Britain is backing a no-proof bet against the dominant AI playbook

Britain is using state support to take a side in one of AI's most important unresolved fights: whether the current path of training models on human-written data can keep carrying the field, or whether a more self-taught approach will have to take over.

That is the real meaning of the government's support for Ineffable Intelligence, the new startup from former Google DeepMind researcher David Silver. In a post announcing its support, the U.K. government's Sovereign AI initiative said it is backing the company with access to the country's largest AI supercomputers, visas, and what it calls "the unique levers of the British state." The company has no public product, no benchmark, and no paper showing the approach works outside bounded simulations.

Silver is not pitching a better chatbot. In an interview with WIRED, he argued that the current large language model path, where systems learn from vast stores of human-written text, will fail as a route to superintelligence. He compared human data to fossil fuel: a powerful but limited shortcut. His alternative is reinforcement learning, the training method where a system improves through trial and error by learning which actions lead to better outcomes. It is the same family of ideas behind AlphaGo and AlphaZero, the DeepMind systems that learned to beat world champions in Go and master chess from self-play.

Reuters reported that Ineffable raised $1.1 billion in seed funding at a $5.1 billion valuation, with Sequoia Capital and Lightspeed leading the round and Nvidia, Google, and British state vehicles also participating. CNBC reported that the company is only months old, which helps explain how much of this bet rests on Silver's standing rather than public proof. Ineffable's own website says it is building a "superlearner" that can discover knowledge from its own experience "without relying on human data." Sequoia's investment post makes the thesis even cleaner: "No pre-training. No imitation." Just an agent learning from the consequences of its own actions in a world built to teach it.

That is the pressure point. The market is not just rewarding a famous researcher with a contrarian view of AI. The British government is now attaching industrial policy to a claim that the dominant pretraining-heavy AI playbook may top out before it reaches something more general.

There is a case for taking Silver seriously. According to his public biography, he was the former lead of the reinforcement learning team at DeepMind, and his ideas carry more technical weight than the usual anti-LLM startup manifesto. If anyone gets to argue that trial-and-error learning could outrun today's AI stack, it is someone with Silver's track record.

But the clean "reinforcement learning versus large language models" framing is too neat. Modern frontier AI systems already use reinforcement learning in important parts of training and post-training. The real unsolved question is not whether reinforcement learning matters. It is whether a system that learns mostly from experience and simulation can keep generalizing once it leaves elegant closed environments and enters the mess of the real world.

So far, Ineffable has not shown that publicly. There is no released product, no technical paper, no benchmark result, and no demonstration that the company's approach survives outside mission statements and investor enthusiasm. That does not make the thesis wrong. It does mean the evidence is still thin compared with the size of the bet.

The next thing to watch is whether Silver can produce public evidence that his post-LLM path works somewhere harder than a simulation. Until then, Britain and Silicon Valley are front-running a theory.