48.920 | And so the main argument has been AI is dangerous, maybe you can wipe out humanity, sort of put
00:22:55.160 | in place your regulatory licensing requirements before you build AI, you need to report to
00:23:01.200 | the government and maybe even get a license and prove a save and basically put in place
00:23:06.240 | these really heavy regulatory burdens, in my opinion, in a false name of safety that
00:23:13.480 | I think would really risk squashing open-source.
00:23:16.360 | It turns out that if these lobbying efforts succeed, I think almost all of us in this
00:23:20.360 | room will be losers, and there will be a very small number of, you know, people that will
00:23:24.800 | benefit from this.
00:23:27.400 | So there's actually a large community here in Silicon Valley that's been -- and around
00:23:31.560 | the world that's been actively pushing back against this narrative.
00:23:35.120 | I think to all of you having the ability to open-source components to build on that lets
00:23:39.960 | you control your own stack, it means that some vendor can't deprecate one version, this has
00:23:45.040 | happened, right?
00:23:46.040 | And then you have your whole software stack needs to be re-architected and so on.
00:23:50.640 | And then to answer the safety thing, I feel like, you know, to me at the heart of it is,
00:23:58.080 | do we want more or less intelligence in the world?
00:24:02.200 | So recently, our primary source of intelligence has been human intelligence.
00:24:06.840 | Now we also have artificial intelligence or machine intelligence.
00:24:10.440 | And yes, intelligence can be used for nefarious purposes, but I think a lot of civilizations'
00:24:15.840 | progress has been through people getting training and getting smarter and getting more intelligent.
00:24:20.120 | And I think we're actually all much better off with more, rather than less intelligence
00:24:24.160 | in the world.
00:24:25.160 | So I think open-source is a very positive contribution.
00:24:28.080 | And then lastly, as far as I can tell, a lot of the fears of harm and affairs actors, it's
00:24:33.000 | not that there are no negative use cases.
00:24:35.000 | There are a few, but when I look at it, I think a lot of the fears have been overblown
00:24:41.520 | relative to the actual risk.
00:24:43.720 | So I want to get to questions in a moment, but just to follow up on that, I interviewed
00:24:48.440 | you something like seven years ago at a WHO conference and asked you about the singularity
00:24:54.000 | and those concerns, and I think you said that worrying about evil AI robots is equivalent
00:25:00.960 | to worrying about overpopulation on Mars.
00:25:03.760 | We're not even there yet.
00:25:05.480 | Are we on Mars yet in this metaphor?
00:25:09.800 | Where are we in that progress?
00:25:10.800 | Yeah.
00:25:11.800 | At this point in time, so I feel like that super-intelligence singularity is much more
00:25:15.760 | science fiction than anything that any of us AI builders know how to build.
00:25:18.760 | So I still feel that way.
00:25:20.720 | You were saying that you've seen less of that type of talk, like you were just at Davos.
00:25:26.400 | In the regulatory discussions, there's less of this like, "Oh my God, we've got to stop
00:25:30.640 | this.
00:25:31.640 | We're building this thing that's so amazing, it might take over humanity," is not as much
00:25:37.440 | part of the discussion now?
00:25:38.600 | It's really dying down.
00:25:40.800 | Last May, there was a statement signed by a bunch of people that I think made an analogy
00:25:44.640 | between AI and nuclear weapons without justification.
00:25:47.240 | AI brings intelligence to the world, nuclear weapons blows up cities.
00:25:50.760 | I don't know why they have anything to do with each other, but that analogy just created
00:25:54.800 | a lot of hype.
00:25:56.200 | Fortunately, since May, that degree of alarm, when I speak of people in the US government
00:26:02.400 | about AI human extension, I'm very happy to see eye rolls at this point.
00:26:09.320 | I think Europe takes it a little more seriously than the US, but I just see the tenure dying
00:26:15.160 | down to talk more about concrete homes.
00:26:18.240 | We want self-driving cars to be safe, we want medical devices to be safe.
00:26:21.200 | Instead of worrying about the AI tech, let's look at the concrete applications.
00:26:25.520 | Because when we look at general purpose technology like AI, it's hard to regulate that without
00:26:29.840 | just slowing everything down, but when we look at the concrete applications, we can
00:26:33.920 | say what do and don't we want in financial services?
00:26:37.440 | What is fair and what is unfair in underwriting?
00:26:41.040 | What standards should medical devices meet?
00:26:43.960 | With regulations, the application layer would be a very positive thing to even unlock innovation,
00:26:49.120 | but these big fears say intelligence is dangerous and AI is dangerous.
00:26:53.680 | That just tends to lead to regulatory capture and lobbyists having very strange agendas.
00:26:57.720 | Do we have any questions in the audience?
00:26:59.640 | I see some down here, can we get a microphone?
00:27:06.520 | The gentleman and then the lady.
00:27:09.840 | Hi, thank you for all you do for this community.
00:27:13.320 | I think your online courses are amazing.
00:27:18.680 | All innovation follows some kind of an S-curve and we're in this rapid acceleration of innovation
00:27:25.960 | around generative AI and machine learning.
00:27:29.800 | Where do you think the plateau is and what are the rate limiters to drive us towards
00:27:34.120 | the plateau?
00:27:36.120 | How much farther can this be pushed before we start to see ourselves hitting a plateau
00:27:40.520 | and what's going to limit that?
00:27:42.600 | Yeah, so I think large language models, they are getting harder and harder to scale.
00:27:47.880 | I think there is still more juice in that onion, but the exciting thing is the core
00:27:52.600 | innovation of large language models, we're now stacking other S-curves on top of the
00:27:56.800 | first one.
00:27:57.800 | So even the first S-curve's plateaus, I'm excited, for example, about edge AI or on
00:28:02.440 | device AI.
00:28:03.440 | I run an L on my laptop all the time.
00:28:05.160 | If you don't yet, it's easier than you might think and keeps all your data confidential
00:28:09.000 | with open source AI.
00:28:10.000 | You can run on your laptop.
00:28:12.280 | I said about agents.
00:28:13.280 | Instead of you prompting AI, it responds in a few seconds.
00:28:16.520 | We now see AI systems where I can tell it, dear AI, please do research for me and write
00:28:21.000 | a report.
00:28:22.000 | It goes off for half an hour and brows the Internet, summarizes all things, comes back
00:28:25.880 | in half an hour with a report.
00:28:27.840 | This is kind of, you know, it's not working as well as I just distracted, but it's working
00:28:31.840 | much better now than three months ago, so I'm excited about AI autonomous agents, goes
00:28:36.240 | and works for an extended period of time.
00:28:39.280 | We saw the unlock of text processing with large language models, with large vision models,
00:28:45.720 | which at a much earlier stage of development, I think we're trying to see a revolution in
00:28:51.240 | image processing in the same way that we saw a revolution in text processing.
00:28:55.320 | So these are some of the other S-curves being stacked up on top, and then some are even
00:28:59.640 | further out, so I'm not seeing an overall plateau in AI yet.
00:29:04.720 | Maybe there'll be one, but I'm not seeing it yet.
00:29:07.000 | >> Do you have a very quick question?
00:29:09.520 | >> Yeah.
00:29:10.520 | >> Okay.
00:29:11.520 | Thank you.
00:29:12.520 | >> Thank you.
00:29:13.520 | It's a great dialogue, and our sophomore at Berkeley spends more time watching your videos
00:29:15.800 | than taking courses, so thank you again.
00:29:18.840 | So you mentioned automating tasks and also human intelligence.
00:29:23.800 | The knowledge of the task are still owned by the humans.
00:29:26.920 | In your dialogues with clients, are you seeing resistance to unpack the tasks that humans
00:29:32.160 | do accurately so that you can apply AI to it?
00:29:35.080 | And if you are seeing resistance, what is the solve for that?
00:29:39.680 | >> Let's see.
00:29:41.480 | So I feel like -- I find that when we have a realistic conversation -- so let's see,
00:29:47.240 | when we work with corporations -- so AI, we often work with corporations to brainstorm
00:29:51.200 | project ideas and figure out what can we help build.
00:29:53.960 | That's actually -- as an AI person, I learned that my slim line is AI, but all of these
00:29:58.760 | exciting businesses apply to it that I just don't know anything about, so a core part
00:30:02.520 | of what our strategy is to work with large corporate partners that are much smarter than
00:30:07.640 | I am about the business domains to apply to.
00:30:11.360 | So what I'm finding is that at the executive level, which probably, you know, who we work
00:30:15.440 | with the most day-to-day, there's not resistance at all.
00:30:19.160 | There's just enthusiasm.
00:30:22.240 | Maybe one unlock that I found is I teach a class on Coursera, a janitor AI for everyone.
00:30:28.600 | It was the fastest-growing course on Coursera last year.
00:30:31.560 | But I did that to try to give business leaders and others a non-technical understanding of
00:30:36.760 | AI and what they can and cannot do, and we found that when some of our partners take
00:30:40.360 | gen-2 AI for everyone, you know, that non-technical understanding of AI unlocks a lot of brainstorming
00:30:45.440 | ideation.
00:30:46.560 | So that's the executive level, kind of learn about what gen-2 AI, brainstorming, execute,
00:30:51.120 | lots of exciting projects.
00:30:52.880 | And then many businesses are sensitive to the, you know, broader employee-based concerns
00:30:58.560 | about job loss.
00:31:00.600 | And I find that when we have a really candid conversation, the fears usually go down.
00:31:07.600 | I don't want to pretend there's zero job loss.
00:31:10.920 | That's just not true.
00:31:11.920 | But when we do the task-based analysis of jobs, you know, pay if AI automates 20 percent
00:31:17.160 | of my job, to a lot of people, that's great.
00:31:20.620 | I can be more productive, focus more on the other 80 percent of tasks.
00:31:24.200 | So on average, once we have that more candid conversation, you know, I'm thinking of this
00:31:29.840 | one time the union stopped us from even installing one camera.
00:31:33.040 | So there are some of that, but most of the time, it's a pretty rational and okay conversation.
00:31:38.320 | [BLANK_AUDIO]