MIT Tech Review: LLMs still don't truly reason like AlphaGo did

An MIT Technology Review piece reflects on AlphaGo's famous 2016 victory over Go champion Lee Sedol, particularly its unexpected move 37, and argues that today's large language models still don't replicate the kind of reasoning that made that breakthrough possible.

The author, who worked on building the AlphaGo system, contends that despite a decade of progress, modern AI models marketed as capable of reasoning are not actually using the same underlying mechanisms that let AlphaGo discover a winning but counterintuitive strategy.

The piece raises questions about how AI companies describe the reasoning abilities of current large language models compared to specialized systems like AlphaGo.