AI expert says large language model development nearing plateau – WisBusiness

This post was originally published on this site.

An AI specialist with the Milwaukee School of Engineering predicts models like ChatGPT will soon hit a plateau, meaning a lot less short-term impact on jobs and society than the hype suggests.

Jeremy Kedziora, endowed chair of artificial intelligence at MSOE, this week addressed members of the Milwaukee Rotary Club. He explained the “deep learning revolution” that kicked off around 2012 is now running into constraints, which have implications for the future of AI development. 

The factors that have driven the explosive growth of AI — particularly large language models, or LLMs — include greater computing power for less money, advances in the underlying mathematics and the internet putting out massive amounts of human-generated information. 

But while the internet itself is growing exponentially in size, Kedziora noted it’s “not necessarily growing at a great rate” when it comes to new, valuable information. He displayed a graph showing that very soon, models such as OpenAI will have essentially caught up with the growth rate of the internet and won’t have enough new information to sustain their rapid evolution. 

“The internet is growing, it’s getting big. However, that doesn’t mean we’re getting a lot of new information,” he said. “What is the informational content of the ten billionth cat picture, right? Probably a lot less than the first. So there’s just less and less available for these models to Hoover up and use.” 

As a result, he expects a “super intelligence” form of AI called artificial general intelligence isn’t likely to emerge in the near future. That means AI scientists will need to leverage the already maximized models better to do end-to-end tasks, giving them the ability to plan, set “subtasks” to support overarching goals and “maybe even access actual levers to pull to affect things in the real world,” Kedziora said. 

This form of AI, which he called an “agentic” system, could automate complex, time-consuming endeavors such as scientific research, according to Kedziora. 

In late 2024, a team of researchers created an AI model based on LLMs that can come up with ideas on its own, review literature, gather data, generate experiments, write the code to run them and write up a study for a peer-reviewed journal. What’s more, the cost of running this entire process is just $15. 

“So if you can automate science, and discover new things with AI, what can’t you do?” he said. 

Kedziora expects “you’re going to see a lot more attempts” this year to build AI systems like this that can handle complex workflows from start to finish. Still, he said concerns being raised about the technology’s impact on the workforce are largely overblown, as AI is expected to affect jobs related to about 5% of the economy over the next decade. 

“We might be heading for pretty modest AI-related impacts, despite the hype,” he said. 

Watch the video.