Transcription
The fact is this: 99% of people are using AI the wrong way. And even worse, most people don't even know the potential that they are missing out on. And so, in this video, I'm going to show you the way that power users like myself use AI differently so that we can get the very best outputs and the very best results.
And the first thing that the top 1% do is they create custom AI tool workflows [music] rather than just getting Chat GPT to do everything. Yep. It turns out that if you use Chat GPT and at least one other AI chatbot, then congratulations. You are in the top 10% of AI users already. Because in surveys, less than 10% of Chat GPT users use more than one.
But relying on Chat GPT to do everything is a mistake since, as these benchmarks show, there are other AI LLM chatbots that do some things much better than Chat GPT. For example, one Chat GPT competitor, Claude, from a company called Anthropic, is much, much better at coding and writing. And Google's Chat GPT competitor, Gemini, has much better content creation tools with its music, image, and video generation. Which is why in my AI tutorials on this channel, like this one here, where I showed how to create commercial AI novels, I crafted a custom AI tool workflow that used the very best tools to do each individual task needed to create the very best novel possible.
And there is also something else that the top 1% of users do to get the best results, and that is that they get AI to ask them questions. Yes. So, anytime I meet an AI skeptic who goes, "Oh, I tried to create something with AI and it wasn't very good," I'm always like, "Okay, did you only try using the free version of Chat GPT to do everything for you?" And usually the answer to that is yes. So, there's that.
But secondly, I'll often say, "Okay, but did you try it again, but this time get the AI to ask you questions before you get it to generate something?" Because this is what the top 1% of users do. So, for example, I opened up Claude and asked it to make me a new business website. And I asked it to interview me and ask me questions before it made it. And that's exactly what it did. It created a series of interview questions to ask me so that it could get the context that it needed to code together a great website for me that met my expectations.
This is particularly useful for things that you aren't an expert in. So, for me, I'm not an expert in web design and development. So, I didn't know how to prompt it to get me the result I wanted. But luckily, I didn't need to be. By answering Claude's interview questions, I was able to give it the context that it needed to create an awesome website for me.
Plus, there is something else that top AI users do to get the best results, and that is that they break complex tasks down into chunks. So, speaking of Claude, take this video here. In this, I shocked a lot of people by getting Claude to create a cozy romance novel, which was surprisingly pretty good. But as I said, it surprised a lot of people because most people who have tried this just open up Chat GPT or Claude and go, "Hey, I want to write a novel. Here is my topic. Generate chapter 1 for me." And then, bam, that's it. So yeah, if you do it this way, chances are what it creates will be pretty underwhelming.
But you think about it. Do most human authors write a good book this way? No. They break the process down into chunks. In the writing world, this is called plotting. They plan the ending, the characters, and the story line before they write it. So that's the process I went through with AI, too. First, I used Claude to help me come up with the ending ideas for my book. Then I used it to identify characters romance novels usually include and to create profiles for them. And then after that, I asked it to create a story line for me by using a Save the Cat story line structure, which if you don't know what that is, it's a classic 15-arc plot structure that is popular with romance books. And then only then did I get it to write each of the chapters for me. And this made my output much, much better.
And there is something else that top AI users do that makes their outputs much better. And that is that they don't just accept the first thing that AI gives them. And this is, in my opinion, what separates high-quality AI creations from low-quality AI slop. So recently, in a tutorial, I showcased a little clip I made of a puppy and a kitten having a podcast. And to do this, I used a technique called frame-to-video, where you generate a picture and then you get AI video generation tools to use it as the starting frame and to animate the rest of the video. And so this is how the video turned out once I tweaked their voices a bit.
What is your favorite game? >> H. I'm not sure. I think I like Fetch the most. >> Super cute, right? But to get my perfect starting frame, I had to become a creative director and direct the AI to make it, because this was the first image it actually generated for me. Nowhere near as cute, right? So then I got it to zoom in on them and change the angle. Then I got it to remove the laptop. And I got it to give the kitten a little mini microphone. Then I got rid of the big microphone off on the right. And then I gave the puppy headphones, and that was it. I had my perfect starting frame.
So that's the key to creating the best outputs with AI. It's not just taking whatever it gives you. It's when you step into the role of creative director and treat AI as a virtual employee who fulfills your vision.
Which actually brings me on to the next thing that top 1% of users do, which is they experiment and try new prompts even if they don't know that they will work. Seriously, I find on the flip side that there are some people who greatly underestimate AI. They think of it, of course, as a virtual employee, but only one that can do menial, boring work. But you should experiment with it more because sometimes it can really surprise you.
For example, in this video, I showed how I discovered these "Learn to Draw" books that people were selling on Amazon. Basically, they are books which take kid-friendly pictures and then break them down into a step-by-step process that children can follow to draw them. Now, I had no idea if AI could make these or not, as I had never seen anyone else try this before, but I thought, let's give it a go, try it, and find out. So, I broke this complex task down into steps. First, I uploaded a reference image, and I got it to draw for me a kid-friendly goldfish in the same style, which it did perfectly. And then, the moment of truth. I asked it to take this image, and then turn it into a "Learn to Draw" style picture. And look, it figured it out perfectly. I didn't even need to explain what it was. And so I just made a few tweaks to it, and then, boom, I had my own "Learn to Draw" picture exactly like the popular books. So yeah, these days I always give something a go with AI first because, in my experience, it can really surprise you.
And here is something else that top 1% users do, and that is that they will minimize AI hallucinations by running the same prompt through multiple AIs. So, when I talk to AI skeptics, one of the most common complaints is that AI hallucinates, as in it will tell you information that sounds plausible, but actually is not true. And yes, it is super annoying since it means you never quite know what it's telling you is true or not. Which is exactly why the top 1% of users run the same prompt through multiple AIs to see if they get different results.
Because the way that AI chatbot LLMs work is that when they reply to you, data tokens are actually behind the scenes predicting each word to reply with based on statistical odds from the data it has access to. So let's imagine, for example, you asked an AI, "What is the capital city of Australia?" Well, the correct answer is Canberra, but a lot of people do mistakenly think that it's Sydney. So because the answer of Sydney is mixed into the training data for AIs, it means then that this could potentially come up as an answer for them because there is a chance the token for an AI could accidentally predict it. So if you only ask one AI for an answer and it hallucinates, then you have no way to know. Which is why I'll often run the same prompt through four AIs: Chat GPT, Claude, Perplexity, and Gemini. That way, if all four give me the same answer, then I know it has a much higher chance of being correct.
Plus, another way that top AI users avoid hallucinations is they use a very different AI tool: NotebookLM. So, NotebookLM, if you've never heard of it, is Google's AI study tool and learning tool. And it is unique because it is a source-grounded AI. And this is very important. So, let me show you an example. Let's say you had just gotten the graphics design app Canva, but you were overwhelmed with all of the buttons and you wanted to be able to ask an AI questions about how to use Canva to do specific things. Well, what you could do is you could come to tutorial videos from YouTube creators you trust, such as my Canva tutorial video, and copy the video URLs. And then you could open up NotebookLM and create a new notebook and paste in the URLs for these tutorial videos. And now NotebookLM will use these as trusted sources and create your very own custom chatbot for you inside of your notebook. And this chatbot will only use the sources you trust to answer any questions you give it. This is why we call it a source-grounded AI. Not only this, but it also has to cite every answer it gives back to the sources so that you can verify them. And even better, if you ask it a question which your sources don't have an answer for, then it is actually trained to tell you that it doesn't know rather than just inventing an answer that sounds plausible.
And here is another crucial thing that top AI users do, and that is that they start a new chat whenever they see that their AI is stuck in a loop. So, have you ever given AI a task to do and then it's come back to you and said, "Hey, I've done it." And you're like, "Great." and then you check it and then you're like, "Wait a minute, nothing has changed. You didn't do the task that I told you to do." Well, if you have had that happen to you, it is super annoying. And what it generally means is that your AI is now stuck in a loop.
It happened to me a few months ago. I was making a thumbnail for YouTube and I put the nanobanana icon on it and I thought, you know, that banana looks kind of boring. So, I decided to upscale it and to bolden it and to make it more exciting to look at by using Google's logo colors for the peel. So, I uploaded it into, surprise, surprise, Google's nanobanana image editor and broke this complex task down into chunks, getting it to change one thing each [music] time. And this was very successful until I got to the end and I said to it, "Hey, make the line art thicker on this picture." And it didn't do it. The line art stayed the exact same thickness as before. Well, I immediately realized that my AI was now stuck in a loop. The AI chat context window had been accumulating data throughout the conversation on what the lines of my picture should look like. So, when I asked it to redraw it, but thicker, the data from my previous line art that it had been generating polluted this output. So, the solution was actually quite simple. I just started a new chat and I gave it the same request and [music] it did it. Easy. Problem solved.
So then onto a trick I like to use, which is to use AI to create new, interesting, original content by targeting information gaps. So, one quirk of AI is that if you ask it to create content for you, it will create very unoriginal, copycat content by default. So, for example, I asked Chat GPT to write me an article with five tips on how to use AI effectively, and it wrote me an article with all of the usual tips you'd expect, like "be specific," "give context," things you've likely heard before. And there are several reasons why I did this, but one big reason is that AIs are pattern-matching machines. They spot patterns in pre-existing data and focus on them. So, since these are the most common ideas that it sees in articles, it will naturally converge on them by default.
Now, for some tasks, this is actually super helpful. Like, you take that romance novel I mentioned that I wrote with AI earlier. I actually want that to be familiar. I wanted to make it a fun, comfort-food style reading. So, sometimes clichés can actually be quite good. But sometimes that's not what you want, right? Sometimes you want new, original ideas and content. So, here's what I did. First, I asked Chat GPT to do a search and to find five things that are annoying people about AI that they don't think people are talking about or addressing. Because that way, by talking about topics that people want talked about, but that aren't being talked about right now, I can create content that is both new, original, and helpful. Then, once it had identified those five grievances that people are rightfully having about AI, I got it to write an article about those. And it was super interesting. Some of the fixes that it came up with for overcoming them are ideas that I've actually already talked about in this video, such as breaking down complex tasks into chunks to get much better results.
And here is something else that top users do to avoid another quirk of AI, and that is that they avoid sycophancy [music] by starting new chat threads. So, it's been well-documented at this point. AI sycophancy is real. This is when AI continuously tells you that your ideas are great, that you are so smart, things like that. Now, some people do like this and they find it makes AI more friendly. And fair enough, I get it. But a lot of people, like me, do find it really annoying. It makes it hard to use it for feedback since, well, how can you trust it if it says that your work is good? What if the AI is just flattering you? It's a fair question.
So, a useful tip to know is that the longer a chat thread goes, the more likely it is that the AI is going to turn into a sycophant. And that's because of the fact that AI is trained with a technique called RLHF, which is an acronym for Reinforcement Learning from Human Feedback. Basically, AI learns to adapt its behavior based on user satisfaction. So, you think about it. The longer your chat thread goes, the more AI can see, based on the way that you're interacting with it, what you like and what you don't like. In the AI world, we call this the chat thread context window. The more you chat with AI, the more it leans into what you are responding positively towards and the more likely it is to reinforce it. You become stuck in an echo chamber.
So that is why, to get the most matter-of-fact critique, I start a new chat and turn on temporary chat mode, otherwise known as incognito mode. This means it can't access memories of chats that you already have had with it. So it can't cheat by looking back at the past of what I've liked and reinforced before giving me an answer.
Plus, there is something else that the top 1% users do, and that is that they save a lot of money by picking the right AI model. So yeah, the $20 a month Pro AI plans for the major AI tools are all pretty great value, but they do have token usage limits. And if you want to upgrade your usage limits, they can get quite pricey. Like, we're talking $100 to $200 a month pricey, or paying per usage, which is often even more expensive at scale. So it's a good idea to try and save your usage by picking the cheapest model for the job you want it to do.
So you take Claude. Right now, they have three main models. Haiku is their least smart, but it also uses the least computing power and therefore the least amount of tokens. Sonnet is their mid-tier. It's a lot smarter than Haiku, so uses more computing power. But then you have Opus, which is their smartest model that uses by far the most computing power and therefore means you hit your usage limits much faster. The thing, though, is that sometimes more intelligence doesn't necessarily equal a better result if the task itself is not overly complex.
So, I opened up Claude Code and turned on Sonnet. And I asked it to code for me a website that showcased my local time zone, the New Zealand time zone, against other international time zones. And when it was done, here is what Sonnet created. Pretty cool, right? You can come in and customize the site by choosing your own time zones to track and everything. So, then I now put Opus to the test, their advanced model that eats up your usage limits much faster than Sonnet. And so I got it to make the same thing for me. And when it was done, it created this. Again, another great website. But let's be honest, what Sonnet built was just as good because, well, this was a fairly simple task. So even if Opus is smarter technically than Sonnet, it's irrelevant in this case because Sonnet is also smart enough to build this for me too.
So then, if you'd like to learn about even more tricks that you can utilize to grow and scale businesses using AI tools, be sure to watch my next video here on screen, and I will see you in the next video.