AlphaGo creator launches unicorn startup to build AI "superlearners"

David Silver, the scientist who created the AlphaGo algorithm, which in 2016 became the first system in history to beat a world Go champion, has announced a new company valued at around $1 billion. Its goal is to build what Silver calls superlearners: AI systems that can master the hardest domains of knowledge on their own, without relying on vast datasets created by humans. It is a direct challenge to the industry's dominant paradigm, in which all the biggest players are betting on scaling language models.

Silver is one of the key architects of modern AI, and his track record speaks for itself. His work at Google DeepMind produced AlphaGo, and then AlphaZero, an algorithm that taught itself chess, shogi and Go from scratch without seeing a single human game. Instead of learning from ready-made examples, the system generated and analyzed millions of positions on its own, uncovering strategies that professional players described as inhuman. That experience shapes his conviction about what next-generation AI should look like.

Silver's central idea is both simple and radical: large language models, including ChatGPT, Claude, Gemini and others, are fundamentally limited because they learn exclusively from text and data produced by humans. This creates an insurmountable ceiling: an AI cannot surpass the cognitive abilities of its creators if it feeds only on their knowledge and their misconceptions. Simply adding more parameters and more training data, he says, does not solve this fundamental problem; it only scales it.

The alternative is reinforcement learning (RL). Unlike supervised learning, where a model learns to reproduce correct answers from a pre-labeled dataset, RL lets an agent explore the space of possibilities on its own: try actions, receive a reward signal and gradually build a strategy. This is how AlphaGo worked, and Silver is convinced this approach opens the way to AI that surpasses humans across a broad range of tasks, not just in predefined games.

The position has serious arguments in its favor. OpenAI is partly moving in this direction with its o-series reasoning models, which use elements of RL to check their own answers. Google DeepMind continues fundamental research in the area. Still, the bulk of the industry's resources is focused on scaling language models, and Silver is openly taking a contrarian stance against that mainstream.

The main difficulty with RL lies outside narrow, clearly defined tasks. For chess, the reward function is easy to define: win and you get a plus. For writing a persuasive text, making a sound business decision or conducting original scientific research, the reward function is not obvious. This problem of ineffable intelligence (Silver's term) is what the new company will have to solve.

A $1 billion valuation with not a single product on the market reflects the weight of the founder's reputation. In the current investment climate, where every AI startup claims historic significance, the name of AlphaGo's creator is both a ready-made proof of concept and insurance for investors unwilling to wait years.

If Silver is right, the next phase of the AI race will look fundamentally different: less human data, more autonomous self-learning, less imitation, more discovery. Systems that can form knowledge on their own beyond what humanity knows are his superlearners.

Whether the idea becomes a real product, time will tell. But the fact that one of the main architects of modern AI is making a public bet against the dominant paradigm is in itself a significant signal for the whole industry.