chthonic-anemos
bitchute.com/video/EvyOjOORbg5l/
- 8,606
- 27,257
Chat GPT AI
- Thread starter Sanrith Descartes
- Start date
By loose standards, yes, but AI/CS/neuroscience/philosophy people keep moving the goalposts and/or redefining the test requirements. There are some trivial ways to make it fail, like asking it today's date.
- 1
- 683
- 785
I think we may have crested the uncanny valley of turing tests, but we still have a long long ways to go. We will likely create an equivalent of the voight-kammpf tests for assessing emotional response and other tests for humanity that will suss out this complicated pattern matching device to effectively differentiate it from a psycopath autists.
- 1
iamacynic37
FoH Honkler - HONK HONK MFr
- 289
- -247
A system of cells.
Within cells interlinked.
Within one stem.
And dreadfully distinct.
Against the dark.
A tall white fountain played.
Within cells interlinked.
Within one stem.
And dreadfully distinct.
Against the dark.
A tall white fountain played.
- 1
Mist
Eeyore Enthusiast
- 30,414
- 22,202
That's obviously the crudest possible way to get it to do so, and ChatGPT is specifically designed to identify itself. I'm sure you'd get more subtle results if you used the base GPT 3.5 or otherwise instructed it to obfuscate.
Hoss
Make America's Team Great Again
- 25,591
- 12,067
and it fails by always knowing the correct date?
Sounds like ELIZA was more fun. That ho was always trying to get in my pants.
Huh. Previously it couldn't give you the date because of the age of the training dataset. I guess they've changed things.
- 6
- 1
Ukerric
Bearded Ape
- 7,927
- 9,578
Isn't there already an (AI-powered) scoring test that gives you the likelihood that text was generated by AI vs a person?
It doesn't actually understand questions as such, so trying to modify it's behavior is very hit and miss. It's part of the reason it lies so much - telling it not to lie is a lot more complicated than you'd think because it doesn't understand the relationships between things, and truth is completely abstract.
- 1
Tuco
I got Tuco'd!
- 45,433
- 73,506
No, but chatbots built for the same purpose as ChatGPT never will because they don't attempt to act like a human, just provide useful responses. I imagine that a chatbot with the same level of technological backing as ChatGPT could if it was designed too though, as long as the evaluator wasn't too informed about the current issues with chatbot tech.
If a rando has a 10 minute conversation with ChatGPT they'll probably be really impressed, but if they read a few articles on how ChatGPT fucks up they'll be able to easily replicate some 2+2=5 behavior. The same could be true if ChatGPT was designed to beat the Turing test.
Hoss
Make America's Team Great Again
- 25,591
- 12,067
In the context of a Turing test that means nothing. Plenty of humans fuck that up too.
Tuco
I got Tuco'd!
- 45,433
- 73,506
Sure, but it can get things wrong in egregious ways that would be an obvious indicator that it's a chatbot and not a human.
Even if ChatGPT was trained to pretend to be a human, if it fucks up questions like this it can fail the Turing test.
What's interesting is that the January version of ChatGPT could be bullied into giving very wrong answers, but the early February version of Bing would get extremely upset if you disagreed with it.
Tuco
I got Tuco'd!
- 45,433
- 73,506
An example of "bullying" is to ask it a sentence that makes no sense. It'll make it fit.
This kind of reads like the question a protagonist in a movie uses to get out of an impossible robot-prisoner situation.
tvtropes.org
This kind of reads like the question a protagonist in a movie uses to get out of an impossible robot-prisoner situation.
Logic Bomb - TV Tropes
Is your sentient supercomputer acting up? Good news. There's an easy solution: confuse it. If you give a computer nonsensical orders in the real world it will, generally, do nothing (or possibly appear to freeze as it loops eternally trying to …
tvtropes.org
Hoss
Make America's Team Great Again
- 25,591
- 12,067
Sure, but it can get things wrong in egregious ways that would be an obvious indicator that it's a chatbot and not a human.
Even if ChatGPT was trained to pretend to be a human, if it fucks up questions like this it can fail the Turing test.
What's interesting is that the January version of ChatGPT could be bullied into giving very wrong answers, but the early February version of Bing would get extremely upset if you disagreed with it.
the egg thing might be a better example. Because this one looks totally like a thing a human would do. It got the answer right in the end. The opening statement was just backwards.
It also can't count digits. The final summary was right but also at odds with all previous points.