« Featured blog | Main | Eleanor Arnason has a blog »

How's that Singularity coming along?


In Vernor Vinge's seminal 1993 article on the Singularity, he wrote:

Within thirty years, we will have the technological means to create superhuman intelligence.

Later in the piece, he elaborated:

Progress in computer hardware has followed an amazingly steady curve in the last few decades [...]. Based largely on this trend, I believe that the creation of greater than human intelligence will occur during the next thirty years. (Charles Platt [...] has pointed out that AI enthusiasts have been making claims like this for the last thirty years. Just so I'm not guilty of a relative-time ambiguity, let me more specific: I'll be surprised if this event occurs before 2005 or after 2030.)

Ray Kurzweil's 2005 book The Singularity Is Near: When Humans Transcend Biology says several times that we'll have human-level artificial intelligence by 2029; see also "Why We Can Be Confident of Turing Test Capability Within a Quarter Century," excerpted (or maybe partly excerpted) from that book.

So Kurzweil doesn't contradict Vinge's more specific claim (that the Singularity will arrive before 2030). Still, it's been 13 years since Vinge's essay, and the countdown to strong AI has apparently moved from "[w]ithin thirty years" to within twenty-five years.

None of what I'm saying is new. I don't mean to say "Look at me, I'm so clever, I found the flaw." Kurzweil knows a lot more about the current state of AI research than I do, and the above-linked article provides a lot of info about the current state of various steps on the road to strong AI.

But I can't shake a certain dubiousness, based almost entirely on the "AI enthusiasts have been making claims like this for the last thirty years" thing. Strong AI has been just around the corner for a long time, and I suspect its proponents have been making strong arguments for it being just around the corner for much of that time. (Preemptive response to obvious next question: I see no particular flaws in Kurzweil's argument, and I personally don't believe there's anything going on in a human brain that's impossible to duplicate by other physical processes. Which I guess means I don't see any reason we can't achieve strong AI eventually. But whether we're anywhere near actually achieving it is another question.)

So I guess what I'm really saying is that I think it'll be interesting to check back in ten years and see whether Kurzweil still thinks we're on target for 2029.

Kurzweil and Mitch Kapor made a Long Now bet on this topic in 2002, btw. So Kurzweil does apparently still think we're on target at least four years after setting his target date. See their wager on the Turing Test for details; see also Kurzweil's Why I Think I Will Win and Kapor's Why I Think I Will Win. And Kurzweil's responses to Kapor's argument.

You may also be interested in Jan-Willem Bats's Singularity FAQ for Dummies (predicts strong AI "well before 2030") and Charles Stross's entertaining Singularity! A Tough Guide to the Rapture of the Nerds, which tongue-in-cheek predicts human-level AI "some time just after 11:14am on Sunday, July 16th, 2017."

And this entry started out intended to be about three lines long, and I'm now running late, so I'd better run.


Did you see the thread on Ben Rosenbaum's blog here?

The boys in Ben's thread did get quite involved with the question of "it's just a trick" -- and myself, I can't help but wonder if part of the "every thirty years" effect comes from the bar constantly moving up.

Take Deep Blue beating Kasparov, for example: I remember lots of people catching their breath at that. "A computer beat the world's greatest chess master! Does that mean computers really are smarter than humans?"

Now, not so much. It's just chess, right? Just a strategy game, right? If there's one thing computers should excel at, besides numerical integration, it's strategy games. That's a far cry from passing a Turing test, eh?

So maybe the singularity really will happen in 2029; maybe the really significant event will be some chatbot passing a Turing test, and everybody shrugging and going "So what? It's just the Turing test."

Hee--I totally missed that thread on Ben's blog (been out of touch the past couple weeks); thanks much for the link! Great discussion.

It touches on a bunch of things I thought about saying here but didn't, too. Like the fact that some strong-AI proponent are still saying that sufficient computational power will automatically result in strong AI.

But one of the reasons my comments about the Kurzweil piece weren't more negative is that he's talking about some other avenues as well. For example, he suggests that in the future we'll be able to "create highly detailed scans of human brain functioning and thereby [reverse engineer] the human brain" and notes that "every aspect of understanding, modeling, and simulating the human brain is accelerating." Brain simulation (whether electronic or physical) always seemed to me a more promising avenue for strong AI than just throwing MIPS at the problem--though scientist friends of mine who know something about the subject have indicated that a whole lot of how the brain works is as mysterious to us as ever.

Kurzweil also quotes Rodney Brooks, director of the MIT AI Lab, making basically the same point as Jackie: "Every time we figure out a piece of it, it stops being magical; we say, Oh, that's just a computation." And there's a cute one-panel comic lower down on the page that says something similar. I think it's a good point--though I think that at the point when a computer can in all ways pass for human (whatever that means; Ben et alia made clear that that phrase's meaning isn't nearly as clear as it initially seems to be), if we're still saying "we understand how that works, and it's just a computation," then we may have to say the same about human brains.

...One thing that I rarely see discussed in these discussions (and it was touched on in the thread at Ben's blog, but I don't think anyone quite said this explicitly, though I may've missed it) is that Eliza has been passing a low-end Turing test for decades; there are people who come away from a conversation with it convinced that they've talked with a human. On the flip side, to anyone who knows something about how Eliza works, it's transparent. One of the things that makes me dubious about near-term strong AI is reading Loebner Prize transcripts; last time I looked, a couple years back, the phrasings and strategies and grammatical problems of the winners still read an awful lot like Eliza.

And Eliza only passes even the low-end TT in certain contexts. Ben indicated that jenny18, a sex-oriented chatbot based on Eliza, regularly passes the TT, but in fact all of the transcripts there that I looked at involved the human saying, fairly quickly, "are you a bot?" Even by the low standards of low-end cybersex chat, Eliza's randomness and clunky responses sound pretty non-human.

Similarly, a commenter on the original Aaronson blog entry (that sparked David's entry that sparked Ben's) points to AOLiza, essentially the 1966 version of Eliza unleashed on IM chats, but the best of those were (the site owner suggests) probably conversations with people in on the joke, and the rest aren't very convincing. It's a cute and fun experiment, but I'm not seeing much evidence that chatbots are passing even a weak TT.

Jenny18 certainly passed this one(NSFW). The trick in finding one she passed is to look at how many KB the file is! Though also looking at the top 4, no one says "are you a bot?" One guy says "you're fucked up", but that isn't exactly a Turing Test failure -- more a sign of cybersex incompatibilities of taste.

If anything, the prevalence of the remark "r u a bot?" lends credence to the argument that this really IS the Turing Test. The principal objection to the "trolling chatbot" variant of the Turing Test, after all, is that, unlike Turing's original Imitation Game, participants do not know that they are trying to unmask a computer; they are unawares. But "r u a bot" belies this, implying a degree of sophistication in jenny18's playmates. At some point, though, when half the entities trolling IRC and AIM are bots and this is well-known, it becomes essentially identical to what Turing proposed (particularly since a large number of others are simultaneously playing the gender-passing Imitation Game he proposed as an introduction to his idea).

Let's beware also of conflating strong AI (passing a much more robust Turing Test -- say, being a valued and beloved conributor to the philosophical debates on David Moles' blog for several years before coming out as an AI) and the Singularity. The Vingean Singularity is a much higher bar -- it is the point where a) an AI is better at designing AIs than we are and b) there turns out to exist in the universe a linearly increasable property "smart" such that "things of smartness N can design things of smartness N+k for any smartness greater than S".

I don't actually believe such a propery exists; I think it misconstrues the nature of "smartness". So I guess I disbelieve in the Vingean Singularity. I do believe we are in the midst of a Strossian singularity (irreversable change in the nature of human life and consciousness, driven by technology) but then, we have been in the middle of one or another Strossian Singularity since the Neolithic, so that's not really news...

say, being a valued and beloved conributor to the philosophical debates on David Moles' blog for several years before coming out as an AI

What are you trying to say here, Benjamin?

Also, is the pronoun "she" the standard way of referring to jenny18? That's a curiosity in and of itself.

Disbelief aside, the "smartness" you describe will require a combination of "evolved" intelligence and "learned" intelligence -- at least at the lowest, human value of N -- because the creators of the first N+k AI will be drawing on the accumulated knowledge of all computer scientists since 4000 BC.


So maybe the singularity really will happen in 2029; maybe the really significant event will be some chatbot passing a Turing test, and everybody shrugging and going "So what? It's just the Turing test."

I take this prediction to mean that in 2029 everyone will have forgotten the real meaning of the Turing Test. If I come to think of meaningful conversation as something that can be achieved by "just a trick," I hope someone will remind me of my posts in these threads.


The principal objection to the "trolling chatbot" variant of the Turing Test, after all, is that, unlike Turing's original Imitation Game, participants do not know that they are trying to unmask a computer;

This may be some people's principal objection to chatbots, but it's not mine. My objection is that IMing is an extremely impoverished mode of interaction; as I said in the other thread, the fact that modern life offers many such modes should not lead us to underestimate the Turing Test.

Let's beware also of conflating strong AI [...] and the Singularity.

I completely agree. Alas, however, most proponents of the Singularity seem to think the two are identical.


By what measure do we assert that these conversations are "meaningful"? Conversely, by what measure do we assert that a "tricky" chatbot wouldn't be?

Jackie wrote:

By what measure do we assert that these conversations are "meaningful"?

Oh, hey, that's a thorny one.

I pause a moment to anticipate the coming avalanche your couple of tossed pebbles should set off.

The general idea is that all (or at least most) meaning is cultural, right? From language-acquisition on up? That makes the question of who gets to decide what is meaningful pretty fraught. And so for me, the temptation is to try to approach the question from an acultural point of view. Is there some way to describe the characteristics or effects of a meaningful exchange without having access to understand the meanings themselves?

But that gets pretty tangled. Maybe it's more about behavior: do we keep coming back to talk to Eliza once we know there isn't a human being on the other end of the line? Sometimes, just for fun, but not with any real investment in the conversation.

Well, in retrospect a.) I'm pretty sure the "acultural measure" ground was well-covered by Ben's discussion, and b.) there were two ways of reading Ted's comment, and I'm not sure I picked on the intended meaning.

But in keeping with my (potential) mis-reading of Ted, I was thinking of a hypothetical, more-advanced, "tricky" chatbot of 2028. Something that's close enough to passing that everybody sits around and argues about the "real meaning" of the Turing test for literally years.

And yeah, culturally-impacted definitions of "meaningful". Hey, did you see the "sociologist lost in a see of physicists" thing Susan was linking to? And the related blow-up which David linked?

Jackie, I took you to be positing a chatbot that could, for example, participate in a lengthy blog discussion the way Ben, David, Jed, you, and I sometimes do, but if we propped the hood open we'd see it was just some conventional algorithms running very very fast, and everyone would agree that it wasn't remotely conscious. I don't believe this will ever happen.

how about: positing a chatbot with "conventional" algorithms circa 2028? But non-existent in 2006.

I don't know what people will be calling "conventional" in 2028. Maybe by then software that everyone agrees is conscious will be extremely well understood; maybe the mechanism by which consciousness arises in the human brain will be so well understood as to be considered "conventional," too. But I understood you to be positing software that everyone would agree is not remotely conscious -- the way that everyone agrees that Deep Blue is not remotely conscious -- and yet can carry on a conversation as well as you or I. That is what I don't believe will happen.