Hacker News Reader

Comment by GolDDranks | Hacker News Reader

Indeed, the virtue of the Turing test is that it's a very general framework – the freedom of the framework allows us to test the AI using any communicational method at hand. One of such is Winograd Schemas [1] – a test of commonsensical reasoning, in which the current conversational AIs spectacularly fail at.

[1] https://cs.nyu.edu/faculty/davise/papers/WinogradSchemas/WS....

View on HN

Btw. here is a sketch of an embedding like that:

[Session starts]

<Challenger> Hi.

<Judge> Hi. Now, listen to me. My job here is to judge whether you are a human or an AI. I also know that your job is to try and persuade me to believe you are a human, no matter which one you really are. I happen to have a few questions that are easy for humans and hard for AIs. So if you are a human, answering them shouldn't be a problem to you, but if you avoid answering, I'm going to judge that you are an AI, no excuses. Here is the first question: ...

[A pattern of Winograd schema challenges ensues.]

So as we can see, the Turing test doesn't have to be any random chitchat nor it doesn't have to pretend to be "normal human-to-human-like discussion", as the Judge has the ability to participate and steer the conversation and zero in on the areas that are most likely to expose the AI.

However, if the human participants don't have any penalty from being labeled as AI, this doesn't obviously work, as they don't have any pressure to prove anything. So this kind of an embedding wouldn't work with "imitation game" -like rules, but it would work with "prove your human-ness" -like rules.

If you classify say 50% of people as AI's and 50% of AI's as people then those AI's passed the Turing test.

So, you are limited to questions that most people will answer correctly. Further, if you find some unusual question that works today someone can just add it to the program for next time.

Did you check the link about the Winograd schema challenge? The questions test common sense reasoning, and are very easy for humans to answer. An example:

The trophy doesn't fit into the brown suitcase because it's too large. What is too large? A: The trophy B: The suitcase

Coding a specific solution to questions in exactly that format is not inherently difficult.

If you keep writing specific solutions to this class of questions it's eventually had to come up with a simple question that's still unknown.

Further you can't get around that by making ever more complex questions because humans will eventually start messing them up.

>If you classify say 50% of people as AI's and 50% of AI's as people then those AI's passed the Turing test.

And if your version of the test is "heads it's an AI, tails it's human" then any AI's that are classified as human will have "passed the Turing test."

I don't think you understood.

The original test specifically had exactly one human and one AI. So, if the judge is forced to do a coin flip that really is success. If the judge does a coin flip because they are lazy then that's not a Turning test.

Then you can just drop "imitate human" part. We don't need that anyway, except for some niche applications.

But the "imitate human" part, as you put it, provides a simple framework & easy to validate testing protocol which is flexible enough to cover the above example and many others.

That's why its useful.

Not to detract from the logic, principle, or arguments here. But if someone said that kind of thing to me I would most likely respond with "Shutup-you talk too much!" I feel like a threat would need to be included to make a rational human incentivized to respond to such an arcane and winding dialogue, albeit a request..

Reply on news.ycombinator.com