Powerful AI Use Cases That Work Today

It's been another week and AI is not slowing down. There's a whole batch of new AI powered tools and LLMs. But there's also new ways people are using these, including some of the world's biggest artists. But seriously, this show is becoming so multifaceted because we gotta cover multiple backers. But then on the LLM side, we have things like the GPT2 chatbot that is back performing better than GPT4. It's a bit tricky to access. I'll show you how we have a new chatbot, Speed Leaderboard, Creative workflows, a new tool to scrape websites that there's so much. All right, so now that you know what's coming, it's time for this week's AI news you can use. And let's just start with GPT2, because this has been the topic of the week. It got people super excited. And a lot of that excitement comes from people's anticipation for GPT5. Oh, GPT5. And as the God King stood before his people, currently, that is the holy grail of upcoming AI releases. People hope that it's going to have agentic frameworks or that it's going to be significantly better than GPT4, more reliable for they expect multimodality and search. Ah, nah, I could make a whole separate video that and maybe I will. But right now, GPT2 is what was released. And if you're not up to date, here's the 22nd summary. Last week, a mysterious chatbot named GPT2 hit the Internet. It was only available through lmsys.org and you could only send eight messages. Then it mysteriously disappeared. And Sam Altman went on to ride the hype wave, tweeting several messages that Fanta fire. But he didn't confirm that it's an OpenAI model at all. This week, the chatbot reappeared under a new name in two different forms, again accessible from chat.lmswise.org and it appears to be superior to GPT4 in several capabilities. One of them would be coding, but in general, people report that it's an incremental improvement over a lot of results you get in GPT4. And that's the summary. A lot of this remains subjective, but you can test it over on chat.lmswise.org, you go to arena, then you type in a prompt. And this is the funny part, you kind of just gotta hope that it's gonna give you the chatbot. So pretty much run anything and then it's going to pick two models for you and you're kind of rolling the dice and hoping that you'll get the GPT2. If you want to test it out, once you do, you're going to see it here at the top, and then you can simply continue the conversation with this bot. If you refresh the page or start a new round, you're going to lose access to it and you're going to have to reroll the dice again until you get the GPT2 chatbot. All right. Honestly, personally, from playing this, yes, I see some differences. I don't think it's a massive game changer here. And the user interface to access this right now is just so unfriendly that yet again, GPT 4 and Opus 3 prevail. And as you can see here, on my first try, I just did a quick vote and it shows you which one is which. So I'm also a good GPT2 chatbot. That's one of the names for the GPT2 here. I could just keep prompting and then it continues the conversation here, very limited and a bit cumbersome. And talking about Claude, this is a very quick one. Claude has a brand new app in the Apple App Store, but this is not accessible in Europe. And as somebody who's sitting in Portugal right now, I can't download it because as opposed to computer, I would need to change my Apple account country. So, you know, if you, if you're outside of Europe and you want to be using Claude, either the free or the paid models, you can get the app on the App Store and that's pretty much it. Okay, so I gotta say, lately I've been loving this weekly series because I do get to spend a lot of time exploring all of these new AI tools and testing them for you. As you might have noticed, I'm a big get a little better every single day type of guy. And if you share that drive, it's really good to get some help along your learning journey. And that's why you need to check out Brilliant.org, the sponsor of today's video. So this really depends on the person, but I personally can only really learn by doing. Memorizing has never, ever worked for me. Just ask my middle school teachers. And this is what Brilliant is actually really good at. It makes you try out some of the content that they teach in their courses. Their interactive learning platform has thousands of lessons in math, data analysis, programming, and yes, even AI. If you're the type of person that hates sitting in a classroom listening to lectures and taking notes, you're really going to enjoy learning with Brilliant. As mentioned, their lessons are not just about memorizing. You learn critical problem solving skills while building your knowledge up with interactive learning lessons. Since you're watching this channel, I would recommend the How LLMs Work course. In seven lessons, you'll learn everything about the inner workings of modern AIs like ChatGPT. This deepens your knowledge when it comes to using these tools and unlocks paths for you to build with AI yourself. If you want to try this and everything that Brilliant has to offer for a full 30 days for free, visit brilliant.org the AI Advantage or click on the link in the description. If you decide to subscribe after that, you'll get 20% off your annual subscription. Thanks again to Brilliant for sponsoring this video. And now let's move on to the next piece of AI news it's that you can use. And while we're on the topic of LLMs, there's actually two more interesting things that came across my radar this week that I really, really wanted to share. One of them is this leaderboard for the speed of different large language models. So whether you're a user or you're trying to build something with some of these models, this might be really useful to you. Look, they're ranked by inference speed. Inference meaning how fast it generates. And look at that. The Mixtral 7B run on Grok chips is the fastest at 48 milliseconds. That's incredible. If you compare that with the fastest GPT4 API on Azure with 230 milliseconds, that's a drastic difference. I don't know about you, but I was roughly aware of these, but never this precisely. Good link to bookmark, I suppose. But inference speed is not the only limitation that LLMs currently face. Another one, and arguably maybe even bigger one, is the fact that they sometimes reply with unpredictable responses. And this especially happens if you attach external files via rag. But attaching a lot of files, a lot of context is one of the best use cases of these systems. So what do we do? There isn't a definitive solution as of yet, but different companies have different approaches, and this one I found particularly interesting. Look at that. It's called Cleanlab AI, and essentially what it does is, along with the answer from the LLM, it gives you a score on how much you can rely upon that answer. So let me just go in here and click one of these presets in here. Let's say we're trying to identify the source of the first following data document. And let's say in the company you'd be retrieving some data inside of a document. Now when you get that, you always have to double check. But here you get this trustworthiness score along with the result. And that's the whole idea here. It gives you a trustworthiness score along with the answers. So if something is clearly labeled and it has confidence in the fact that this is represented in the data that you attached, and this is clearly accurate, it gives you a high score. Other than that, you get something like this. So look, if you use some of these other examples here, by the way, this doesn't even take a sign up. You can just try it. It gives you this high trustworthiness score because it lines up with something in rag. Now look, this is just the first demo. I don't know how well this works in practice, especially when you give it large data sets. That's where this is most useful, but surely also most challenging. But I really like this idea. And their claim is that with this technology they turn a large language model into a trustworthy language model. Nice, interesting stuff. An approach like this might be the future of large language model usage for businesses where you need to attach a lot of context. And answers need to be accurate too. Okay, so we need to talk about what happened in the AI audio space. And before you skip the segment because you feel like, hey, I don't work with audio, I don't care about the creative stuff that much. Well, let me say you might just care about this because this is one of the most popular AI use cases we've seen yet when it comes to public perception. Drake used it at one of his new tracks, namely in a diss track where he used AI to recreate Tupac's and Snoop Dogg's voice without their consent. And he just included in his track that was all across the Internet Spotify at this point that has been taken down because Tupac's estate pushed back, as you can see in this article. And I thought this was so interesting, especially if you contrast it with the fact that Drake was really hating on the AI generated songs that were created with his voice just a year ago. So this is what some people are doing with AI voices, but then others like Randy Travis here are using it for or something that is undisputably incredible use case. So I didn't notice either. But Randy Travis is a singer that suffered from a massive stroke and unfortunately that damaged a part of his brain which is responsible for speech and language. Now, he used AI to recreate his voice from his old recordings and now the musician can create music again. And that's exactly what he's doing. So we had this fantastic team discussion around how funny this is. This is twice two sides of the same coin. Drake using it without consent to diss his competitors, and then Randy Travis using it to recreate his own voice. So just interesting use cases that I really wanted to point out here, as you could be doing this too. I have my very own voice clone that we actually use in some productions and I don't have to rerecord stuff if you want to change a word or three. And you could be doing that today. And while we're on the topic of AI music, I also want to highlight the updates that came to Yurio, because if you're not familiar yet, this is the best in class song generator for AI generated songs. But they had some serious limitations. The songs were only 30 seconds long, then you could extend them 30 seconds at a time. With this new update that changed and they implemented probably the most important feature there, which is song extension, while considering the entire song, not just the last 30 seconds. Beyond that, if you're on a paid plan, they now added inpainting, which if you're familiar with inpainting in something like Photoshop or Midjourney, it's essentially selecting just the part part of the track and replacing that, not the whole thing. So if you don't like a five second sequence, you don't have to regenerate the entire track over and over again. And this is the biggest thing that has been missing up until now. If you didn't like a single second of the entire thing, you had to regenerate it from the ground up and everything was different again. Now you get to do it selectively for just a little part. So this has become a really capable tool very quickly here. Plus they also added plans. So before you only had the free tier and now you can subscribe to these different, different plans. This is on May 9th at the time of this recording. But these usually change over time, so just check for yourself. Alright, so let's talk about something a little more practical and that's this brand new Python package called Scrape Graph AI. But before you dismiss it because you're like, hey, I don't know Python, let's move on to something that I can actually use. Hey, they have this streamlit space that you can actually use and what this does is very easily described. It scrapes a website in combination with a prompt. So when I insert the API key and let's just use GPT 3.5 Turbo here. I can. You could do something like take the new iPad Pro product page on the Apple website, insert the link here, and then what it's going to do is a two step process. One, it collects all the data that it sees on this website. Two, it runs a prompt on top of it and organizes it for me into whatever I want. In this case a list of all features in a bullet point list. And there you go. Here's my answer in something that's called a JSON format. Now I could save that as a JSON or a CSV and keep working with it. Obviously this is meant to be implemented into another workflow with code, but you also have this web interface where you can kind of just try it and see what scraping can do for you in the combination with prompts. Useful little library and tool that I thought might be valuable to share whenever you need to gather some data off the Internet. Even if you're non technical, this might be worth a shot. Okay. And to round things out, I actually have a use case here for you that is universally applicable to everything we talk about here every single week. And that's the fact that you can combine some of these. Okay, we go tool by tool. I give you a first look at what it does and some comments. But just because we look at it tool by tool doesn't mean you can't take multiple. That's why I always promote the playlist with all of these videos. I think it's really worth to go back, review some of the things that have come up recently. So then you could combine them because that's exactly what a lot of people are doing and they're getting unique results that you won't see anywhere else. So what Jared Lu did here and shared on Twitter is very interesting. He uses a combination of two apps. One project Neo by Adobe. I didn't know this. This is a new super simple 3D designer from Adobe where you can create little scenes like this. Okay, look, he just drags in cubes and it's all drag and drop. Everybody can do this. And then he takes that image and brings it into Adobe Firefly, which we covered a few weeks ago, and it turns it into all these different styles. And as you might remember, Adobe Firefly has this reference image feature where you give it a certain reference structure and then you run a prompt on top of it. You can apply all these different styles and it uses the composition of the reference structure to apply the prompt on top of it. So look at that. I'll change the reference from this flowing river to the street. Then I'll remove the hyperrealistic style and maybe I'll add something like line drawing and I'll leave the art filter and slightly modify the prompt and just see what we get here. And there you go. As you can see, it maintains the composition of the original image and does all of this. And this goes for every tool we use. So I don't know, you could maybe scrape a website and use the results with the GPT2 chatbot and prompt on top of it. Or you could scrape a website and use the results to generate a brand new song or a stand up comedy set, whatever it might be. And that's the magic of AI. Transformers are the tech behind a lot of these gen AI models, meaning they can transform one thing into another, but it doesn't mean that you're limited to one transformation. Matter of fact, some of the best use cases are multiple transformations, but the first step is being aware of what's possible, and that's the goal of the show. So my hope is that this was helpful, maybe even a bit inspirative to what you do here. You can check out a playlist of past episodes for even more inspiration. And with that being said, I'll see you next Friday on the next episode of AI News. You can use.