MIT professor: Only 1/4 of AI-related tasks can be automated with cost-effectiveness in the next 10 years.

Only a quarter of AI-related tasks will be able to achieve cost-effective automation in the next decade, says MIT professor Daron Acemoglu.

Acemoglu said on the Goldman Sachs Trading podcast that even if AI makes significant breakthroughs, its impact will take years to show.

Goldman Sachs said that means AI will have less than a 5 per cent impact on all tasks in the next decade, raising US productivity by 0.5 per cent and contributing 0.9 per cent to GDP growth cumulatively.

He said: “The fact is that the current large language model architectures are proving to be more impressive than many people predicted, but I think that just by predicting the next word in a sentence, we can get something as smart as HAL in 2001: A Space Odyssey, and that requires a lot of faith.”

Acemoglu said: “Our current LLM architectures are likely to be severely limited.”

He also doubted whether AI could reach its goals faster by simply throwing more GPU capacity at it.

He added that it would need increasingly high-quality data, not capacity, and that it was unclear where such data would come from.

Jim Covello, global equity research head at Goldman Sachs, said AI would have to solve complex problems to deliver a return from companies such as Nvidia (NVDA.US), Microsoft (MSFT.US), Google (GOOGL.US), Meta (META.US), Amazon (AMZN.US) and AMD (SMCI.US), which are expected to spend billions of dollars on AI in the coming years.

He said: “We’ve been studying this for years, and nothing is cost-effective on that point. I think people have an incredible misunderstanding of what this technology can do. It’s not solving big problems. It’s not doing any cognitive reasoning.”