Ex-DeepMind researcher wants AI that checks its facts before answering

Thore Graepel, who worked on the AlphaGo system at DeepMind, has started a company called Metis Reasoning. It aims to build AI that checks its claims against evidence before it answers. He laid out the plan at Nvidia's GTC conference in Berlin. The catch is that Metis has no working system yet, only a white paper.

The idea

AlphaGo beat Go professional Lee Sedol in 2016. It did so by pairing a network that suggested moves with a systematic search through possible game sequences. That search found the famous Move 37, which the network alone would have given roughly a one-in-ten-thousand chance of being played. Metis says today's chatbots are like the move-suggesting part alone. It wants to add the search part to everyday reasoning.

How it would work

At the center would be what the paper calls a knowledge ledger. It is a record kept outside the AI model, playing the role the board plays in a game. Metis says the ledger would hold three things:

• What the system knows
• How confident it is in each item
• Which questions are still open

A separate module would check new entries against calculations, formal proofs or real-world evidence. If an assumption turns out to be wrong, every conclusion that relied on it would be re-evaluated. The system would end with either an answer or a justified decision not to answer.

Metis says this should mean fewer made-up answers, explanations with source citations, and knowledge that can be edited or deleted without retraining the model. Breaking reasoning into small steps could also suit smaller models running on devices and in robots.

Why it matters

Metis argues that bigger models will not fix the core problem. Chatbots are trained to produce plausible text, not true text, and their step-by-step explanations often do not show how an answer was really reached. The paper says this is why AI has had so little economic impact so far. Software code is easy to test and undo. Medicine, finance, chip design and robotics, it argues, need AI that can be verified from the start.

What is not known yet

The white paper describes a research program only. It has no working system, no benchmarks and no comparisons with current reasoning models. The team admits the real world gives only partial information, evidence can be costly to get, and some questions have no answer. A big open question is how to build a trustworthy ledger in messy fields. Older symbolic AI systems failed partly because people had to enter all the knowledge by hand.

Who is behind it

Graepel also worked on AlphaZero and MuZero, and helped build the matchmaking system behind Xbox Live. He left Google DeepMind over the summer. According to Bloomberg, he is raising a low tens-of-millions round, with a follow-on of several hundred million dollars already in the works. He joins other DeepMind veterans who have left to build elsewhere.