Thore Graepel，曾帮助打造谷歌DeepMind的AlphaGo，已在伦敦启动Metis Reasoning。该公司的创始主张是大型语言模型在推理方面不可信。创始人希望将AlphaGo的经验应用于通用推理。
Thore Graepel, who helped build Google DeepMind’s AlphaGo, has launched Metis Reasoning in London. The company’s founding claim is that large language models cannot be trusted to reason. Its founders want to apply the lessons of AlphaGo to general reasoning instead.
该公司在一篇题为《推理时代》的白皮书中阐述了其论点。Tech.eu 报道，Graepel 于今年夏天离开谷歌DeepMind。
The company set out its thesis in a white paper titled “The Age of Reasoning”. Graepel left Google DeepMind this summer, Tech.eu reported.
他的联合创始人包括技术和生物科技企业家 Piers Ingram，已筹集超过3.5亿美元；以及曾在波音工作的机器人专家 Michael Warren。第四位创始人是剑桥顾问公司（Cambridge Consultants）人工智能负责人 David Berman，他还为英国政府和北约提供人工智能咨询。Tech.eu 将团队规模定为12人。白皮书列出了伦敦、新加坡和洛杉矶的办公地点。
His co-founders are Piers Ingram, a technology and biotech entrepreneur who has raised more than $350m, and Michael Warren, a robotics expert who worked at Boeing. The fourth is David Berman, head of AI at Cambridge Consultants. He also advises the UK government and NATO on AI. Tech.eu puts the team at 12 people. The white paper lists offices in London, Singapore and Los Angeles.
Metis 对大型语言模型的批评 白皮书列出了三个问题。大型语言模型会产生幻觉，且没有人能够审计其推理过程。用户也无法检查答案背后的知识，因为这些知识隐藏在模型的权重中。
What Metis says is wrong with LLMs The paper lists three problems. LLMs hallucinate, and nobody can audit their reasoning. Users also cannot inspect the knowledge behind an answer, because it sits inside the model’s weights.
白皮书称：“然而，存在一大批问题，无论规模如何扩大都无法解决。”
“There exists, however, a significant set of problems that no amount of scaling can solve,” the white paper says.
它指出，模型的逐步‘思考链’往往并不反映其得出答案的真实过程。Ingram 更直白地阐述了商业案例。
It argues that a model’s step-by-step “chain of thought” often does not reflect how it reached its answer. Ingram put the business case more bluntly.
Ingram 对 Tech.eu 说：“当你考虑工程、科学、医学或法律推理等高风险场景时，模型给出的论证基本上是不能信任的。”
“When you think about higher-stakes cases in engineering or science, or medical or legal reasoning, the arguments that the models come up with basically can’t be trusted,” Ingram told Tech.eu.
AlphaGo 的经验 白皮书以2016年 AlphaGo 对阵李世石的第37手开篇。专家最初认为这是一次失误。AlphaGo 自身的策略网络将人类专家走出该手的概率评估为约万分之一。创始人认为，这一步来自搜索与评估，即他们所称的推理，而非单纯的预测。
The AlphaGo lesson The paper opens with Move 37, which AlphaGo played against Lee Sedol in 2016. Experts first took it for a blunder. AlphaGo’s own policy network had rated the chance of a human expert playing it at about one in 10,000. The founders argue the move came from search and evaluation, which they call reasoning, rather than from prediction alone.
Metis希望AI能明确记录自己所知道的内容以及确信程度。一个独立的判断系统只有在证据支持时才会接受新的主张。随后，系统会根据每一步能在多大程度上减少不确定性来选择下一步。论文以哲学家雅科·欣蒂卡（Jaakko Hintikka）关于推理是探究者与自然之间一场博弈的思想为模型来构建这一体系。
Metis wants AI to keep an explicit record of what it knows and how certain it is. A separate judgment system would accept a new claim only when the evidence supports it. The system would then pick each next step by how much it cuts uncertainty. The paper models this on philosopher Jaakko Hintikka’s idea of reasoning as a game between an inquirer and nature.
如同AlphaZero的自我对弈一样，系统随后会从自身记录的探究过程中学习更好的策略。Metis表示，这意味着更少的幻觉，以及任何人都能逐步核查的推理。用户还可以在不重新训练的情况下更新或删除知识。较小的模型就能处理许多步骤，该公司称这可以降低成本，并将推理能力部署到设备和机器人上。
As in AlphaZero’s self-play, the system would then learn better strategies from its own recorded inquiries. Metis says this would mean fewer hallucinations and reasoning that anyone can check step by step. Users could also update or remove knowledge without retraining. Smaller models could handle many steps, which it says could cut costs and put reasoning on devices and robots.
英格拉姆对Tech.eu表示，这一设计比前沿模型“在计算效率上要高得多得多”。
Ingram told Tech.eu the design is “much much more compute-efficient” than frontier models.
“我们绝不是要走模型越来越大、训练预算越来越大的这条路，”他说。
“We are very much not going in this direction of ever larger models and ever larger training budgets,” he said.
尚无模型 这份白皮书描述的是一种方法，而非一款产品，也没有公布任何测试结果。英格拉姆称，Metis以极低的成本在前沿推理问题上取得了“令人鼓舞”的结果，但未给出具体数字。Metis尚未表示何时发布其首个模型。
No model yet The white paper describes an approach, not a product, and it publishes no test results. Ingram said Metis had achieved “encouraging” results on frontier reasoning problems at a fraction of the cost, without giving figures. Metis has not said when it will release its first model.
对于Metis正在融资数千万美元的报道，英格拉姆拒绝置评。他说Metis“迄今为止的融资已经足够充裕”。
Ingram would not comment on reports that Metis is raising tens of millions of dollars. He said Metis had raised money “more than adequately to date”.
Metis加入了一批由DeepMind前员工及其他知名人士领导的新兴伦敦AI实验室之列。这些实验室包括由格拉佩尔昔日AlphaGo同事戴维·西尔弗（David Silver）创立的Ineffable Intelligence，以及理查德·索赫尔（Richard Socher）的Recursive Superintelligence。英格拉姆称Metis的方法“确实是一种非常不同的、近乎哲学性的方法”。
Metis joins a cluster of new London AI labs led by DeepMind alumni and other big names. They include Ineffable Intelligence, founded by Graepel’s former AlphaGo colleague David Silver, and Richard Socher’s Recursive Superintelligence. Ingram called Metis’s approach “really a very different, almost philosophical approach”.