A news summary based solely on the provided article body: Since the 2017 introduction of the transformer, AI developers have improved large language models primarily by scaling them up—using hundreds of billions of parameters and scraping vast data from the internet and books. This approach, rooted in the connectionist deep learning that powers every LLM today, has proven effective, with current models capable of executing many long, complicated tasks without human oversight. However, some leading researchers believe scaling is a dead end. Before leaving Meta last year, Yann LeCun stated, “We are not going to get to human-level AI just by scaling LLMs.” OpenAI co-founder Ilya Sutskever, whose 2012 work on AlexNet demonstrated scaling’s power, noted that while the period from 2020 to 2025 was “the age of scaling,” the scale is now so big that “it’s back to the age of research again, just with big computers.”