How productive can generative AI be?
In an interview with Goldman Sachs, MIT Institute professor Daron Acemoglu and author of several books, including “Why Nations Fail: The Origins of Power, Prosperity, and Poverty” and most recently, “Power and Progress: The Millennium Struggle for Technology and Prosperity,” argued that the positive impact of generative AI technologies on U.S. productivity and growth over the next decade, and perhaps beyond, is likely to be more limited than many expect.
Acemoglu estimates that only a quarter of AI-based tasks will be cost-effective to automate within the next decade, meaning that AI will impact less than 5% of all tasks. He also takes little comfort from history showing that technology improves and costs fall over time, arguing that advances in AI models will not happen as quickly or as impressively as many think.
Acemoglu also questioned whether the introduction of AI would lead to new tasks or products, saying such an impact is “not a law of nature.” He estimates that the impact on total factor productivity over the next decade will not exceed 0.66 percent, and will be even lower at 0.53 percent when accounting for the complexity of tasks that are difficult to master. This equates to a GDP impact of roughly 0.9 percent over the decade.
“All human inventions should be celebrated, and generative AI is very much a human invention,” Acemoglu says. “But too much optimism and hype can lead to premature use of technologies that are not yet ready for full-scale use. This risk seems especially high today when using AI to drive automation. Pushing automation too early can create bottlenecks and other issues for companies that are missing the flexibility and troubleshooting capabilities that human capital provides.”
Return on Investment
Jim Covello, head of global equity research at Goldman Sachs, argues that to fully benefit from expensive AI technology, the technology would have to solve incredibly complex problems, which it currently can't do and may never be able to do.
“My main concern is that AI applications must solve extremely complex and important problems for companies to see a decent return on investment (ROI) because of the enormous cost of developing and operating AI technologies,” he says. “We estimate that building out AI infrastructure will cost over $1 trillion over the next few years alone, including spending on data centers, utilities and applications. So the key question is: what is the trillion-dollar problem that AI will solve? Replacing low-wage jobs with very costly technologies is essentially the opposite of previous technology transitions I've witnessed in my 30 years of following the technology industry closely.”
“Many people try to compare AI today to the early days of the internet,” Covello says. “But the early internet was a low-cost technology solution that allowed e-commerce to replace expensive existing solutions. Amazon was able to sell books at a lower cost than Barnes & Noble because it didn't have to maintain costly brick-and-mortar stores. Thirty years later, Web 2.0 is still offering cheaper solutions that are disrupting more expensive solutions, just like Uber replaced limousine services. While the question of whether AI technology will deliver on the promise that many people expect today is certainly debatable, what is less debatable is that AI technology is very expensive, and to justify its cost, it must be able to solve complex problems. But AI technology is not designed to solve complex problems.”
Covello doesn't believe the lack of competition will dramatically reduce technology costs as it evolves, because today only Nvidia can make the GPUs that power AI, and the starting point is so high that even if costs do come down, they will need to come down substantially to make automating tasks with AI affordable.
Read the full report here.
(You can now subscribe to the Economic Times WhatsApp channel)
