Is there a reason for initial overconfidence of scientists and engineers working on artificial intelligence in the 1960s?

Question

I just started an AI & Data Mining class, and the book. AI Application Programming, starts off with an overview of the history of AI. The first chapter deals with the history of AI from the 1940s to present. One particular statement stuck out at me:

[In the 60s] AI engineers overpromised and underdelivered...

What was the reason for the overconfidence? Was it because of mathematical prediction models showing that a breakthrough was around the corner, or due to the ever-increasing hardware capability to take advantage of?

Answer 1

My personal opinion is that it was due to hubris. There were some mighty big egos walking the halls of MIT, Stanford, etc. back in the 60s and 70s and they just knew they had cracked this problem. Right.

Although I wasn't part of that universe in those days, in the mid-to-late 80s I was working with similarity searching. Our work was initially based on research done by Gerard Salton at Cornell in the 60s, which used weighted attribute vectors to represent documents and queries. It actually was a useable approach, but when neural nets went down in flames (at least until they discovered back propagation), Salton's work was included with it because of similarities (pun intended) to neural nets. He was trying to do something different, but there were several years where he was lumped in with the rest.

Every time someone comes up with a solution for the Current Brick Wall™ they get very excited and declare AI to be a solved problem. Only it's not. Because behind that brick wall is another one. This cycle has repeated over, and over, and over again, and not just in AI. I firmly believe that all prospective computer scientists and engineers should be required to take a semester-long class in the History of Computing, with special emphasis on the number of Next Big Things™ that went up like rockets ... and then made a very large crater in the valley floor.

Addendum: I spent the Labor Day weekend with an old friend and we talked a little about this. Context — figuring out what that means, how to represent it, and then how to use it — emerged as possibly the single biggest hurdle to be cleared. And the longer you look at it, the bigger a hurdle it becomes. Humans are capable of amazing, near-instantaneous partial-pattern matching of "what is happening" against a vast store of "what has happened before," and then combining that knowledge of the past with the present situation to create a context in which understanding can lead to action. For example, we can use it as a powerful filter of "things we can/can't ignore" as we whiz down the Waldo Grade at 60 MPH with traffic 4 lanes abreast and separated by only 3 or 4 feet (or less!).

On the spectrum of stuff > data > information > knowledge > understanding > judgement we are still straining to get to the information/knowledge steps, and even that is limited to highly constrained domains of discourse.

Answer 2

Quite simply, they massively underestimated the scale of the problem at hand, especially where combinatinatorial explosion is concerned. Many AI solutions work fine for "toy" samples, but fail hard when they scale up to human-level problems.

Arguably, they were also simply inexperienced. AI as a field had (relatively) only just been invented in terms of practical applications, so nobody had significant experience applying theory to anything.

Answer 3

I can think of a couple of reasons.

AI experienced such rapid success with some of the toy problems tackled in the late 50s and early 60s, that they overestimated what they had accomplished. ELIZA and SHRDLU stunned folks despite being relatively simple programs. Unfortunately, a large part of what made those programs stunning was really just novelty. Nobody is very impressed by a conversation with ELIZA today, but at the time folks thought it was near miraculous.

Also, as problems are "solved" or at least become tractable, folks no long think of them as AI. Code optimization used to be an AI problem. Statistical learning budded out of AI into its very own speciality, and took speech recognition with it. As data mining becomes mainstream it will loose its association with AI. Over time AI forgets its successes and gets stuck holding on to the intractable and insoluble problems, and it ends up looking like a flop.

Answer 4

I think people in the 60's used their own human experience to divide problems into "hard problems" and "easy problems": Things like winning chess, solving logical riddles, solving mathematical equations seem hard to us humans. Things like understanding natural languages or finding the outlines of objects in an image seem easy, because our brain does all the work without conscious effort. When we try to explain how we do those things, we come up with simple explanations like "English sentences always have the structure subject-predicate-object where subject can be a simple term or a phrase...", or "I'm looking for edges and connect them to object boundaries". Today we know things aren't that simple, but only because all the simple (and many not-so-simple) solutions have been tried and didn't work.

Besides, this fallacy didn't start in the 60's: There's centuries of research on how to solve those "hard problems" (heuristics, game theory, decision theory, mathematics, logics, etc.) but I'm, not sure anyone ever bothered to research how natural languages might be parsed before the 1950's.

And even today, you can regularly find questions on stackoverflow, where people ask how they can parse English sentences, estimate the age of a person in an image, judge if an image is "safe for work" or if two images show the same thing. I don't think the people who ask these questions suffer from too much hubris or arrogance: These problems just seem so simple, it's unbelievable that there is no simple algorithm to solve them.

Answer 5

AI has a long history of disappointments, but I think many critics often over-simplify what happened, such as with your quote "1960's engineers overpromised and underdelivered".

In the 60's, AI was the domain of a relative handful of researchers (the field wasn't really sufficiently developed yet to call it engineering), mostly at universities, and very few of them were accomplished programmers.

The sudden availability of computing machines in the 1950's had led to great expectations for automation, particularly in machine translation of natural language, playing chess, and similar problems. You might find some actual predictions of success from those days, but the promises inevitably came BEFORE anyone tackled one of those problems in depth. (Or, they wrongly assumed one success guaranteed another, such as expecting to be able to implement good chess playing after Samuel had so much success with checkers.)

Also, be wary of any claims of "they said", "they felt", "they thought", etc.; retrospective opinions (like this one!) are easy to throw around, while documented evidence of actual predictions by "experts" (those who actually tried solving a given problem) can be much harder to find.

Overpromising and undelivering has always been a symptom of software development, regardless of the specific field where the programming is applied. A major difficulty with AI is that non-trivial problems are beyond the capabilities of most engineers. For example, although Charles E. Grant's answer categorizes ELIZA and SHRDLU as "relatively simple", I'd say that's true only of ELIZA (which most first-year programming students could probably implement without much difficulty). On the other hand, SHRDLU is a large, extremely sophisticated program that most programmers would have a very difficult time inventing, let along implementing. Indeed, two teams of university students couldn't even get the source code fully running again, and SHRDLU-like abilities are still hard to find nowadays, over 40 years later.

Since AI is probably one of the least understood and most intractable problems where computers can be applied, overall I'd say progress in AI has generally been at par for the course. There are still high expectations, and our hardware speed and capacities have increased tremendously since the 60's, but I'd say engineers' abilities and understanding of AI aren't improving all that much, so a holy grail like passing the Turing test is still probably a long way off, and overpromising and underdelivering will probably continue for some time.

Answer 6

I think the reason was arrogance. Had I been an Engineer in the 60s working on AI, I would have been pretty arrogant myself.

I think in order to accomplish great things, you have to reach for great things. So overpromising is not necessarily a bad thing as long as you don't exceed the limit. Scientist today are promising things I don't believe will be possible, but if they don't reach for that, we'll miss out on what will be accomplished as a result.

Answer 7

It can be very hard to get somewhere when you don't know where you're going.

If we had some sort of reasonable explanation of what intelligence is and how it works, maybe we'd have a shot at effectively mimicking it. The Turing test is fascinating and useful, but is probably not sufficient to help us model true intelligence. For all we know, a "model" of intelligence might not be sufficient for true intelligence either.

Answer 8

Well, I would say it's more or less the same thing that is happening with OWL right now. Look around, and try to draw parallels.

Sounds good on paper, seems to work well on toy problems, gets incredibly complicated on most real data.

Answer 9

In addition to the good answers given, two observations:

Some quotes of the day seem to imply that many of the researchers were thinking that the trivial solutions could be scaled up once faster computers were designed. For some kind of learning systems this was very true, but for the kind of thing I think the OP is referring to it really didn't get any better at scale.

Researchers at the time had a very low estimate of the complexity of the human mind (focus on ideas like Turing test, the idea that people only use a small percentage of their brain ever, etc). AI at the level of a simple animal has been achieved by some measures as things scaled up, but the jump to a human level AI was a lot bigger than expected. This has lead some researchers to try learning baby systems and other growth/evolution based simulations as an attempt to bridge that gap.

Answer 10

One reason was the success that we were having ELSEWHERE in the 1960s. We had just launched into space, and would soon land a man on the moon. We had just discovered cures for polio, and other major diseases.

But "artificial intelligence" was a different animal from the "engineering" problems we faced then. It was a "reasoning," rather than "mechanical" problem.

In short, AI (in the 1960s) was an idea "whose time had not yet come." It took more development, in subsequent decades, before it became as accessible as the other problems.

Answer 11

Another reason may be that mastering a computer / writing computer programs gives us little control freaks a feeling of omnipotence - in fact, one creates little universes, albeit closed ones.

This plus the lack of philosophical/epistemological education and naive trust in simple explanations like "Intelligence is nothing but ...." can lead to hubris.