Speaker A
I could definitely see a future, very scary future in which a lot of people would not have meaningful work.
MIT economist Daron Acemoglu, 2024 Nobel laureate, discusses AI's uneven productivity impact, why coding automates fastest, and why customer service and healthcare lag.
Key Takeaways
- AI's economic impact so far is modest relative to its raw capabilities, and automation is slower than popular narratives suggest.
- Tasks with verifiable outcomes, like coding, are automating fastest, while personalized and social tasks lag.
- The best path is to build AI applications that complement humans and create new productive work rather than only cutting labor costs.
- Individuals should invest in strong fundamentals and adaptable skills to remain valuable.
What the video covers
- Acemoglu warns of a future where many people lack meaningful work if automation displaces jobs without creating new high-productivity roles.
- He revisits his 2024 estimate that AI adds about 1% to US GDP over 10 years with 5% of tasks automatable, saying he would now put the figure somewhat higher but still considers automation slow and hard.
- Coding advances fastest because it is verifiable: code either runs or doesn't, enabling reinforcement learning with a clear ground truth.
- Customer service is harder to automate because queries are personalized, contain gray areas and edge cases, and require human oversight or collaboration.
- Diffusion of new technology is slow, especially where it requires complementary investments and organizational adjustments.
- Measuring AI's economic impact is difficult, and current capabilities are advancing faster than observed economic effects.
- Healthcare offers opportunities for AI to make nurses more involved in diagnosis and care, but quality control and information barriers persist.
- Acemoglu argues the focus should shift from pure automation toward applications that create new tasks and opportunities for workers.
- He advises people to build strong foundations and study fundamentals, since human creativity depends on them.
- He criticizes media coverage of AI for weak information debate and discusses the risk of AI going badly.
Chapters
- 00:00Introduction: Fears of Meaningless Work
- 02:30Revisiting the 2024 GDP and Automation Forecast
- 05:45Why Coding Is Automating Fastest
- 08:44Why Customer Service Automation Is Hard
- 13:20Productivity, Measurement, and Healthcare Opportunities
- 20:00Automation vs. New Applications and Worker Skills
- 28:20Risks, Media Coverage, and Closing Thoughts
Full Transcript — Download SRT & Markdown
Speaker A
This is Darren Asimoglu, an MIT economist and winner of the 2024 Nobel Prize [music] in Economics. He thinks we're overlooking some of AI's biggest opportunities. So, where should we be looking and what will keep us valuable?
Speaker A
This is Daron Acemoglu, an MIT economist and winner of the 2024 Nobel Prize in Economics. He thinks we're overlooking some of AI's biggest opportunities. So, where should we be looking, and what will keep us valuable?
Speaker A
productivity jobs to create new opportunities for people. So, what should people be doing? What skills should they be acquiring?
Speaker A
I've seen a lot of posts where people say, "Oh, instead of hiring someone, I deployed an agent." Automation is great. The problem is the only thing we do is automation.
Speaker A
So, you said something in 2024 that AI would add about 1% to US GDP in the next 10 years with only 5% of tasks being fully automatable. What do you think about this now in 2026?
Speaker A
That will eliminate too many jobs without creating meaningful, high-paying, high-productivity jobs to create new opportunities for people. So, what should people be doing? What skills should they be acquiring?
Speaker A
to bring a little bit of basic economic calculus to the discussions of productivity. And this was more than two years ago.
Speaker A
If everybody wants my opinion, I would say go and study. Thanks to HubSpot for sponsoring this video.
Speaker A
Probably now I would have put that number a little higher than 5%. But I still think that a lot of the discussion exaggerates how quickly we're going to see automation because automation is hard.
Speaker A
So, you said something in 2024 that AI would add about 1% to US GDP in the next 10 years, with only 5% of tasks being fully automatable. What do you think about this now, in 2026?
Speaker A
to be grounded in thinking about how quickly and how uh thoroughly the economy is going to change. And we still still have the same discussion that we should have about every technology. What's the best way of using it? How
Speaker A
Well, first of all, it's my pleasure to be here, and I do not claim that I have any unique insight into AI. AI is like an elephant. Everybody is feeling a different part of it. But I think I try to bring a little bit of basic economic calculus to the discussions of productivity. And this was more than two years ago.
Speaker A
of AI, I think the advances in coding have been really really rapid. What is it about coding that makes it so conducive to have the most rapid advances of any occupation?
Speaker A
There's a huge amount of uncertainty, and there is even more today. But there have also been some very rapid developments, especially this year, with agentic AI.
Speaker A
this is a very important part just like math just like chess coding has rules something either runs or doesn't run reinforcement learning is very effective when you have a ground truth that says no that's wrong I think in many
Speaker A
Probably now I would have put that number a little higher than 5%. But I still think that a lot of the discussion exaggerates how quickly we're going to see automation, because automation is hard.
Speaker A
difficult and most of the organizations that I deal with their AI system is so bad it's like a nightmare when I call customer service and I get AI oh my god it's going to be another 10 minutes before somebody can actually help me and
Speaker A
Integrating automation into organizations is hard. Diffusion of new technologies, especially when they require lots of complementary investments and adjustments, is hard and slow. That doesn't mean it's not a transformative technology. It doesn't mean that we cannot get benefits from it. But we have to be grounded in thinking about how quickly and how thoroughly the economy is going to change. And we still have the same discussion that we should have about every technology. What's the best way of using it? How could it be misused, and who will benefit from it?
Speaker A
cases. So that makes it much messier and it will require humans to either either oversee it or work with it or ideally a much better collaboration so that AI could be complimentary to humans. So all of those are in play. Which ones we
Speaker A
You mentioned agentic. Is there anything you're when you see it, you'll think okay, now I can expand my figure to, like, two digits?
Speaker A
With coding, we have deterministic outcome that we can check easily with customer service. It's much more personalized as you mentioned. Do you think in the future with stronger models we'll be able to achieve it or it's just the bottleneck of AI in general that it
Speaker A
Uh, when it comes to impact of AI, I think the advances in coding have been really, really rapid. What is it about coding that makes it so conducive to have the most rapid advances of any occupation?
Speaker A
data can make progress on many things. But the question is how fast and [snorts] can it do so across all the domains. If if I step back and I say if you ask me why so little productivity improvements and what can we do about it
Speaker A
I think it's the fact that coding is obviously done in offices. You don't have any interaction with the real world. The people who are supervising coding are very, very fluent in AI and in technology in general, but also, and this is a very important part, just like math, just like chess, coding has rules. Something either runs or doesn't run. Reinforcement learning is very effective when you have a ground truth that says no, that's wrong. I think in many occupations we lack that.
Speaker A
this this may be incorrect. Look I'm not in anthropic inside the organization. I'm not an open AI. I don't know what they are learning in real time from their experiments and the enormous amount of resources and they're very
Speaker A
Customer service, I think most people would have said customer service is going to be the first occupation to be fully automated by AI, and we're seeing some of it, but it's turning out to be very slow and very difficult. Most of the organizations that I deal with, their AI system is so bad it's like a nightmare. When I call customer service and I get AI, oh my god, it's going to be another 10 minutes before somebody can actually help me.
Speaker A
organizations. So when I when you know we're we're told all the time US is ahead of China and indeed in large language models that seems quite clear.
Speaker A
That's because there's a social element. Many of the things that you ask from customer service change over time. They are person-specific. Different people use different language for asking the same questions, and there are a lot of gray areas and edge cases. So that makes it much messier, and it will require humans to either oversee it, or work with it, or, ideally, a much better collaboration so that AI could be complementary to humans.
Speaker A
they're putting a lot of emphasis in getting robotics, getting AI into the production line. And one thing that they are throwing at that problem is a huge amount of engineering talent. The United States doesn't seem to have that much
Speaker A
So all of those are in play. Which ones we choose is going to have a very important impact. How many jobs are going to be displaced? How much productivity effects are we going to get? And who's going to benefit from it?
Speaker A
as oh no that's the only way to do it I think here is one thing one other thing I'll I'll say I don't want to make you know overgeneralizations but I think as a signpost if there is one thing that's been not
Speaker A
With coding, we have a deterministic outcome that we can check easily. With customer service, it's much more personalized, as you mentioned. Do you think in the future, with stronger models, we'll be able to achieve it, or is it just the bottleneck of AI in general that it can't function without a human in the loop?
Speaker A
completely reject that there. I think AI is a very versatile technology as are all general purpose technologies, GPTs, what economists call things like electricity, uh printing press, etc. which spawn a lot of other applications. One thing we know from history, one thing we've
Speaker A
I would definitely not say there's this one thing that AI will not be able to do in the future. I think the last 10 years have given us enough examples of AI: if you put enough resources and enough data, it can make progress on many things. But the question is how fast it can do so across all the domains.
Speaker A
just thinking that whatever we're doing is the only way or by far the best way I think is a mistake. And I think another problem that comes with it is measurement, right? So how do you measure the impact of AI on your
Speaker A
If I step back and I say, if you ask me why so little productivity improvements, and what can we do about it, I would have just one word for my vision of how we can improve the economic effects of AI applications. We just need much better and more applications.
Speaker A
So the raw capabilities are skyrocketing, but the economic impact is pretty muted right now. that gap between the capabilities and what's actually happening is a big opportunity I think.
Speaker A
In some sense, one can make the argument, and again I don't know, this may be incorrect. Look, I'm not inside Anthropic. I'm not inside OpenAI. I don't know what they are learning in real time from their experiments and the enormous amount of resources, and they have very, very talented people in all of these organizations. But you can make the case that the United States, Silicon Valley, is putting too much emphasis on foundation models, and not enough on applications, and not enough on integrating applications into organizations.
Speaker A
effects in 2035, 2034, which is what I was writing about at the time, he would have put it somewhat more aggressively.
Speaker A
So when we're told all the time that the US is ahead of China, and indeed in large language models that seems quite clear. Chinese models are catching up sometimes, but they are using the lead of US models. There's some distilling going on, probably. But one thing that I see Chinese companies do that I don't see as much in the United States is they're putting a lot of emphasis on getting robotics, getting AI into the production line. And one thing that they are throwing at that problem is a huge amount of engineering talent. The United States doesn't seem to have that much engineering talent, and whatever we have is all going into foundation models.
Speaker A
than in my opinion than what Eric would say. But but I I would be very very happy if he's right and I think he's got a much better chance of being right than many other people in the industry who
Speaker A
You touched upon such an interesting subject, because as an immigrant, when I see things how they work in the US and ask why is it so, the reply is, oh, historically, or, oh no, that's the only way to do it.
Speaker A
Productivity has always been difficult to measure and there's a quality issue. I think one place where uh Eric has put a lot of emphasis rightly is that [snorts] in the digital age more broadly over the last 40 years we don't know how
Speaker A
I think here is one other thing I'll say. I don't want to make overgeneralizations, but I think, as a signpost, if there is one thing that's been not good for the US AI industry, in my opinion, it's the inevitabilism: the view that whatever we're doing is the only way to do it. There's no other way, and if you're raising that question, you're an idiot. So I completely reject that.
Speaker A
the per the quality improvements coming from uh better cars or uh better highways either. So there there always been an issue of measuring quality.
Speaker A
I think AI is a very versatile technology, as are all general purpose technologies, GPTs, what economists call things like electricity, the printing press, etc., which spawn a lot of other applications. One thing we know from history, one thing we've learned from a lot of economics and looking at previous examples, is there are so many different ways of developing these general purpose technologies, and different ways are not neutral. They bring different types of benefits. They have different types of costs. Just thinking that whatever we're doing is the only way, or by far the best way, I think is a mistake.
Speaker A
It is I mean that's that's also great. I think well-being is a very important thing there. I would be more negative about AI and all the associated technologies. is I think social media has been horrendous for us. It has
Speaker A
I think another problem that comes with it is measurement, right? How do you measure the impact of AI on your personal productivity, on your company's productivity? When I talked to Erik Brynjolfsson when they were doing this study, the economic impact of AI was basically zero. We talked to him in May. So the raw capabilities are skyrocketing, but the economic impact is pretty muted right now. That gap between the capabilities and what's actually happening is a big opportunity.
Speaker A
It has made our worst instincts more salient because algorithms go after rage, emotions, envy. So, I think it has been horrible. I hope AI will not repeat social media's mistakes.
Speaker A
Let's take a quick break here because what Darren just said about technology really stood out to me. AI can be used simply to replace tasks and cut cost or it can help people become more productive and make better decisions. As
Speaker A
I see this all the time running two companies and working with a big team, every new AI tool promises to save time.
Speaker A
But adding more tools doesn't automatically make your business more efficient. If your marketing data is in one place, sales is working somewhere else and customer information is spread across different systems, your team ends up spending more time managing the tools
Speaker A
instead of actually using them. That's what I like about HubSpot. It brings marketing, sales, customer service, and content together on one AI powered customer platform. Your teams can work from the same customer information at every stage of the journey instead of
Speaker A
constantly switching between disconnected systems. Each HubSpot product is powerful on its own, but they work even better together. So whether you're building a company or trying to make your existing team more effective, you can start with the tools you need
Speaker A
and grow from there. You can explore HubSpot and get started for free through the link in the description. Thanks to HubSpot for sponsoring this video. Let's talk about workers sharing gains of this new technology.
Speaker A
My favorite topic. Yeah, you're particularly vocal about it. So when you say we have these strong models but we lack applications, is that the education problem? Do you think we still need a stronger model or what do we need to see the impact of AI on our
Speaker A
day-to-day positive impact? I think existing models are already very capable as a foundation upon which to build applications.
Speaker A
There are so many things we can do that would be quite feasible with the existing infrastructure. [snorts] Let me mention a few examples.
Speaker A
Much better quality control very easy with existing models. Tools for nurses. We have a real health care crisis in this country. We spend almost one out of every $5 on healthare. And look at the outcomes. There are real public health
Speaker A
challenges. There are real bottlenecks in the healthcare industry. Imagine nurses can be more involved in the care process in diagnosis, prescription, decisions about patient well-being. And one big barrier for that is quality control and information. those can be
Speaker A
very very easily supplied by existing AI models with the right applications built on top of our current uh models. Yeah, I think you probably need to train nurses a little differently. So there's a human capital aspect that's going to go complimentary to that. But
Speaker A
the bottleneck isn't the human capital of the nurses. The bottleneck is the application. The bottleneck is the tech industry not doing it. Teachers, same thing. electricians. We can give so much better tools to electricians. It's just like a such an easy lowhanging fruit.
Speaker A
Many blue color tasks can be made much more effective with AI. We're seeing some of them in China. Again, why are we not doing that in the United States? So, this is the bottleneck isn't the foundation models. The bottleneck I mean
Speaker A
even last year's models would have been good enough for most of the things that we could do.
Speaker A
A lot of it sounds like a great business opportunity. It is. It is AI for electricians.
Speaker A
Electric. This is such a huge market. Again, discovered after I moved here, how much those people are making, how much you pay uh for a small and there's such a shortage of electricians.
Speaker A
Oh, yeah. And uh you wait three 3 weeks for the most basic task and then electrician comes and says, "Oh, no. This new machinery, I can't deal with it." Why not? Because I've never seen it before.
Speaker A
And this is where AI comes in. This is where AI comes in. It's just such an easy problem. When it comes to teachers and healthcare, do you think it's a government problem or there's still opportunity for companies?
Speaker A
Well, I think healthcare and education is a two-way street or three-way street because most healthcare dollars have involvement from the government one way or another. Teachers too. So, so yeah, I think the government could play an important role. But for electricians,
Speaker A
blue collar workers, for customer service representatives, for scientists, that's the private sector, mostly the private sector. So I said three-way because even when you take the government out, there are two parties, the tech companies and the management of
Speaker A
the business that we're talking about. So both need to [snorts] be on board. In my own conversations with CEOs of leading companies or high level managers of leading companies, when I mention what I call proworker AI, AI that will
Speaker A
make workers more productive, they say, "This is exactly what we need. We're on board with it." Do they mean it? I don't know. I can't test because the tools are not being supplied by tech companies.
Speaker A
When I mention the same thing to people from tech companies, they either nod and turn their head or they say, "You live in a different world. We're going to create AGI. What are you talking about?" They're worried about something that's
Speaker A
way ahead versus something they can fix now. When you mentioned something for bluecollar workers and you mentioned China, what are the good examples?
Speaker A
[snorts] Several manufacturing companies I've spoken to in China are integrating both robotics and AI uh Foxcon BYD uh their other uh AE EV producers. They're all very teched up and one thing I see from their experience is that it's a slow
Speaker A
process and they require more engineers than they always think. Fortunately for them, China has an enormous supply of engineers.
Speaker A
What would be your advice to entrepreneurs who are watching who are choosing what to work on? How do they choose the right niche? Because I think from all of the podcasts I've been doing recently, everyone keeps saying me
Speaker A
riches are in the niches, especially with AI because opportunity lies in this boring business which hasn't been automated in the past 20 years. And maybe still pen and paper, but if you come in and you tackle that problem,
Speaker A
this is where the opportunity. Absolutely. And uh the only thing I would not have I would not have used the word boring. I think it's very exciting.
Speaker A
Imagine boring from outside sounds a lot more I know. [laughter] I know. But that's that's that's that's so wrong. First of all, it is pie in the sky. Nobody can even agree on the definition of AGI.
Speaker A
It's always in the future and it is frankly speaking not a future that we can fully imagine either. I mean if if you really mean by AGI and then the ASI artificial super intelligence machines could do everything [clears throat]
Speaker A
everything that humans can do better that seems quite far to me but it seems also very dystopian I mean there won't be even any jobs for the best physicists the best mathematicians the best social communicators AI will do everything
Speaker A
better so what I don't understand there's so many like inconsistencies in the public debate on AI one of them is people they say, "Oh, AGI is great. It's going to come soon, but we're going to also have lots of things for humans to
Speaker A
do." That can't be. If you mean anything serious by AGI or ASI, there can't be that many things left for humans.
Speaker A
I wanted to dig a little deeper into the practical use of AI today. So, when you mention uh that business opportunity, the researchers, how do you use AI in your daily research and what can be done?
Speaker A
I must say I am behind the frontier. I use AI for background research and I have increasingly over this year started consulting AI for criticism of uh things I write and ideas I have which is very interesting because when I
Speaker A
first experimented with this with chat chat GPT 3.5 and then chat GPT4 I thought it was very disappointing and now it's so much better. I also encourage all of my students and research assistants to set up an AI
Speaker A
infrastructure for their research and uh and sort of as support for what they do, never to replace them. I discourage people uh even people who have English as a second language from using AI to write for them because I think people
Speaker A
should be in charge of their own writing. But I encourage every student who's a second not a native speaker of English to use AI as as a correction like run it through AI and see where your sentences don't make sense or are
Speaker A
too long or are non- grammatical. So I think there's a lot of mundane uses of AI already as very very good and there are lots of things but I have I very much worry about a different kind of
Speaker A
loss of control which is that we collectively but especially as students and learners delegate too much to AI so that we don't flex our mental muscles and we don't engage in learning. That is I think a first order issue. I see that
Speaker A
a lot with students from studies that I think the use of AI has been pretty disastrous in K through 12 education because we've done it blindly without guardrails. I said that actually in 2024 also or 2023 that it was a dangerous
Speaker A
huge social experiment to unleash generative AI on schools without any guard rails all at the same time.
Speaker A
because we didn't know what it was going to do and I don't think it has my worst fears have come to pass but it's been pretty close to that. When do you need to say to yourself, okay, I'm going to
Speaker A
stop here and I need to think by myself when it comes to this issue and just use AI maybe to yeah I mean because for me it's like different first of all I'm from a different generation unfortunately and
Speaker A
second I just love thinking I love taking pen and paper and sketching things out so in fact I have to prompt myself run this through AI I have never used for example AI to do math I should I mean AI is much better than
Speaker A
me in math. But I like doing the math myself. When I write down models or when I do calculations, I do them by hand because that's what I'm used to and I can do it fast and I can learn
Speaker A
something. I learned something from my mistakes. But I'm sure if I step back, I could do more.
Speaker A
When you do math by yourself, do you ever think about this? It was the calculator problem because teachers were protesting against the calculator. Now we still use it everywhere and it sped up the complexity of things that we can
Speaker A
solve. Do you think AI is going to enable us to do that or because it's taking away in general our critical thinking we won't be able to progress further?
Speaker A
I worry about the latter. I think the analogy to the calculator is useful but you have to do it carefully. The calculator could only take over some very very very small part of it. And I know of nobody who used the calculator
Speaker A
to sort of do thinking for them. And I know of no school ever that said to students, "Now you have a calculator, you don't learn how to multiply." I mean, there were calculators when I was a student.
Speaker A
We all learned how to multiply, how to divide. And I can't imagine not knowing that. That is such an important building block for many other things. Same thing for differentiation. Like if you don't know calculus, how could you even take the next step?
Speaker A
Yeah. It's how far the skill goes. With calculator, it's basically just this function. with AI is the whole critical thing.
Speaker A
It's the whole critical thinking and and and the boundaries are very gray like where do you stop in using AI? How do you still [snorts] engage in the key steps of human innovation and human creativity together with AI?
Speaker A
One thing before we keep going. If you want to stay updated on all the things these founders say here, we take the use cases and prompts from every conversation and put them into my newsletter called future proof. And of
Speaker A
course, we also put my own beautiful use cases there. For example, we'll put together the seven skills that came up again and again, the ones people will need in the next few years. These are the skills that my team and I are also
Speaker A
using daily. Subscribe through the link in the description. You'll get that one. And you can also read our previous newsletters and see what you missed.
Speaker A
Going back to businesses. So when you look at all those posts on X where someone's saying, you know, I haven't seen a lot of like I fired my whole team and hired agents, but I've seen a lot of
Speaker A
posts where people say, oh, instead of hiring someone, I deployed an agent. What do these companies and people miss in the long term when they do that?
Speaker A
Because I know you're against marketing agents as a replacement for a person. Let me actually clarify my stance on this. And this is based, you know, on work I've done on robots, [snorts] work I've done on early 20th
Speaker A
century technology, work I've done on the industrial revolution, and theoretical work I've done. I am not against automation, nor do I think it's possible to stop automation. We want automation and anybody who wants to reverse automation, I'm against that. We
Speaker A
don't want to go back to the days where we carry heavy things on our backs, automation is great. The problem is if we the only thing we do is automation.
Speaker A
A, we're not going to get the full productivity because it's not that easy to do everything with automation. and B, we're not going to create shared prosperity because that will eliminate too many jobs without creating meaningful highpaying high productivity
Speaker A
jobs to create new opportunities for people. So my point has always been use technology both for automation and other things have a balanced approach and digital technologies I have argued have provided a lot of opportunities for generating new things for humans to do
Speaker A
and it's doubly true with AI. So I want AI to be used for automation so long as we also use AI for creating new things like for nurses, for electricians, for many occupations we can't even dream about right now.
Speaker A
How do you explain it to a small business owner whose only metric is uh increasing profits? What is the economic downturn?
Speaker A
Productivity. Productivity I think again and in my experience again it may be cheap talk. People may [snorts] be nice to me when I talk to them or I might have fooled them but most businesses actually get that. Not everybody, but
Speaker A
most of them get that your workers are your greatest asset. Imagine making them more productive. That's fantastic for the business. And in fact, it's a better calculus. Like imagine you have 100 employees and I tell you, you can
Speaker A
automate 5% of those jobs and reduce the cost by half in those 5%. Or you can make your workers 5% more productive.
Speaker A
The latter is going to get you much more. Oh, and your competitors are going to be automating their workforce. And if you're getting rid of people then your productivity is going to eventually right. I mean it's a it's
Speaker A
it's just your workers have a lot of experience have a lot of some of them if you have been a good at pass they have commitment to the work to the to the to the business. If you can make them more
Speaker A
innovative better at problem solving better at doing new things that's the kind of business that's going to succeed. Let me ask you a question. What are the companies? Which ones are the companies from the past that you remember? And you'll give me familiar
Speaker A
answers like Ford Motor Company, General Motors, General Electric. The first comes to mind, you know, uh IBM, Microsoft.
Speaker A
What's common about all of them did something new. They didn't just cut costs and take took what others were doing. They imagined new markets, new technologies, new tasks. Henry Ford himself captured that. He said, "If I had asked the consumers what they
Speaker A
wanted, they would have told me foster horses." So that's what we need from successful companies and cutting your cost by 5%. That's not going to make you go down the in the history books.
Speaker A
When it comes to automation, do you have a rule on how do you know which task you mentioned 5% of tasks are automatable.
Speaker A
How do you know when to automate something or when it's not worth automating? If you cut your cost by 5%, GDP is typically going to go up by 5% times your share of GDP. That's not hard. The hard part is A, if you're
Speaker A
changing quality, that's a little bit more complicated. And B, actually, even if you're going to cut costs by 5% in a number of years, it's going to take you a lot of work to get there. And and what I'm seeing from a few businesses
Speaker A
that I know is that they take the automation plunge sometimes because they think it's going to be easy, sometimes because they think they're under pressure, and then they can't make it work because automation is not that easy. You have to integrate into the
Speaker A
company. Now, many people have started using the word AI has a jagged frontier. I I like that image, but it's also a code word for saying, oh, we overestimated AI. It's much harder to do it. And we thought we could lay off
Speaker A
workers because of AI. No, we can't. But if you take that really seriously, it means, oh, it's so jagged. I'm going to keep more or less the same number of workers and now I'm doing AI and it's just a milange.
Speaker A
Yeah, it's just a lot more work from businesses moving on to education. As someone who's in academia, is university still going to be worth a lot of money in 10 years?
Speaker A
I think so. But let me preface that with a few caveats. First, I have never believed and I'm even more convinced about this both for economic and social reasons that everybody should go to college. People have diverse skills. They have different
Speaker A
interests. We should be engaged in lifelong learning. But the traditional liberal arts type education is not for everyone. Nor do we want everybody to have exactly the same perspective in society. Second, a lot of what university does, I'm afraid to say, is
Speaker A
also a signal. When you go to Harvard, MIT, Stanford, you're also buying the name. And that's good for you, not necessarily that good for society. And third, every university is understanding now that they have to change their business model. They have to change how
Speaker A
they deliver education. They are in search of what types of skills they should focus on. And my very very superficial opinion is that universities understand that. But you know what? The same actually applies to high schools and middle schools and they are not.
Speaker A
They're they're even further behind the curve. One thing I've experienced in the United States and is that many students when they come to college are unprepared and that is going to get worse and that really will make colleg's business even
Speaker A
harder. I've seen that because I've you know written a textbook for first years in economics how often weekly prepared students are and how that's actually getting worse over time even before AI. how many colleges are finding it necessary to do more and
Speaker A
more remedial stuff [snorts] because high schools aren't doing well enough and the evidence we have I mean it's it's we are in early days things can change but the evidence is that AI has made learning harder in middle schools
Speaker A
and high schools and so that problem is going to get worse totally if you start using AI for to solve all the problems and then university is only going to get worse then so what should people be doing?
Speaker A
What skills should they be acquiring? And do you agree with the notion that we need even more academic knowledge to acquire those skills? Cuz I'm thinking about taste and strategic thinking.
Speaker A
Absolutely. Because I don't believe true AGI is just around the corner. I think we need human creativity more than ever.
Speaker A
And human creativity without strong foundations is an oxymoron. If anybody wants my opinion, and I'm not the most qualified to give you this opinion. I'm not a mathematician and I'm not a computer scientist, but if anybody wants my opinion, I would say go and study
Speaker A
math. Don't be put off by Astra's solving difficult math problems. We need human mathematical knowledge. We need human mathematical maturity. That's going to be so important. Go and study physics. Go and study economics. Go and study engineering. I think those are
Speaker A
going to be great. But also learn flexibility. I think we're going to need a lot of flexibility because even within the same occupation, we will often need to change what we do as the technology changes, as the social context changes.
Speaker A
So, I think those are really, really important things. But human ingenuity and human creativity, I think, are going to be critical in the next 10 years.
Speaker A
Yeah. And normalizing nonlinear career is something that resonated with me a lot. Someone who changed careers throughout our life. I think this is going to be our new normal.
Speaker A
But then two other principles that I think are very important and we have they we should impart starting from middle school. Learn how to work alongside and with AI absolutely but also even more emphasis on civic responsibility. We have lost in this
Speaker A
country and in most of the rest of the world a true sense of community and a true sense of caring for the common good and working for our collective good and this is even more important in the age
Speaker A
of AI and that starts with being informed being engaged that starts with elementary school really it starts with elementary school I I yeah I could have extended to elementary school absolutely we may have encouraged too much selfishness We may have encouraged people to draw
Speaker A
into their cocoons via social media and other things. We may have denigrated some aspects of community that are really important. That's also like a major theme of my new book on what happened to liberal democracy that we really need community to be at the
Speaker A
center of a liberal democratic society. You know, here's what I am not pessimist. I'm not an [clears throat] optimist. I'm not but I'm definitely not a pessimist either. I try to remain hopeful and two things I find hope in
Speaker A
are that humans are infinitely adaptable and second even in this polarized divided world we have very similar notions of fairness moral shared understanding on both sides of the aisle. They just need to come up more clearly in our communication. And we
Speaker A
need to see the other side not as the enemy. Even the deepest disagreements [snorts] should be situated within a context in which we recognize we're on the same boat.
Speaker A
We're talking about all of this, but then there comes Dario and says there's a 25% chance things go really, really badly with AI. I'm relatively an optimist. So I think there's a 25% chance that things go really really
Speaker A
badly and a 75% chance that things go really really well with with not much uh not much space.
Speaker A
What do you think when you hear this especially coming from someone who has to IPO soon and prove that their technology is very transformational?
Speaker A
I think if people in the leading AI labs are worried about some of the implications of AI, of course they should voice that. I welcome and I applaud their courage to speak, but I am myself much more worried about a general
Speaker A
sense of loss of control of society over AI than existential risk as articulated by Dario or by Elon Musk or uh by Claxon. Loss of control I'm talking about is that the general public, the workers, students, parents lose control
Speaker A
of their own agency because of AI. Not necessarily because AI has become super intelligent and an entity onto itself but partly because of the way that other companies or frontier labs are developing AI are using AI and are rolling out AI. So I am
Speaker A
very worried about loss of control. We should have that conversation [clears throat] but we should not ignore other risks.
Speaker A
You have this rating for AI going from minus 10 to + 10 and you said it was minus 6. Did it get worse from then? I think it was 2025 when you gave this review.
Speaker A
No, I actually don't think so because in 2024 2025 I thought there was a lot of damage that AI could do and I was not clear about what bottlenecks would [snorts] be facing us in improving it. I think the
Speaker A
improvements have been fast. So the dangers have increased but I think the good things we can do with AI are also becoming clear. But we just need to redirect it. We need to steer AI and AI cannot just be the decision of a handful
Speaker A
of people very smart many of them very well-meaning people but it's too important to be left to just a few people how do you foster that on an individual level information I think we have to start with an information and a narrative and
Speaker A
let me tell you where I think and this is very risky thing to say because I'm going to say something negative about the biggest gatekeepers in our world. But here is my opinion of which institutions have failed us worst. The media. Absolutely.
Speaker A
The media. The media has been awful in cultivating a true sense of information debate about AI.
Speaker A
For 10 years, it's starting to change slowly, but very slowly and very unevenly. for 10 years either Elon Musk and others like that could do no wrong and every criticism of them was just elevating them and everything about AI was just
Speaker A
fascinating, bewildering, wonderful or we just talked of like AI as like the enemy as stoastic parrots or killer robots. the middle ground where people could get engaged, could understand the issues and could actually exercise their role of stewards of our future as
Speaker A
consumers, as democratic citizens that was not cultivated. And this is true for pretty much every organization in the news media.
Speaker A
Let's wrap up with some questions that are super practical because I want people to walk away from this with an action plan, what they can do. First of all, what do you do with your savings when you hear all the news or nothing
Speaker A
has changed? Nothing much has changed. I try to have a balanced portfolio. I think it's very very difficult to beat the market. I'm not saying it's impossible, but you need to be either extremely skilled, spend a lot of time, or have inside information.
Speaker A
I don't have that much time. I have zero inside information. I wouldn't want to have inside information. So, I try to have a balanced portfolio. Let's jump into 2035.
Speaker A
If everything goes in a positive way, what is a typical workday of a knowledge worker who spends 8 hours by their computer? Now, great question. I don't know. I can definitely see a future in which people would work a few less hours. I could and
Speaker A
still be very productive. I could definitely see a future, very scary future in which a lot of people would not have meaningful work. And I could definitely see a future in which people still go to the office or do some
Speaker A
combination of office work and work from home but differently because of true complimentary AI but still human innovation and human talent are central. I think we live even leaving AI aside we live at an uniquely turbulent time. Climate change, aging
Speaker A
population, changing geo geo uh geopolitical balances, a huge middle class emerging from the developing world. All of these require new goods, new services, new tasks, new ways of approaching the market, new ways of organizing companies. And who's going
Speaker A
to do that? Not AI. Humans. And if we can help those humans be more creative, amazing. What would be your advice for someone who sees this amazing wealth being created by at least what we see in company's valuations? Uh we still have
Speaker A
to see the ROI on the actual uh the actual economy. But when you say, you know, we want to participate in this goods creation. How can we do that practically in the next?
Speaker A
Look, I I actually don't see how foundation models are going to make a lot of money despite all of the investment going into frontier labs. So, we'll see some failed IPOs.
Speaker A
Not failed IPOs right away because I think the enthusiasm is there at the moment, but I really don't see how you're going to make trillions of dollars from foundation models because of potential bottlenecks either in the development of the models or in their
Speaker A
diffusion. But even if those don't happen, the next layer of model open source models are going to be very good as well. So, it's just not the the monetization opportunities are going to be limited. So I think the money is in
Speaker A
highquality data and in applications. That's where I see the edge. If your company, your line of business has a lot of data, cultivate it, use it. Even build better data, find ways of capturing that data while also compensating your workers who are
Speaker A
generating that data. That's fairness. That's also a good business decision. Think of applications. That's where going the edge is going to be. applications are going to be much harder to replicate by open source uh if they are really
Speaker A
creative and fill a niche. So I think those are the places and as a worker have that flexibility have that understanding of AI but maintain your creativity and maintain your base.
Speaker A
Yeah. Also for a worker I think it's also very important to realize that they are the source of data. They've been working with customers in the product for years.
Speaker A
And and by the way, this is another important thing. If you're going to develop applications, a lot of the tacet knowledge is with the workers. You can't just sit in your office in front of a computer and have been put a very
Speaker A
powerful AI and come up with an application. You need to go there, get your hands dirty and talk to workers.
Speaker A
Well, hopefully this conversation sparks some bright ideas and entrepreneurs who are watching who want to make practical applications of AI. I want to finish with the last question. You said students going to work at AI companies need a strong moral compass. I think in
Speaker A
general like even as entrepreneur member of society strong moral compass. Can you explain what you mean by that?
Speaker A
Look, let me give one example from a different field which is gene editing. Another amazing technology that has seen a lot of big advances. But I would say most people who go into medicine or gene editing adjacent occupations
Speaker A
recognize that this is a very morally ambiguous technology. and many organizations that have oversight role from NIH to the surgeon general and the uh medical association have put out general uh ethical guidelines for gene editing like germline editing is off right now
Speaker A
because it's just too dangerous we don't understand it too many risks and etc. So that's the kind of moral campus that I think we need in AI as well and we don't have it. People think of oh I'm doing it
Speaker A
for the good and I completely believe it. Almost everybody I know who goes into the field of AI is well-meaning and they they they want something good for not just themselves but for society but there is no clear moral campus and
Speaker A
that's what I'm talking about. It's another interesting conversation that might hopefully to continue at some point again. Thank you so much.
Speaker A
Thank you so much. It wasn't my pleasure.
Topics:Daron AcemogluAI economicsautomationproductivityGDP impactcoding AIcustomer service AIhealthcare AIlabor marketagentic AI
