Two days ago Tesla Motors Inc (NASDAQ:TSLA) CEO Elon Musk unintentionally made news when a private email of his was published warning that the possibility of something very dangerous happening because of artificial intelligence. The email was taken down almost immediately, but a screenshot was captured and multiple people say that they had seen it on edge.org before it was taken down. This is the first time that Musk has warned about the dangers of AI, but it’s the first time that he gave such a specific timeframe.
“Unless you have direct exposure to groups like Deepmind, you have no idea how fast-it is growing at a pace close to exponential. The risk of something seriously dangerous happening is in the five year timeframe. 10 years at most. This is not a case of crying wolf about something I don’t understand,” said Musk.
But his reference to the AI research team Deepmind, which Google acquired for $400 million earlier this year, reveals why he thinks AI could accelerate so rapidly.
Elon Musk: Five years out from artificial general intelligence
The Deepmind website is incredibly vague about what the company is actually doing, but they have talked about their work in the past. For example back in 2011, cofounder Shane Legg did an interview with LessWrong about the possible dangers of AI. LessWrong asked him, “Can you think of any milestone such that if it were ever reached you would expect human-level machine intelligence to be developed within five years thereafter?”
And Legg answered, “That’s a difficult question! When a machine can learn to play a really wide range of games from perceptual stream input and output, and transfer understanding across games, I think we’ll be getting close.”
Now fast forward to April of this year at the FDOT14 (First Day of Tomorrow) conference, where fellow Deepmind cofounder Demis Hassabis showed one of the company’s recent breakthroughs. An algorithm that they had developed was not only able to play Atari games competently, using perceptual stream input as Legg had mentioned, but he was able to devise strategies that took advantage of each game’s idiosyncrasies. In Hassabis’ words, “It sort of ruthlessly exploits the weakness in the system that it’s found.”
Elon Musk: Neural Turing Machines
Watching the algorithm crush these games (without being programmed to do so) is impressive, but it’s also not what Deepmind expected to happen.
“When our researchers… saw that, that actually shocked them because we haven’t added in the sort of capabilities like long-term memory at that point, that we had thought would be needed to solve that kind of game in that way.”
In other words, Deepmind had stumbled on to more capabilities than they had realized only using a combination of a neural net and short-term memory. Then, last month Deepmind researchers released a paper titled Neural Turing Machines, arguing that the addition of short-term memory to neural nets effectively creates a new approach to artificial intelligence because the algorithm can use the patterns stored in that short-term memory to apply what his learned in solving one task to a seemingly unrelated task, but we normally referred to as general intelligence.
So it seems like the five-year-out mark set by Legg three years ago was passed relatively recently, which probably explains why the dangers of artificial general intelligence has been on Musk’s mind.