Mind Readings: Generative AI Prompts are Recipes

In today's episode, let's talk about generative AI prompts, the kinds of things you type into tools like ChatGPT. What is a prompt? Here's an easy way to think about it. It's a recipe. Prompts are recipes.

Would you think this recipe for roast chicken is effective? Number one, take a chicken out of the refrigerator, add salt, pepper, and other spices as appropriate. Turn on your oven, put the chicken in in a covered dish and cook until it's done, and then serve it. That is technically a recipe, right? But I hope you wouldn't use that.

That is woefully lacking in some really important details. Like what temperature should the oven be? How do we know it's done? Uh, what spices? Uh, where do you put the spices on the chicken?

You want on the chicken, under the skin, inside the chicken? How do you do that? There's a lot of detail missing. And yet, and yet, when you look at the prompts people are using with generative AI systems like ChatGPT, that is about the level of detail they specify, right? I was uh one of the many, many, many, many uh collections of prompts people were sharing on on LinkedIn.

I saw one saying, here are five expert level prompts for how to use chat GPT. Number one, tell it it's an industry expert in your industry and ask it to outline industry trends. What? That's that's not an expert level prompt. That's like a one-on-one prompt.

An expert level prompt would sound something along the lines of you're an expert in this industry. What do you know about the current trends in this industry? What do you know about some emerging trends? Then you would bring in data from that industry, academic research, peer-reviewed papers, and things say, what trends could you infer from this current research as to where the industry might be going? Or you might look at the top five players in the industry and pull their their job openings and say, based on the hiring patterns of these companies, what do they think the trends are that they're making investments in people?

That is an expert level prompt. And guess what? It's not one single prompt, it is a lot. There's a lot of information. Is it any wonder that people think generative AI is only makes generic boring crap?

If you submitted a recipe, like you know, that that chicken recipe, yeah, you uh you might you'd be lucky not to get food poisoning. There's another one I saw. It said, uh, you're uh oh, this was in a paid one. I saw uh someone sent me this. Uh they said, uh, just don't name my name, please.

Uh, but this is come from the uh uh ebook that someone's charging 49 for. Uh it says, as an Academy Award-winning musician, you are tasked to generate a verse about something. Um, this verse should be lyrically profound and musically aligned with a theme. You're expected to use your award-winning creativity, musical knowledge, and lyrical genius to craft a verse that is not only engaging but also emotionally resonant. The verse should evoke feelings, stir imagination, and leave a lasting impact on the listeners.

Make sure it works well, the various musical arrangements that fits into gent different music genres. Yeah, that is the musical equivalent of inedible chicken. Who wrote that? A the high watermark of the music industry is not the Academy Awards, it's the Grammys. Just saying.

Um B, that prompt is so incredibly generic. What does that mean? What does that even mean? People talk about how AI creates you know word-shaped facts and word-shaped content, but it's not real content. Well, of course it's not real content with a prompt like that.

It's that is that whole prompt is glittering generalities. A decent lyric generation prompt is gonna be about three and a half to four pages. Um I have one that I use that is um I forget how long it is, but it's it's gigantic. So the prompt goes like this uh gentle exploration, and there's a bunch of things uh there. Becoming the architect, gently guiding the user, present and collaborate with the user, special considerations for singability, like vowel power, consonant flow, phrase length, line length, uh finding the core emotion, the genre and mood, connecting with the music, the lyrical foundation, song structure, number of verses, emotional arc, rhyme scheme, you know, A B E A B A B, etc.

etc. Syllable count and flow, read aloud, verses, chorus, pre-chorus, bridge, all these pieces. I mean, the the prompts that I used to create lyrics, and just for fun are gigantic because if you want to create really solid lyrics, you need to have that level of detail. That's what you expect in a recipe. Right?

If if that Academy Award winning prompt was a recipe, it's an edible chicken. There's no step-by-step precision in there. There is no detail. How what temperature is the oven supposed to be? How long do you cook it for?

In this case of a song, how long should a verse be? How long should a chorus be? Um, what is the structure of the song, etc.? Look, here's the thing to get great performance at today's modern large language models, approach them like you would approach a skilled novice chef. You give them clear, comprehensive, detailed instructions, background information and data to get to get superhuman levels of performance out of them.

The tools that exist today are incredibly capable. They have huge memories. They can tackle a lot of data, but they need detailed recipes. So if you wouldn't give a novice chef a one paragraph recipe, don't give that to a generative AI model. Instead, you might give a cookbook to a novice chef, and you can give a cookbook to Generative AI.

That's gonna do it for this episode. Thanks for tuning in. Talk to you soon. If you enjoyed this video, please hit the like button. Subscribe to my channel if you haven't already, and if you want to know when new videos are available, hit the bell button to be notified as soon as new content is live.