AI Realism Breakthrough & More AI Use Cases
Flux, an open-source image model, is generating photos so realistic they're blurring what counts as real. The video demos virtual try-ons and deepfakes, reviews new LLM updates from Grok and Google, and shows Anthropic's prompt caching cutting costs by 90% and latency by 85%.
Chapters
Okay, listen, so this week and news you can use is quite different than the usual weeks. As you know, every week me and the team go in and pull together all the new AI releases, research them, test them for you. And then in this video I present you all the results. And usually we start with ChatGPT upgrades, LLM upgrades, but this week I want to lead with Hyper Realistic image generation because I think we literally had a breakthrough in this space and the first actual use cases like E Commerce are already popping up. So I'm very excited to bring you a pact we week of news you can use. Although it is mid August, we're going to be covering Hyper Realism and its use cases, what happened there since last week. But there's also a new ChatGPT model that ranks number one above everything else now, which is already inside of ChatGPT. And we have Google releasing a voice assistant of their own that you can actually use on your phone. The top comment on last week's video was already saying that it feels like we're beginning to accelerate from Frisco fatsies. And I can only agree last week was intense, but it feels like we're entering a whole new era. And I'm not just saying that lightly. Let me prove my point here by showing you this week's AI news. And you can actually use starting out with the Grok release. And this is linked to this Hyper realism story because Grok 2 has released and it includes image generation 2 and the image generation is from the Flux 1 model by Black Forest Labs that we covered last week. And that's where I want to start. We're going to talk about Grok 2, the LLM and how it compares to other LLMs later on when we talk about that. But the fact is that because Flux is open source, we covered that last week. If you haven't seen that, check it out. It's a legit breakthrough to get a mid journey level model that is open, open source and people can build upon. And in the next few minutes I'll show you why. But the point is that it's integrated into Grok and Grok already shipped. It's here, this is Grok 2 Mini. There's the larger model again, we'll talk about that later. But you already have this Flux integration in here and it's quite unhinged, like not completely uncensored. Okay, so you can't do nudity and things like that, but you can generate political figures and compromising situations and you can also generate all sorts of copyrighted materials. Like company logos. This is AI generated right here. I made it live, so fair enough. Now one of the most popular social media platforms on planet Earth can generate copyrighted materials or political images like this one, and all sorts of other weird stuff that is related to politics and tragedies and sometimes combining them both. I don't even want to show that stuff in this video. The point is, it's quite unhinged. But that's where the story only begins. Because Flux is open source, so people can do all sorts of stuff with it. And if you've seen last week's video, my review of it was, wow, it's really good. It's pretty Best in class at text generation. In Hyper realism, it's quite good. But midjourney is still king. But that was last week because people have done a lot of work since then. And the fact that it is open source allows for something that is called a Lora. And if you're not familiar, let me introduce you to the concept of a Lora for a second. LORA basically stands for low rank adaptation. And what that means in human terms is that you can add extra data to the imaging model. In practical terms, you can add images of yourself and then train the model to generate images of you. Or you could add a whole bunch of hyper realistic images that look really crisp and super realistic real photos, and then the model will be able to pick that up. And that's exactly what people have been doing. And that's why we have various offshoots of this Flux model now, because it is open source and you can do things like combine it with Loras and we get something like Flux Dev Realism, which is basically the Flux model with a realism lora attached to it. Now running this is not free. It costs a few cents. You need to sign in with GitHub on this replicate space. I'll just briefly do that. And also I should note that what we learned since last week's testing is that the prompting is a little more intricate we with Flux. So you need to be using a prompt generator or be very detailed in your promptings. A lot of the simple prompts that might reduce stunning images in midjourney won't work as well in Flux. But before we even get into this app, I want to address the question of like, okay Gorsa, like hyper realistic images. Why should I even care? Like, image generation is really good, but I have no use case for it either in my work or my everyday life. And to that concern that is very common by the way, these days. I Would say, fair enough. I for myself, found this use case of creating these amazing custom thumbnails of me in various situations a lot of the times. But most people won't really have a use case. But what you do have is the fact that the word photo is kind of a term that everybody uses, and everybody has a fixed definition of that now. And the point of this might not even be a use case. It might be the fact that you need to change your vocabulary or change the definition of what you consider a photo. Because what we're about to generate with this flux dev realism model here is indistinguishable from real life. Like literally. And I don't mean sort of indistinguishable. If you would see this image and let's say National Geographic, nobody would be able to tell, not even a trained eye. The fingers are perfect, the skin texture, the beard, the focal plane, it's all just like a real photo, just like this other images here. And what I hope that this segment here in this video does is that you might start questioning what even a photo is. Because up until now, a photo is a moment in real life that was captured through a camera. Whether that was done back in the day with film or through digital. Everybody agreed on what a photo is. But now also, this is a photo and this is not real life. And sure, you might argue that Photoshop took us in that direction already, but that was still a skill that was hard to access. Now it really gets democratized. Like heck. If you're watching this video, you can just log in here, add a few cents to your replicate account, and go ahead and run this yourself, like so. And all of a sudden you can generate all sorts of fake images. But that's only my first point. The second point is actually use case related, because, okay, sure, we might have to redefine what we perceive as real when we see digital imagery from here on out. And there you go. This is the generation. So the eyes a little weird. No problem. I'll just rerun it and in another four seconds we'll have another alternative, because that's how simple this is. But then certain small companies and indie hackers already found use cases for this in the real world. And they're first because they're the most agile. Right? A big corporation is going to take 12 to 24 months to actually implement this meaningfully. But this is the moment where that process begins. This is sort of the tipping point of realism, because this model is open source. Look at that. This one is flawless, except of Maybe this little text piece. The text is right, the background. This could be from any conference. So what did these small teams or individuals find? Well, I have two examples here. One of them is called let's AI and they basically plugged in Flux into their product that allows people to try on various clothing in an online setting. And keep in mind, this is just the first version of it. Look, this is Linus trying on Ray Bans from some e commerce store without actually trying them on. Same example with a Moncler jacket like so. So it's quite easy to imagine a future where online shopping turns into, hey, upload five images of yourself and then here's the product catalog with you actually wearing the products. I mean, that will convert so much better than you just seeing an image of some random model wearing it that might have a completely different body type than you. So that's one very interesting use case. And the second use case actually relates to what Peter Levels here on X has been experimenting with. He's a popular indie hacker that is always up to some new project. And right now he's playing with Flux. And he added his own Lora to the model. And here you can see he generated himself in four different generations, which is interesting, but I think even more interesting than that, he actually did a little pipeline where he generated an image with Flux and then fed it to Clink to generate a AI YouTuber that looks hyper. And the video aspect here is really the next step, but we'll cover that once it's relevant. For now, character consistency and the lip syncing is just not there yet. But hyper real images are with these Flex models that we just covered here. And just to round out this segment, I want to just point your attention towards this GitHub repo that popped up over the last week. It's called Deep Live Cam. And in case you haven't seen this yet, it's very simply described. You basically can install this locally and with one image, it creates deep fakes of anybody and it creates a webcam image that you could then feed to Zoom or Google Meets or whatever you might be using. And all of a sudden you could potentially get fooled by somebody using something like this into thinking that you're talking to somebody else. So this is why I wanted to feature this first, because these incremental AI advancements often seem meaningless. Like, okay, new model, who cares? I'm not going to be using it. But in a case like this, I want you to think about what this means for the current digital world and for things that we take for granted. Like if a family member sends you an image, you don't question if that's a real image or if they photoshopped it, right? With technology like this being accessible inside of WhatsApp Instagram, their models are not so good. But now with Twitter or X integrating Flux into their platform, it's just a question of weeks or months until this is widely available to billions of people. And not just people who watch his videos and use something like Replicate or premium subscribers on X as it is now. And one more thing before we move on, consider sharing this video with a loved one. Because no matter how I look at this, education is the only way that I can see on how to protect yourself from these technological advancements and these potentially malicious use cases. And then on the bright side, that will probably transform E Com very soon here. And that should be relevant to everyone involved with marketing or entrepreneurship in any sense. So more and more AI tools are becoming incredible at generating code, which is great unless you don't know what to do with it. We've seen a lot of people recently hop into something like sonnet 3.5 by anthropic that is really good at generating code, and they ask it to generate something like a Snake game just to get an error which they don't know how to resolve, and they completely hit a brick wall. And that's why having at least a little bit of understanding of how code works is really beneficial while trying to utilize the latest AI tools. And one fantastic resource that you can use to get up to speed on how to get these basics under your belt is brilliant.org the sponsor of Today's video. They have beginner level courses to teach you all the basics, but then they also have more advanced courses like this one called Designing Programs that can really take your coding skills to the next level. Here you can actually learn how to build games and apps that respond to live user inputs. It also teaches you how to properly check for errors and debug if problems come up. And by the way, that's a skill that's really useful when working with AI tools because they do a lot of the writing for you. It's just that bugs make their way into the code sometimes and you need to know how to deal with that. One thing that I really like about Brilliant is that you're always hands on building something or interacting with exercise. You're never forced to sit for an hour long lecture on something that you really don't care about traditional education anyway. If you really want to level up your own skill set and take full advantage of the tools available to you today. Head on over to brilliant.org or click the link in the description to try it for free for a full 30 days. If you decide to stick with it, you'll get 20% off an annual subscription. A big thank you to Brilliant for sponsoring this video. And now let's get back to some AI news you can use. Okay, now it's time to talk about LLMs. I'll put my headphones on here to get a little more serious about this because there has been quite a few updates and I'm going to keep it short. I'm not going to go too deep. I don't think we had anything that is like a complete game changer this week, I would tell you that. But we did have various releases from well on one side x Twitter with Grok tool and then we have a brand new model out of Chat GPT, the ChatGPT 4o latest 20240808 release. You can find that in the new ChatGPT app too. And then there was this entire story that unfolded with this new model called SUS Column R that popped up on LMSYS and Arena and it ranked really high. Nobody knew what it was. People were rumoring that It's a new ChatGPT model, but now it has been revealed that it was actually the Grok 2 beta release. Okay. And this was the proper Grok 2 model as of right now, at least for me and the people that I know. When you go to X you can only access the Grok 2 mini model which is sort of like GPT 4 on mini. So what is unique about Grok 2 and what is new about the new ChatGPT model? Well, first of all, the story really begins with this Chatbot arena here because as I told you, this new model sort of just popped up out of nowhere and was ranking really high. And this seems to be the new default way like a lot of these companies like OpenAI now also X test their new models because it's a great way to get it into users hands and to get some feedback on how users actually use it and enjoy it without revealing the model. So they release under anonymous names. And this is also how a few recent OpenAI models were introduced. And one comment on Chatbot arena actually made a mistake last week that I want to correct this week. Thank you so much influential studio for this comment here on the video pointing out that when you vote on Chatbot arena and you can see which model you're voting on those votes don't actually count. It only counts the anonymous votes. That makes a lot of sense as this ranking here is fully user voted. So for example, in this view where you can actually see what you're comparing, these votes do not count. Only the ones from the arena here actually count where the models are. Anonymous Just wanted to correct that. But back to Grok 2 so it has released and what's the story here? Well it's a top tier model, it's a GPT four level model that is not quite best in class at anything in particular, but it's really well rounded and the biggest selling point here is the following. It's plugged into all of the Twitter data, the full Twitter firehose, all of the news story, all of the opinions that go down on Twitter day to day they are being infused into the model so you can use it for some use cases that require browsing and that don't work as well with other models. Like what are the top news stories relating to AI for today? And keep in mind this is the mini model. This is not the GROK tool main model that we are looking at here. By the way, while this generates, this will take a few seconds because it does need to look at the Twitter API and all the data there. But as of these released benchmarks it's interesting how they structured it and it's a bit deceptive. So I want to clarify this because it compares it to the Turbo model or Gemini 1.5 Pro and some of these more competitive models like Claude 3.5 Sonnet or Llama 3405B are all the way on the right. And the reason I say that is because for example GPT4 Turbo, this is the release that happened right after the voice assistant announcement back in May. And back at that point a lot of people argued that GPT4 is actually worse than GPT4O. The benchmarks were slightly better than GPT4 but the point is this is not the fairest comparison and they're right next to each other. So don't take this delta too seriously. What you really want to compare is Sonnet over here with Grok2 because what you really want to be looking at is Sonnet over here and llama 405B. Those are the most up to date versions. Again, this GPT 4.0 and Turbo models are back from May. It says that down here and if you compare it to something like Sonnet it actually loses out on all benchmarks closely. But it's worse except of Math Vista over here. But again, as I always say, these minimal differences in benchmarks are not a game changer. What matters is how it performs in practice. And oh, okay, right now it's actually still loading, which is a little weird. I'll regenerate this. I gotta say, during my testing of this, before the recording of this video, this went actually super smoothly. Ah, there you go. On second try, it just gets the stories right away and as you'll see at the bottom, it will reference the tweets it pulled it from. So this is actually a fantastic Twitter search engine and I think that is the main use case here. This combination of the Twitter data firehose and having an LLM that has access to all of it is actually quite powerful. And as you can see, it talks about the Search GPT prototype and Google's AI integrations with the Pixel 9 and of course the Grok 2 launch. So this is a fantastic use case that you can be using today if you're subscribed to X Premium, which I believe in Europe comes in at €10amonth. It's actually 860. Yeah, that's correct. If you go to monthly and then premium plus is 20. But Grok already comes with this. So yeah, this is a paid thing, but it's just a brand new way to use Twitter. And then of course it also has the image generation features with Flux that we talked about in the first part of the video. But there's even more here because as it showed, what's coming down the line is also some multimodal capabilities where it has vision capabilities. They'll be offering enterprise API, which could be an interesting way. I mean, they'll give you access to LLM that has all of the world's knowledge that is in Twitter. Time will show what that will be used for. And what's my personal first impression of Grok and its outputs? Well, it's good, it's certainly usable and if you only want the text generation features, then it can certainly act as a replacement to ChatGPT. Now I personally and many others still prefer Anthropic's voice. It's just more human and less robotic like ChatGPT. But this is decent. But again, I didn't get my hands on the full Grokto version, so I can't really give my full opinion. But I also say this, it lacks the tooling. Just some of these other tools also do file uploads, functional mobile apps, GPTs. I use these things all of time. Now, I do consider myself a power user. But still, if you use those features, maybe code interpreter or image inputs, then you won't have those in here. So what should you use? Well, let me sum it up briefly. As of 15th of August 2024, as a General Purpose AI Assistant, ChatGPT still is best because of all the functionality that I just named. When it comes to writing tone though, Anthropic Sonnet 3.5 is my go to. When it comes to code generation specifically. Also Sonnet 3.5 head and shoulders above everybod else right now, but when it comes to research, perplexity is your friend. And when it comes to actually using LLMs with live data, well, I think Grok actually sort of takes the crown here because it is plugged into all the Twitter data and it references all of that. And as Twitter is the place where news breaks first, I mean heck, a lot of this video is just me and the team spending every single day on Twitter and pulling everything together and then me digesting it for you. Well, Grok can sort of do that already too. So that would be the one use case where this, this really stands out. And also one more thing is that Grok is actually sort of uncensored. And by sort of, I mean again, it's the same thing as with Flux. It won't produce R rated content, but it doesn't have problem with cuss words or things that are in like an ethical gray area. Anthropic is on the other side of the spectrum. They're extremely strict and ChatGPT is quite restrictive, but not as much as Anthropic. Anthropic is really extreme and Grok on the other hand doesn't have a problem with profanities. And now moving forward, there's actually also a brand new ChatGPT model and this is not just the API, this has actually been integrated into chat that you might be using every single day. Because if you look at this tweet from the official ChatGPT app, there's a new GPT4 model out in ChatGPT since last week. Hope you are all enjoying it and check it out if you haven't. We think you'll like it. And the funny thing is, nobody really noticed. That's how minuscule the differences between the models are these days. They ship a new thing and they have to announce that there's some update because nobody has noticed otherwise. But yeah, there's a slightly upgraded model. Apparently the biggest difference is in how it handles chat conversations, so it has been optimized to interact with users in a dialogue. And there's also a brand new API endpoint for people to use. But it's funny because the dev account still says hey, if you're a dev, you probably still want to use the 0806 API endpoint, not this latest one that was released for ChatGPT. That one is just best for chat use cases. So there you go, a minor update on that front. If you were confused by this, I will be reporting back once full Grok 2 comes out and I'll get to test it a little bit more for my personal use cases. For now I only have the mini version. So moving on to the next story, here is a few releases out of Google. One of them is Imagen Free. And you know, I'll keep this as short as possible. It's a good image generator. It's their best image generator, but compared to something like Flux or Midjourney, it just doesn't hold up. It does text well, but so do others and they're open source. But yeah, it's better than anything that Google has done before with image generation and it will be introduced into their hardware and their software offerings. Just like this second announcement, which might be more interesting here. This is Gemini Live. Okay, and this is the voice assistant that OpenAI promised, but for Google. Or is it because the reality of this product is probably the biggest delta between what some people hyped it up to be and between what it actually is. Because the reality of it is yes, it is a voice assistant and yes, it also already shipped. Android users already have this on their phone. I'm an Apple user, but I'm lucky enough that team member Daniel actually went ahead and gave this a shot and tested it and I'll just quote some of the points he forwarded here to me. Keep in mind that this is coming from an angle where we're comparing it to what the voice is saying assistant promised and what is available in the OpenAI app today. Because if you're not familiar, there's a voice assistant already. It might not be the sophisticated one with the voice changes that you can interrupt and the multimodal capabilities, but there's a voice assistant you can use the voice function to talk to. ChatGPT. Right? Quick spoiler. That's what this really is. Google shipped a voice input and output function that you can also interrupt, but it's not great. Okay, so what's the review? Well, apparently the Gemini Live voice assistant feels more like a beta release than something that is actually on the level of the voice assistant teased by OpenAI. The voices are good. How can I help you today? But so are ChatGPT's voices. Today there's no voice modulation and no multimodal capabilities like using the camera to actually infer context and to use it as this advanced voice assistant. What it does have is the ability for you to interrupt it, which is actually my biggest gripe with the current version of the ChatGPT voice features. But the problem is it's not great. And Daniel reported back that if he has a speaker volume on the phone over 75%, the voice assistant actually starts interrupting itself because it hears the output and then it stops. I think you get the point. ChatGPT never does that. And because of this, it concluded that the interrupting feature is something that's currently more annoying than useful. Now, sure, they can fix that over time, but again, it just goes to underline this point that it feels like something a little half baked that was maybe rushed out, but not to bash it too hard, what it does have is access to integrations like your Google Calendar or your Gmail, and then you can interact with those on your Android phone. That is absolutely, absolutely fantastic and something you cannot get inside of ChatGPT as of yet. So there you go. That would be my first little look at the voice assistant feature. I do have to add that at the end of their presentation, they showed that they're looking at this advanced multimodal voice assistant in the future. Right. But as of what's released today, it's just voice input and output with Gemini, which is a nice to have, but that doesn't mean they're pulled ahead of OpenAI. They just caught up and that's fine. But let's call a spade a spade. If you have a different experience, by the way, please leave a comment below. Would love to hear about it. All right, this is, this is going to be a quick but fun one. Viggle, the app that came out a few months ago that lets you put yourself or somebody else into a dancing type video, has a new update where you can actually do it for two people. And I think that's sort of fun because you can use it as a fun way to communicate with friends or family. I just briefly want to show it to you. So if you go in here, you can see the new update right here. I just logged in with Google on a free account, by the way. You can just try this right away. And if you head on over to this multitab, then you can pick a template. I'm going to simply take A Matrix fight. That sounds perfect. All right, use the template. And now I can pick the two characters, right? So one character right here. I'll just use the camera real quick. Okay. Ideally, it should be a full body photo, but I'll just do selfie here and then as a second one. How about a picture of Tyrion Lannister here? Because you guys seem to enjoy the Game of Thrones clip we did with the the sponsor last week. And that's it. I'll just go ahead and generate. And by the way, this is not sponsored. It's just an interesting release of a cool app. Let's see what we get here. Okay. No. Hate Tyrion. This is epic. Well, this is as I thought. This is really sort of funny and quirky, and you could just download it with the watermark light, so no problem. You could send it to somebody. Again, this is just possible. The free plan, sure, they have other tiers. I haven't even looked into them so far, but there you go. A fun little use case. Because AI can be that too. It doesn't always just have to be useful and productive. And then last but certainly not least, there's a very, very interesting release out of Anthropic coming this week. And this was sort of shocking to me. Prompt caching with claude and what this essentially is that it saves context into a cache memory that goes along with the API, with the practical result of reducing costs up to 90% and latency by up to 85%. Meaning you can integrate really complex Personas into Claude and then call the API. It will cost 90% less, and it will be roughly five to ten times as fast. I mean, that's a bold claim. And reading through this, this sounds really impressive. All of these use cases, conversational agents, coding assistants, anything that needs a little bit more context. Like, for example here, if you upload a book of 100,000 tokens into one of Claude's models, the latency without the caching would have been 11 seconds to get a response. Okay, so you give it all this context, then you ask something about the book, and then it takes 11 seconds to reply. That's what this means in practical terms. And with the caching 2.5 at a 90% cost reduction, honestly, this sounds a little too good to be true. So I had a closer look at this, and I have, to be honest, I have one single gripe with this, which is like, what is the downside here? What is the negative part? Honestly, this looks too good to be true. They say it's still In Beta they give you explanations on how it works and how it's priced. One limitation is that it doesn't work on the OPUS model, but the Sonnet model is best right now anyway. There's even a prompt caching cookbook on their GitHub if you want to check that out. And hey, let me tell you what, this just came out. I didn't really have time to dive deep into this, but over the weekend I'll be having a closer look at this and experimenting with this because again, it just sounds too good to be true. This is sort of like having RAG up to a certain context limit I suppose, but without the downside of the long loading times and the embeddings being created and retrieved and compared to just adding a lot of context, which was the RAG alternative. It's way faster now too, so that's amazing. But I want to know what's downside here and how does this compare to for example fine tuning? Because they do say that one of the best use cases here is actually giving it multi shot prompts into this cache and then it can consider that as extra context for the generations. So I have to run a few experiments and I'll report back. And if you're interested in this sort of topic, I do want to point out the fact that I'm actually restarting my LLM Innovations event series. Before this was called ChatGPT Innovations and I used to do it every two weeks for a year straight for all course members and now I hold it in the community and it's a part of the membership and I'm going to hold it once a month. We went with this image of Igor Einstein because this is going to be where I'll be presenting some of the experiments that I run to the community. It's a long format, usually the lecture takes about an hour and then we do a Q and A afterwards and the next session in September we'll be looking at when you should be using prompts versus using a GPT versus using fine tuning. And apparently now I'll have to extend this with when should you be using prompt caching? And all of the results I'll be showing there will be evidence based and include results of the experiments we run internally with the team and then moving forward I'll do one of these a month as this has always been the most popular format within the community. So we do a lot of things there. But I thought I'd tell you about this one and as I pointed out many times, that's the whole idea behind the community. We can go deep on single topics rather than doing what we do on YouTube, which is brushing over many different topics. And that's because this will be assuming that you already took my prompting course and the GPT building course that is also accessible within the community. I cannot make that assumption in a video like this, but what I can do is test prompt caching more and then come back with a video on that. So there you go. AI has been wild lately and these are some very exciting developments. We'll be playing with all of it then if we find something interesting. I'll be reporting back next Friday in our weekly show AI News you can use and that's all I got for today. See you soon. We'll be playing with all of it and if we find something interesting, I'll be reporting back next Friday in our weekly show AI News you can use and that's all I got for today. See you soon.
