Michalowski, Jennifer. “The Cost of Thinking.” MIT News, Massachusetts Institute of Technology, 19 Nov. 2025, news.mit.edu/2025/cost-of-thinking-1119.
In this latest article, researchers at MIT’s McGovern Institute led by Associate Professor Evelina Fedorenko found a remarkable parallel between how humans and modern reasoning large language models take on difficult problems. They tested both people and AI models on various types of reasoning tasks, from arithmetic to more abstract spatial-transformation problems, and found that the problems that take humans the longest also require the most internal computation from the models. According to the researchers, this suggests a kind of convergent thinking cost even though these models weren’t built to think like humans, they end up resembling us in how much work certain problems demand.
It’s interesting that within this article it reveals a deeper connection between artificial and human intelligence. It’s not just that AI gets smarter, in some sense, it thinks in a way that mirrors us, at least when facing challenging problems. This raises big questions about what thinking even means, and whether understanding that similarity could help us build better, more efficient models. For our class, this article is a powerful illustration of how AI research isn’t just about performance, it’s also about drawing on insights from cognitive science and neuroscience, which brings together technical and philosophical dimensions of AI in a very human-centered way.
Leave a Reply