Artificial Intelligence at Human Level by 2029?

For those who don't follow Slashdot:
http://hardware.slash...
Anybody here think this is a mite conservative? I might also argue with the choice of the 14 challenges, but that's a separate discussion....

Ray and others use Moore's law and estimates of human brain power such as those done by Hans Moravec to arrive at these predictions. But as Eliezer points out at

http://www.accelerati...

"Suppose you want to factor a 75-digit number. Would you rather have a 2007 supercomputer, IBM?s Blue Gene/L, running an algorithm from 1977, or a 1977 computer, an Apple II, running a 2007 algorithm? And Geordie Rose calculated that Blue Gene/L with 1977?s algorithm would take ten years, and an Apple II with 2007?s algorithm would take three years."

He continues

"In artificial intelligence, this sort of thing is harder to calculate and graph. AI breakthroughs usually let you do things that previoulsy would have been outright impossible because you just had no clue how to do them."

I think this is an excellent way to point out that advances in software will have a larger impact on the arrival of anything worthy of the name "AI" than advancements in hardware. These advances often take the form of major saltations; they are much harder to predict than the hardware advances. You cannot schedule major breakthroughs.

And advances in theory of what intelligence really is will have a greater impact than other kinds of software advances. We can reach these insights through advances in brain scanning (unlikely in my opinion, but a favorite of Ray's) or through new insights in epistemology supported by experimental programming (my personal favorite).

Advances in CPU and memory technology are done by large corporations like Intel. In contrast, both algorithm and epistemology advances can be done by individuals working on personal computers or just by following "The Feynman Problem-Solving Algorithm":

(1) write down the problem;
(2) think very hard;
(3) write down the answer.

It is therefore important that we encourage many independent trains of thought about these issues. Like Herb Simon has said about this very topic, "Let a hundred flowers bloom". Independent thinkers and "Small AI" efforts (companies, non-profits) are more likely to invent these saltations than established entities such as universities, larger AI companies, and Three-Letter Agencies.

Alternatives to the western worldview might even give thinkers in places like India and Russia an edge; there are indications that our cultural biases have been preventing progress in certain directions. Increased access to more powerful computers worldwide opens the field to many new players. The internet also provides worldwide access to books and papers in epistemology, systems biology, philosophy of mind, complexity theory, and other relevant fields. Even a decade ago it would have been much harder to find this literature unless you had direct access to a major university library.

In short, I think timeframes are unpredictable because advances in epistemology will completely overshadow advances in hardware technology. Without advances in epistemology it does not matter how fast our hardware is in 2029. But a significant epistemological breakthrough could give us something worth calling "AI" (or at least very valuable) in a couple years.

- Monica

It's funny, my 2ghz Macbook Pro with 2 gigs of RAM isn't much faster then my 1.2ghz Celron with less then a gig of RAM. For well-tuned algorithms, it flies; but for day-to-day stuff, it's marginally faster.

Assuming that faster hardware would magically bring us human-level AI; we would have human-level AI now, except that it would merely run outside of real-time. Can I upload a copy of War and Peace into my computer and have it write a book report, even if it takes a decade?

What's clear is that faster hardware isn't going to magically bring us smarter software. Smarter software requires smarter programmers and better programming techniques.

Yeah, it's a pretty well known fact that today's software usually takes full advantage of today's hardware. Instead of creating programs that are designed better, companies usually opt for the bells & whistles that attract consumers at first glance. So we get fatter & fatter bloatware because new advances in hardware will let it run as if everything were OK under the hood.

However, i imagine this is not an issue for those developing AI-related software. Most of the computations needed are pretty intense, so any optimization would be looked for to make the program run that much faster.

-jim