For years, the AI industry has scaled one formula: more human data, more computing power, larger models and more products. That formula created the AI boom. But systems that learn mostly from what people have already written, played, coded and explained stay tied to the limits of those human examples. Researchers are trying to build AI that learns without relying mainly on human data.
David Silver, the researcher behind the Go-playing program AlphaGo, has raised $1.1 billion for Ineffable Intelligence. His London company, valued at $5.1 billion with backers including Nvidia and Google, builds AI that learns by trying, failing and improving on its own.
Silver’s systems learn inside simulated test settings with goals, feedback and repeated attempts, instead of from records of human work. A system that learns this way can search a larger range of actions, including paths missing from the human record.
AlphaGo is why the idea is taken seriously. In 2016, it defeated Lee Sedol, one of the world’s best Go players, by four games to one. Go is a 2,500-year-old board game with an enormous number of possible games, which makes it one of AI’s hardest tests. In the second game of the match, AlphaGo played a move so unusual that expert commentators thought it looked wrong. That move, later known as Move 37, worked. The program had found a valid path outside established human judgment.
A cleaner proof came a year later. AlphaGo Zero started with only the rules of Go and improved by playing against itself, without training on human games. After three days, it beat the version of AlphaGo that had defeated Lee Sedol. That is the pattern Ineffable is pursuing.
Silver makes the broader case with researcher Richard S. Sutton in a paper called Welcome to the Era of Experience. They argue that future AI may develop stronger abilities by learning mainly through experience. Human data is also expensive to collect and messy to clean, while a system that learns through trial and feedback can generate its own lessons.
The hard part is building test settings where those trials are useful. Bad goals produce bad learning, and weak feedback creates false progress. AlphaGo Zero learned inside a board game with clear rules, and Ineffable is betting the method can work outside such narrow settings.
