Oct 2, 2026, 10:48 PMDon’t be fooled—LLMs don’t reason | MIT Technology Review
Thore Graepel is chair of machine learning at University College London. He was a core member of the AlphaGo team at DeepMind.
https://www.technologyreview.com/2026/10/02/1145639/dont-be-fooled-llms-dont-reason/
@teledyn A lot of people don't reason either. It's not the flex you think it is.
Oct 3, 2026, 2:00 AM@skewray to go forward from intuition often involves reason
Oct 3, 2026, 12:06 AMI like the idea but it seems like he's talking about a state table that might be impossible to accurately model in a generic way. Chess and Go have established rules and a predefined win state.
Of course, open-world reasoning is harder than playing a board game such as Go or chess. In the real world the current state of affairs is only partially known, the set of available actions is large and variable, and the consequences of actions are stochastic or unknown.
What's the win state for "how do I make lasagna from scratch?"
The article ends right when it was getting interesting.
It offers no real path forward. I handwaves that LLMs need their training data reinforced with things that are proven to be correct but doesn't begin to explain how that might work or how it overcomes some of the fundamental aspects of how the next token prediction works.
Oct 3, 2026, 1:59 AM@jrdepriest points taken, not that anyone has any answers, but I do know lasagna and, as is said, the proof is in the pudding 😅
Oct 3, 2026, 6:57 PMYes, and then how to define "harder" well (it isn't one-dimensional) in
«open-world reasoning is harder than playing a board game such as Go or chess»,
never mind quantifying it?