Human Level AI

Some of the fellows at the AGI Roadmap working group are debating:

A:

> p. 9, Artificial Scientist Test: Almost every human can't do this. Are you suggesting that most humans don't have human-level intelligence?? (None of us have gotten a Nobel Prize!) It is important to set the bar at a reasonable level. For example, an IQ of 85, say, should be good enough to have human-level intelligence. We needn't set the bar at the genius level!

B:

This can be a touchy subject. I think that a lot of people in our field aim for an AI that is at least capable of AI programming.

Personally, I think that A is putting the bar too low: an AI that could only score 85 on an IQ test would almost certainly fall short of human-level intelligence, because an AI is likely to be better at the kind of abstract symbolic manipulation that IQ tests measure than humans are, compared to its other abilities; IQ tests will have to be recalibrated, and possibly re-written entirely, to give an accurate measure of artificial intelligence.

On the other hand, B is surely putting it too high. Only people at the IQ 140 (or so) level are typically capable of writing the kinds of AI programs we have now; and nobody has written one - so far - that is actually successful in capturing human intelligence.

It is clear that there is a level of intelligence and competence that the average human has, that these descriptions bracket. And there are two things about having AIs at that level that will be critical in shaping the future when human-level AI is achieved.

First, AIs will be able to do the jobs that average humans do, and in general do them better. Better because of special additions: AI taxi drivers will have built-in GPSs; AI accountants built-in calculators; AI librarians built-in Google access. Better because they will work 24/7, not get tired or sick, not make mistakes, not lose their patience. And most importantly in an economic sense, better because they will get cheaper along a Moore's Law curve. AIs like this will clearly revolutionize the economy.

But, secondly, what they won't do is improve themselves at an accelerating rate. An average human is incapable of that, and an average human-level AI, while being a perfectly good policeman or tax accountant, will not be able to any more than a human policeman or tax accountant would be. They will improve themselves the way humans do, by practice, study, and exchanging information, albeit probably somewhat faster than humans do this.

So we can build a world full of human-level AIs, capable of completely revolutionizing the economy and giving all humans a lifelong vacation/retirement, without ever building one capable of ``recursive self-improvement'' or turning itself into a superintelligence overnight.

Whether we will or not is of course another matter.