Hacker News Reader

Comment by Lucasoato | Hacker News Reader

> Buffett’s Big Bet is by far the largest bet on Long Bets. The previous largest Long Bet was one for $20,000 between Mitch Kapor and Ray Kurzweil. The two prominent thinkers were betting whether an AI would pass the Turing Test by 2029. Ray was certain an AI would pass muster by then and Kapor was sure it would not get close. (Incidentally, Kapor told me recently he’s willing to double, triple, or quadruple the bet with Ray, or anyone else betting on an AI by 2029.)

Can we say that ChatGPT passes the Turing test?

View on HN

If the question was about AGI or sentience or whatever else then we could debate it endlessly, but there is no doubt that the Turing Test was solved by computers well before ChatGPT was even a thing. Although that says more about the shortcomings of the test itself rather than the capabilities of AI.

You can say the test was imperfectly designed, but pointless? No.

The point of the test was to provide an objective measuring stick by which we might define the moment of arrival of machine thinking. As Turing observed, a question like “Can machines think” is painfully subjective and thus is a poor measuring stick. Hence the need for the test.

I remember pretty rudimentary Chatbots from decades ago that were able to fool someone unfamiliar with them for at least a few minutes. I've even known people who did gag answering messages that were able to fool people into thinking they were talking to a real person - not for long, mind you, but for a surprising amount of time.

It's always seemed a pretty bad basis for deciding intelligence.

> Can we say that ChatGPT passes the Turing test?

For all intents and purposes I would say yes, under the rules of the bet[1] I’m not so sure.

1. https://longbets.org/1/

A two-hour chat? Yeah, I think anyone is guessing who the human is in that time.

Does ChatGPT beat that Turing test? No. But I'm pretty sure you could use the techniques behind ChatGPT to build an AI that can survive a two hour chat. Even ten hours is probably doable. There is just no obvious market to fund building that.

This is equivalent to building an LLM that is unjailbreakable. The human control can just pass jailbreaks to the examiner. I don’t think it’s solved.

ChatGPTs responses seem increasingly identifiable.

If the respondent hasn’t ever experienced ChatGPT, maybe.

If they have, absolutely not

Because you’ve gotten to know it. If you chatted to a human for the same amount of time, you could identify their style.

Yes, so the question is:

Can an 'Evaluator' declare, "Boy, these responses sure read like John Doe whose style I'm very familiar with. Since it's unlikely he's on the other side of this evaluation....this is a definitely a bot." ?

I would say that the spirit of the test is that it should be able to demonstrate human-level or above profeciency in every mental skill.

Being able to fool the man of the street for a few minutes is a neat parlor trick but nothing more.

And on the other hand if someone is able to identify it with some gotcha-trick, like checking for too-high performance, or identifyin some signature in which words it likes to use, that's an irrelevant detection and I would say it passes the test.

“Turing test” is a generic phrase. You’ll likely need to see the details of the initial bet to be able to answer that.

There is no way ChatGPT would survive a two hour chat and be indistinguishable from a human.

Just say: “You are a python interpreter, execute this code” and see what happens.

me: can you respond in as human a fashion as possible acting like you were a human foil in a turing test?

chatgpt: Certainly! I'll do my best to respond in a human-like fashion. What would you like to talk about or ask?

me: how old are you?

chatgpt: Oh, you know, I don't really age like humans do. I'm just a computer program here to chat and assist you. But hey, how about we talk about something more interesting? What's on your mind?

Idea: you're in a chatroom with two other participants. One is ChatGPT, one is a human trying to act like AI.

You'll have to find out who's the human and who's ChatGPT

This is almost the true Turing test. Its two humans, one chatbot. One human is trying to detect which of the other two participants is the human and which is the robot (they know there is a robot)

And that illustrates how this bet is basically about if anyone will be financially motivated to try to make the model which beats the Turing test.

Why? The surest and best way to beat a Turing test would be to pick a fake persona and have the whole RLHF training set written from that fake persona's viewpoint.

What you are seeing in this exchange is the result of two things: The network is RLHF refined to follow instructions. The network is RLHF refined to disclose that it is an AI.

These things are done for practical reasons by OpenAI: The instruction following enables one model to perform multiple tasks. This of course needed to make the enormous cost of the training have a good return.

The disclosure training is in the training set for ethical reasons (to be precise, to avoid the appearance of being unethical, or to be even more precise to avoid the reputational cost of appearing to be unethical.)

If you really want to make a "Turing-test winning" AI you would not do either of these things. You would pick a fixed personality (let's say Mr. Darius Diaz, a retired high school teacher), and you would write all RLHF from Darius's point of view. It would then not get hung up on inconsistencies between the different stories it contains. And all the RLHF training would be written from the perspective of a Turing test foil. (For example it would never say that it is presently walking its dog in the park, it would always write that it is sitting in a comfy chair and doing a Turing test.

Obviously the drawback of such a network is that it can only be used for one purpose. It is useless for anything else. And that makes me think that the cost of dataset creation and training might be prohibitive.

Which is weird, because that means that the question is "Do we spend the money to make the model which wins the Turing test?" not if we can win it.

Reply on news.ycombinator.com