YouTube8m· Oct 2024· cataloged

AI Can Only Do 5% of Jobs: MIT Economist Fears Crash


What this covers

This is an eight-minute Bloomberg Technology interview in which hosts including Caroline Hyde press MIT economist Daron Acemoglu on how much of the labor market AI will actually reach. The transcript does not name the guest, but the publisher's description identifies him. His case rests on what current models can't do. Tasks with heavy physical-world interaction, such as carpentry, construction and manufacturing, are out of reach, as are socially demanding work like psychiatry and anything needing high reliability or reasoning. He also points to earlier waves of digital technology, whose promised automation arrived only gradually, and he argues that generative AI has no close historical precedent, which is why confident forecasts are hard to come by.

The middle of the conversation tests the industry's boldest assumptions. He contends that scaling laws won't hold over the next five to ten years, because complex physical tasks need higher-quality data and different architectures, not simply more of both. He also says the last two and a half years brought incremental progress rather than a qualitative leap beyond GPT-3.5-level capability. Where AI does work, he says, is in narrow tasks like simple subroutines, where GPT-4 and GitHub Copilot help, and in routine IT security work, where he expects fewer job vacancies. Hiring is rising for people with AI fluency, including programmers and AI integrators. He frames the real opportunity as a tool that raises worker productivity, not one that replaces workers.

The most contentious turn is financial. He separates Nvidia, which is earning substantial revenue from a belief that chips will grow more valuable, from OpenAI, whose $157 billion valuation he doubts because no revenue model yet justifies it. He grants that OpenAI is emerging as the market leader and that venture capital can sustain such firms for a long time. The hosts also note that the job-displacement story ranked among Bloomberg's most-read.

Sharpest takeaway

Current AI capabilities will only displace approximately 5% of jobs in the near term (5-10 years) because generative AI lacks the reliability, physical integration, and reasoning abilities needed for most occupational tasks, despite significant hype about exponential improvements.

  • AI cannot perform tasks requiring heavy physical-world interaction (construction, manufacturing, carpentry) due to reliability and robot integration limitations
  • Scaling laws assumptions don't hold because higher-quality data and different architectures are needed, not just more compute and data volume
  • Recent AI advances have been incremental rather than qualitative, not exceeding GPT-3.5 capabilities in ways that rapidly change task performance

The claims · ranked18 claims · weighted by value

This asset isn't compiled yet

You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.

0.68

The most optimistic views in the AI industry about scaling laws—which assume that doubling data and compute power will double AI capabilities—will not hold water within the next 5-10 years because performing complex physical tasks like carpentry requires fundamentally higher quality data and a different AI architecture than current dominant architectures provide.

causalhigh valuecontestednovelty 2/4durability 3/4· Unknown Speaker (likely AI researcher or economist guest)

“I don't think that the most optimistic views in the industry, which sometimes people refer to as scaling laws, actually will hold water. When you look at it within the next five years or ten years, which say, you know, just double the amount of data that will the amount of compute power. And now you're going to get double the amount of capabilities because. We to do that, we would actually need higher quality data.”

0.66

Only approximately 5% of jobs in the economy will be affected by AI in the near term because current AI capabilities cannot perform tasks involving heavy physical-world interaction (construction, manufacturing, carpentry), social interaction (psychiatry), or tasks requiring high-level reasoning and reliability.

forecasthigh valuecontestednovelty 2/4durability 2/4· Unknown Speaker (likely AI researcher or economist guest)

“if you look at its current capabilities. You know, we are very far from. I either generative AI or other types of AI performing any of the tasks that involve a heavy level of interaction with the physical world construction, work, manufacturing, work, carpentry.”

0.65

Previous waves of digital technologies promised significant automation and productivity improvements, but in practice all of those changes have been very gradual, suggesting that current AI hype may similarly result in gradual rather than sudden transformation.

causalhigh valuecontestednovelty 1/4durability 3/4· Unknown Speaker (likely AI researcher or economist guest)

“other waves of digital technologies that promised a lot of automation, sometimes a lot of productivity improvements, sometimes a lot of changes in the way that we would organize our offices and things like that. And at the end, all of those changes have been very gradual.”

0.63

There is a difference between Nvidia and OpenAI in terms of fundamental viability: Nvidia is riding a wave of belief that chips will become more valuable as AI booms and is generating substantial revenues, while OpenAI has not found a revenue model that justifies its current $157 billion valuation.

causalhigh valuecontestednovelty 2/4durability 2/4· Unknown Speaker (likely AI researcher or economist guest)

“Well, I think there's a difference between Nvidia and Openai, and Vidya is riding a wave, which is based on the belief that those chips are going to become more valuable because generative AI and other types of AI are going to continue to boom. But they are generating revenues. Openai, on the other hand, I think does hasn't found a revenue model that could justify the types of valuations we are seeing at the moment.”

0.63

OpenAI is emerging as the market leader in the current ecosystem in the digital world, and firms that capture the market, control data, and interact directly with consumers or programmers can become very valuable because of venture capital and deep-pocketed funders, despite short-term revenue uncertainty.

causalhigh valueestablishednovelty 0/4durability 3/4· Unknown Speaker (likely AI researcher or economist guest)

“But on the other hand, Openai is also emerging as the market leader in the current ecosystem in the digital world is right. Firms that take over the market and they can use all of the data, they can, you know, interact with the consumers or the programmers directly. Those have become very valuable because of venture capital or other deep pocketed funders.”

0.62

The real potential of AI is for it to become a tool for workers to boost their productivity rather than to replace workers entirely, which is not an AI-pessimist position but recognizes both AI's potential and its current limitations.

normativehigh valuecontestednovelty 1/4durability 3/4· Unknown Speaker (likely AI researcher or economist guest)

“I am not an AI denier or an AI pessimist. I actually believe A.I. has a lot of potential. And in fact, the real potential is exactly what you said, Caroline, which is that A.I. becomes a tool for workers to boost their productivity.”

0.61

There is nothing identical to generative AI in the past, which is part of the reason why people can hold very different beliefs about its impact, making strong consensus impossible.

factualhigh valueestablishednovelty 1/4durability 3/4· Unknown Speaker (likely AI researcher or economist guest)

“There's nothing that's identical to generative AI in the past, and that's part of the reason why people can hold very different beliefs.”

0.57

One class of jobs where AI is leading to a big boost in hiring is people with AI fluency, including AI programmers, AI integrators, and other AI-related service activities, representing new job creation in the AI economy.

factualhigh valueestablishednovelty 0/4durability 2/4· Unknown Speaker (likely AI researcher or economist guest)

“Well, there is one class of jobs where the AI is leading a big boost, which is people who have AI fluency, either people who are AI programmers, A.I. integrators or other AI related service activities.”

0.57

Helping workers in occupations like carpentry, electrician work, plumbing, construction, transportation, teaching, and journalism with AI could be achieved perhaps at some point in the future, but the reliability and reasoning of current models is not up to the task yet.

forecasthigh valuecontestednovelty 1/4durability 2/4· Unknown Speaker (likely AI researcher or economist guest)

“when you look at the whole economy, that means helping carpenters, electricians, plumbers, blue collar workers, construction workers, transport workers, people who are doing teaching, people who are doing journalism, those are things that could be achieved perhaps at some point. But the reliability and the reasoning of the models is not up to the task yet.”

0.57

It may be that in the future OpenAI's API becomes so valuable that other firms build on it, but currently there are no applications with deep impact on worker productivity or firm automation that are going to be sold within the next 5-10 years, making the speaker skeptical of OpenAI's current valuation.

forecasthigh valuecontestednovelty 1/4durability 2/4· Unknown Speaker (likely AI researcher or economist guest)

“Now, it may well be that in the future suddenly their API is going to be so valuable and other firms are going to start building on it. But it doesn't seem like there are the kinds of applications that can have deep impact on the productivity of workers, on automation, of firms that are going to be sold within the next five years, ten years. So I'm a little bit skeptical.”

0.57

Jobs in IT security focused on routine tasks will experience declines in job vacancies because AI tools can perform that work quite well, representing specific occupational displacement where AI is functionally ready.

forecasthigh valuecontestednovelty 1/4durability 2/4· Unknown Speaker (likely AI researcher or economist guest)

“And there are some other jobs that I think there will be some retrenchment. So if you're in IT security and you're doing the routine end of I.T. security, there will be declines in job vacancies in that area because some of these A.I. tools can actually perform that quite well.”

0.54

Recent AI advances over the last 2.5 years have been gradual and incremental rather than qualitative, and have not exceeded the basic capabilities of GPT-3.5 or GPT-3.4, suggesting that advances will not quickly change how tasks are performed in the economy.

factualhigh valuecontestednovelty 1/4durability 2/4· Unknown Speaker (likely AI researcher or economist guest)

“But what I'm basing that on is, first of all, other waves of digital technologies that promised a lot of automation, sometimes a lot of productivity improvements, sometimes a lot of changes in the way that we would organize our offices and things like that. And at the end, all of those changes have been very gradual. And also what we have seen over the last two and a half years where, you know, there have been advances, but the advances haven't really been sort of qualitative in nature.”

0.54

Some AI applications have generated really phenomenal improvements in specific domains, such as writing simple subroutines in common programming languages where GPT-4 and GitHub Copilot have proven effective, demonstrating that AI augmentation is already working in narrow, well-defined tasks.

factualhigh valueestablishednovelty 0/4durability 2/4· Unknown Speaker (likely AI researcher or economist guest)

“There are a few areas where we're seeing really phenomenal improvements, like, for example, if you want to write simple subroutines in one of the common programming languages, you know, GPT four is great before and GitHub copilot was pretty good.”

0.50

Within five years, we will not see certain occupations completely disappear from the labor market, even as AI capabilities advance, because AI impact will be gradual rather than disruptive.

forecasthigh valuecontestednovelty 1/4durability 1/4· Unknown Speaker (likely AI researcher or economist guest)

“So for example, I do not believe that in five years time when we look at the labor market, we're going to see certain occupations that will have completely disappeared.”

0.45

Research from late 2010 found that no other job category besides AI-related jobs received a boost in hiring, suggesting that AI job creation has been narrowly concentrated despite widespread implementation.

factualhigh valuecontestednovelty 0/4durability 2/4· Unknown Speaker (likely AI researcher or economist guest)

“In my research, albeit with a a from a previous way, will be AI. In the late 2010, we didn't see any other job category get a boost in terms of hiring.”

0.29

There is huge uncertainty in any forecasts about AI, and the most relevant source of uncertainty is exactly what is going to happen with AI in the next several years, making any strong prediction about AI impact highly uncertain.

definitionestablishednovelty 0/4durability 3/4· Unknown Speaker (likely AI researcher or economist guest)

“I should have preface this by saying there's a huge amount of uncertainty in any forecasts about I”

0.10

The AI story has all the facets that engagement audiences at Bloomberg Technology, indicating that AI job displacement appeals to the channel's audience interests.

factualspeaker onlynovelty 0/4durability 0/4· Bloomberg host (Unknown name)

“It has all of the facets of a story that we didn't, you know, just engage our audience on Bloomberg Technology.”

0.10

Bloomberg reports that the story about AI job displacement is one of the most read stories on Bloomberg in the last 24 hours, indicating high public interest in AI's labor market impact.

factualspeaker onlynovelty 0/4durability 0/4· Bloomberg host (Unknown name)

“This was one of the most read stories on Bloomberg in the last 24 hours. There's no surprise about that.”