Will AI pass the Longbets version of the Turing test by the end of 2029?
684
3.4K
1.5K
2030
48%
chance

This is based on the inaugural longbets.org bet between Ray Kurzweil (YES) and Mitch Kapor (NO). It's a much more stringent Turing test than just "person on the street chats informally with a bot and can't tell it from a human". In fact, it's carefully constructed to be a proxy for AGI. Experts who know all the bot's weaknesses get to grill it for hours. Kurzweil and Kapor agree that LLMs as of 2023 don't and can't pass this Turing test.

Personally I think Kapor will win and Kurzweil will lose -- that a computer will not pass this version of the Turing test this decade.

((Bayesian) Update: But I admit the probability has jumped up recently! I created this Manifold market almost a year before ChatGPT launched.)

Related Markets

UPDATE: Originally I was confused about the resolution date (beginning of 2029 vs end of 2029) and have now adjusted this to match Longbets. See discussion in comments.

Get Ṁ500 play money

Related questions

Sort by:
justifieduseofFallibilism avatar
Timothy Curriebought Ṁ0 of NO

Manifold normally:

Manifold whenever something is in the distant future:
Maybe it is 50/50?

AngolaMaldives avatar
Angola Maldivespredicts NO

@justifieduseofFallibilism I think there's an economic reason for this; markets that won't resolve for a long time need a high expected return to beat shorter-term investments, so people who believe in YES will bet up to 50% and then say 'further betting wouldnt give me a good enough return (better than double)', and people who believe in NO will bet it down to 50 and say the same. Not a robust thesis by any means and there's no real reason for 50% to be special as a betting threshold other than psychology (although if both sides mostly demanded, say 2:1 odds it'd still anchor the market in the broad vicinity of 50), but i think it's a factor

ErickBall avatar
Erick Ballbought Ṁ30 of YES

@AngolaMaldives It could very well work that way psychologically, but in terms of the actual rate of return needed to justify the investment everything longer than about 6 months is roughly equivalent. Actually, maybe it would help to have an easily accessible calculator/formula that shows people what their effective rate of return is after accounting for loans? I'm not sure exactly how it would work though.

DanielParker avatar
Daniel Parkerbought Ṁ7 of NO

My bet is that the AI will easily lose the more stringent Turing Test because the AI will be too competent. It will be able to tirelessly answer questions with immediate access to much more background information than any human. If it is aligned properly, it will also readily admit that it is an AI. I don't know whether AGI will be achieved, but there are certain predictable areas where even now an AI will be consistently more competent than an average person (or even an expert) so it should be fairly easy to tell the difference between an AI and a human. Maybe these shortfalls could be partially addressed by certain workarounds (e.g. instructing the human participants to also claim to be AIs so that honesty is not a mark against the AI, adding time delays, allowing human participants to research things they didn't know about, and so on), but my guess is that an AI expert would still be able to consistently tell which one is the AI.

dreev avatar
Daniel Reevespredicts NO

@DanielParker I think it would be poor sportsmanship for Kapor to accept a victory on that basis. The spirit of the wager is to predict AGI. And I expect it not to be too hard to train an LLM to, y'know, tone it down and emulate human stamina/competence. Or, like you say, factor out the honesty factor and add time delays.

Another idea that should work without retraining an LLM: pass all responses through an anonymizing proxy. It could both strip admissions of being an AI and equalize throughput. Also the judges could simply agree not to use questions like "please translate this into 20 different languages" that are prognostic by being harder for humans than machines.

So I don't agree with betting on NO based on an expectation of the AI having superhuman competence.

PS: I just tried the 20 languages thing with GPT-4. It can do it fine! But then I said "now role-play a typical human with a technical PhD and answer again" and it nailed that too (it did 3 languages and then was like "and that's about as far as my language proficiency goes!").

PPS: Ironic that I'm making this whole argument against NO even though I'm a huge NO bettor here. I guess I only want to win for the right reasons!

DanielParker avatar
Daniel Parkerpredicts NO

@dreev The problem is that the Turing test isn't a great test for AGI (no matter how fair the attempts are to administer it). LLMs may look human on the surface, but underneath it they are quite alien, and it isn't too hard to detect that alienness if you know what you're looking for. Competence in things like language translation is part of that, but there are other areas that could probably be thought of as "style" rather than "competence", per se. If AGI is ever achieved, I expect that it will be roundly denied by most people until long after the fact, and tests like the stringent Turing Test could be part of that denial.

IMHO, the original Turing Test (which was surpassed long ago) is a more useful metric. It doesn't test for AGI, but it does tell us that humans are gullible and can easily be tricked by a computer; or to put it more positively - computers can cheaply do work that otherwise would have required expensive human labor (in general this is not "stealing jobs" because computers generally perform labor which otherwise would not have been done at all). The development of computer capabilities is an advancing process and there probably won't be any one point on the line that you could point to and have a general consensus to say "that's AGI". If there ever is a consensus that AGI has been achieved, it will only be long after the fact.

DavidBolin avatar
David Bolinpredicts NO

@dreev It would NOT be "poor sportsmanship" for Kapor to accept a victory given AI's inability to imitate a human.

Turing directly admitted that this test is far more stringent than testing for intelligence; he was fine with that. Because given that you have the far more stringent test, if it is passed, then you know AT LEAST that the thing is intelligent; it does not mean it could not be intelligent without passing.

The test is perfectly clear, and Kapor will win the bet even if that is the reason.

dreev avatar
Daniel Reevespredicts NO

@DavidBolin If it's fair to characterize it as "the AI was unable to imitate a human" then sure. But, like, if I'm trying to mimic a toddler and keep forgetting to make the adorable grammar mistakes they make and give myself away that way, have I demonstrated that I've not yet reached toddler level intelligence?

We could say it depends what the spirit of the bet is. How human-like AI will be or how transformative it will be.

In any case, I think there are plenty of strategies available to patch the problem where the AI is detectable by virtue of being hypercompetent. And I'd expect Kurzweil and Kapor to be happy to employ such strategies because I expect they care about AI capabilities and transformativeness more than human-mimicry specifically.

DavidBolin avatar
David Bolinpredicts NO

@dreev "have I demonstrated that I've not yet reached toddler level intelligence?"

Of course not. But you also did not demonstrate that you did reach it, given the method of demonstration being that you plan to imitate them.

jojomonsta avatar
Jojopredicts NO

funne market

ErickBall avatar
Erick Ballbought Ṁ60 of YES

I'm confused about the win condition. Longbets says:

>Each Turing Test Session will consist of at least three Turing Test Trials.

>For each such Turing Test Trial, a set of Turing Test Interviews will take place, followed by voting by the Turing Test Judges as described below.

...

>The Computer will be deemed to have passed the Turing Test if it passes both the Turing Test Human Determination Test and the Turing Test Rank Order Test.

So does it need to "pass the Turing Test" in 2 of the 3 Turing Test Trials for Kurzweil to win? Or just 1 of 3?

Jai avatar
Jai D (Jai)bought Ṁ30 of NO

There's a one year difference between the deadline in this bet (Dec 31 2028) and the deadline in all other linked versions of this bet (Dec 31 2029), meaning that it's likely that there won't be any evaluation of the bet before this market's deadline is up.

Jai avatar
Jai D (Jai)predicts NO

@Jai from the terms:

Ray Kurzweil, or his designee, or another member of the Turing Test Committee, or the Long Now Foundation may, from time to time call for a Turing Test Session to be conducted and will select or provide one Computer for this purpose. For those Turing Test Sessions called for by Ray Kurzweil or his designee or another member of the Turing Test committee (other than the final one in 2029), the person calling for the Turing Test Session to be conducted must provide (or raise) the funds necessary for the Turing Test Session to be conducted. In any event, the Long Now Foundation is not obligated to conduct more than two such Turing Test Sessions prior to the final one (in 2029) if it determines that conducting such additional Turing Test Sessions would be an excessive administrative burden.

Frogswap avatar
Frogswappredicts YES

@Jai I suspect this is a mistake. The short form of the bet just says "by 2029", but they mean "by the end of 2029". The spirit of this market seems to be "who will win the bet", based on the references to other markets in the description and that @dreev hasn't made any markets with the correct year. But, as written, it's explicitly not reflecting that, so I'm not sure whether it's intentional or whether it should be changed if it isn't.

ShadowyZephyr avatar

@Frogswap Every time someone mentions this. THATS NOT A MISTAKE. By 2029 can mean either beginning or end, and it’s a roughly even split among markets.

Frogswap avatar
Frogswappredicts YES

@ShadowyZephyr Has this come up before? I did skim over the comments here before I wrote the above, and I looked a little more thoroughly now with no luck

Frogswap avatar
Frogswappredicts YES

@Frogswap This comment seems to suggest that it is a mistake, as the longbets version states, "If a Computer passes the Turing Test, as described above, prior to the end of the year 2029, then Ray Kurzweil wins the wager. Otherwise Mitchell Kapor wins the wager."

https://manifold.markets/dreev/will-ai-pass-the-turing-test-by-202#ewwnwoKUqOwiaTO8wB7S

dreev avatar
Daniel Reevespredicts NO

Crap, yes, this must be my error. I remember seeing "January 1st, 2029" somewhere and so I made this market be "by the end of 2028" but I've now scoured both Longbets and Metaculus and it's all consistently December 31, 2029. I'm inclined to just quickly fix this here and presume that that one-year difference won't have influenced anyone's trading meaningfully. But please holler if that's not the case.

IsaacKing avatar
Isaac Kingpredicts NO

@dreev My NO position was largely based on the ending date.

Jai avatar
Jai D (Jai)predicts NO

@dreev Me too. I noticed the discrepancy, made a bet based on it, and then immediately posted about it. I figured that the price would quickly correct, at which point I would sell.

dreev avatar
Daniel Reevespredicts NO

@IsaacKing @Jai What's your personal subjective probability that this happens by the end of 2028? And same question for end of 2029?

Jai avatar
Jai D (Jai)predicts NO

@dreev Subjective probability that this particular bet is settled in the affirmative in favor of Kurzweil:
- by Dec 31 2028: 3%
- by Dec 31 2029: 37%

dreev avatar
Daniel Reevespredicts NO

@Jai Well that's fascinating. Now I'm dying to hear your world model that has so much probability mass concentrated on the specific year 2029 for an AGI breakthrough.

Jai avatar
Jai D (Jai)predicts NO

@dreev My probability mass is very specifically about this bet resolving, which requires a specific test to be performed which is not expected to be performed until 2029.

Jai avatar
Jai D (Jai)predicts NO

@dreev I think your underlying question is "What's your subjective probability that a publicly-available AI will exist which is able to pass this test by <date>, regardless of whether or not this test is actually performed?". On that my 2028 answer would be considerably higher - actually probably about the same, accounting for the possibility that there exists an AI that can pass this a different but failing AI is chosen for the 2029 test anyway (due to not having perfect predictive accuracy for model performance on this precise test ahead of time).

But since this is talking about an outcome which we probably will not be able to observe directly, we can't directly bet on it.

dreev avatar
Daniel Reevespredicts NO

@Jai Ah, I see. In any case, is it ok to just say that end of 2029, matching Longbets's end date, is what was intended all along and that that's when we'll resolve this market?

(In general I think if you're trading on a technicality, that's totally fine but it needs the clarifying question. In this case that might've been, "What happens if we have AGI or something that could pass this version of the Turing test but it hasn't technically passed because the test hasn't been conducted yet?" (Answer: "We defer to Longbets's verdict.") Or perhaps, "Is this market asking if Kurzweil will have officially won his bet with Kapoor a year before the deadline?" (Answer: "No, I just screwed up the date!"))

Jai avatar
Jai D (Jai)predicts NO

@dreev That's valid.

As for my shares - well, as the saying goes, you lose 100% of the shots to win fake internet money via exploiting technical loopholes that you don't take.