I wrote a prompt asking gpt4 to emulate the style of a comedian and write a message based on a short input. Initially, the outputs were very good, and there were a bunch of genuinely funny jokes, but now I can’t get anything funny, even using the same inputs that it worked for before. All the outputs are just cheesy or cringe-worthy. Even when I regenerate 20-30, there is no good joke. Any idea what might be the cause and how could I fix this?
If you haven’t heard these zingers yet, then you haven’t been using ChatGPT long enough.
Why don’t scientists trust atoms? Because they make up everything!
Why did the scarecrow win an award? Because he was outstanding in his field!
What do you call a fake noodle? An impasta!
Why don’t we ever tell secrets on a farm? Because the potatoes have eyes, the corn has ears, and the beans stalk!
Why did the bicycle fall over at the party? Because it was two-tired!
That’s with my comedy writer AI being asked “I’d like five funny jokes that might go over great at my friend’s birthday party.”
I have a feeling that OpenAI realized that the AI pretraining that came out of books and from the darkest parts of the web was way unsafe to let people just ask for jokes, and pounded these kiddy jokes into the AI’s brain.
Is it really worse though? Since you indicate API as a tag, we’ll try the previous GPT-4 version tagged -0301:
Why don’t scientists trust atoms? Because they make up everything!
Why did the scarecrow win an award? Because he was outstanding in his field!
Why don’t some couples go to the gym? Because some relationships don’t work out!
Why did the tomato turn red? Because it saw the salad dressing!
What do you call fake spaghetti? An impasta!
Fresh cornyness right off the cob.
While you can use other models and get different results, all the way to gpt-3-davinci showing why joke training was necessary, I’d try a significantly different input to an API model change the AI’s attitude:
system: Act as Andrew Dice Clay the comedian, portraying accurately his style.
A: not quality, but different:
- My buddy Tommy got himself one of them smart cars. You know those tiny things? I said to him, “Tommy! Why’d you buy a car you gotta put on like a pair of pants?” He said it’s good for parking in tight spots. I told him if he wants something that small that can fit anywhere - just take his wallet out of his pocket! Oh!
My prompt was similar to what you suggested; I asked gpt to emulate the style of a comedian. It did a decent job initially and gave some genuinely funny responses. But after the first few, they just got bad and stayed bad. The prompt did not change; the input did not change. What gives? (this happened over the course of ~24h)