David Silver walked out of DeepMind and raised $1.1B to skip human data entirely

David Silver, the guy who helped shape reinforcement learning at DeepMind, just pulled off one of the biggest funding rounds I’ve seen for a lab that’s barely a few months old. Ineffable Intelligence, his new outfit, raised $1.1 billion at a $5.1 billion valuation.

That’s a lot of money for a company with no product, no revenue, and arguably not even a clear demo yet. But Silver’s pitch is compelling enough to open wallets: build an AI that learns from scratch, without any human-generated data.

If you’ve been following the AI space, you know the current paradigm is all about scraping the internet clean of human text, images, and code. GPT, Claude, Gemini — they all rely on vast datasets of human output. Silver wants to skip that entirely. His bet is that true intelligence doesn’t need to mimic humans. It needs to interact with the world directly, through reward signals and trial and error, the way AlphaGo learned to beat the world’s best Go players without ever seeing a human play.

This is the reinforcement learning approach Silver pioneered at DeepMind, just scaled up to absurd levels. The question is whether it works at the scale of language and reasoning, not just board games or robotics simulations.

Ineffable Intelligence is being deliberately vague about their actual architecture and training methods. The name itself — Ineffable — suggests they’re going for something that can’t easily be described or predicted. That’s either visionary or a convenient way to avoid hard questions.

The valuation is interesting. $5.1B for a pre-revenue, pre-product company is aggressive even by AI standards. For context, that’s more than many publicly traded AI companies with actual customers. But investors are clearly betting that Silver can replicate his DeepMind magic outside Google’s walls.

Silver left DeepMind in late 2025, and rumors had been swirling about what he’d do next. The man literally wrote the textbook on reinforcement learning (the one everyone in the field cites). His credentials are impeccable. But credentials don’t guarantee breakthroughs.

The big challenge here is sample efficiency. Reinforcement learning works great when you can run millions of simulations in a game environment. For language and general intelligence, the space of possible actions is enormous. Without human data to constrain the search, training could be prohibitively expensive, even with $1.1B in the bank.

Silver’s team is reportedly a mix of DeepMind alumni and fresh PhDs from top universities. They’re based in London, which is quietly becoming a serious AI hub outside the Bay Area. That’s a smart move — talent is cheaper there, and there’s less competition for compute resources.

I’m cautiously optimistic about this. The field needs diversity of approaches. Right now, everyone is chasing the same scaling laws with transformer models trained on human data. If Silver can make RL-from-scratch work at scale, it could unlock capabilities that current architectures can’t reach. Or it could burn through a billion dollars and produce nothing useful.

Either way, it’s going to be fascinating to watch. And it’s a reminder that the AI race isn’t just about who has the most GPUs — it’s about who has the most radical ideas about what intelligence actually is.