AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Former Deepmind researcher thinks AlphaGo's playbook can cure AI hallucinations

Collected Oct 27, 2026

Thore Graepel, a former DeepMind researcher who contributed to AlphaGo, AlphaZero and MuZero, has founded a startup called Metis Reasoning and presented its approach at Nvidia's GTC conference in Berlin. The company's white paper argues that scaling language models with more data and compute will not fix their core problems, and proposes borrowing AlphaGo's structure — a prediction network paired with systematic search — as a blueprint for general reasoning. Graepel left Google DeepMind's post-AGI team over the summer and also holds the machine learning chair at University College London. According to Bloomberg, he is raising a round in the low tens of millions, with a follow-on of several hundred million dollars at a higher valuation reportedly in preparation.

The central idea is an explicit knowledge ledger that plays the role the game board played in Go. Instead of knowledge being baked into model weights, the ledger records what the system knows, how confident it is, and which questions remain open. A separate module checks new entries against calculations, formal proofs or empirical evidence. A reasoning step only changes the system's knowledge state if that independent check deems it sufficiently supported; when an assumption falls away, every conclusion that depended on it is re-evaluated. A "move" can be a question, an experiment or a challenge to an existing assumption, and the system picks whichever reduces uncertainty about the original question most. The output is either an answer or a justified abstention.

As in AlphaGo, two neural networks suggest and evaluate moves. They are trained on logged prior investigations but cannot override the verification rules. Metis says this should yield fewer hallucinations, explanations with source citations, and knowledge that can be edited or deleted without retraining. Because reasoning decomposes into small steps, the company suggests the approach could suit smaller models on edge devices and in robots. Metis frames current language models as roughly equivalent to AlphaGo's prediction network alone — the part that proposed possible moves — without the search that produced the celebrated Move 37 against Lee Sedol in 2016, a move the network would reportedly have given a human player about a one-in-ten-thousand chance of making.

The paper's diagnosis of hallucination is that models are trained to produce plausible text rather than true text, and that even chains of thought in today's reasoning models often fail to show how an answer was actually reached. It argues this limits economic impact: when an agent acts for a company, errors are hard to trace and wrong knowledge cannot be removed from the model. Software engineering adopted AI earlier, the paper contends, because code is easy to test and roll back, whereas medicine, finance, chip design and robotics need verifiability built in from the start.

The concept has recognizable precedents. Microsoft's GraphRAG stores knowledge outside model weights in a knowledge graph, and Google's AlphaGeometry2 pairs a language model with a symbolic engine that verifies mathematical proofs. Metis's ambition is to extend such principles to open-ended real-world questions, which is where the difficulty lies. In practice, a ledger only helps if entries can be checked cheaply and reliably; where ground truth requires a wet-lab experiment, a market outcome or a physical trial, verification becomes the bottleneck rather than the reasoning. The team itself acknowledges that the real world, unlike Go, offers only partial information, that evidence can be expensive to obtain, and that some questions have no answer. A likely trade-off is latency and cost: searching over possible investigative moves is far slower than a single forward pass, even if each step is smaller.

The most consequential open question, in my reading, is how a verified ledger gets populated in messy domains beyond textbook examples. Symbolic AI systems once faltered precisely because their knowledge had to be entered by hand, and a ledger that depends on manual curation risks the same ceiling. Automated extraction from documents plus independent checking could ease this, but the paper does not describe a working mechanism; it describes a research program, with no system, no benchmarks and no comparisons against current reasoning models.

Graepel is the latest in a line of DeepMind alumni building outside Google. David Silver founded Ineffable Intelligence and closed what the article describes as the largest seed round of any European startup; Tim Rocktäschel co-founded Recursive, now valued at $4.65 billion. According to the Financial Times, DeepMind dissolved the team behind AlphaFold, and Nobel laureate John Jumper left for Anthropic with two colleagues. CNBC attributes part of the talent flow to limited internal access to Google's TPU chips and to bureaucracy, while DeepMind itself is losing independence and founder Demis Hassabis has stepped back from day-to-day operations. Graepel's own track record includes TrueSkill, the matchmaking system behind Xbox Live, built during his time at Microsoft Research; he earned a PhD in machine learning in Berlin in 2001.

Why it matters: Teams deploying agents in regulated or high-stakes settings — medicine, finance, chip design, robotics — are the intended beneficiaries, since the pitch targets exactly the auditability and knowledge-removal problems that block production use. If the idea works, developers could get conclusions tied to citable evidence and editable knowledge without retraining, which is a meaningful alternative to retrieval-augmented generation and graph-based approaches. For now, though, this is a thesis and a fundraise rather than a tool; my inference is that the near-term signal to watch is whether Metis publishes benchmarks showing verification holds up outside toy domains.

Read at The Decoder

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Former Deepmind researcher Thore Graepel wants his startup Metis Reasoning to bring AlphaGo's search principle to general reasoning. Instead of locking knowledge inside model weights, an explicit knowledge ledger would make every conclusion checkable against evidence. So far, though, there's no working system and no benchmarks. The article Former Deepmind researcher thinks AlphaGo's playbook can cure AI hallucinations appeared first on The Decoder .