There is at least one AI-related topic that Google's chief scientists are trying to avoid.
Focusing on Google DeepMind and Google Research's AI Advances, Jeff Dean explained why he avoids artificial general information or conversations about AGI in an episode of “The Moonshot Podcast,” released earlier this month.
“The reason I tend to move away from AGI conversations is that a lot of people have very different definitions of it, and the difficulty of the problem depends on a trillion factors,” he said.
He said today's AI models are “probably already” better than the average person on most non-physical tasks.
“Most people aren't that good at random tasks they've never done before. Some of today's models are pretty reasonable in most cases,” he said. But he added, “They will fail in a lot, they are not at the level of human experts in some things, so that's a very different definition from the world experts in every task.”
When asked how far AI can make breakthroughs faster than humans, Dean said, “It's actually probably already close to that in some domains.”
“There are a lot of domains where automated search and calculations can actually accelerate progress: scientific advancements, engineering advancements,” he said. “I think all of this will be important to help people move forward with what they can do over the next five, ten, fifteen, twenty years.”
Related Stories
Business Insider tells the innovative stories you want to know
Business Insider tells the innovative stories you want to know
The definition of AGI depends on who you ask.
Many AI labs and researchers define it more or less as a form of AI with intelligence like humans and the ability to understand and learn in a way that humans can. Others define it as the point in which autonomous computer systems can do better than humans in most economically valuable jobs.
Top AI researchers have a variety of predictions about when AGI is expected. Google Deepmind CEO Demis Hassabis, Openai CEO Sam Altman and humanity CEO Dario Amodei hope for that in a few years, but others like meta scientist Yann Lecun and AI researcher Andrew Ng say it could be decades away.
