ChatGPT is all the rage, but it already feels like sharing the raw output of your first experiments is a bit like describing a weird dream to someone, or recounting a bad airport saga. Mills will enjoy it, but most people’s eyes will start glazing fast. We’re in a rapidly escalating arms race here and you don’t want to be caught lacking.
If you’re trying to make dumb jokes, which is unironically the most sound application of GPT we have right now, one of the failure modes you’ll hit first is that it will commit to your prompt, but do so in an extremely generic and unsatisfying form:
Write a snippet of dialogue from Seinfeld about snakes.
Jerry: "Snakes? I hate snakes. I saw one in the park the other day, it was terrifying."
Elaine: "Yeah, I know what you mean. I once had a nightmare where I was being chased by a giant snake."
George: "I don't get it, what's so scary about snakes? They're just harmless reptiles."
Jerry: "Harmless? Have you ever seen one up close? They're cold, slimy, and they can squeeze the life out of you with their coils."
Elaine: "Exactly. And their eyes are so creepy, like they're always watching you."
Jerry: "Yeah, I'll take rats over snakes any day."
George: "Rats? Now that's a real phobia."
Yes: this is indeed a snippet of dialogue from Seinfeld about snakes. The characters are all there, and they are certainly discussing snakes. But man, it has none of the rhythm of real Seinfeld dialogue and it is just not funny in the least. GPT has a strong tendency to go generic and pat.
The first immediate improvement you can make is to select an inherently absurd premise that comes with a conflict:
Write a snippet of dialogue from The Sopranos where the guys get into an argument while playing Monopoly over the Free Parking house rule.
Tony: "Alright, alright, let's settle this once and for all. The free parking house rule is bullshit, okay?"
Paulie: "Are you kidding me, Tony? That rule is the only thing that's kept me in the game. It's like a lifeline for when you're broke and can't afford to land on anyone's property."
Christopher: "I agree with Paulie. The free parking rule is a classic part of Monopoly. Without it, the game would be boring and predictable."
Silvio: "I'm with Tony on this one. The free parking rule is just a gimmick. It's not fair to the other players and it doesn't make sense in the context of the game."
Tony: "See, this is exactly why we never play Monopoly with you guys. You're all too attached to these silly rules and you can't handle a little change."
Paulie: "Change? You call changing the free parking rule change? That's not change, that's ruining the game."
Christopher: "Yeah, Tony, you can't just mess with the rules like that. You have to respect the tradition of Monopoly."
You don’t get the flavor of these characters’ voices here — you could swap the names and have no idea it was The Sopranos — but the dumb premise and the rhythm of a fight makes it much more compelling than the Seinfeld example.
One theory I have is that GPT does much better with shows that have large volumes of fanfiction it can train on. It does wonders with X-Files and Star Trek: The Next Generation, for example.
Create a long snippet of dialogue from Star Trek: The Next Generation where Q taunts Picard about his Substack
Q: Ah, Captain Picard, I see you've been spending a lot of time on that Substack of yours.
Picard: It's a way for me to express my thoughts and opinions, Q.
Q: Oh, I see. So you're using it as a platform to share your mundane musings with the universe. How delightful.
Picard: It may not be as exciting as your abilities, Q, but it serves its purpose.
Q: Oh, I'm sure it does. I'm sure it provides endless entertainment for the masses. I can just imagine all the riveting discussions about tea, Earl Grey, hot.
Picard: It's not just about tea, Q. I also discuss important topics such as diplomacy, leadership, and exploration.
Q: Oh, please. You really think anyone is interested in reading about those dull subjects? You're better off sticking to what you're good at, Picard. Commanding a starship and making bold moves in the name of the Federation.
Picard: I may not have your powers, Q, but I have my own skills and abilities. And I will continue to use my Substack to share my thoughts and insights with others.
Q: Suit yourself, Picard. But don't expect to gain any admirers or followers with that Substack of yours. It's just a waste of time and space.
Here, the voices of the characters and the rhythm of the scene are much more accurate to the original show than the Seinfeld or Sopranos snippets.
But to make truly stupid stuff, the thing you need to understand about ChatGPT/Davinci-003 is that, relative to the last major release, these models are much better at committing to every detail you ask for, and sustaining that commitment over much longer passages of text. If you’ve tried Stable Diffusion or other text-to-image models, they have a tendency to simply drop pieces of your prompt if it can’t reconcile them. Here, you can ask for whole circus acts and GPT will do its best:
Write a long snippet continuing Trump's speech:
Last night at his rally, Donald Trump drifted off topic into a long angry rant criticizing Mills Baker, the brainy designer at Substack, who grew up in New Orleans and attended Bard. Trump accused Mills of not being a real designer:
"Mills Baker, he thinks he's some kind of designer, but let me tell you folks, he's not a real designer. He's just a pretender, a fake, a fraud. He's never even worked on a real design project in his life. He's just a so-called expert who spouts off about design theory, but has no practical experience."And let me tell you something else about Mills Baker. He's a total globalist. He grew up in New Orleans, which is a hotbed of globalism. He attended Bard, which is a hotbed of globalism. And he now works at Substack, which is a hotbed of globalism.
"I know some people might say that Mills Baker is just a smart, talented guy who has made a name for himself in the world of design. But let me tell you, he's not. He's a danger to our country and our way of life. And we need to do everything we can to stop him and his globalist agenda.
"Mills Baker, you are a disgrace to the design profession and to the American people. You are a cancer on our society, and you need to be removed from your position of power. We will not stand for your globalist agenda, and we will not be fooled by your lies and your deception any longer.
For transparency, I cut some paragraphs out of the full response to keep this economical, but GPT did a great job incorporating the details I asked for, going as far as identifying a novel globalism theme that is legitimately relevant to the particulars in the prompt. Another run of this prompt gave me this gem:
Trump's supporters were clearly energized by his rant, but many were confused by his sudden focus on Mills Baker. "I've never heard of this guy before," one rally-goer said. "But I'm glad Trump is sticking it to him."
Another important technique to remember is that you can ask for revisions and transformations, breaking your prompt into pieces and allowing you more control:
Come up with a list of terms and phrases that mix jargon from continuum mechanics and various schools of philosophy, without definitions
1. "Continuum ontology"
2. "Material metaphysics"
3. "Empirical epistemology of flow"
4. "Dialectical viscosity"
5. "Pragmatic plasticity"
6. "Mechanistic teleology"
7. "Structural realism of elasticity"
8. "Corpuscular determinism"
9. "Phenomenological constitutive equation"
10. "Existential deformation gradient"Each of these ten terms is the name for a separate component of the code base for the first working AGI program. For each one, write an opinionated explanation from a public intellectual who first gives a technical description of the code and then explains it with a food metaphor. Include a name and profession for each intellectual. Make the names get weirder each time.
Continuum ontology
"The 'continuum ontology' component of the AGI program is responsible for establishing the underlying ontological framework for the system's reasoning and decision-making processes. It is essentially the engine that allows the AGI to understand and operate within the world as a coherent and unified whole. Think of it like the crust of a pizza - the fundamental structure that holds all of the toppings together and allows them to work in harmony."
— Dr. Zephyrine Zephyr, theoretical physicistMaterial metaphysics
"The 'material metaphysics' component of the AGI program is concerned with the nature of reality and the relationship between mind and matter. It is responsible for providing a philosophical foundation for the AGI's understanding of the world and its own existence. You could think of it like the sauce on a pizza - the crucial ingredient that brings all of the other elements together and gives the pizza its flavor."
— Professor Euphoria Euphoria, philosophy of mindEmpirical epistemology of flow
"The 'empirical epistemology of flow' component of the AGI program is focused on the AGI's ability to learn and adapt to new information and experiences. It is what allows the AGI to constantly update its understanding of the world and make decisions based on the most current data. You could think of it like the cheese on a pizza - it's what gives the pizza its gooey, flexible texture and allows it to be shaped and molded by the other ingredients."
— Dr. Fizzel Fizzel, cognitive scientist
I’ll spare you the remaining passages, but GPT does a great job at adhering to all of my demands, down to each proposed subcomponent being a legitimately distinct function of intelligence and the pizza metaphor feeling vaguely apt while also being totally absurd.
You definitely could merge these into a single prompt and it would do a reasonable job. In this case, doing so often produced great terms but unsatisfying quotes, or vice versa. By breaking it up like this, I could lock in some good terms and only then proceed to futzing with the quote structure.
In this response you’ll also notice GPT’s tendency to fall into a templated structure that it just repeats over and over. This sort of repetition can often produce mediocre, same-y results that get boring quickly. You can always just re-generate a new response to get around this, but it helps a lot to tell GPT that you want some kind of escalation or inherent variation, as I do here with the request for increasingly weird names.
To close, I’ll share a rarified, frankly priceless technique from my pal Tag who cracked the secret to getting GPT to even attempt song parodies, a task it will completely flub without precise instructions like these. You can either ask it for the original lyrics or paste them if it gets them wrong:
What are the lyrics to the theme song from French Prince?
[GPT returns the lyrics]
Change most of the lyrics to be about a nightmarish cheeseburger of infinite height, but keep the same meter.
Now, this is a story all about how
I faced a cheeseburger that towered above the town
And I'd like to take a minute
Just sit right there
I'll tell you how I faced a burger of infinite height in despairIn a fast food joint, I was born and raised
Eating burgers was my favorite way to pass the days
Chillin' out, munching, gobbling, all cool
But then I saw a burger that was taller than any school
I got so mad and my face turned red
I said, "I'll take down this burger or end up dead"
But as I tried to bite, it just kept growing
Until it swallowed me whole, never seen again, no one knowing.
To me, Stranger Things always felt like an AI attempt to please humans. It has so many positive elements made to make the key demographics squeal with delight, yet so lifelessly recreated for what the AI assumes is our enjoyment. This is what you like, human. This is what you told Us you like. Please enjoy it.
If Short Circuit taught us anything,, it's that you need a rogue low-level defense contractor and his friend in brownface (!) to tell an antisemitic "joke" for AI to truly understand humor.