Google AI chief: Computers aren't as smart as children
Google executive John Giannandrea opined a couple years back that artificial intelligence had reached the level of a four-year-old child. But Giannandrea said on Tuesday, "I think I was over optimistic," noting that a four year-old child can do lots of things a computer can't.
Key points: Even young children can learn from just one or two strong examples while computers need lots of data for machine learning. Computers are also bad at transferring knowledge to adjacent tasks: if a kid is good at Pong, they will probably be good at Breakout — not so for computers.
Concerns about an AI apocalypse are overblown
- A lot of people are "unreasonably" worried we are about to be taken over by general purpose AI, he said.
- "I just see no evidence we are on the cusp of this," he said, arguing that sound bytes on the subject (he's talking about you, Elon Musk) have been "unwarranted and borderline" irresponsible.
2 key areas of machine learning where concern is warranted
- Codification of bias: Google, he said, has released some tools to help analyze data and look for bias, but he said more attention to this topic is needed.
- A lack of transparency around what the system is using to make decisions. "If somebody tries you sell you a black box and doesn't explain how it works I'd be very concerned."
Google product roadmap
- Pervasive computing: There is a shift away from putting all of the smarts on the phone, he said. Google Home is a good example, with more to come over next couple years, Giannandrea added. (It was reported Tuesday that Google plans to introduce a smaller version of Google Home at an Oct. 4 event in San Francisco.)
- Computer vision: Computers are getting very good at seeing and recognizing objects, he said, noting that it is a bit like speech recognition where it was predicted for years, but is finally a viable reality.