This famous roboticist doesn’t think Elon Musk understands AI
techcrunch.com
techcrunch.com
- People who don't understand AI are afraid of it; those who do know how fragile and limited it is.
- The call for "regulation" of technology that doesn't exist is too vague to be useful.
- It's the AI in self-driving cars which has the most potential to immediately kill/save thousands of people, and it's telling that it's not this technology that Elon seems to be calling to be regulated. Whether or not regulation is the right thing to do, any argument for/against regulation of self-driving cars could be applied just the same to a hypothetical super AI, but the former is tied to real, practical problems which exist today.
Brooks clearly knows what's up.
Also, to add my own commentary:
- The dystopian robot future we should all be afraid of is not the [paperclip maximizer](https://wiki.lesswrong.com/wiki/Paperclip_maximizer) Musk and friends wave their arms about, but marketing/business algorithms that have ripple effects at the scale of societies-- the Facebook, YouTube, and Google ranking algorithms are examples of this. We could shortly be in a place where large scale human behavior is shaped by algorithms with more data and insight about collective human behavior than any single human could have, and it will be used to optimize for money making instead of stability, fairness, or cultural values. Some society-shaping decisions/policies could even be made without any human awareness of the reasoning behind them. This is not less scary if they're being made by fragile/flaky algorithms.
The call for regulation also need not be specific to be useful, though I suspect it would help. The greatest hurdle with AI risk is getting acceptance that this is a problem that needs to be dealt with ahead-of-time. Even if Elon does nothing but improve awareness, he has been useful.
Self-driving cars are actually the least in need of extra regulation. They are already regulated, their effects are observable, there is market pressure for them to perform along lines beneficial to humanity, there is little to no incentive to extend their AI to anything more general, etc. My expectation is that the AI that is most dangerous is the AI developed behind closed doors for purely private interests on broad domains.
I personally dislike the paperclip maximizer analogy; although it serves as a meaningful explanation of the AI alignment problem, people take the literal meaning too seriously, and the absurdity of it discredits the actual risk.
AI isn't terrifying.
AI built with our current values is an horrific prospect.
On the other hand, Brooks doesn't show any indication he knows what Musk is talking about and throws out a bad summary of his position. I got nothing from this article except a slightly lower respect for Brooks.
The real minds to watch IMO is the Hinton + Deep Mind crew, and I think Yudkowsky and the fearmongers are largely correct, or at least correct enough to be taken seriously. I don't think people following the meme 'real AI researchers know that AI is limited and fragile' are on the right track. So that's my bias.
Not quite true: "Tell me, what behavior do you want to change, Elon?"
Its inevitable we will eventually solve AI. And when that day comes it will be dangerous. How easy do you think it is to control a being thousands of times smarter than you? If it was invented today we would have no ability to control it. Our best AI control mechanisms are just pressing a button to reward or punish it for it's behavior. You can't imagine any way that would fail?
Our slightly larger brains made the difference between swinging in trees and walking on the moon. But we are only the very first intelligence to evolve. Its unlikely we are anywhere near the peak of what is possible.
And this will likely happen in our lifetimes. The median expected date estimated by AI researchers is in the 2040s. Sure they can't possibly predict it very well, but who else can? And there is something to the wisdom of crowds.
Do you have references for this 2040 date? I'd love to see who's making this prediction.
Our current transistor tech is already orders of magnitude smaller, faster, and more efficient than neurons. There's no reason to expect the software stats of the brain are much better.
Here's one survey: http://www.nickbostrom.com/papers/survey.pdf
Considering that we still have many issues with basic computing - not least working out how to write code that works reliably without constant manual updates - the suggestion that we're going to start building self-improving crash-proof hyperbrains any time soon is wishful thinking.
Realistically, we don't have the first clue how to start solving that problem. As of 2017 "Throw deep learning at it" is more of a fashion statement than a practical solution.
When I say "massively parallel" I mean that a human brain (forget rest of nervous system) has about 80 G neurons. Each of which has a fan out between 10^4 and 10^5.
Architecturally this is quite different from a transistorized device with a much faster clock speed but a paucity of interconnect. Now few people say that we need to duplicate the brain (Numenta excepted -- although I have been half of a team that made a new brain and it was fun). After all an automobile does not look like a cheetah nor a plane look like an eagle. But this should give you some idea of the complexity of the problem.
(At Leela we are building something that approaches a more "brain" like computation when compared to, say, deep Q learning, and even for us, the equivalent of 2 M neurons requires about 16 TB of RAM).
I don't see any mention of confidence intervals mentioned in the survey, so what are you talking about? Means and standard deviation are given. But they are mostly useless as the distribution is very skewed. They would be infinite if the "never" people were included of course.
Perhaps not "magical", but the human brain exists in a very unique situation because it's the only one we're trying to understand with other instances of itself.
Over the past seven years most of my time has been spent in robotics research labs, and I really struggle to reconcile the state of research with the concerns of those like Elon Musk. I think a series of discussions between the major figures on each side of this would be really valuable.