The AlphaGo Architect Who Says LLMs Can't Reason - And Is Building Something Better

Ten years after AlphaGo stunned the world in Seoul, the researcher who helped build it is making a pointed case: the AI systems dominating headlines today still lack the core reasoning capability that made AlphaGo's legendary Move 37 possible - and until that changes, claims of "reasoning models" should come with a heavy asterisk.

Who Is Making This Argument, and Why Now?

The author is Thore Graepel, writing in MIT Technology Review on October 2, 2026. [Graepel is chair of machine learning at University College London and was a core member of the AlphaGo team at DeepMind.]1 He recently departed Google DeepMind - [he left Alphabet's research lab in summer 2026]2 - and is now raising funds for a new startup called Metis Reasoning, [seeking tens of millions of dollars from an initial group of backers]2 to pursue what he describes as one of AI's biggest unsolved problems.

The timing is deliberate. This year marks the ten-year anniversary of AlphaGo's 4-1 victory over Lee Sedol, and Graepel is using the milestone to draw a sharp distinction between the kind of AI that beat one of humanity's greatest Go players and the kind powering today's chatbots and coding assistants.

What AlphaGo Actually Did - and What LLMs Don't

[Move 37 in game two of the five-game match looked so absurd that some commentators thought it was a programming glitch.]3 It wasn't a mistake - and it also wasn't pure intuition. Graepel's core argument is that most popular accounts of Move 37 get the explanation wrong.

[AlphaGo is made up of two systems.]3 The first, its policy network, was trained to predict what move a strong human would play. [This "intuitive" part regarded move 37 as nothing special - a play that had a roughly one in 10,000 chance of being made by an expert human.]3 What elevated the move was AlphaGo's second system: a value network paired with Monte Carlo Tree Search that could simulate thousands of future positions, reason about long-term consequences, and identify the move as strategically brilliant even though it looked wrong to human pattern-matching. That deliberate, forward-looking search process is what Graepel calls genuine reasoning.

[It was actually AlphaGo's powers of reasoning that made this creative choice - and these are powers that today's AI lacks.]3 Large language models, by contrast, predict the next token from learned statistical patterns. [A large language model predicts the next word from human-written text.]2 There is no internal search tree, no model of future states, no mechanism for weighing possible outcomes before committing to an answer. Current "reasoning models" that use chain-of-thought are, in Graepel's framing, a workaround - the model reads its own scratch-pad notes back - rather than a fundamental redesign. [Many of the recent advances in LLMs, such as reasoning models and large-context handling, are not neat extensions of that core technology but workarounds that patch over some of its fundamental flaws.]4

The Gap Between Fluency and Competence

The practical consequence is what Graepel describes as a gap between fluency and competence. LLMs are genuinely useful for tasks where language patterns correlate strongly with correct answers: summarization, code generation, research synthesis. But they struggle when conditions change unpredictably and a system needs to plan across many steps rather than retrieve a well-worn answer.

[If we want future AI systems to produce trustworthy results and really novel insights in fields like science and medicine, we need to equip them with genuine reasoning capabilities.]3 That is Graepel's stated motivation for leaving DeepMind: the conviction that pursuing better LLMs at scale will not bridge this gap, and that a fresh architectural approach is needed.

What Metis Reasoning Is Proposing

[Graepel's new venture is working on artificial intelligence that can respond to unfamiliar problems and choose an action, with uses in robotics, science, and engineering.]2 The technical thesis centers on [bringing AlphaGo-style reasoning to frontier AI so machines can plan and act under uncertainty.]2

[His architecture reportedly emphasizes structured search methods and probabilistic planning.]5 Rather than generating the most statistically likely next token, [systems built on this framework evaluate possible futures and assign probabilities to different outcomes.]5 The distinction is roughly between memorizing every answer in a textbook and learning how to solve novel problems from first principles - [LLMs are exceptionally good at the former; Graepel wants to crack the latter.]5

[Graepel is raising capital around an alternative: redesigning how systems evaluate possible actions before they commit to an answer.]6 The startup is still early-stage, and most commercial details remain undisclosed.

Why It Matters

Graepel's argument is not new - Yann LeCun has made related points for years - but it lands differently coming from someone who was inside the room when AlphaGo defeated Lee Sedol, and who subsequently returned to DeepMind before choosing to leave the scaling consensus entirely.

The deeper issue he raises is architectural: can you get to trustworthy AI reasoning by scaling token prediction, or does genuine deliberation require something structurally different? For engineers building systems that need to plan, adapt, and act reliably - autonomous agents, scientific discovery tools, robotics - that question is not academic. Current benchmarks reward fluency; they are much weaker at distinguishing pattern recall from structured search over future states.

[We need to go back to the architecture and fundamentally redesign it so that it does proper reasoning.]7 Whether Metis Reasoning delivers on that ambition is an open question. But the fact that one of AlphaGo's architects is staking his post-DeepMind career on it is worth taking seriously.