Hacker News Reader

Comment by Sharlin | Hacker News Reader

No article about vintage AI is complete without mentioning the Dartmouth summer workshop of 1956, which is widely regarded as the event that kickstarted the entire field of AI research. Participants included people like John McCarthy, Marvin Minsky, Ray Solomonoff, Claude Shannon, and John Nash, among others. In retrospect, they might have slightly underestimated the nontriviality of the problems involved:

> An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves. We think that a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer.

View on HN

It is not dissimilar to how many people feel today about ChatGPT et al...

I mean, it took a bit more than eight weeks, and a few doublings of hardware capacity, but these days we have obviously made tremendous progress in making machines "use language, form abstractions and concepts, [and] solve kinds of problems now reserved for humans". The self-improvement part is – perhaps luckily – trickier, although what else are generative adversarial networks if not that? AlphaGo rapidly self-improved itself to superhuman levels without any human help after a certain point!

Stuff like checkers was a "a kind of problem reserved for humans" but was solved decades ago; you can see why the situation is similar to the context of the above quote. Time and time we have been through claims of "X solved, we are just a summer away from thinking machines" ... yet after the smoke cleared we only had a better solver for X, and nothing else. We may have a ridiculously better solver for chess, but that's about it -- it's as far from "abstracting and forming concepts" and "self-improving" as everything from the 60s was (in retrospect).

For me the biggest difference is that the current days we are happy with more black boxy methods where it's unclear if they can lead to actual progress or applicability.

LLMs reached super-human ability on their task long ago.

Everything else is a nice side quest. https://arxiv.org/abs/2212.11281

Or it’s that as computers learn to solve problems previously reserved for humans, people tend to shift the goalposts of what counts as real intelligence accordingly, similar to the so-called "God in the gaps" fallacy. (Incidentally, something very similar has been going on as our understanding of animal intelligence has improved.)

That said, I agree that however impressive narrowly intelligent tasks computers have learned to excel at, most of the research involved has not helped us make much (if any) headway towards figuring out how to create a fully general machine intelligence despite the early optimism.

But that said, for the past several decades the mainstream AI research community has mostly not even attempted to pursue AGI research, being entirely content with working on different types of narrow AI and indeed trying to dispel any illusions that their work might have anything to do with "science fiction AIs". This, of course, exactly because of those broken early dreams and the AI winter that followed.

So, at least within the field itself, I don’t really agree that there has been any tendency to assume that narrow AI breakthroughs imply that AGI is just around the corner. The current hype around GANs and LLMs is an exception, not the rule.

Amazing team! John McCarthy was the guy! Invented Lisp around this time!

Ah, the eternal summer.

Reply on news.ycombinator.com