Huge Week for AI Use Cases (Google, Apple, NBA)

This week we had a plethora of new AI use cases pop up. I collected them all and we're going to go through them one by one. Some of these are major advancements of how AI is going to be used today and in the future. Like this interface that is more than 10 times faster than ChatGPT, or this open source image generator by Apple. Then we have NBA coming out with the NBAI great name, I guess, or Gemini advance already pushing highly requested features. There's a lot to go through and, and everything we cover here is either usable today or used by others today. So as per usual, these are all the new use cases in AI that came out over the course of the last week. Let's get into it. So first things first, we need to talk about grok, because this was all the rage this week all over social media. And the bottom line here is it's more than 10 times faster than ChatGPT. I'm not talking GPT 4. GPT 3.5, okay, which is already super fast. Check this out. Here's GPT 3.5. Not bad, right? And then here's grok. So as this is super popular right now, there's a bit of waiting time up front, but as soon as the generation starts, you can see the real speed of this. And the whole point of this is this is a brand new hardware architecture that was introduced by this company. So what they're using in the background is referred to as LPUs linear processing units as opposed to GPUs. Graphics processing units. You can see the speed comparisons here, and it's just wonderful to see alternatives emerge to Nvidia's GPUs. Now expect this to be mainly used by developers moving forward, but this does unlock a whole new set of use cases, as when you're wearing something like a smart device, you want the answers instantly, right? You don't want to delay between you speaking and AI generating answers. That's where this is going to be super useful. But you as a consumer can also use it if you're in a rush. Grok dot com. There you go. Okay, moving on to the next one. This is a Gemini Advanced update. So this is the GPT4 competitor from Google, and what they included here is the ability for it to run Python in a sandbox. So, as you know, this is something we have in GPT4 already with the advanced data analysis tool. And for most people this was at the top of the list when it comes to things missing in Gemini Advanced. So less than two weeks after the release, they added it in. And I think it's interesting to point out that the implementation is slightly different. Different here. So as you can see here, it composes the email and then it counts the characters in Python. Okay, A very simple operation, but if you've used the data analysis tool a lot in GPT4, like me, you will know that it just does it. Okay, and here it rather writes the email and then it sets up the python code, but you have to run it yourself. Okay, so you have to go ahead and say run code in here and then it counts the characters. And there's even more explanation here. So everything that I outlined in my original video about Gemini advance holds true. This is more of a helpful assistant, whereas OpenAI is going more into the direction of an action space model that just gets stuff done for you. Here you have more explanations and you have to run the code yourself, just a different product. So there you go. Now you can execute code inside of Gemini Advanced. Nice. Moving on. Okay, for the next few use cases, we're looking at various implementations and new releases in the image generation space. And it's actually crazy what's happening because some of this stuff is getting implemented by the NBA and players like Apple are entering the scene here now. Okay, to be completely fair and objective, what Apple is doing here is less than impressive. And that's being very soft with my judgment here. They basically released this new open source model, which is like what Apple open source in one sentence? That doesn't make much sense either way. You could try it out over here on Replicate. But look, let me save you some time. If I upload an image of myself and say, make him look happy or anything else, really just the results are not there. It's not good. This is just very far behind the competition. But I did want you to be aware of that because you know it's Apple and they released that AI thing. So let's move on to the next one, which you actually might want to use, and that's stable Diffusion XL Lightning. And this one is released by the company behind TikTok, ByteDance. If you've been following closely, they're releasing more and more AI research, so they're very serious about their moves in this space. You can try this one out over at this hugging face space. So let's try the AI Advantage benchmark with a cat with a hat and just run this. And the interesting part about this is that it actually generates the image four times and it's very cool. Quick at that. So a new approach to image generation. And the claim here is that they become more realistic. And I gotta say, look, when I compare the same prompt, and look, this is one prompt. I'm testing this a few times, so don't take this too seriously. But just by my judgment, the results do come out better and they are more realistic. Look at that. Let's generate another one. So this is just base Stable Diffusion XL on the right side and Stable Diffusion XL Lightning from ByteDance on the left side. Yeah, all of these are just straight better. So if you're using Stable Diffusion XL regularly, this, this might be worth considering. And they also offer this as a comfy UI workflow where you see and get to adjust a lot of the details that are happening under the hood. I like the fact that it includes four steps. Okay, so for our next use case, we'll be looking at a tool that will allow you to edit the text inside of AI image generation really easily. This is obviously fantastic for marketing and I'm happy to say that this tool is actually the sponsor of today's video. And they built something that legitimately answers a request that I receive regularly. Like, hey, go. You show us all these techniques with image generators, but every time I create something with text in it, it's 90% of the way there. And that's just not good enough. And usually my recommendation there is just, yeah, no problem, open up Photoshop, mask out the part with the clone stamp tool and reinsert the text. But what if you don't use Photoshop or you don't have the time for that? Well, that's where a multi tool like Storia comes in. You can try it for free. And as you can see, if you come in, you get 10 credits. And what I want to show off today is their Textify tool. So they have more here. But this is what we're going to be looking at today. So if I just drag over our image that I generated with Dall e here, right. It's a cat with a hat that holds a sign that says AI Advantage. Now this is the typical thing that you get whenever you have text in your shot. It's going to look something like this. The AI is great, but the advantage is just not there at all. So all you need to do here if you want to replace the text, is just pull out a text box over the text and it's important that it covers the entire area that you want to replace. Okay, so I'm going to say AI Advantage, going to change the font size a little bit and then you can manipulate this text box so you can rotate it like so, resizes it a little bit. And what I'm going to do is actually pick these fonts. Now you could let it do it automatically. Okay, that looks good. So all you need to do is say apply and then the Textify tool here analyzes the image and replaces the text for you. No Photoshop editing skills needed here. And that's pretty great. This was actually the first take in the recording that I'm doing here. So good job story. Honestly. Now you could do more operations with this, like send this to the editor, or you could clean it up, remove the background. And even if you want to generate images, they have an image generated in here where you can switch between stable diffusion XL, Dall E3 or Leonardo. So whenever you're creating AI art and you need the text to be right and you want the fastest workflow possible, well, here you go. The free version has 10 credits and then you do need to upgrade to remove the watermarks. But if you want to try Storia today, go to the first link in the description of this video and thanks again to Storia for sponsoring this video. And I gotta say, great job on making this super simple to use. Okay, and then I have two more image generation related ones. One of them is this new GitHub repo which is a one click installer for a brand new way to run some of these models. Now some features that still stand out to me here is first of all, it has an easy setup. So you literally just download the compiled package and run the executable and you've got this interface on your computer, which is amazing. Usually these take a little more effort to install. Secondly, the requirements on this are not crazy, especially compared to some others and what they require. So with 10 gigabytes of RAM and a CUDA capable GPU, you're already in the game. And then you unlock a whole set of custom features that you won't be getting in a lot of these web interfaces like inpainting, outpainting, Control Net or NSFW Filter Toggle. Right? These things are not going to be available if you use it for the web. So that's an interesting one. And here's another really fun one that I just ran into while researching new tools. And this has really interesting and kind of crazy image generation workflows. Okay, and what I mean by workflows is that it ties multiple models and generations together. Sometimes it takes a bit. But for example, this crashed dall e1 is really interesting to Me. So I don't know if I just say spongebob, King of Sponges. And King, keep in mind these are crazy, quirky, weird. Okay? So this is not going to be high quality stuff. It's just going to be interesting and it's going to be wild. If that's your thing or if you have a use case for that, then this site is exactly for you. But I did want to feature it. This is kind of cool. So as you can see, it always does multiple steps and in some other examples here it does like eight to 10 steps. So there you go. All hail SpongeBob. So obviously generated a wilder prompt and you can regenerate this multiple times. There's various models here and here you can see some results that other users just recently generated. I mean, look at that. This is not your typical Dall eject generation, right? And you get that by combining a chain of multiple steps. So if you're into image generation and playing around, and if you're into weird results, which I really am, like this is really cool, then I guess you would enjoy doing this. This was completely for free, by the way. Look at that. SpongeBob, king of sponges. Much more on here. Some oddly specific, but what I did find to work as this site is not that popular yet is that if you look at the play numbers, the ones with a lot of plays are usually the ones that you want to be looking at. Okay. And to round out the section on AI image generators, actually the NBA itself announced something interesting that they called mbai and rather than me just explaining, let me show you what the NBA is using AI for or what they're envisioning. It's kind of interesting and weird and just have a look. So nbai, show me the Pacers game as if it were a Spider man movie. Generates, generates in real time with the game. Now that is pretty cool, right? So where are they heading with this? I guess during live broadcast they will have segments that are stylized like this and generated in real time. So if a replay is happening, it might just happen in a Spider man style or whatever you prefer. Now it's just interesting to see organizations like this implement this technology where a lot of times you look at these things and like, okay, you know, this is really great, but what am I going to be using this for? Well, big players are already coming up with different use cases and this is what one of them looks like. Just thought it was an interesting use case. I wanted to share, but now let's move on over to the world of Large language models because this week has been generous in terms of large language model releases. Let's start by talking about the big one. Gemma by Google. These are two small models that are fully open source. So wait a minute, there's an Apple open source model and a Google open source model. What is going on here? I'm not exactly sure, but we did get two open source models from Google's DeepMind and they say these are based on the Gemini research that they've done and what they're exceptionally good at is code assistance. And first of all, as per usual, they're claiming these are best in class. Now keep in mind, these are best in class at the small sizes. And right here you can see the comparisons to llama 2. So they perform better on all the popular benchmarks than Llama 2 at the 7B model. Now mind you, the parameters are important because these couldn't match up to the 70 billion models. But the cool thing about these, if you didn't know already, is that if they're this small, you're going to be able to run them on your computer. Even though you don't have a powerful computer, you're going to be able to use them on your phone and that makes it very versatile and people can build apps on top of this now and they'll be getting way more quality and capabilities. Now I do want to point out that all the benchmarks they're selecting here are mostly coding and math related and that's what these models really shine in. Okay, so if you check out their technical report, you'll find that yes, the Llama models that were released by Meta perform very similarly when it comes to answering questions and reasoning. But when it comes to math, science and coding, Gamma just outperforms it and it's not even close. Now, without even diving into the technicalities, if you just look for the word safety, as this is a Google model, this will be very censored. Looks like the word safety appears 24 times in a 16 page paper and it performs better on all of these safety benchmarks too. In other words, it's going to be quite censored. But hey, that's fine if you're using it for coding rather than answering questions. Right there, it doesn't matter as much. So if you want to try building something with these, you can head on over to this URL and right here you have a quick start guide. And if you want to just try it, there's this hugging face chat where you can try out the model. Let's have A code of Snake game. Now, I need to point out this. At this point, it feels like all the new models that are coming out include training data for some of these snakes. Standard tests like coding a Snake game or playing Tic Tac Toe, I don't think those are good benchmarks anymore because they know everybody's testing the same thing, so they just include that. It's just ridiculous how quickly these models got better at these standard use cases, as all people on the Internet keep testing the same things over and over again. So just be aware of that. But hey, that's not it for large language models this week because we got another one from Mistral and this one is called Next. And you can also try this out right away. Okay. If you head on over to chat.lmsys.org you can chat with the Mistral Next model right here. In typical Mistral badass fashion, they just dropped the discord message and thereby announced the whole thing. No announcements, nothing. And this one also just came out. It performs well on some of the standard benchmarks, like generating Snake games and stuff. And what people basically say about this one is that it's really good at logic, but many people also claim that it's somewhat weird and way weaker in some things. So I'll withhold my opinion on this one. For now, I'll be following the leaderboards. What they say this Chatbot Arena ELO leaderboards have, in my opinion, really turned out to be the best evaluation of how good a model performs. But both Gamma and Next are not on there yet, so we shall wait and see. All right, and there you go. Those are all the new ways you could be using AI. This week we had a really colorful mix. A lot of big releases, as per usual. If I missed anything or you found one particularly interesting and you want more of a deep dive rather than a first look and overview, leave a comment below and if you enjoyed this. I make a video like this once every week, and there's a playlist where you can check out what came out over the course of the last few months. All right, and that wraps it up for now. I'll see you next week.