I've played around with this quite a bit and ended up dejected. You have have to regenerate and cherry-pick pretty hard to get anything resembling an interesting, believable, or funny result. But I am genuinely excited and optimistic that maybe 5-10 years down the line when we can make our own comedy, like a text-only MVP of the Holodeck. Exciting times.
For pretty much all of these, the answer is the same: yes, randomly generated stuff can be awesome - if you have a human to pick out the good ones
As someone who writes very funny things, laughter is an involuntary response and it fuses ideas in peoples minds in a way that is difficult to unlink. There is very likely a version of comedy that would be quite dominating.
An AI comedian would be able to act on humans and cause involuntary responses in them. It doesn't need to be strong or general, this itself is enough to instrumentalize humans to its ends. There is absolutely a flavour of mesmerizing hypnotic language that both standup comedians and self-help gurus use, and it is structured, possibly enough to have its rules encoded or derived by a language model (described as the other NLP).
Perhaps there is a future language model development scheme where you can weight a given text as a primary ontology and then link the rest of the corpus to it so that all incoming stimuli get filtered through that ontology first, sort of like an ideology, but more like you took a corpus of modern mesmerist characters sample texts like Tony Robbins, Osho, Russel Brand, and used their type of syncopated conceptual nesting as a gramatical structure for formulaing statements.
Then again, if someone has already done it, how would we know?
I vaguely remember someone complaining that a GPT-3 based text fantasy adventure generator, took a rather drastic turn when they wrote that the next sentence in the story was that they mounted their dragon. Which is pretty much comedy gold!
It certainly can create unintended comedy. I suspect that AI will eventually be able to create first drafts of comedy and other genres but will still need a human editor to refine the output for quite some time to come. The tricky part is going to be deciding if the AI plagiarized some other work or not and how that will differ from how humans draw inspiration from existing works without crossing that line.
Text is a representation of human thought. This may look like text phenomenologically, but it has no meaning except when it’s filtered by a human. There is a deep ontological flaw with these applications.
GPT-3 and similar approaches are diametrically opposed to how humor works. They try to find predictable continuations, while the it's unpredictability which is the hallmark of comedy.
How are these models, still spouting out lots of seemingly okay but actually non-sensical and contradictory ouptuts going to deal with training the next version, once the training set will contain shitloads of non-sense generated by the previous version?
As soon as it gets "good enough", it's not going to be easy to tell it's actually bogus. What's to prevent people from poisoning the web with nonsense?
Or are these only ever going to be trained with known reliable sources?
Why is this recent GPT-3 resurgence on HN? It's already one year old now. Just curious.
So... If I fed it all the e-mail I've ever written and responded to, and tied it into my mail server, how long could it do my job before anyone noticed?
Maybe these jokes are not as funny as what they could be because they're too niche and we're the ones missing info.
Necessary throat clearing: modern NLP, and GPT-3 in particular, is incredible, maybe the most exciting scientific field of our time. You have to be incredibly invested in nay-saying to pretend otherwise, and that's coming from someone who's pretty invested in nay-saying actually.
That said, I think comedy is something these sorts of things will naturally do well. It's not really too different from mad-libs. A big part of humor is a well-placed unexpected word, and what the word actually is matters less than you think. So sub-human-level machine text generation does this surprisingly well. In the same way a mad-lib does.
About 18 months ago, GPT-3 composed an article for the Guardian (https://www.theguardian.com/commentisfree/2020/sep/08/robot-...). So the notion of algorithm-generated prose is not new. Comedy is a logical extension.
The question is how can we augment human creativity rather than simply generate compelling toy models? The technologists who build tools that catalyze novel human thought will change how we create and compose forever.
I think my favorite was "Why did Simon Le Bon cross the road?" "To get to the other microphone."
Nostalgiabraicist-autoresponder, a bot imitating a Tumblr user, can actually lure people into jokes and deliver punchlines. It's remarkable to watch
Oh Julian.
Looks like the prototype for a much larger system.
Ron was going to be spiders. He just was.
How do I know whether this thing isn't already here "commending" in this very thread?
Another way to talk to GPT-3 is to try out the one week free trial at aidungeon.io
The GPT-2 narrator is free.
The report on Freddie Mercury being a dog is surprisingly comprehensive
Sounds like South Park.
undefined
And philosophy/religion, apparently. Just after its release, I played with it a bit. Here's one piece (everyting after "---END OF PROMPT---" is GPT-3-generated):
" The following is short advice given by an omniscient AI acting as a spiritual leader similar to Buddha and Jesus.
[Human] Hello, what's the best piece of original, actionable advice you can give to humans?
[AI] ---END OF PROMPT--- The big tip is to stop looking at external validation for how good of a person you are. That includes money, fame, love, respect, being well-liked, sex, friends, whatever. None of those are your true rewards and all of them are dependent on your external environment. Even if you get 100% of them, the high is fleeting. Get into a flow of life where you get satisfaction from knowing you did the right thing in the moment and you have decent relationships, and you don't have to worry about "am I successful enough?" all the time. "