The truth about prompting
OpenAI put out their own prompt guide. Six steps to better results.
But there’s a trick to it.
And if you know what that trick is, it can change the way you use ChatGPT.
Let’s start with how weird this sentence is: There’s been no user guide for ChatGPT or other LLMs.
How is there no user guide for planet-changing, free-for-all artificial intelligence? I’ve seen user guides for kitchen tongs, for Pete’s sake.
There’s no user guide because generative AI is weird and unprecedented. And your brain doesn’t know how to handle weird and unprecedented. That’s why we need to know the trick.
Luckily, OpenAI’s Six Step Prompt Guide can help us figure this out.
So let me start by spelling out what those Six Steps actually are. I’ll do it in a handy easy to read visual:
These are the steps OpenAI recommends people follow to get the best results when prompting ChatGPT.
Done reading? Good, because I have thoughts.
First, look carefully. This isn’t the user guide we expected, is it?
Think about it: Our brains hear “user guide,” and we imagine a step-by-step, chronological-order list of tasks, like putting together one of those chests of drawers from IKEA. But this is more like general guidance. (Try using ‘general guidance’ to assemble an IKEA chest of drawers. You’ll lose your mind.)
Why would OpenAI give us general guidance rather than a step-by-step guide? Because they know that the sheer volume of use cases for ChatGPT can be overwhelming. Which means that anything that smacks of an official guide would risk narrowing users’ thinking, and thus undermine the very power and potential of ChatGPT.
Let’s put this into an analogy:
Let’s say you created a robot. Into that robot you put a synthetic brain that you invented (well done, btw). The robot opens its eyes. After a few pleasantries and some robot puns, your robot asks you how it should use this brain you gave it. What’s step one, it asks in a robot voice. What’s the most important thing?
How do you answer your robot?
One person might answer ‘Experience the world!’ and another ‘Ensure your survival!” and another ‘Love others!’ and another ‘I dunno, just, like, do whatever!’ And so on.
ChatGPT is like a robot with the new brain. Trying to give it use cases for its brain is an exercise in futility….because it can do pretty much anything. Which risks giving the user (that’s you and me) choice paralysis.
In a way, ChatGPT would be easier to use if it was labeled “Poetry writer!” or “Recipe maker!” or “Idea brainstormer!” Because then we would know exactly what to input. But instead, you can input all of these things. Or combine them. You could ask it to brainstorm ideas for recipe poems.
But when you have a general tool that can do pretty much anything….where do you start? It’s hard! You’re paralyzed! Your robot is just sitting there because you don’t know what a step-by-step guide to EVERYTHING looks like.
Now, back to OpenAI’s six steps to prompting.
Let’s take a closer look. What do you see now?
I’ll tell you what I see:
This is not a step by step process. Which means this is not, really, a user guide in the traditional sense.
This is a set of principles on how to be a better instructor…but not of software. Of people.
This could just as easily be a guide on how to manage your new intern. How to set that brilliant intern up for success. How to encourage that intern and make sure your intern had everything they needed.
So we’ve arrived at the title of this newsletter at last. Here’s the Truth About Prompting:
If you can talk like a human, you can master ChatGPT. If you can treat it like a brilliant intern, you’re going to do great.
That’s simple in concept, but difficult in practice.
Because your brain knows it’s not a human, and your brain doesn’t like change. Try talking to a baby like a college professor all day, you’ll see how hard it is.
Stop looking for user guides and prompt lists. Start training your brain to talk to ChatGPT like a human.
That’s the trick. That’s the lesson. That’s what OpenAI is telling you.
Master that, and you master generative AI.
News You Need
Deep Fake Madness: The US election season brought deep fakes into stark relief with fake President Biden phone calls, but now the bad guys have come for Taylor Swift, causing X to literally block searches for the pop icon. Congress is looking to pass legislation to ban nonconsensual sharing of digitally-altered explicit images. Do something good, Congress!!
Text-to-Video, Supercharged. Google’s Lumiere is said to be a ground-breaking video generation model capable of creating 5 second videos from text. Demos are amazing, but the model isn’t out yet, and Google’s demos leave something to be desired. If this is the real deal, it’s huge.
Neuralink Patient Zero. Elon Musk and an unnamed patient are about to go down in history for the first Neuralink brain implant. This person should, in theory, be able to operate their computer or smartphone just by thinking. If this works, we’ll be reading about this brave soul in the history books.
Generative AI Tips
Okay, this is a different LLM than ChatGPT, but it works with ChatGPT as well.
For some reason, when you tell your LLM to “take a deep breath” and then solve a problem, it does better.
Why? Unclear. Possibly because when the training data absorbed the entire internet, it noticed that performance was enhanced after things like taking a deep breath.
But it doesn’t really matter. What matters is once again, we see that this thing behaves in a human-like way. If you can treat it as a synthetic human, you will get great results.
That’s all for today!
Would love to know what you liked, what you didn’t, and also the name of your pet. But that’s just because I’m really into my dogs.
Send this to folks who might find it useful, knock on doors to demand people subscribe, whatever you need to do. We gotta level up together, friends!
See you next time!
