AFAIK the AI winter was more related to the Minsky paper, and the hype and no delivery of the over-promisses made by the industry.
The way that I heard it, it was the fact that Lisp environments on Sun workstations were able to outperform Lisp machines at a much better price point. And just like that, a significant AI specific industry collapsed, and its other promises came into question.
That said, all three versions are consistent. The fact that researchers thought that they were closer than they were caused them to overpromise and underdeliver. Then when the visible bleeding edge of their efforts publicly lost to a far cheaper architecture, their failure became very visible.
Which we call cause versus effect almost doesn't matter. All of these things happened, and lead to an AI winter. And we continued to get incremental progress until the unexpected success of Google Translate. Whose success was not welcomed by people who had been trying to get rule-based AI systems to work.
Jesus. I remember when statistical translation was considered "AI".
Fun fact: One time I put "trompe le monde" into Google Translate, and it came back with the inspired mistranslation, "doolittle"
(computer science) Anything that performs better than whatever we called “artificial intelligence” a few years ago.
M-W has (IMHO rightly) these two senses:
artificial intelligence
noun
1 : a branch of computer science dealing with the simulation of intelligent behavior in computers
2 : the capability of a machine to imitate intelligent human behavior
And that's it. There's no plural "artificial intelligences" or singular "an AI" because this term never refers to a specific system, it may refer to the field or the property but not to the specific machines which (perhaps) possess some artificial intelligence as the attribute/capability. Even if you'd have a system with fully superhuman capabilities, it wouldn't be "a artificial intelligence" because you simply don't (or at least shouldn't) call things or systems "artificial intelligences", just as you don't call people "natural intelligences".
a) Intelligence that is produced "artificially, meaning by computer programming
b) Intelligence which is not intelligence but artificial so, thus not "real" intelligence.
Despite the documented failure modes (and they were many), suddenly it was possible to read articles in other languages, and it was likewise possible to make yourself understood in other languages using it. I personally know a lot of people who speak multiple of those languages. And they all agreed that it was a giant improvement. And the fact that it WAS a giant improvement was why they got rid of the previous translator.
I understand that it was terrible with Chinese. But I never used it for that.
Part of the problem was that there is a lot less grammar in Chinese than in Indo-European languages. So there are many ways to translate a given Indo-European sentence into Chinese, and you need to understand context on a Chinese sentence to properly translate it into an Indo-European one.
The many ways to translate to Chinese is a problem because Chinese flexibility in word order means that there are many choices of reasonable next word, and they didn't have enough data to tell the difference between a reasonable next word and an unreasonable one.
Going the other way Chinese may not care whether you have one apple or 10 apples, or whether Xi is a man or a woman. But Indo-European languages generally do care. So Google Translate has to guess, and often gets it wrong.
I think it was just that it became clear the projects didn't deliver anything very useful. You can't keep the hype up very long if it can't be backed up by real applications.
But some good stuff that got started then prevailed, like speech understanding and language translation. But it didn't come usable overnight.
Classic AI was a reasonable research program, but research takes time. Think of nuclear fusion.
There was little point investing money into a hardware market which did not produce cheaper and/or faster machines, given the small market.
There were a lot of interesting applications development on Lisp Machines, but there was no point to deliver them on that expensive hard- and software. Development environments were catching up. Common Lisp was actually designed to be able to deliver applications on many different platforms, even though its main influence was Lisp Machine Lisp.
So a $50k ART expert system development system was replaced by a low-cost CLIPS on machines with less hardware/software costs. It also was moved away from Lisp, as Lisp was extremely unpopular (and with almost no funding left) in the 90s.
Nowadays a native Lisp on a M2 processor from Apple is 1000 times faster than on the Lisp Machine from 1990. That's just a single CPU core, we are not even talking about GPU or Neural network functionality. Expensive 40 MB main memory from then is now 8 GB entry level.
AI was a tiny field in those days. Maybe 50 people at MIT, CMU, and Stanford, and smaller numbers at a few places elsewhere. No commercial products that were any good.
It seems like, AI research produced some fantastic results, but those systems were quickly relabeled to not be AI. Like, win at chess.
Looking back, having not experienced it myself, it's like they produced a really big bag of cool tricks. But you're not going to be doing much searching in 640k of ram. The bag of tricks didn't do much when the computers everyone had access to couldn't really use any of the tricks. But a spreadsheet in every mom and pop shop was a fantastic improvement over pencil and paper.
Right, some things came out of it but progress was slow in the more general areas. Still there was progress but hyper had to die, it was time to come back to the planet and do some database-stuff.
What was the limiting factor of the expert system approach? Something like the size of the search space or that the number of rules you’d have to write was just infeasibly large?