Google’s DeepMind scientists say the future of artificial intelligence lies in letting machines teach themselves through real-world experience, not just human data. In a new book chapter titled Welcome to the Era of Experience, researchers Richard Sutton and David Silver argue that AI must move beyond mimicking humans to unlock breakthroughs we can’t yet imagine.

　Take AlphaZero: The chess and Go master taught itself by playing millions of games against itself—no human coaching needed. It even invented moves experts called “creative,” outperforming earlier models like AlphaGo, which relied on human strategies. “Human knowledge isn’t the key,” said Sutton, a reinforcement learning pioneer. “Raw computing power and smart algorithms are.”

　The shift marks a new phase in AI development:

　1. Simulation Era: Systems like AlphaGo mastered games using human rules.

　2. Human Data Era: Models like GPT-3 trained on vast text archives.

　3. Experience Era: AI learns by doing, like humans—but faster.

　DeepMind’s latest projects, including health-care AI that tracks sleep or heart rate data, hint at this future. CEO Demis Hassabis claims general AI (AGI) could arrive in 5–10 years, focusing on scientific discoveries like fusion energy. “We’re building tools to solve humanity’s toughest problems,” he told TIME.

　But risks loom. Self-taught AI might outpace human values, and mass job automation could disrupt economies. Hassabis admits the road is uncertain: “We might need a new political philosophy.”

　No dollar figures were shared, but one thing’s clear: The race to build thinking machines just got a lot more interesting.