A public exchange among AI researchers kicked off after former Google DeepMind researcher Thore Graepel used an X post tied to a new MIT Technology Review piece to argue that today’s most advanced AI systems still lack something essential. Graepel wrote that large language models are “remarkably capable,” but that “generating longer chains of thought is not the same as genuine reasoning.”
That post laid out Graepel’s core complaint: current systems, in his view, usually do not keep an explicit, inspectable record of what they know, what remains uncertain, what evidence supports a conclusion, or whether real progress has been made. He argued that if AI is going to produce “trustworthy and genuinely novel insights in science, medicine and beyond,” its conclusions need to come from “an auditable process of evidence, inference and belief revision.”
Graepel tied that argument to AlphaGo. In the same post, he wrote that the system’s famous Move 37 was not the result of intuition alone, but of being able to “search possible futures, test its instincts and reason about what would happen next.” He also said that view helped drive his recent departure from Google DeepMind, writing that he left because he believes AI needs “a fresh approach to machine reasoning,” drawing on “some of the architectural lessons from AlphaGo.”
The argument turned into a debate about search, tools, and structure
The discussion quickly moved from Graepel’s critique to a broader question: what should count as reasoning in AI, and can LLMs get there by adding the right machinery?
Yann LeCun responded in clear terms: “Good piece. True reasoning must involve a search. LLMs don't possess this capability.” That lined up with Graepel’s AlphaGo comparison, which emphasized exploring possible futures rather than just producing fluent explanations.
David Duvenaud pushed on a different point. In a reply, he asked whether “we might simply be able to provide similar tools to LLMs to achieve true reasoning” by Graepel’s definition, and whether that was “more or less” what Graepel planned to do.
Graepel’s answer was that the issue is not just output length or raw verbal fluency, but disciplined structure. In one reply, he argued that reasoning improves when the process is carefully organized rather than left as a “stream of consciousness,” adding that otherwise humans fall into familiar cognitive biases. In a later follow-up, he said, “Definitely the tools, but more importantly, the rigorous application of the scientific method at scale!”
Gary Marcus also endorsed Graepel’s quoted claim that trustworthy AI requires “an auditable process of evidence, inference and belief revision,” adding, “It is madness to believe otherwise.”
Comments