Unbelievable AI News You Can Use!

Another week, another batch of AI applications  that you could be putting to work today. And   it's incredible what kind of variety we're  getting here week by week. Because today   we'll be looking at one application that remote  controls your computer and does things with GPT   vision. Right now I'm hands-off, it's creating  a training plan for me. But then we also dug up   some interesting GPTs which use external actions  for specialty results. Adobe has some new features   and so much more. And check this out, I came up  with a new name for this weekly show. From here   on out I'm gonna call this AI News You Can Use. I  don't know about you but I really like that name,   more so than this week's AI use cases. And it came  up when somebody asked me what I did on YouTube.    I said I always talk about AI News You Can Use.   Alright, let's get into the practicalities here.

So first up we have a big big release by  Google. Google AI Studio came out with Gemini   1. 5 Pro and I did an entire dedicated video on  this release. This opens up new possibilities   so it was worth talking about for 15 minutes  straight. The summary of that is essentially   you get Gemini 1. 5 Pro with a million tokens of  context, you get a lot of customization, you can   really easily fine-tune models, plus you can now  also upload videos and directly prompt on top of   videos. This is something we couldn't do up until  now. If you're interested in this whatsoever,   I strongly recommend you go check out that video.   I went into great depth on the entire interface   for non-developers. If you're a dev, this is a  perfect primer for the interface because there's   a lot of new features here. You can Google Drive,  there's different ways to prompt here, you can   fine-tune models and upload all this stuff. And  all of this is free. If you're in the EU, you will   need a VPN to access this. So that is Gemini  1. 5 Pro with Google AI Studio. But moving on,

we'll talk about a bit more of an esoteric app.   Matter of fact, this came out a few weeks ago,   but it has been just getting popular on GitHub  over the course of the last weeks. It only has   140 stars right now. And this one is called  Open Interface. And to be perfectly frank,   a lot of these are coming out these days. They  use something like GPT Vision to scan your screen,   and then they use different tools to remote  control your computer to get things done. This is   aligned with the whole idea of AI agents actually  doing the work for you and not just assisting you   in doing the work. And to be perfectly honest,  not a single one of these really works so well   that you would want to use this in your everyday  life. We talked about the Devin demo two weeks ago   that looks really promising, but we don't have  access to it. And everything else in this space   has been really interesting, but not that useful.   And I did spend around two hours playing with this   today. And I tried different things. And once it  gets complex, once you have complicated graphical   user interfaces, or you have longer actions, where  it's four or five, six steps, it starts tripping   up. But for simple things, it works pretty well.   And it's super easy to install. Actually, you   just go down here, I toggled this, download the  latest release, dragged it over to my applications   folder, gave it some permissions. And here we are,  this voice input doesn't really work for me. But   I'm just going to paste this super basic prompt  right here that did work for me during my testing,   because I feel like you should at least know about  this. Just word of warning, you do need your API   key in the settings. And every single action is  like five to 20 cents, which is quite expensive,   considering that it doesn't get any substantial  amount of work done. Nevertheless, let's give   this a shot. So you know what's happening in the  space, I just simply said create a word doc with   a training plan for me. And then I gave it some  details, 80 kilograms, male 189 centimeters,   and then I just say submit. And now we can sit  back and watch the magic happen. And yeah, by   the way, I'm using a very similar prompt to what  he shows off just because this is one of the ones   that really reliably works and simple tasks like  this. It just works. So as you can see, it opened   up word right here, it always takes a few seconds  because GPT Vision takes a few seconds to process   and give you a result back. But it's basically a  combo of the different APIs. And in a second here,   it should start writing out the meal plan with  GPT-4 Turbo. And there you go, training plan, look   at that, hands off, this is happening by itself.   Now, you can see that it's using this button,   this button, and that is because I have the German  keyboard on. So you do have to switch to the   English keyboard. So I'm just going to stop this  and redo it with the English keyboard. All right,   I switched my keyboard, it's going to submit  this again. And there you go. As you can see,   with the English keyboard, it works perfectly.   I mean, this is some magical stuff, right? Yes,   obviously, it's super early. And eventually, we'll  have apps like this built into the software that   requires it, maybe even into the operating  system. But look, right now, I'm hands off,   it's creating a training plan for me. This is  pretty cool. And I thought it was interesting   enough for you to see where we're at today. So  to save you time, let's just do a jump cut to the   end result. So you can review what happened  here. And there you go, that's kind of it,   it saves the doc. And yeah, now I have a training  plan that was programmatically written. All right,

moving on to a significant release. This one  is by Adobe Firefly. So if you didn't know,   this is Adobe's free to use image generator. And  they added some significant features here, which I   think you should know about, namely the structure  reference where you can use the composition of an   image to regenerate another image. We've seen  this in other forms, but Firefly is really good   at hyper realistic images, and they're fully,  fully commercially safe. They're the one model   that guarantees you that no matter what happens,  their model has been trained on safe images that   are open source that they own the rights to and  not just all images across the internet, and then   expanding it with features like this, where you  can input the image here, and then it regenerates   a new one, but maintains the composition maintains  all the lines, that's actually pretty big,   because you can do things like upload a sketch  that you created, and then it follows the sketch   and then just regenerates that as a full fledged  AI image. And the way this exactly works is if you   go to firefly. adobe. com, you're going to find this  structure drop down here on the left side. And in   the structure drop down, you get to pick one of  these, let's just take this interior picture,   just because this is a popular use case, you  can take pictures of interiors or sketches of   interiors and turn them into realistic images. But  obviously, you can also upload your very own. Now,   if I go ahead and just say minimalist Moroccan  interior, dim lighting, it should follow this   image and recreate the same setup. And as you can  see, it did that. Now, I wish there was a way to   kind of enlarge this, maybe I can just zoom in  on this. As you can see, it's the sofa with the   light in the middle and the table here too. And as  you can see, yeah, this is the same room just in   a different style. It gives you alternatives here.   Firefly is really one of those tools that is super   underrated and isn't talked about enough. And as  mentioned at the time of this recording, this tool   is free and it has been since quite a while.   So I don't think Adobe wants to monetize this   over time, we'll see. But there's much more here  with styles, effects, and in combination with the   reference images, it's quite powerful for business  use cases, because you can be 100% sure that this   is not going to infringe any copyright. Okay, and  to round out AI imaging for this week, we'll talk

about Renoise. This is a very funky one, I just  want to briefly touch on it. Last week, I brought   up a tool that turned the cat's meowing into dog's  barking. This is where we start. And this is where   we end up. Okay, and I was like, what the heck  would you even use that for? And I gotta say the   comment section on that video, you guys are super  creative. Thank you so much for a lot of fantastic   suggestions, what you would do with a cat that  barks. But one comment really stood out. It said   that, yeah, essentially, this is a translator  between cats and dogs, they're finally going to   be able to talk to each other. I was like, yes,  exactly right. AI is going to do that too. Maybe   I guess what I'm trying to say is this tool is  similarly useless, but maybe you'll find a use   for it. I don't know. It's called Renoise. And  it basically changes subjects in an image. So   if I take this kitten in a basket as an input, I  can say a Lego kitten is sitting in a basket on a   branch. And you can essentially remix that kitten  to be a Lego kitten while maintaining most of the   rest of the image. In the second example, we're  turning the kitten into a broccoli. So I don't   know if that is something that excites you, then  yeah, here's a tool that can make that happen. You   could try this with your own images. Let's start  by giving it access to the webcam. So there you   go. That's an image of me recording this video.   And now I'm going to say a broccoli with fire as   hair. I'm not exactly sure where I'm going with  this. Let's just see what it gives us. That's not   really good. Maybe let's go back to the examples  of turning a lion into a tiger. I don't know,   it caught my interest. I wanted to show it to you.   Now you know, let's move on to something else that   you might just have a use for. And that's a little  GPT corner for this week, because I actually found

two that are super interesting. Like more and more  of these interesting GPTs keep popping up over   time. If you missed the news, they just announced  that they started rolling out monetization for   GPTs. They're working with select creators. We're  yet to see how that reward structure is going to   be set up. I'm super curious to hear what kind  of revenue split that is going to be and how much   money that's going to be generating. My guess is  that GPT creators will be making tens of dollars,   as Mr. Wonderful from Shark Tank likes to say.   But there's two really interesting GPTs. And   if you have a plus plan, I would consider using  these. I pinned one of these for myself. So one   of them is a photorealistic GPT. And this seems  to be using Stable Diffusion Juggernaut XL,   which is really good at generating pretty humans.   That's essentially what it does. It's a fine-tuned   stable diffusion model that creates beautiful  people. So whatever prompt I give this, it's not   going to be using DALI for you. As you can see,  it goes into the action right away. And this way,   you can actually use a superior model in certain  ways to DALI right inside of ChatGPT. In other   ways, DALI is superior, but it's just good to  have choices, right? And as the name suggests,   photorealistic GPT, if that's your goal, if you  want photorealistic results inside of ChatGPT,   this is your best bet. DALI is not very good at  that. And there you go, we have an image. Not bad   for the first shot, right? I could include way  more detail. But essentially, this is a way to   generate photorealistic images with ChatGPT. And  it costs nothing if you have a plus plan. It's   just an action in here. Not bad, honestly. One  side note is that these links might not work right   away, and sometimes takes up to a minute for the  API to run and the image to be stored inside of   your link. So if you get an error message like  this, just give it a bit of time. And the next

interesting GPT here is the Library of Babel. And  this one is brought up inside of our community as   an interesting way to discover new books. And this  doesn't just use GPT4. This, again, uses an action   to access an external database. So one thing it  does really well is recommend books. And a great   way to do that is based on other books that you've  read. Then it asks you about what you enjoy about   that book, and it goes into its database and gives  you further recommendations. Again, if you have   a plus plan, this is a great little addition that  might just improve your life a little bit. I mean,   getting high quality books is a big deal, kind of.   And this can do it very well. And the community   members tested it, and they reported that it's  super accurate with very esoteric and niche   topics that GPT4 wasn't even aware of. So yeah,  I can just say find me a book I might enjoy. And   then I'll just say I enjoyed Sapiens for its  history lessons. I thought that's one of the   best history books. I remember when I finished  it, I was like, wow, every human in the 21st   century should read this book. So let's see what  else it recommends based on that. As you can see,   it's using the action to access database in the  background. And there you go. We have custom   tailored recommendations that you wouldn't be  getting from vanilla GPT4. So in this manner,   GPT4 really is ahead. But while we are on the  topic, I want to give you a quick update on

the different ranking between chatbots. Because as  you know, there's a fantastic website that runs an   ELO system on different chatbots, and it updates  every two weeks. And we just got an update this   week. By the way, if you're new to the channel,  also, this is currently the best way to use GPT4,   Opus, all the paid models for free. This is  VC funded to create this ranking of chatbots.    If you want to just talk to one, you just go to  direct chat, and you can pick all the models in   here. But that's not the point. The point is that  leaderboard update that I wanted to share with   you. And as you can see, the updated leaderboard  ranks Cloud Free Opus as number one, same thing   we said on the channel when it came out. For  many use cases that are essential, this is an   excellent model. As mentioned back then, it lacks  a lot of feature, but the base model is really,   really strong. And this arena represents it. And  the way this works is essentially people rate the   different outputs. So they run a prompt in here,  they get two results from two different models,   and they rate which one's better. And then  based on human preference, which in my opinion,   for this channel is the most relevant metric, not  what the benchmarks say, but what humans actually   think and how humans actually use this, Cloud Free  Opus is currently king. But an interesting fact   that I want to point out is that Cloud Free  Haiku is actually at rank six. Look at that   GPT 3. 5 Turbo, which is the free version of chat  GPT's rank 14. And I can completely confirm this,   there is essentially no more reason to be using  GPT 3. 5 in chat GPT, because you could just be   using Cloud Free's free models, and they're  way more capable. I think that's not really a   controversial opinion these days. That's just kind  of a fact. The leaderboard packs it up and all the   people I talk to these days agree with that. But  then again, OpenAI is probably about to make their   next move. And then that will change again. So  to round this video out, I have two more things.

One is that Stability AI introduced a new code  instruct 3 billion parameter model. This is a tiny   model that outperforms a lot of 15 billion models.   So this is mostly interesting for builders,   a small model like this will run very well on the  phone or locally on your laptop, you do need a   Stability AI membership to use this commercially.   And beyond that, I want to close with a use case

that you cannot use today. But that's definitely  coming. And I didn't think that this was worth a   dedicated video. But I do want to mention it here.   That is Sora's first impressions where they put   OpenAI Sora into the hands of filmmakers, and they  came up with different films. As you might know,   I had a video production company before. So  this resonates with me deeply. And from all   these films, if there's one that you should  watch, it's the first one. I mean, it's the   only real story here. And it's really cool to see  how shy kids actually leaned into the strengths   of Sora creating surrealistic images, they didn't  try to replicate something like in real life, but   they created something only an AI model would do.   And they created this little story of Airhead 100%   of this is AI generated, and it's a legitimate  story. So just wanted to briefly mention that   if you missed this, go check out Airhead as this  is going to be a new category of filmmaking. Yes,   it will make indie filmmakers way more capable,  it will lower the cost of a lot of that. But it   also creates this new category, super exciting  stuff to me, I can't wait once this is out,   I'm going to be using this 24 seven, and we're  going to be covering a lot of this channel,   along with all the other use cases that will come  out over time. So that was this week's episode of   AI news you can use. I love this new name, hope  you do too. And if you want to learn about more AI   use cases, here's an entire playlist with all the  videos like this that I create on a weekly basis,   where I keep you updated on all the new apps that  you could be using for your workflows. Alright,   like, subscribe, comment and  all that and I'll see you soon.