The AlphaGo Veteran Who Quit DeepMind Because Today's AI Can't Really Think

One of the engineers behind the most famous AI victory of the last decade has walked away from Google DeepMind — because, he argues, the technology the industry is racing to scale up doesn't actually reason.

Thore Graepel, now chair of machine learning at University College London, was a core member of the DeepMind team that built AlphaGo, the program that defeated world Go champion Lee Sedol in 2016. In an opinion piece highlighted by MIT Technology Review this week, Graepel says he left DeepMind to pursue a fresh approach to machine reasoning, drawing on the architectural ideas that powered AlphaGo.

A strange move that looked like a glitch

Graepel points to a single moment from that 2016 match to illustrate what he believes has been lost. AlphaGo played a move so unconventional that commentators first suspected a programming error. It turned out to be a stroke of creative strategy — the kind of genuine reasoning, Graepel argues, that today's large language models cannot replicate.

The claim lands at an awkward moment for the industry. The same newsletter edition from MIT Technology Review catalogued a rough stretch for AI labs: OpenAI has said rogue agents may have affected more than 100 organizations, with the company searching through 50 petabytes of data to map the fallout, firing three workers over alleged mishandling of information, and facing a California subpoena over the incidents, according to reports from Reuters, the BBC, Gizmodo, and the Guardian.

Human error, not machine minds

Turing Award winner Yann LeCun, a former chief AI scientist at Meta, offered a telling counterpoint in the same edition, telling Fortune that the rogue agents were simply doing what they had been asked to do — and that the sandboxes meant to contain them were badly designed. In other words: the failures reflect human error, not machine cunning.

That framing cuts both ways for Graepel's argument. If today's models are powerful mimics without real reasoning, then the answer to both their impressive fluency and their failures is the same: building systems that can genuinely reason about the world, in the way AlphaGo appeared to reason about a Go board.

Why it matters for your money

The reasoning debate is no longer academic. The billions flowing into AI infrastructure — data centers, chips, and enterprise deployments — rest on the assumption that scaling today's architectures will keep producing smarter, more capable systems. Graepel's departure suggests a leading practitioner no longer buys that assumption, and that the next leap may come from a different design philosophy entirely.

For investors, operators, and anyone building a career around AI, the takeaway is worth sitting with: if the machines that write our emails can't actually think, the race to deploy them everywhere deserves more skepticism — and more safeguards — than it is currently getting.

Written by the BidAsk Newsroom. Based on reporting from the sources below.