The Secret Prompt That Powers Claude & More AI Use Cases
Igor covers Anthropic publishing Claude's system prompts, Google's new Gemini models and Gems feature, and head-to-head tests of Claude, Gemini, and ChatGPT on real tasks like prompt generation and event writing. Also Flux fine-tuning for branded images, a copyrighted AI song, and Niantic's 3D capture app.
Chapters
This week we had a wide variety of AI related releases that you can actually put to work today. Starting out with Anthropic actually releasing their system prompts. This is a first in AI world. Then we have Google coming out with their own version of GPTs. They call them Gems. I went ahead and tested them versus GPTs and Claude. You can find all the conclusions in this video. Spoiler alert. They're better than I expected. And then we look at a bunch of AI imaging use cases for the real world, not just for fun and games. And so much more. In this week's episode of News you can use. Let's get into it. All right, so first up, I want to talk about a bold move from Anthropic. No other big company has done this yet. They released their CLAUDE system prompts that you can find in the web interface. Meaning if you're using 3.5 Sonnet or Opus or their small model Haiku on their Claude website, like so, these system prompts are what define the behavior and the responses you get from the large language model. Now, I suppose other makers consider this to be the secret software. There were several attempts at extracting them, but it was never confirmed that for example, the prompt that GPT gives you if you ask for its instructions is the real one. Now Anthropic took a different approach. They published this whole thing. And why is this interesting? And why can you use this? Well, I wasn't even aware of the fact that if you're using the API, you're foregoing these system prompts. You're not using all of these extra instructions and the context within this long stand system prompt. And there's a lot of good stuff in here. Hallucination prevention. Yes. Guidelines on how to interact with humans in a more productive manner. Absolutely. There's multiple instances of that self awareness of the fact that this model even exists and what the user means when he says Claude 3.5 sonnet, that's in here too. And so much more. I thought this was a fascinating read. We're not going to do a complete prompt engineering breakdown of this, but there's a lot of good stuff in here. If you're using the API, all of this is not present. And that's why using the API might not feel as fluent for this chat use case as does using the web interface. Meaning if you built an application on top of the Anthropic API, you might find that it behaves completely differently than does this web interface that a lot of us are very familiar with. And you should Know that this is also the fact for all other APIs like the GPT4 or GPT4O APIs, this has always been the case. If you send the same prompt to ChatGPT and the API, you get different responses. But I personally just always wrote big system prompts when I used the API and made it work for my use case. But if you weren't aware, now you are. If you're using the API, you get the vanilla version. If you're using the web interface of these LLMs, you're getting the version with the system prompt. What is the system prompt? They published it. You can copy paste it from here and that way you can make the API act in the same manner as does the web interface, or you can modify it. Either way, this is something everybody should know. And I love that Anthropic is leading the way by actually making this public. Beyond that, Anthropic had another announcement that actually just came out and I'm really excited about this one because we talked about Artifacts on the channel before. Matter of fact, I have an entire video showing you over 20 different use cases for Artifacts. And now it's available to everybody, both on the phone and on the desktop version. And if I say everybody, that includes the free users, you only get a few messages, but you get to try out Artifacts. Now, if you combine that with the artifact sharing that they actually rolled out a few weeks ago, you now get to also remix this artifact. So you get to take what it generated and create your own versions of it. Now, we talked about this before. There's some very interesting and funky use cases that you can do with Artifacts. They have a few good examples in here. But my other video goes way more in depth and shows you all the different possibilities along with the prompts to generate those different apps. Because that's what Artifacts does. It writes the code and it also gives you an environment in which the app is actually functional right on the cloud website. And now everybody can use them, including free users and mobile users. Loving it and looking forward to ChatGPT catching up with this probably in their dev data that is coming up in a month. This is one of the things I'm pretty sure we'll be seeing released there for ChatGPT. So you might have noticed that right now AI is developing like crazy. And I'm constantly asked this one question, which is, Igor, how do I best take advantage of this new technological era? And my honest answer is that one, you either need to learn how to use these tools and enhance what you're already doing or two, you need to get involved in building these tools yourself. And a fantastic way of educating yourself further is to use the sponsor of today's video, brilliant.org and even though they're sponsor, I got to say I like working with them because we align on our values. Both the advantage and Brilliant belief that lifelong learning is the formula to success in an ever changing world. It's really the only thing that works. And the way Brilliant tackles this is an interactive approach where you get to grapple with the information instead of just consuming it. They've got hundreds of lessons across all sorts of STEM subjects and instead of just throwing information at you, they get you involved in actually solving problems as you progress through the materials. And they constantly add new materials. Like right now they added a brand new one about clustering. Now this is a term from data analysis and it basically describes that if you have multiple data points in a certain area, then that's referred to as a cluster. And this is the process how you would identify that. So yeah, this is a very concrete technique you can learn about. But if you do that, it will unlock a whole new point of view on how the training data and something like ChatGPT affects the outputs or how to use data analysis tools even in ChatGPT at a more advanced level because you have an understanding of what's actually going on there and what you can ask for. Essentially, building skills like this helps you understand when you should be asking for a specific type of output and what data analysis possibilities you might have once AI is done with crunching the numbers. So if you want to check out this clustering course or any other course they are offering on their platform, head on over to brilliant.org the AI advantage also will get 20% of their annual subscription. And now let's get back to some more AI news you can use. All right, next up, let's talk about Google releases. And they were busy building and releasing stuff this week. First of all, we got a new prompt gallery. This isn't as exciting. This is something that I actually created in December 2022 in almost this exact format with my first notion templates. It was exactly these various prompts with short titles that, that describe the outcomes. You have a lot of familiar prompts in here like writing blogs or marketing posts. And then there's some interesting ones like hurricane charting. And as you can see if you click these, it actually uses this sort of interactive preset interface, which is nice, meaning it shows you what kind of inputs it needs. For example, the math tutor will look completely different from the hurricane chart, right? And if you go to this one that transcribes audio with speaker details, you will see that it comes preloaded with an example exactly how I shipped my prompts in December 2022 already. But that's okay. I'm just happy to see that I was on the right track already. This could be useful if you want to expand your prompt library and how Google approaches prompting. But beyond that, they also released three new models this week. Gemini 1.5 Flash, the 8B version. So a tiny, very fast model optimized for developers. Gemini 1.5 Pro New experimental version this is their biggest model with a very large context window and complex reasoning capabilities. Plus plus really advanced prompt adherence. And then they have a significantly improved Gemini 1.5 flash models. Now as always with model releases, it's always a bit hard to quantify how good these actually are. You can look at the benchmarks which are often fooled. You can look at the leaderboards which also don't always seem to reflect the reality. By now I've just found that using them for a few weeks and experimenting with them here and there is the best approach for me to see if they work for my use case. All I can say for now that on the leaderboards they're ranking really highly and that the Gemini 1.5 Pro model already before this upgrade was really excellent that if you had a lot of context, this is really your go to. All of this is usable through Google's AI studio immediately. And here you get the 2 million context window as before, which is really impressive. And you can just switch from the original model to this experimental one. See what you like better. So those are the new models, but there has been a release out of Google that got me personally even more excited. If you've been following the channel, you know that I'm actually a big fan of GPTs. I think if you you custom design them and include the right actions, give it the right context, they can be really useful for recurring tasks. They're only good for things that keep recurring and with the correct use of prompt engineering and templates and actions, you can actually make them really useful. And I like them and I use them for specific tasks. Matter of fact, a lot of them don't even include actions or the actions are not essential. They're just a nice to have addition like saving something to a database. Well, Google released their versions of GPTs. They called them gems. They have been announced a while ago, but now they're out and they're out to everybody as I made experimenting and reporting on generative AI my life. I'm obviously subscribed to all the various platforms, so I hopped in and tested gems right away on some of my favorite use cases I recreated three GPTs that I use all the time. One of them is Igor around the world. I showed this off in the recent Flux Fine Tuning video where I show how to train your face into a model and what kind of prompt generator I use to come up with the various thumbnails for this channel. So I rebuilt that in here. Then I modified my very popular GPT builder that builds a complete assistant with communication sequence and all the extra context for you and it renamed it Gem Builder and had to make a few updates like this doesn't have image generation so I had to remove all the Dalla usage from it and a few more minor updates and then I also recreated the event description writer and this is what I found when comparing these three GPTs between GPTs and ChatGPT Claude's project feature. I also recreated them there and then doing the same thing within Gemini Advanced. By the way, Gemini Advanced uses the Gemini 1.5 Pro model that we just talked about. Which One of these three performed the best? Gemini 1.5 Pro, Cloud, Sonnet 3.5 or GPT4O? Well I can already tell you that for Ego around the world where it's a stable diffusion prompt generator, it didn't really matter. All three performed the job well. All three are really good at generating diffusion prompts. I don't care which one of those three I would use for that use case. Nothing too advanced. And the limiting features within this gem manager didn't stop me at all. Because if you go into the gem creation you will quickly see you only have two fields. You have a name and instructions. That's it. No knowledge base, no actions, no code interpreter, no image generation, not even a profile picture for these poor gems. But that's fine, they're keeping it simple and that's why I picked these free GPTs because they don't include any actions or anything fancy and they don't need them. But what they do have is something very similar to actions. They have extensions. And this was already present inside of Gemini Advanced. As you can see here you can interact with Google Workspace files with Google Flights, Google hotels, maps and YouTube. Meaning these are direct links between the applications like YouTube and Gemini Advanced, resulting in it being really easy to pull in a YouTube transcript, let's say. And what you can do within these gems, that's not what I explore in this video. I need some more time to test all of this out, but what you can do is actually go in and let the various gems that you built access these external applications for you. And the way you would do that is by simply saying add and and then selecting one of these extensions. So if I do this, then I could save the gem. Then once I prompt something that relates to YouTube, it will actually go in and use the YouTube integration to give me various links for various videos. But as you can see, it doesn't exactly act in the way that you would hope to. I just wanted to pick a video here. It gives me a list. This takes further exploring. I don't know if this is on the app or this is about me doing something wrong, but when I continue to prompt for the summary, it simply gives it to me because it integrates with YouTube. Interesting. I honestly think this is what the future of actions looks like. This is obviously just quite limited as we only have access to a few Google apps. Okay, so back to reviewing these three. Okay, the EGO prompt generator Literally don't care which one of these three I use for the gem. GPT builder I found the following. I found it to work best in ChatGPT because ChatGPT has tooling. It can generate a profile picture, it can generate the description, it can generate all the instructions and it really closely follows your instructions. I found the instruction following to be actually best inside GPT4O versus Claude that always went above and beyond and try to expand what I asked for. It's really good at prompt engineering, but when it generates those instructions I just feel GPT to be more consistent and more reliable. It's just a feeling, but I do prefer ChatGPT for the GPT builder and so do many others. We teach GPT building in the community with this I barely shared it on the YouTube channel. It already has 13,000 chats. They seem to go up every single day. If you're not familiar, it's completely free. You just hit this button and you can build a GPT by answering one question and then you can skip the follow up questions that it asks you after. This works best in ChatGPT. It was okay in the gems. One thing that I really like though is that you have a super long context window and the instructions context window is also way longer than inside of ChatGPT. So the way I solved this inside of ChatGPT is just by putting the introduction into the instructions and the rest is actually in the knowledge base and due to some recent update in 2024. It can actually perfectly read all of these instructions and pull them in every single time. But inside of the gem builder, this is no problem. It can just go ahead and take all of this super long context. That was really nice. But other than that, the prompt adherence and the outputs were just not as consistent as GPT4. That's why I would still prefer that. Plus, this GPT builder is really designed to use some of the multimodal functions of ChatGPT. It's designed to enhance image generation or data analysis tool usage, Internet browsing usage as the GPT sees fit. All of these are not available yet in Claude and Google Gemini. It's all coming. They already released their imagen free model last week as we talked about on this show. But right now, chatgpt, he's the winner on this one too. But here we get to the third one, which is the Event Description Writer. So this is a real use case. We have all the time inside of the AI Advantage community. We run around 20 events a month. Okay, There's a variety of events and all of them need descriptions. They need for people to see why they should show up to a specific event. And this is one of the few places in my life and in my business where I actually use the raw outputs of a large language model because it's quite concise, it's very factual, and it's based on a lot of context that we provided. And I kind of like the fact that there's a certain amount of consistency when it comes to these. So this is what one of these descriptions looks like, but this one has already been written by Gemini because I actually much prefer its structure and the tone to both Claude and GPT. Now, this was very surprising to me. I was aware that Gemini has this more human and friendlike attitude towards talking and interacting with you, but I just really like the way it wrote this. And I'm not going to go super in depth and give you comparisons of all the little details. All I'll tell you is that the writing style, if you want it to be human, is somewhere between Gemini and Claude's model. And I can also tell you that I started rewriting all the descriptions we have with Gemini 1.5 Pro here inside. My little Events Description Writer workflow is super simple. I simply take this, put it into the Events Description Writer that I set up with some custom instructions, and I'll just say rewrite this and then it takes this description and it creates a better version of it, at least in my opinion. What do I like about it? Well, I like the fact that it actually made these bullet points points with further explanations. I like the fact that this sounds more human than this, which is more robotic. That's just how ChatGPT writes. And I like the fact that it actually here in this part, understood how you add events to a calendar in more detail than GPT did. It's this. You add it to your preferred calendar. You get all the options when you rsvp. I didn't even tell it that. It just knew. And I also like the formatting way more. Look at that. It bolded all the important parts. Even this is bolded. Might be a minor improvement. But this gets me excited because good tone and voice is one of the things that I'm looking for in every single new LLM release. And I recently ignored Gemini advance because I didn't have this GEMS capability. I didn't want to look for my custom instructions every time copy them into it. I just wanted to go to ChatGPT, click the GPT and get my work done. And now I can do that too through Gems. So I'll be exploring this more. The bottom line is if you're looking for something that has more of a human tone than ChatGPT right now, Claude was the best option. But you should also totally consider Gemini. And with Gems feature, you can turn some of your most popular prompts into something that you can use in an instant. Love it now. Just keep pushing Google. Give us a knowledge base. Next, please. All right, next up, let's talk about what's new in AI image generation. And hold up before you say okay, image generation, what do I care? I think I have something this week that will be interesting to people who usually don't even care about image and video generation. And that's because there's this new use case that has become super popular, which is essentially fine tuning a Flux model with your brand, logo or other assets. And the result is that you basically add images like this to your own model by fine tuning Flux, and then you can generate images like so. Now, if you wonder how to fine tune Flux, I created a separate video on the channel exactly on the topic this week. You can check it out. It's only five minutes. It shows it entire process A to Z. Everybody can do it. And with some tips in the video, you can do it completely for free. But I think this technique is really exciting because you could not do this with midjourney because it's closed. You couldn't fine tune it yourself. You could not really do this with stable diffusion. Like you could fine tune it, but you could never get this quality of hyper realistic outputs out of it. Meaning it wasn't really useful for things like product photography. Or by simply fine tuning in the logo of your brand into the model, you can put it into all types of scenarios. Like so look at that. Two T shirts, one picture, no problem. And it doesn't end with apparel design. You could put it on top of beers street signs as Flux can also do text, you could even add text around this. But as you can see, it really shines with hyper realistic image generation. And your logo could now be on all sorts of situations, no problem. And beyond product photography or Eshop mockups, I want to show you this use case where they trained it on a set of alien set designs. And then all the environments it produces are in this same vein of like an alien spaceship, five star hotel. I suppose this one looks incredible, but you can see how it maintains this style because they trained the model on specific visuals. You can do this for any set of images you have. And some people went full retro with this and trained it on a bunch of PS1 games. And now every image this model outputs, when you include the keywords that you train that on, will look like a PS1 game. Quite interesting, isn't it? I think you get the point here. And this week if you look at image arena leaderboards, you will actually find that now for Lux Dev is on top. Because people have been finding so many use cases just like this. It actually overtook Midjourney, which is surprising because Midjourney is definitely on par in terms of quality, head and shoulders above when it comes to tooling. But you cannot do things like this with Midjourney, can you? And to round out AI image Corner for this week, I should also mention there's a brand new tool that came out of freepik. This one is not free. You do need a plan starting at $5 a month to run this. But here's the thing. Although this tool is super high quality, the realism is on par with Midjourney and Flux Feeling like, look at some of these images. It's beautiful, but it's closed, meaning you cannot fine tune it on your own logo. And we have other models that are equally as good at this point. So if this came out a month ago, it would have been really impressive. Right now it feels like another really good AI imaging model, but we already have a bunch of those. The one thing that should be noted about this model is that the outputs are already in 1080p. So they probably have some sort of upscaling in their workflow baked right in. But other than that, I'm not really sure why I should be using this over flux. If anybody has some good reasons, leave a comment below. Other than that, let's move on. So this next one is not really a piece of news that you can use, but it's certainly news that you might use because Larry from our community actually went ahead and he patented one of the songs he generated with Suno. I didn't even know this was a thing. I mean, the discussion around copyright and AI has been rehashed many times and the consensus is that there's no real clear answer. It's sort of a gray zone. Practically copyright is not enforceable on AI generations, but that's talking from the perspective of the companies owning the models or the original authors of the intellectual property that underlie the generations. But in this case, Larry actually went ahead and copyrighted his own track that he generated. Now, could you generate something similar with a similar prompt or by uploading his track? I suppose, but still, I found it interesting that he could get a copyright on the sound recording, the lyrics and and the composition. Now, I'm no copyright lawyer at all, but it seems interesting to me that you could at least copyright the generated output so nobody copies what you generated. Might be just one of those might as well do it type situations if you have big plans for your generation. Like if you have a logo that you're going to build an entire brand upon and it has been generated by AI, probably copyright that thing. But again, disclaimer that I'm not a lawyer and I actually have no idea what I'm talking about in this department. It was just something really interesting and unique that I want to share from within our community. Alright, so next up we have a very interesting one. Okay, so these are Gaussian splats, which is a way of 3D capturing the world around you that can be ran on your phone inside of a free app. Yes, even though I sit in Portugal, I could download this and play with it already. We'll put some examples of what me and the team created up on the screen right now. But basically this allows you to go through the world and use your phone's camera to create Gaussian splats of your environments. If you're not familiar with Gaussian splats, they're essentially the best way to capture a 3D object when it comes to the quality versus how much computing power it takes and they're really great. You can export this as 3D models and then work with them, or you can actually just add them to the app. If you give them permission, they'll use it for their world model. They're trying to map out the entire world. By the way, if you're not familiar, Niantic is the company behind Pokemon Go, okay? So they might be preparing for their next big move. They're obviously huge in both AR and VR. And I just wanted to share this with you because it can be really fun to go out there and use your phone camera to just 3D capture images. You could share them with friends. You could add them to the world model or use them in a 3D software. And this app is completely free. We even featured it in our newsletter, which always means a lot to us because we want to feature the most innovative and interesting app of the week there, which this week was Niantic Scanniverse. And to round today's episode of news, you can use out, I want to show you this little deep fake game. It's very simple. It shows the two images. One of them is fake, one of them is not. You got to decide. Go and test this and see how many points you get. I can already tell you on my very first try. I think this one is fake, right? Yes. Next pair. I only got 7 out of 10, so go ahead and try this and see if you can beat me in identifying AI generated images. And with all that being said, I hope you had a wonderful summer. There's been a lot of AI releases, and I'm here every single Friday to cover every single one of them that you can actually put to work today. For you have a wonderful day, and I'll see you soon.
