Sign in to continue learning
AI Prompting for Everyone
Beginner
3h4m
Topics
Prompt Engineering
One of the most helpful uses of AI is as a thought partner. When I'm trying to think through a complex problem or make a complex decision, it's nice to have a human expert as a thought partner, that is, someone to talk things through with. But if there isn't a human expert readily available, AI, which actually knows a lot about a lot of things, can be a really good resource for this. We'll go through together multiple examples of this, but to get started, brainstorming is one great such use case. Now, I know a lot of people ask AI to help brainstorm lists of ideas, but there are more effective ways to use it as a brainstorming partner than just having it generate a list. Let me show you what I mean. According to data released by OpenAI, analyzing ChatGPT conversations, about half of ChatGPT chats are asking for writing and practical guidance. And in fact, creative ideation accounts for 3.9%, almost 4% of all chats. I found using AI to help me brainstorm to be really valuable. Let me share with you some ways to do so. AI can be pretty good at generating options. There's a common creativity test, which asks people to name 200 potential users for a brick. So given a brick like this, how many users can you think of for it? This is actually pretty difficult. Some people think, oh, it could be a paperweight, maybe a planter, and oh, it could be used to build a house too, I guess. But to come up with 200 examples is not that easy. But if you ask an AI model, there's a good chance it can come up with a long list of ideas. And your role, if you're actually trying to use a brick for something, would be to evaluate these options to pick out which ones are the ones that you like. In brainstorming, common guidance is the more ideas, the better. And so sometimes having AI generate a lot of ideas for you to pick from can be a powerful way to find one or two good ideas. So this is the maybe more common use of AI as a brainstorming partner. I want to show you a different form of brainstorming in which you give it more context, and then also iterate with the AI longer, meaning have a longer back and forth conversation to help get you to better options. So if you tell it, help me build a workout plan, I'm 38, beginner level, have 10 pound dumbbells and 15 minutes a day. Then the AI may give fairly generic answers like here are three workout plans, start with 10 squats, 10 push-ups, pretty reasonable, very sensible, common sense answer. But if you want more creative options, giving it more context can be helpful. So if you say, I can't stick to these, give me hacks to stay on track, I have a trampoline and a cat. By encouraging it to give you trampoline and cat-related workout options, which is an unusual way of approaching workouts, it may ask you to consider trampoline breaks or cat-triggered micro-workouts where maybe every time you see your cat wag its tail or something, go do a tiny little workout. But these are certainly more creative ideas. AI models have some inherent creativity because they've trained on a lot of texts on the internet, which covers a lot of very different ideas, including some creative ones. And AI's output is a little bit random. So if you ask it multiple times, help me build a workout plan, it'll probably give you slightly different answers. But if you give the AI basic questions, then common sense, relatively generic responses, like do squats, push-ups, and so on, are more likely. Let me plot a conceptual diagram where on the horizontal axis, I'm going to plot how unique a response is, how creative a response is. So on the left we'll have responses like normal weightlifting exercises, like bicep curls, which is very common sense. To then maybe slightly more unique things, like standing on one leg with a yoga block on your head, to the really creative ones like cat-triggered micro-workouts. And on the vertical axis, I'm going to plot the probability of AI giving these different responses. And it turns out that it's much more likely to give a common sense response than a highly unique creative response. There's a reason for this. Namely, it was trained on internet texts, and there's a lot more internet texts talking about dumbbell curls than there are cat-triggered micro-workouts. And for most questions, this is actually okay, because the average information on the internet is probably decently factual. So when you're seeking information, such as what's the tallest building in the world, it's actually the Burj Khalifa, most internet text will say it's the Burj Khalifa. There are smaller amounts of text that will name other buildings, but the average response, the most common response on the internet, is usually the factual one for questions like what's the tallest building. But if you're brainstorming, then giving the average information, the most common response, ends up with squats, push-ups, and almost never trampoline breaks, and pretty much never cat-based sessions. Which is why if you ask an AI model to brainstorm with you, you get a lot of common sense ideas, rather than the more creative ideas, which depending on your goal, may or may not be what you want. So what do you want to do if you want to get high quality, more creative ideas from AI? We've seen with a basic prompt, you get responses from the common sense space. But if you give the AI model more context, so give your age, your level, but also tell it you have a mini trampoline and a cat, trouble staying motivated, and no squats, then this context pushes it into the more relevant and creative space, and it's more likely to give a custom answer rather than to generate common sense answers. Now, one problem that we face when brainstorming is if you're trying to come up with creative ideas, there's so much context you could potentially give the AI model, what should you prioritize telling the AI model? It turns out there's a technique that is very helpful for driving what context you decide to give the AI model, which is to iterate with the AI. Let me show you what I mean. If I want to ask AI to help me brainstorm plans for paying off my debt, I have $1,100 of credit card debt at 19% interest, monthly minimum payments of $40, a student loan, 8% interest, and a family loan, $900. So this gives decent background context. Then one thing you could do is ask the AI not to give you one option or tell you what to do, but to give you multiple options to choose from. I'll often ask it to give me three to five options. And so the AI may come up with a few different plans. Plan one is liquidity first to preserve cash. Plan two is eliminate the highest interest loan. Plan three is prioritize paying your family back first. So these are all actually reasonable ideas, and I've not yet given it enough context to know which of these plans it should favor. And it turns out that one of the really good ways to figure out what additional context to give to AI is to give it feedback on the options it presented to you. Highly relevant feedback that allows it to then give you the next set of options. I don't like option one. It's too passive. I do like the idea of paying off the 19% interest loan. Oh, and I forgot, I actually have $450 cash coming, and I'm also moving house soon. And then with this additional context, it now knows among plans one, two, and three, maybe what you like and what you don't like, and you can ask it to create three new plans. And then once again, by giving it feedback on these plans, you are giving it additional context that will help shape the AI model's thinking. And you can keep on iterating like this for a while until it comes up with a plan that you do like, and then maybe have it flesh out the details of the one plan or two plans that you like the most. I've found that giving feedback to the AI on what it thinks are good ideas is just a very useful mechanism for very efficiently figuring out what's helpful context to give to the AI. To summarize, if you're brainstorming, consider giving AI as much of the relevant context as you can in advance, and then ask it for a handful of options. Then give it feedback on the different options, and ask it for more options, and iterate multiple times, get more options, get feedback, get more options, get feedback, and do that a few times until you have one or more ideas that you're satisfied with. If you follow this recipe for brainstorming, I think you'll find you get consistently more useful and creative ideas. Now, you've heard me use the word context quite a few times. It's important to give your AI model the right context so that it knows enough to do what you want it to. Let's take a look at the next video at how context works and how it is used to produce a response.
