But, like you point out, it fails to grasp the subversion.
The surprising thing is how much of human conversation is successfully approximated. The bullshit fails when we ask the software to attempt to perform inductive or deductive logic or to derive information from a theory of mind.
What I meant is that GPT's attempts at bullshitting its way through a pseudoanswer looks a lot like what a highschooler (who also doesn't understand the topic) would attempt. That alone is amusing enough to me!
I think many comedians can tell you why a joke is funny. They know, many of them spend their entire career crafting jokes and refining them to be funny. It’s just not funny when they tell you why. Even so I think it still is more art than science.
A lot of what also makes a joke funny or not is the delivery and who’s telling it. I think a lot of Norm MacDonald’s long, rambling jokes, for instance, wouldn’t be funny at all if most other people told them (and some people will say they’re not funny even when he told them).
The US version had the same text, with minor changes to account for e.g. the different way to call back to the last incoming call on a landline. The US version had reasonably good actors.
But … somehow, it was just not funny.
I think if you took the script for Peep Show, for example, and set it in the US with American actors it would probably be pretty bad.
> The joke is funny because it relies on a play on words and a surprise ending. The first two lines set up the expectation that something is going to happen to the two men who are walking into the bar, but the punchline subverts that expectation by introducing a third man who ducks, which is a unexpected and humorous twist.
I asked what wordplay was involved, and it told me this:
> The word play in the joke involves the use of the word "walk" in the first two lines and the word "duck" in the punchline. In the first two lines, "walk" is used in the literal sense, as in moving forward on foot. However, in the punchline, "duck" is used in the sense of avoiding or evading something, rather than its literal meaning of a type of bird. This sudden shift in the meaning of the word "duck" is what creates the surprise and humor in the joke.
More nonsense. But then I told it "It's actually the word "bar" that's being used in the wordplay. The word "duck" there just means evading something" and asked it to explain the joke again, and it returned this:
> In the first two lines of the joke, "walk into a bar" is used in the literal sense, as in two men entering a place where drinks are served. However, in the punchline, "bar" is used in the sense of a barrier or obstacle, rather than its literal meaning of a type of establishment. The punchline introduces a third man who "ducks" under the perceived obstacle, which is a unexpected and humorous twist. The joke also relies on the idea that walking into a bar is a common and mundane activity, so the unexpected ending adds a element of absurdity to the joke.
Though the initial response was nonsense, talking to it as if it were a human and giving it a limited hint (telling it the wordplay involved "bar" and not "duck") was enough for it to give the correct answer.
I think this is probably why so many people say they're underwhelmed. A lot of the initial responses to simplistic questions aren't so interesting. But the degree to which it is able to correctly adapt and respond to conversational input is fairly astounding.
>In the first two lines of the joke, "walk into a bar" is used in the literal sense, as in two men entering a place where drinks are served. However, in the punchline, "bar" is used in the sense of a barrier or obstacle, rather than its literal meaning of a type of establishment.
The literal sense of walking into a bar isn't walking into a drinking establishment, it's walking into a actual bar. The punchline reveals the first 2 walked into a bar and the third one avoided it, no?
If it develops connections between words like "bar" => "sometimes an obstacle" => "can be ducked", that could explain it.
Like it would have all these attributes for each of those words, or concepts that relate to that word. "bar" has properties like "means multiple things", "often used in jokes", "could be an obstacle". But it must have quite a bit of properties.
I'm not sure how this AI is built internally, but if it works something that has this network of words and concepts that it can have weighted connections to each other, then you could retroactively come up with how the AI could have got to that result.
So user first says that they want a joke to be explained. It adds "joke" as a concept.
Joke has strong network connections to something like "puns", "unexpected punchlines", "absurdity", etc. And then when a joke is presented it will try to find connections with strongest weights and least concepts travelled. Like if it's a pun joke, it will go through all the words and will see that "bar" => "multiple meanings" => "pun" => "joke".
But then is it pre programmed to consider both meanings when it tries to analyse a joke that is a pun or has it somehow learned to come to it on its own?
> But then I told it "It's actually the word "bar" that's being used in the wordplay. The word "duck" there just means evading something" and asked it to explain the joke again
It gave a couple wrong answers and then was told bar is the wordplay. It didn't figure out bar has two meanings on it's own. It was also told duck is meant to mean evade and to try again. I don't agree that it came up with these associations by itself when was fed the knowledge bar has multiple meanings in the joke and duck in this context means evade.
Making that conclusion by itself would have been the only impressive and creative thing I could think of there.