Nobel Laureate Busts the AI Hype (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:00) - KAUSHIK VISWANATH: AI is poised to transform everything, (00:00:04) or is it? From agentic AI to instant cures, (00:00:07) the hype around AI can be deafening. (00:00:10) But what's the real economic impact, (00:00:12) stripped of the speculation? (00:00:14) Today, we cut through the noise with MIT economist (00:00:17) and Nobel Laureate Daron Acemoglu, (00:00:20) whose data-driven research reveals a surprising reality. (00:00:23) Forget overnight transformation, (00:00:25) Acemoglu's research projects that AI will automate (00:00:28) just 5% of all tasks and add just 1% (00:00:32) to global GDP this decade. (00:00:34) So why the massive disconnect? (00:00:36) And what should smart business leaders be doing (00:00:38) with AI right now? (00:00:39) I recently interviewed Acemoglu (00:00:41) and asked him these questions and more. (00:00:44) (bright upbeat music) (00:00:50) KAUSHIK: Thank you so much for being here with us today. (00:00:53) I have a few questions for you about generative AI (00:00:56) and AI in general and its impacts on the economy. (00:00:59) So chat GPT came out in November 2022, (00:01:02) and since then we've seen generative AI (00:01:05) go through a lot of developments. (00:01:07) It has observers, I think, excited (00:01:09) and a little bit worried about what it means for their jobs (00:01:11) and for the economy in general. (00:01:15) Last April, you published a paper called (00:01:18) "The Simple Macroeconomics of AI," (00:01:20) in which you estimate that over the next 10 years, (00:01:24) only about 5% of all tasks will be profitably automated (00:01:28) by this technology, and that it's only likely (00:01:31) to contribute about 1% to global GDP. (00:01:34) That's a stark contrast (00:01:36) to what some other analysts have said. (00:01:39) You know, people have been predicting that this will be (00:01:42) a truly transformative technology to the labor force (00:01:46) and to the economy in general. (00:01:48) Can you explain why your estimates (00:01:50) are different from these others? (00:01:51) And and since you published that paper last year, (00:01:56) have you seen anything that either confirms (00:01:58) or makes you question those estimates you made? (00:02:01) - DARON ACEMOGLU: Well, well, thank you, Kaushik. (00:02:02) Well, look, I said one other thing in that paper, (00:02:05) it's hugely uncertain and these are just guesses. (00:02:08) I think it's very difficult to know (00:02:10) because it's a very rapidly changing technology, (00:02:13) and over the last year we have seen even more advances. (00:02:17) So we don't know where we're going. (00:02:19) But the basis of my prediction, (00:02:23) uncertain though it may be, still remains. (00:02:27) The industry has not produced applications (00:02:32) that are critical for the production process (00:02:37) or for generating new goods and services (00:02:39) that are gonna be hugely valuable. (00:02:41) So if you compare AI to the internet, (00:02:45) I think from the very early days of the internet, (00:02:48) even when there was hype and a boom, (00:02:51) it was clear how the internet was gonna change everything. (00:02:54) The way that we communicate has been completely transformed (00:02:59) by the internet. (00:03:00) It was very clear at the time, it was also very clear (00:03:03) that the internet would introduce a lot of new goods (00:03:05) and services and provide platforms for people (00:03:07) to come together in various ways for production, (00:03:10) for recreation, and other things. (00:03:12) I think those things are not clear yet for AI. (00:03:17) Of course, if you're a believer that AGI (00:03:19) is just around the corner, you think somehow (00:03:24) in the next few years, somehow we're gonna get such amazing (00:03:29) machines that they can start performing (00:03:31) all the cognitive tasks. (00:03:33) But even that scenario is not so clear. (00:03:35) You know, how are you gonna actually get (00:03:39) AI tools into the production process? (00:03:42) And I think the current approach is well targeted (00:03:47) for dealing with cognitive tasks that are performed (00:03:52) in predictable environments in offices, (00:03:56) and don't require much social interaction (00:03:59) and very high levels of judgment. (00:04:01) So if you are a software engineer (00:04:04) that does some very basic routines for your work, (00:04:09) or you are in IT security or you're in accounting, (00:04:12) those are things that I think there will be applications (00:04:16) based on AGI and some other AI tools (00:04:20) that will be able to perform these tasks. (00:04:22) If you're a CEO, if you are a CFO, if you're an entertainer, (00:04:26) if you're a professor, if you are a construction worker, (00:04:31) or a custodial worker, or a blue collar worker, (00:04:33) I think those things are beyond what AI can perform (00:04:39) or AI can indirectly contribute (00:04:42) to by being bundled with flexible robotics (00:04:45) because we're not there in terms of those technologies. (00:04:47) So when you do that calculation, (00:04:51) you end up with about 20% or so of the economy (00:04:55) that is either at the cross hairs of AI to be automated (00:04:59) or could be majorly boosted by AI input. (00:05:02) Things that are feasible, they take, takes a long time, (00:05:05) many of them are performed in small companies, (00:05:07) it's not gonna be profitable to do them. (00:05:09) So that's how I arrived to the 5% number, (00:05:11) based on these inputs and a lot of detailed material. (00:05:16) But it may may turn out to be wrong. (00:05:18) - KAUSHIK: Last year, I wouldn't have expected (00:05:20) to see the kinds of leaps and bounds. (00:05:22) - DARON: Yeah, I mean the leaps and bounds (00:05:24) are really inspiring at some level. (00:05:26) So I'm pretty impressed by those. (00:05:31) The question is, with these leaps and bounds, (00:05:35) do you still think that in two, three, four, years time (00:05:43) you can have an AGI with no human supervision that can do (00:05:48) all of your accounting (00:05:50) or all of your marketing? (00:05:51) And I think that is a much higher bar. Why? (00:05:55) First of all, because every single occupation (00:05:58) has so many complex tacit knowledge parts (00:06:03) and requires a lot of checking (00:06:05) and a lot of different types (00:06:06) of intelligence being applied to it. (00:06:09) - KAUSHIK: And does that tie into the distinction (00:06:10) you make in the paper between what you call easy to learn (00:06:14) and hard to learn tasks? (00:06:15) And should that distinction inform how executives study (00:06:21) or decide what business processes (00:06:23) are most amenable to automation? (00:06:27) - DARON: Look at the domains in which we have truly (00:06:30) inspiring achievements from AI (00:06:34) such as AlphaGo, AlphaFold, or answering some complex, (00:06:41) but knowledge-based questions. (00:06:44) Those are all domains in which there is a ground truth (00:06:48) that everybody can agree on. (00:06:51) You either fold the protein or you do not. (00:06:53) AI is capable, there's no doubt about that. (00:06:55) That's why we're talking about AI. (00:06:57) And it is capable of learning that knowledge (00:07:00) if it's in its training data set. (00:07:03) So once you provide AI with the right powerful algorithm, (00:07:07) for example, reinforcement learning (00:07:08) was very important for the Alpha series, (00:07:11) maybe other things for generative AI. (00:07:13) And the ground truth is there, AI is gonna get there, (00:07:17) but no task that we perform in reality (00:07:21) is just recounting already established knowledge (00:07:25) or playing a parlor game. (00:07:27) They are much more complex. (00:07:28) They involve interactions, they involve a lot of things (00:07:30) that are based on tacit knowledge, (00:07:32) or they are based on matching your contextual understanding (00:07:37) of a problem with the specific task at hand. (00:07:41) For example, diagnosing a difficult ailment (00:07:45) or finding the kind of product that's gonna work well (00:07:48) given the retirement planning that an individual is doing. (00:07:51) With the current architecture, (00:07:52) the best that we can do is we can copy (00:07:54) human decision makers that make decisions. (00:07:56) So we can load in a lot of data from doctors (00:08:01) making diagnoses or reading radiology reports (00:08:06) or from financial planners. (00:08:08) And then AI, generative AI in particular, (00:08:11) has a great way of imitating these human decision makers. (00:08:15) But if you do that, you're not gonna get much better (00:08:17) than the human decision makers. (00:08:18) And especially if you don't know who the very best human (00:08:20) decision makers are, you may not even very easily achieve (00:08:24) the human, best level human decision maker level. (00:08:26) Places where we need a lot of judgment or social interaction (00:08:30) or social intelligence, (00:08:31) I think are still beyond the capabilities of AI. (00:08:35) And on the basis of this, I would say, (00:08:37) my prediction, which again has huge error bands around it. (00:08:42) So may it well turn out to be wrong, (00:08:44) but I don't expect any occupation that we have today (00:08:47) to have been eliminated in five or 10 years time. (00:08:50) So if you are an AGI believer, that you think (00:08:53) that generative AI and other AI tools (00:08:57) are going to completely transform the economy (00:08:59) within the next three, or four years, or five years, (00:09:02) then you must have in your mind a list of occupations (00:09:05) that will completely disappear. (00:09:06) All of this that I have summarized briefly (00:09:11) is predicated on the current approach to AI. (00:09:16) And what I have been arguing, (00:09:19) and this paper was a small part of that bigger edifice, (00:09:23) is that we are not developing AI in the best possible way. (00:09:29) And that best possible way is much more pro-human. (00:09:32) It's much more targeted at working (00:09:35) with human decision makers. (00:09:37) It requires a bigger celebration of the places (00:09:40) where AI is better than humans, (00:09:42) and the places where humans are better than AI. (00:09:45) And once you take that approach, I think the biggest promise (00:09:49) is using AI for providing new goods and services, (00:09:53) new ways of doing things for humans. (00:09:55) We are at the cusp of many major transformations. (00:10:00) We are an aging society. (00:10:02) There are gonna be many, many more people (00:10:04) over the age of 60, many, many, many more people (00:10:06) over the age of 70 in the United States, (00:10:08) many more in Europe, (00:10:09) that they are going to demand new goods, (00:10:13) new services, new accommodations. (00:10:16) Financial industry is at the cusp of big changes. (00:10:19) Again, this is not gonna be on cost saving. (00:10:21) It's gonna be, for example, (00:10:23) what sometimes people call financial inclusion. (00:10:26) Meaning we provide new, better services for people (00:10:29) who are not currently making enough use (00:10:31) of financial services, including banking. (00:10:33) Climate change. (00:10:34) Whether you mitigate it or not (00:10:36) is going to change many aspects of our lives. (00:10:38) Again, new goods and services (00:10:40) and the entire production process requires new tasks, (00:10:44) new ways of increasing the expertise (00:10:46) and sophistication of workers. (00:10:48) All of these, I think, are to play for, (00:10:50) and those are the places where I think AI (00:10:53) could make a big difference. (00:10:54) So my recommendation to business leaders would be, (00:10:56) don't be taken by the hype. (00:10:58) I think the hype is an enemy of business success. (00:11:01) Instead think where my most important resource, (00:11:06) which is your human resource, can be better deployed. (00:11:09) And how can I leverage that human resource (00:11:12) together with technology, together with data (00:11:15) so that I increase people's efficiency (00:11:17) and I enable them to create better (00:11:21) and newer goods and services, not just cutting costs, (00:11:25) but doing new things that are so important (00:11:27) in this changing world. (00:11:28) - KAUSHIK: Business executives should really be thinking (00:11:30) about a much wider scope of possibilities (00:11:33) than simply eliminating costs or finding roles (00:11:37) that they can cut from their organizations. (00:11:40) - DARON: That's my perspective. (00:11:41) Again, you will be hard pressed to find many people (00:11:45) in Silicon Valley who agree with this perspective, (00:11:47) but I've been researching this for quite a while. (00:11:50) I may be wrong, but at least I do have data. (00:11:53) I do have historical knowledge (00:11:55) and I do have some theoretical (00:11:56) understanding of these issues. (00:11:57) And I would say on the basis of those that of course (00:12:00) any business leader should be happy (00:12:02) if they can reduce their costs even by 1%, that's great. (00:12:05) 1% more profits. (00:12:07) But the evidence, as far as I read, is quite clear, (00:12:13) no business has become the jewel of their industry (00:12:18) by just cost cutting. (00:12:19) - KAUSHIK: All good business leaders (00:12:22) are looking for that next big idea, (00:12:24) that next innovation that can turn them (00:12:27) into one of these stars of their industry. (00:12:30) In the meantime, right now (00:12:32) is when they are putting investments into AI (00:12:35) and they are starting to look for a return (00:12:38) on that investment. What metrics do you think (00:12:40) they should be paying attention to, (00:12:41) to know whether those investments are really paying off? (00:12:45) - DARON: Well, I'm not gonna be able to provide a simple (00:12:47) metric for you, but let me give you my perspective. (00:12:49) And the reason why I wrote the paper (00:12:51) that you started with is precisely (00:12:52) because I'm worried about those investments. (00:12:55) I think most business executives, not all, (00:12:57) but most business executives are investing in AI blindly. (00:13:03) They are doing so without understanding how AI (00:13:06) can be synergistically deployed with their workforce. (00:13:10) And they're doing so because they're under (00:13:11) tremendous pressure because every day (00:13:12) they hear from management consultants, from the newspapers, (00:13:16) from podcasts, that your competitors are investing (00:13:19) big time in AI and if you're not, you're falling behind. (00:13:23) That's not a way to create a successful business. (00:13:27) You never create a successful business (00:13:29) because you think your competitors are investing (00:13:30) and you should do it not to fall behind. (00:13:33) And I think the recipe that I would suggest is, (00:13:37) start by thinking about where it is that you can make (00:13:41) a big difference in terms of the new things that you do. (00:13:44) I think for many financial industries (00:13:46) it's quite clear - new financial services are badly needed. (00:13:49) I think if you are producing other services, (00:13:54) health services, education services, (00:13:55) I think a complete overhaul of these things is necessary. (00:13:58) And that's not gonna happen just by buying (00:14:01) more cloud services from Amazon or just introducing (00:14:06) some generative AI tools easily. (00:14:08) It's gonna happen by identifying, with the help (00:14:11) of your most skilled employees, (00:14:13) identifying where these new services can be introduced, (00:14:18) what the demand for them is, (00:14:20) and how that can be made possible. (00:14:21) And AI would then be a great tool (00:14:24) to augment the capabilities of your workforce (00:14:26) and yourself in doing that. (00:14:29) - KAUSHIK: That's fascinating. (00:14:30) Well, thank you so much for your perspective, Daron. (00:14:32) You've given us a lot to think about. (00:14:34) I hope you enjoyed my discussion with MIT economist (00:14:37) and Nobel Laureate Daron Acemoglu on AI's economic impact. (00:14:41) The key insight for leaders: (00:14:43) Rather than following your competitors (00:14:45) into blind AI investments, (00:14:47) focus on how the technology can help you and your team (00:14:50) deliver meaningful innovation. (00:14:52) Are you seeing AI create new opportunities in your industry? (00:14:56) Share your thoughts in the comments. (00:14:58) For more research-based information from MIT SMR, (00:15:01) check out this playlist. (00:15:03) Thanks for watching. (upbeat music)