Feb 20 2008
Café Scientifique: Robots ruling the world
On Monday night the Edinburgh Café Scientifique was about AI and robotics — specifically, “Why Robots Will Never Rule the World”. I also managed to convince a bunch more people to come along, which is always good. (In fact, it was very busy this time, which is quite difficult because it’s held in a small and awkwardly-shaped room.)
The gist of the speaker’s argument was
Computing power, as measured by processing ability, is increasing exponentially (as observed by Gordon Moore). But battling against Moore’s Law is Wirth’s Law: Software gets slower, faster than hardware gets faster. Software must increase in complexity to tackle the latest problems of machine intelligence, and if there’s one thing we’re not very good at it’s minimising the effects of complexity when designing software.
Dijkstra said that sharpening a pencil with ten blunt axes is no more effective than with one blunt axe. If we don’t know how to solve a problem it doesn’t matter how much faster our machines are. The pace of improvement in AI goes at human thinking speed, not silicon fabrication speed.
On the physical level, the square/cube law applies just as much to robots as it does to humans and cockroaches. Anthropoid robots will not have much greater strength and agility than us, because of the limits of materials at their body size.
It was quite an interesting talk but pitched at the wrong level — there were an awful lot of “beards” there (and an XKCD t-shirt) and I don’t think Moore’s law was a great unknown to them all.
But I think a more compelling argument than the ones mentioned here is some historical perspective. If you look at the numbers, bacteria rule the Earth, and they didn’t need intelligence to do it. Our intelligence hasn’t helped us overcome some very real physical barriers preventing us from ruling the planet — harsh climates, limited resources and intense competition.
I think that’s dead wrong.
We’re close to ‘simulating’ a rat brain (there is an excellent article in Edge you should read about that [not online AFAICT]), and there is no reason to assume we won’t be able to similarly simulate human brains.
Whether a perfectly accurate simulation of a brain is a real brain is a stupid question as far as I’m concerned. If it acts in exactly the same way then pragmatically I need to treat it in the same way (if say it is connected to a robot body with lasers or something).
There might not be a singularity of intelligence where computer brains become enormously more intelligent than us, but it doesn’t require someone with an intelligence a billion times larger than ours to take over the world. History has shown it can be done by people with equivalent or slightly higher intelligence.
I suspect but cannot prove that once we have human level intelligence in computers it won’t be that tricky to get super human intelligence. At the very least we could have a human level intelligence in a computer that ran a hundred times faster. OK, so in some sense it’s not more intelligent, but with a hundred times more time to consider what to do (and possibly some kind of VR environment that allows it to interact with plans) I suspect it could outdo us quite easily.
You should read Diaspora by Greg Egan. It doesn’t directly address these issues but it is set in a world of AI intelligences in super computers, ‘flesher’ humans, and AI intelligences in robots. Some of these issues are part of the backstory. What I found so interesting was the exploration of what life would be like as an AI living in a super computer. I strongly recommend the book.
I don’t really understand what you think is wrong. Intelligence isn’t a definable problem and the higher-order, strategic problems that we can solve (chess etc.) are not nearly as useful or complex as the mundane, low-level problems (playing football like Pele).
The connectionist approach you mention, with the rat brain, is as intractable as any other. Modelling neurons isn’t a very difficult problem, but mapping the human brain to the level required to recreate it in computer is much harder.
The question of strong AI is not in doubt in my mind, and I don’t think the speaker was averse to the idea either. But the stumbling blocks are going to be our own intellect.
Given the article in EDGE I’m fairly confident that (baring any serious unforseen problem) we will have a simulated rat brain (that they intend to connect to a real world rat robot) in the next 5 years. That’s not the 50 years you hear when someone talks about human like AI or fusion power, it’s 5 years. So when you say the problem is intractable I believe that is just plain wrong.
I can’t remember off the top of my head how they were going to map the rat brain (I think maybe there is a method where a brain is sliced very finely and each slice is programmed in?), but that problem was more or less solved. The step up from a rat brain to a human brain is fairly trivial (ok, human neurons might be a bit different but it’s not a major problem). The only problem we’re going to run in to are any ethical issues.
I’m not sure why you mentioned higher order / lower level problems as I don’t think anything I said was relevant to that, but I disagree here. A computer system that you can communicate with properly with natural language, and otherwise acts as a human (except having knowledge / control of computer based information or systems) would be an extremely useful thing to have. Who wouldn’t find the ships computer on the Enterprise (only more so) useful?