Every AI tool professionals use today was built the same way: take enormous amounts of human-written text, feed it through a model, and produce a system that can rearrange and reflect that knowledge in useful ways. That approach created ChatGPT, the AI features in your procurement software, and the reasoning tools in legal and finance. It also has a hard limit.
Researchers at Epoch AI estimated the world could exhaust the usable supply of quality public human text for AI training somewhere between 2026 and 2032, and possibly sooner given how aggressively labs are consuming it. This is not speculation at the fringes. It is a structural problem that the entire AI industry is quietly working around.
David Silver spent more than a decade at Google DeepMind, where he led the teams that built AlphaGo, AlphaZero, and AlphaStar. These were AI systems that, crucially, did not learn from human examples. AlphaZero taught itself chess and Go from scratch, with no human game records, and then played moves that human grandmasters had never conceived. That is not a remix of human knowledge. That is something new.
Silver left DeepMind in early 2026 to found Ineffable Intelligence in London. The company raised $1.1 billion in seed funding at a $5.1 billion valuation, the largest seed round in European history, with backing from Sequoia, Lightspeed, Google, and NVIDIA, plus the UK's Sovereign AI Fund. The company has no product, no revenue, and no public roadmap. Investors are betting entirely on Silver's thesis and his track record.
The thesis is this: AI systems that learn purely from experience, rather than from human data, are not limited by what humans already know. They can discover strategies, solutions, and knowledge that no person has ever articulated. Silver calls this a superlearner.
NVIDIA's partnership is where this becomes an infrastructure story, not just a research one. The AI tools businesses use today were trained on fixed datasets: gather the data, run the training, deploy the model. Experience-based learning is fundamentally different. The AI has to act, observe the result, score itself, and update continuously, in tight loops, generating its own training data as it goes. That requires computing architecture designed specifically for that kind of constant, self-feeding cycle rather than for processing large static datasets.
Engineers from NVIDIA and Ineffable are jointly designing that pipeline. The work starts on NVIDIA's current Grace Blackwell hardware and moves to the Vera Rubin platform, NVIDIA's next-generation system launching to cloud providers in the second half of 2026. Vera Rubin is already being framed by NVIDIA as a system built for agentic AI and continuous reasoning, not just one-shot training tasks. The Ineffable collaboration puts that architecture to its most demanding test yet.
For business operators, the short-term implications are limited. Ineffable Intelligence is a research lab with a long horizon. What matters right now is the signal: the world's leading AI chip company and one of its most credible investors are treating experience-based learning not as a distant academic possibility but as the next serious direction for the field. The practical AI tools businesses adopt in 2027 and beyond will be shaped by whether this approach works.
The deeper question is what happens to industries if it does. Today's AI assists with tasks defined by existing human knowledge: drafting, summarising, classifying, retrieving. An AI that can genuinely discover new knowledge would eventually assist with tasks no human has yet framed. Drug discovery, materials science, and financial risk modelling are the obvious first targets, given their reliance on pattern-finding in complex systems. But the implications for any knowledge-intensive profession are real.
Silver's critics note, fairly, that experience-based learning has historically worked brilliantly in environments with clear rules and clear win conditions, like chess or Go, and has struggled in the open-ended messiness of the real world where there is no obvious score to optimise. That problem is not solved. But $1.1 billion and a partnership with the world's dominant AI hardware company suggest that serious people believe it is solvable.