Recently, Baidu CEO Robin Li interviewed Bill Gates and Elon Musk about a range of topics, including machine superintelligence. Here is a transcript of that section of their conversation:
Li: I understand, Elon, that recently you said artificial intelligence advances are like summoning the demon. That generated a lot of hot debate. Baidu’s chief scientist Andrew Ng recently said… that worrying about the dark side of artificial intelligence is like worrying about overpopulation on Mars… He said it’s a distraction to those working on artificial intelligence.
Musk: I think that’s a radically inaccurate analogy, and I know a bit about Mars. The risks of digital superintelligence… and I want you to appreciate that it wouldn’t just be human-level, it would be superhuman almost immediately; it would just zip right past humans to be way beyond anything we could really imagine.
A more perfect analogy would be if you consider nuclear research, with its potential for a very dangerous weapon. Releasing the energy is easy; containing that energy safely is very difficult. And so I think the right emphasis for AI research is on AI safety. We should put vastly more effort into AI safety than we should into advancing AI in the first place. Because it may be good, or it may be bad. And it could be catastrophically bad if there could be the equivalent to a nuclear meltdown. So you really want to emphasize safety.
So I’m not against the advancement of AI… but I do think we should be extremely careful. And if that means that it takes a bit longer to develop AI, then I think that’s the right trail. We shouldn’t be rushing headlong into something we don’t understand.
Li: Bill, I know you share similar views with Elon, but is there any difference between you and him?
Gates: I don’t think so. I mean he actually put some money out to help get somewhere going on this, and I think that’s absolutely fantastic. For people in the audience who want to read about this, I highly recommend this Bostrom book called Superintelligence…
We have a general purpose learning algorithm that evolution has endowed us with, and it’s running in an extremely slow computer. Very limited memory size, ability to send data to other computers, we have to use this funny mouth thing here… Whenever we build a new one it starts over and it doesn’t know how to walk. So believe me, as soon as this algorithm [points to head], taking experience and turning it into knowledge, which is so amazing and which we have not done in software, as soon as you do that, it’s not clear you’ll even know when you’re just at the human level. You’ll be at the superhuman level almost as soon as that algorithm is implanted, in silicon. And actually as time goes by that silicon piece is ready to be implanted, the amount of knowledge, as soon as it has that learning algorithm it just goes out on the internet and reads all the magazine and books… we have essentially been building the content based for the super intelligence.
So I try not to get too exercised about this but when people say it’s not a problem, then I really start to [shakes head] get to a point of disagreement. How can they not see what a huge challenge this is?