I think most of the "magic" starts to fade as soon as you encounter a few bad outputs and quickly become unimpressed.
However, if you retry the same prompt multiple times, one of those is likely to produce a good output. I think it's important to give users of GPT-3 based tools multiple alternatives and let the user decide which of the options they like best.
That's the approach I took with my side project for generating short stories.
For example, with this story [1], not all the options for the progression of the story are great. But if you pick and choose which progressions you like best, you can arrive at a pretty good ending, such as [2].
Oh, yikes! I was following the progression pretty well until that compelling speech... I suppose this example gives more credence to grandparent's post of GPT-3 not being impressive or magical.
Nonetheless, here is a direct link to the homepage if others would like to try it out:
Once upon a time, there was a young man named Jason.
He did what all parents encourage their children to do. He got up early, studied hard, did well in school, participated in extracurricular activities, got great grades, made friends, did not date for several years, was not very attractive, went to a prestigious college, met the girl of his dreams, did not spend much time with said girl until he did, did not engage in any premarital activities, got married at the end of college, got a job, moved to a big house, earned a big salary, told his wife he loved her every day, taught his daughter to not have premarital sex just because she can, paid a lot of money into a pension fund, watched a lot of reality TV, read the news, and on its 20th anniversary took his wife on a modest European vacation.
This story is not just about Jason. It may also be about other people in his generation who all did the same things. What is it like to be like them? Because Jason and his peers did everything right, he does not believe he will ever have to worry.
That is because Jason is not intellectually curious. He is only interested in his own little world.
Jason believes in the retirement system, the company he works for, the banks, the government, the media, the rule of law, the peace of capitalists everywhere that will dominate the geopolitical landscape for another several centuries.
In other words, Jason is a sucker.
He is a sucker because he doesn't understand that to have a retirement account means that you have just contributed to a scheme that when push comes to shove has no real value.
He is a sucker because he need not know that stocks are really shares in nothing other than an intricate system to distract, confuse, or flatter the investor.
He is a sucker because he does not understand that his pension, just like Social Security, is just a joke.
And why does Jason need all these myths?
Because without them he would be frightened to death.
And so Jason, along with everyone else of his generation, contributes to a massive scam on an epic scale.
However, if you retry the same prompt multiple times, one of those is likely to produce a good output. I think it's important to give users of GPT-3 based tools multiple alternatives and let the user decide which of the options they like best.
That's the approach I took with my side project for generating short stories.
For example, with this story [1], not all the options for the progression of the story are great. But if you pick and choose which progressions you like best, you can arrive at a pretty good ending, such as [2].
[1] https://toldby.ai/arK_3OpvpkG
[2] https://toldby.ai/aQAXlq3LNku