David Silver walked out of DeepMind and raised $1.1B to rethink how AI learns

David Silver, the guy who basically wrote the book on reinforcement learning at DeepMind, just pulled off one of the biggest AI fundraises I’ve seen in a while. His new outfit, Ineffable Intelligence, locked in $1.1 billion at a $5.1 billion valuation. And the company is only a few months old.

Let that sink in. A lab that barely has a birthday under its belt is now worth more than most public companies. But Silver isn’t some random founder cashing in on hype. He was the lead on AlphaGo, the system that made Go champions look like amateurs. He co-created AlphaZero, which taught itself chess and shogi from scratch. The guy understands self-learning AI better than almost anyone alive.

What’s Ineffable actually building? An AI that learns without relying on human-generated data. That’s the pitch, anyway. Most large language models today — GPT, Claude, Gemini — are trained on massive piles of text scraped from the internet. Silver wants to go a different route: agents that figure things out through interaction and reward signals, not by memorizing what people already wrote.

This isn’t a new idea. Reinforcement learning has been around for decades. But scaling it to the point where it competes with supervised learning on language tasks has been the holy grail. Silver thinks we’re finally there, or close enough to throw a billion dollars at it.

The valuation is aggressive, no doubt. $5.1 billion for a pre-revenue, pre-product company is the kind of number that makes VCs sweat. But the investor list suggests serious conviction. I’m hearing names that usually only show up for later-stage rounds. The bet is that Silver can replicate what he did at DeepMind — building systems that surpass human performance without human teaching — but on a much bigger canvas.

Ineffable’s approach has some obvious advantages. Human data is expensive, messy, and biased. If you can build an AI that learns by exploring its environment, you sidestep copyright issues, privacy concerns, and the diminishing returns of scraping the entire internet twice. But it’s also harder. Reinforcement learning is notoriously sample-inefficient. AlphaZero needed millions of games to master Go. Scaling that to something like natural language understanding is a whole different beast.

Silver has been thinking about this for a long time. In a 2024 talk, he argued that intelligence isn’t about absorbing human knowledge — it’s about discovering structure through interaction. Ineffable is the practical test of that philosophy.

The name itself is a bit pretentious, I’ll admit. “Ineffable Intelligence” sounds like something a philosopher would name their cat. But Silver has earned the right to be ambitious. Whether he can deliver on the promise without drowning in compute costs is the real question.

I’ll be watching this one closely. If anyone can make reinforcement learning work at scale for general intelligence, it’s probably him. But the AI graveyard is full of well-funded labs with smart founders and good ideas that just didn’t pan out. Silver’s track record tilts the odds, but they’re still long.