Artificial Intelligence and the year 2047

[ciphergoth]

Clarification: By "smart" I mean general smarts: the sort of smarts that allow you to do things like pass a Turing test or solve open problems in nanotechnology. Obviously computers are ahead of humans in narrow domains like playing chess.

NB: your guess as to what will happen should also be one of your guesses about what might happen - thanks! This applies to wriggler, ablueskyboy, thekumquat, redcountess, thehalibutkid, henry_the_cow and cillygirl. If you tick only one option (which is not the last) in the first poll, it means you think it's the only possible outcome.

[Poll #1103617]

And of course, I'm fascinated to know why you make those guesses. In particular - I'm surprised how many people think it's likely that machines as smart as humans might emerge while nothing smarter comes of it, and I'd love to hear more about that position.

Comments

[zwol.livejournal.com]

I marked the "won't happen in 40 years" and "won't exceed human intelligence in 40 years" boxes, with preference for the first one.

I'm currently in a field that came out of, among other things, frustration at the failures of classical AI, and I have friends doing work at the cutting edge of machine translation and decision support. This makes me more pessimistic about human-equivalent AI than I might be otherwise, because I can see the road from here to there, and not very far from where we are now, there's a cliff face, the road goes straight up, and I don't see any way to climb it.

Science and engineering method is to take intractable problems apart into small pieces. Often those small pieces are tractable, and often, when you put the solved problems back together you find you've got an acceptable solution to the original. That has worked spectacularly well for us to date, but it didn't work in classical AI and it's not working in modern AI either. Every decomposition of human-level intelligence that's so far been tried produces a bunch of small problems that we can solve but don't go back together into the original.

To put it another way, to make much more forward progress on any of the things that are generally thought of as subcomponents of intelligence — natural language, speech recognition, spatial reasoning, object identification, decision making — we will have to put them back together and solve it all at once. We don't have any idea how to do that.

It gets worse. Confronted with this problem (it has been foreseeable since the 1970s if not earlier), my field decided to go study real brains for a while. We have a pretty good empirical understanding at this point of how a human child develops, learns stuff, becomes a functioning adult. And it is dependent in detail on the child's brain being part of the child's body. You can cause horrible developmental problems by, for instance, raising a kitten for the first six weeks of its life in a box with no visible detail other than vertical stripes. The cat is ever after unable to recognize horizontal lines.

Thus, the science fiction trope of the disembodied intellect in the computer is never going to happen. (In particular, contra suggestions above, Google is not going to turn into Skynet.) If we want a true AI it's going to have to be in the form of a biomimetic robot. Furthermore, the easiest way to implement the thinking part is going to be with a detailed mimic of the human brain, not necessarily in meat, but including all its limitations. In particular the robot will not be able to learn faster or via a qualitatively different method than human children do. (Biological brains do a whole bunch of stuff, especially to do with memory, with resonance loops at 5-20Hz, and if you mess with the timing you get a fascinating variety of cognitive disorders.)

I think the construction of such a robot is feasible and might even happen in the 40-year timeframe, but I wouldn't call it a sure thing. And I think we will be able to make them very smart, but not in any qualitatively different way from very smart humans.

[meico.livejournal.com]

I agree with almost all of what you are saying above but disagree on a few minor, but important points. One thing I think I should point out is that an embodied intelligence doesn't have to necessarily be embodied in the real world (though that would probably be helpful). It can be embodied in a simulated world.

Embodying it there makes so many things much easier. For example, simulated skin that detects touch becomes a simple by product of your physics contact and penetration solver and not some intractable materials engineering problem...

Anyway, I'm working on the cutting edge of reality simulations (games) and have no doubt that they'll soon be good enough to place embodied intelligences into and have them develop real world knowledge and skills (actually some people are already doing exactly this).