David Silver just landed $1.1 billion for his new startup, Ineffable Intelligence. The round values the company at $5.1 billion. Sequoia Capital and Lightspeed Venture Partners led the deal, with Google, Nvidia, Index Ventures, the British Business Bank, and the UK's Sovereign AI fund all participating.
Silver spent over a decade leading reinforcement learning at DeepMind. He was behind AlphaZero, the program that taught itself to master chess, Go, and shogi through self-play alone. No human games to study. No curated datasets. Just the rules and millions of rounds of trial and error.
Now he wants to apply that same idea beyond board games. Ineffable is building what Silver calls a "superlearner," an AI system that discovers knowledge on its own, with no human training data. On the company's website, he describes it as "his life's work" and compares its potential impact to Darwin's theory of evolution.
That's a big claim. Darwin gave us a framework for understanding all life on Earth. Silver is arguing that learning-from-scratch could do the same for machine intelligence. The bet here is that the current approach, scraping ever-larger datasets of human knowledge, has diminishing returns. You run out of high-quality data. You encode human biases and limitations. You build systems that can remix what we already know but struggle to discover anything genuinely new.
Reinforcement learning has shown it can work in narrow domains. AlphaZero didn't just learn chess. It found moves no human had considered in centuries of play. But scaling that approach to general-purpose AI is unproven territory. The real world is messier than a chess board. Rewards are unclear. Feedback loops are slow. And RL has a history of impressive demos followed by difficulty translating to complex, open-ended problems.
If Silver succeeds, the implications are big. AI that can generate new scientific knowledge, invent novel solutions, and explore domains where no training data exists. If he fails, it's a reminder that board games have clear rules and win conditions. Reality doesn't.
He's also pledged to donate any personal profits to high-impact charities. A noble gesture, though $1.1 billion is a lot of money to prove a point.
Silver isn't alone in this conviction. AMI Labs, co-founded by former Meta AI scientist Yann LeCun, raised $1.03 billion last month at a $3.5 billion valuation. Recursive Superintelligence, from ex-DeepMind principal scientist Tim Rocktäschel, pulled in $500 million with enough demand to reportedly stretch it to $1 billion. London is becoming a serious hub for these star-scientist ventures, fueled by DeepMind's alumni network and the UK's willingness to back them with sovereign capital.
But there's a reason most of the industry still relies on human data. It works. GPT-4, Claude, Gemini. They're all trained on enormous corpora of human writing, coding, and reasoning. They have clear limitations, but they also have clear capabilities. Reinforcement learning from scratch is a harder path with less evidence it scales.
Either Silver is right that the future belongs to self-taught machines, or the data-hungry approach wins again. A billion dollars says we'll find out.