This podcast discusses the timeline and definition of Artificial General Intelligence (AGI). Demis Hassabis estimates a 50% chance of achieving AGI by 2030, defining it as matching the cognitive functions of the human brain, emphasizing its general and consistent intelligence across various tasks, unlike current systems which have "jagged intelligence." He believes true AGI should possess capabilities like invention and creativity, and suggests testing methods like a "brute force test" involving tens of thousands of cognitive tasks, or having top experts test the system. He also looks for "lighthouse moments" such as inventing a new scientific conjecture or a game as deep and elegant as Go. The podcast also touches on AGI's self-improvement, noting that while current systems are good at incremental improvements on specific tasks, they struggle with vague, high-level instructions.
No comments:
Post a Comment