This AI Has 10x ChatGPT's Memory, Here's How to Use it For Free

Welcome to another week with various AI releases. We're going to be looking at new O1 alternatives and new insights on how to get more out of prompting them. A video generator that is set to change the visual effects industry and a clothing try on tool that actually works. All of that and much more. In this week's episode of AI News, you can use the show that looks at all the AI releases that came out over the course of the last week that you can actually put to work today. Let's get into it and I want to start this conversation by talking about ChatGPT tasks. Now we did create a separate video on this showing you the basic prompting formula for this and some use cases. But I do want to add a few comments here and that is that one, the desktop app on Mac actually makes fantastic use of these. So if you're using this, make sure to check for updates and then you can set tasks in here. And I also kind of want to talk about the criticism that this feature has been facing. The how to and the tips and tricks you can get in a separate video on the channel. But the criticism has been that what are they even doing here? This is sort of a useless feature. Isn't this just a glorified notification sending application? And I think largely that criticism is actually relevant. This just feels like a feature that has been shipped before, something else that will be really utilizing this. Nevertheless, I think there's some fun use cases there, so check out the separate video for more details. I think the best use cases really are the ones that relate to your personal context, inside of memories or custom instructions that that really makes sense. You can make it compliment you every morning if you're into that, or you can make it give you a motivational speech every morning that is custom to you. So that does go beyond the notification. But I think ultimately they ship this feature because they're going to be coming out with their Operator product hopefully sometime soon. There's no dates or anything, just several rumors with me giving this one a lot of weight. They actually added a host name which is operator.chatgpt.com this will be their agentic product that will actually do things for you. And in that case, tasks are going to make a lot of sense. And, and if you're familiar with how they work already, you'll have an easier time getting into the various functionalities that this will probably have. Okay, now let's talk reasoning models, because this seems to be the main theme of model releases in 2025. In 2024, everybody was catching up to GPT4 and we were getting these GPT4 level models out of Mistral, Llama, Anthropic, really all the competitors and now it seems like everybody's catching up to their reasoning models, concretely catching up to 01. Nothing is even close to what they showed with O3 and this is one of them. So look at this. Sky T1 was actually trained with under 400 $250, which is remarkable. And while I don't think you'll be using this particular one, it is available in retraining models to do reasoning from what we see here is way simpler than training a model from the ground up. They also open source all the code that goes into it. But in practice if you want a free alternative, the Chinese Deep Seq will be probably your go to if you're on a ChatGPT paid plan, you're just going to be using a one and I want to take this opportunity to talk about the prompting of it. We're not going to spend too much time on it, but I do think that this is important because as people gain more experience with this, more and more of these frameworks have been appearing and I particularly like this one from Dan Mack. Even Greg Brockman retweeted this if you followed my content on O1 before, you will probably have heard me talking about goal based prompting instead of instruction space prompting. Instructions based prompting is where you tell it hey, I need you to do task xyz. Here's a bunch of context. So it's very concrete and goal based prompting is just telling it where you want to go and it has to figure it out. Okay, but that is not enough. And as I also said previously, super short prompts don't work super well with O or other reasoning models. That's why I think this framework here from danmac is fantastic. Yes, you specify your goal, but then you also tell it what kind of format you expect it to return and what to avoid and then you add all the extra context. It's quite simple structure but it can help you thinking about these things. And in case you're not aware, we're actually running a public challenge right now in the public area for community, where in my latest YouTube livestream I asked everybody on this channel to share their favorite 01 use cases. It's completely free, it's in the public area. You can check out the submissions there. There's only a few so far because this is not an easy challenge. But at the beginning of February we'll do another stream and we'll be evaluating all the best submissions on how people are using these models. So here's a little crutch on how to prompt them, how to get better results out of them and to round the segment out. I just want to point out one case that I'll be submitting in a post there, and that is that I took my entire digital to do list from the past six months because I manage my tasks inside of notion and when I finish them I drag them into section where all the finished tasks live. So took all of those, then I wrote a medium length prompt that told it to basically create a categorization system for all of those tasks. And then for me as a business owner, I wanted it to create a delegation plan for all of those tasks and what kind of roles I would need in my company for me to delegate everything that I have done throughout the last six months. Then I ran that for 01 Pro and I was actually surprised by its organizational capabilities and it even created a roadmap on how I should start proceeding with this rather large task. It was really impressive and I'll write up all the details in them into the challenge submission, but for now you can check out some of the other team members or community members submissions there. And I hope that this format will help you get better results out of 01 because everybody that has been using this regularly, both 01 and 01 Pro just shares the opinion that we wouldn't want to go back to a world with just standard models like GPT4O. These do perform better, it's just hard to quantify exactly what better is. But if you're a power user and you haven't given 01 a shot yet, try a framework like this and Compare it to 4.0. If you haven't done so yet, I think you will be pleasantly surprised that a lot of it doesn't feel like the typical AI output. Okay, onto another Release which is Minimax01 actually now has LLMs. And yeah, this is Minimax, one of the top tier AI video generators. That's at least where you might know them from. But now they're also entering the arena of LLMs and they actually released something really capable here that you can use for free on their website Hylo AI now. Now if you look at the benchmarks this ranks slightly above GPT4 all, but essentially with all these LLMs that are GPT4 level practically doesn't make much of a difference for the day to day usage of most people, except for the tonality or specific domains like coding, where certain models are just better. But the big difference with this model is that it now has actually the world's longest context retention with 4 million tokens. If you're using ChatGPT, it's 128,000 tokens. Currently. I think the largest one was Google's gemini Pro with 2 million tokens if you use it through AI Studio. But yeah, most people are used to around 100,000 tokens of memory. That's essentially how much info you can put into it before it gets amnesia and just forgets what happened at the top of your conversation. This model has 4 million tokens. And not just that, it's also exceptional at maintaining all of the context there. Especially some of these models that stretch the context window just start losing some of the info. By the way, all that information comes from their paper that they released. Together with the model. I found one thing particularly interesting and that is a perfect score on the needle in a haystack benchmark, which essentially hides a little piece of information inside of a long document and then tests how well the LLM retrieves this. Usually other LLMs are not perfect at this. Like they're usually very good, but little pieces of context do get lost from attached documents. This one is flawless. And as mentioned also some of the other benchmarks come in above some of the competitions. By the way, this essay on penguins, which is 1,000 tokens, translates to about 5,000 characters, which is 780 words. So this new Hylou LLM can take around 3.1 million words without losing any of them. I'll be testing this more in depth over the coming days and weeks, but that's pretty impressive and you can give it a shot today under Hyliu AI links in the description. Now let's move on to the next one. All right, the next one I have here didn't exactly come out this week, but I really wanted to feature this because I think this is by far the best clothes try on tool with AI that I've seen yet. It's called Leffa and there's a working demo on Hugging Face. And the amazing thing is it doesn't just work with all the presets right here. So I could combine this model with this T shirt and generate. I could also go on Google and I don't know, find some random product like this spongebob hoodie. See, this result was amazing. And now I can switch it out for the spongebob hoodie, because why not? And we just generate this and you will see that this actually performs very consistently. We looked at tools like this before. They weren't anywhere close to this. Look at that. That's legit. It's tucked in and everything. Now the same thing goes for the model. And this opens up a lot of fun possibilities because you could upload your own images to this. In this case, I'm just going to go with Keanu Reeves because I don't know, that's just top of mind for me. And we're going to try it out. We need some sort of, like, headshot like this. This should work well. Change it out. There you go. Let's see how Keanu Reeves looks in a spongebob hoodie. And while it generates, I just want to point out something that you could do. And this came up during our research meeting. Shout out Daniel for this idea. And what he did is he took a picture of his and he put him in the sports jersey of a team he hates and then he gave him that picture. It's a really easy way to piss off one of your friends if you have that type of relationship with somebody in a playful way and only do that if you have that type of relationship with somebody, obviously. But yeah, look at this. This works really well. Keanu Reeves in a spongebob hoodie. And it's free to use. You could use your own phone pictures. So if you're going to do this, leave a comment below who you're going to send, what type of picture, I really want to know because I feel like you could have a lot of fun with this one. And with that, let's move on to the next piece of AI news that you can use, which is something that currently is an early access so you can sign up to the waitlist. I do like highlighting those because this one seems amazing. It's called Daily Listen by Google. And this is sort of a follow up product to NotebookLM. That's where you can think about this. And it basically summarizes all the news that might be relevant to you into a personalized podcast that will get delivered to you daily. I think this is just one of these ideas that just intuitively makes sense. I think a lot of us can get behind this idea of having an AI assistant that just autonomously pulls together information that is relevant for you and provides presents it to you in the formats that you prefer. And if you're like me and you enjoy audiobooks and podcasts and YouTube videos, etcetera, then getting a podcast format is one of the best things that can happen to you now, this is the first version of this, but I can see over time something like ChatGPT tasks being integrated here, where it just automatically generates your Good morning news podcast for you. And that's the form that our AI assistant will take in the short term. So if you're watching this channel, chances are you want to arrive in this future before the bulk of society does. And, and in this case, you can sign up to the waitlist on Google Search Labs. Keep in mind, this only works in the US so you need a VPN to sign up to the waitlist, but it's free. And this is something that I personally am looking forward to getting early access to. And we'll cover it in more detail once I do. Okay, next up we have something that is called Kokoro, at least I think that's how it's pronounced. Kokoro. It's a text to speech software that is fully open source, has a manageable size, so you can run this relatively easily on some machines. And the quality level of it is incredible. I would say it's almost 11 laps level with the difference that this is open source and therefore free to use. And one more thing, the latency on this is unbelievable. But let me just show you, let me put on the headphones here so I can hear this. And we're just gonna go with the preset text here. You can put anything you want into here. But the point is that if you generate the speech, look at the length of this text. This is quite lengthy, right? Literally without even making an edit in the video, this thing generates in what, estimated 12 seconds? It's kind of like halfway done already. There was a task before this and we just get this super long text in essentially no time. And this runs online. Okay, so you're sharing the GPUs with others. So in some cases when the traffic on this is lower, it's even faster. But there you go.

The studio was filled with the rich odor of roses. And when the light summer wind stirred amidst the trees of the garden, there came through the open door the heavy scent of the lilac.

So I would say that's really good quality. Maybe not perfect, but really good. And the thing is, if you run this on a proper graphics card and not on like a shared hugging face space that they provide for free to test it out. You can get something like this where it generates two and a half minutes of speech in about four seconds. A text to speech model named Kokoro. This quality at that speed on an open source project is unprecedented. There's just no other way to put this. And what this opens up is for developers who have applications, any sort of software, literally. I'm not just talking AI stuff. You could build this into your hardware, somebody's making cameras, you can build this as a feature into it because it's fully open source and it sounds really good. So yeah, this is literally what the future looks like. Projects like this being open sourced and then the horde of developers just putting it into all sorts of applications. Pretty great. And if you were looking for a free alternative to 11 labs, well, this demo just works and you get several voices that you get to pick from right here. All right, next up, I quickly want to talk about a video generation model. And this one is very different from the usual ones we look at because it generates videos with translucent backgrounds, meaning you can use it for visual effects compositing. A topic that I happen to know a little bit about from my video editing background. I actually went on like a two month arc where I started learning After Effects, which could essentially be summarized as Photoshop for video. And in there, one of the big things you learn right away is how transparency works and is rendered inside of a software like that. There's a lot of situations where you need the part to be transparent and a lot of tricks inside of the software. But let me tell you that still to this date, separating a foreground object from the background in video is just extremely hard task. A lot of times you have to do it manually. That would be called rotoscoping, where you literally go through frame by frame and move a mask around so it perfectly fits the subject. Now this is where we enter something that Adobe Research calls Transpixar. This generates footage with the background already cut out. And I may be over explaining it a little bit because the point is this is a huge deal for visual effects. A lot of these assets often have to be purchased or 3D modeled from scratch and inserted into here. If you can just whip up exactly the right type of fire or explosions, particles, portals, whatever it might be, just from a simple text prompt. The amount of work this saves in these post production workflows is ridiculous. And they're basically giving us a first look at it. But this is the first really good video model with transparent backgrounds that I've seen yet expect it to to be added to things like comfyui soon so they can build entire pipelines to generate visual effects assets with a lot of control. Now I do have to add that this is the first one and doesn't work particularly well. We ran a few prompts through it here in this hugging face space here we try to generate matrix letters running across the screen. Sort of works I guess, if you colorize it. We also tried balloons with confetti behind them. By the way, in software like After Effects, there's a setting where you remove the black background, AKA the alpha channel really easily. So black equals transparent. That's why it looks this way. I tried a few more tricky ones like burritos raining from the sky, that didn't work at all. What is even this? So it's really hit or miss. But once this technology progresses beyond the stage, trust me, this one is going to be a big deal for video and film production. Final note, if you want to learn about how it works, Adobe Research was kind enough to include their research method. So look at this beautiful flowchart. I don't think I need to explain a lot here, honestly. This is just very intuitive stuff. Pretty self explanatory if you ask me. Me. Okay. And this week I got one more AI video innovation. This one is actually super fun. Probably my favorite AI video release in a minute. And it's a feature from haiuai's Video Generator that allows you to reference subjects. That means you can create AI videos of yourself. Finally. Now obviously we took this opportunity and we started putting me in different scenarios. Look at this one. Mad Max Igor. You can put yourself into movie scenes, you can put yourself into characters. What about this one? John Wick Igor. Not bad. I'll let the video editing team have some fun. Show us some examples with some music behind it. And while you check this out and think about what type of situations you could put yourself into with the new subject reference feature. Pretty cool. Creating custom videos of yourself with AI has never been easier than this. Except maybe something like Viggle. But here you get way more control and all you have to do is upload an image of yourself and you can start generating. Two tries are included with a free account. Alright, that's all I have for today. I hope you find something in here that was useful to you. And with that being said, I'll see you next Friday in the next episode of AI News that you can use.