In a new article, scientists from Google DeepMind predict that limited human data is no longer the optimal material for guiding machines toward generating intelligence. Instead, it is time for humans to allow machines to explore the world independently and learn from experience. Only in this way can AI create knowledge that is truly worthy of human study.

Humanity's quest for "thinking machines" has spanned over seventy years, and a new era may be dawning. Humans will eventually abandon the dream of "recreating human intelligence in machines," realizing that artificial intelligence cannot—and does not need to—replicate human intellect. Allowing machines to learn from their own experiences is the only way to breach the boundaries of human knowledge and solve critical problems that remain beyond our reach.

This is the central thesis of a recent article titled "Welcome to the Era of Experience" by DeepMind scientists. The text serves as a chapter in the upcoming book Designing an Intelligence, co-authored by Richard Sutton, a pioneer in the field of reinforcement learning (RL), and David Silver, a key proponent of the discipline.

The authors categorize modern AI development into three distinct eras: the "Era of Simulation," the "Era of Human Data," and the impending "Era of Experience." AlphaGo is a product of the Era of Simulation, while GPT-3 can be viewed as the representative advancing the Era of Human Data. Meanwhile, AlphaProof, developed internally by DeepMind, is considered evidence of the shift toward the Era of Experience.

Read the full article at Global Views Monthly: https://www.gvm.com.tw/article/121115