3 ms·

So the Turing Test is not dead, we just need to redefine what it entails? That sounds like moving the goalposts.

krapp 3y ago
The Turing Test was never anything but a thought experiment, developed when computers were in their infancy, and has been widely criticized by philosophers and AI researchers. It should never have been taken as seriously as it has, just as Asimov's Three Laws of Robotics should never have been taken seriously. Both presuppose rigidly and objectively defined values for, respectively, self-awareness and morality, where none exist. If anything they reveal more about human intellectual bias, popular cultural assumptions (and possibly hubris) than they do anything about AI. The goalposts must be moved because we still don't even know what the nature of the game is.
fvdessen 3y ago
The Turing test lets the examiner have whatever conversation he wants with the machine. So the examiners can ask to solve math problems, recognise shapes, interpret poetry, perform whatever task. It doesn't matter either if the machine fails or succeeds at the task, it matters if it fails or succeeds in a way that is recognisably different from humans. And if some new task comes up that the machine has a hard time with, it is only natural that the examiners are going to ask that particular task. So yes it is completely allowed to move the goalpost in the Turing test, that's why it is a hard and interesting test to pass.
golol 3y ago
Well yes, thr Turing has to be defined exactly. That is not mvoing the goalposts because I never considered the Turing test to be a quick 10 minute chat. I am sure that Turing in his original publication soecific the conditions manner under which a Turing test should be performed. That was over half a century ago though. What are we more interested in, when talking about the concept of a Turing test, the exact version that Turing described, basically v0, or a test which captures that concept that AGI can be tested via an exchange of language? I think the latter is the much more useful and fundamental concept: The hypothesis that AGI can be tested purely via language has not been disproven for Turing tests of sufficiently large compelxity/length. Hence I find it most useful to think of the Turing test of a family of tests that have in common that they are a purely language based conversation but that may be held on different time, complexity or adversiality scales. Why would I interpret the definition like this instead of sticking to the original paper? Because the article here makes the fallacy that the v0 Turing test not detecting AGI implies that the "AGI can be tested bia language" is in question. This is just not true.

The Turing Test was never anything but a thought experiment, developed when computers were in their infancy, and has been widely criticized by philosophers and AI researchers. It should never have been taken as seriously as it has, just as Asimov's Three Laws of Robotics should never have been taken seriously. Both presuppose rigidly and objectively defined values for, respectively, self-awareness and morality, where none exist. If anything they reveal more about human intellectual bias, popular cultural assumptions (and possibly hubris) than they do anything about AI. The goalposts must be moved because we still don't even know what the nature of the game is.

fvdessen 3y ago

The Turing test lets the examiner have whatever conversation he wants with the machine. So the examiners can ask to solve math problems, recognise shapes, interpret poetry, perform whatever task. It doesn't matter either if the machine fails or succeeds at the task, it matters if it fails or succeeds in a way that is recognisably different from humans. And if some new task comes up that the machine has a hard time with, it is only natural that the examiners are going to ask that particular task. So yes it is completely allowed to move the goalpost in the Turing test, that's why it is a hard and interesting test to pass.

golol 3y ago

Well yes, thr Turing has to be defined exactly. That is not mvoing the goalposts because I never considered the Turing test to be a quick 10 minute chat. I am sure that Turing in his original publication soecific the conditions manner under which a Turing test should be performed. That was over half a century ago though. What are we more interested in, when talking about the concept of a Turing test, the exact version that Turing described, basically v0, or a test which captures that concept that AGI can be tested via an exchange of language? I think the latter is the much more useful and fundamental concept: The hypothesis that AGI can be tested purely via language has not been disproven for Turing tests of sufficiently large compelxity/length. Hence I find it most useful to think of the Turing test of a family of tests that have in common that they are a purely language based conversation but that may be held on different time, complexity or adversiality scales. Why would I interpret the definition like this instead of sticking to the original paper? Because the article here makes the fallacy that the v0 Turing test not detecting AGI implies that the "AGI can be tested bia language" is in question. This is just not true.