Next Generation Uncensored AI is HERE & More AI Use Cases

Another week in Generative AI between the CES and various model releases. This has been a very colorful week. And actually this week I decided to include a few things that look into the future, namely into some things coming up in 2025, because we have some massive news out of Nvidia, a brand new uncensored model that you might want to consider, and a few small applications that you probably miss but might just want to be aware of because they do fun and useful things free of charge. Oh, all right, let's get into this episode of AI News. You can use the show that looks at all of this week's Generative AI releases and presents you with the info that you really need to know. Let's get into it. And I want to start this week by talking about Dolphin. So this is a fully uncensored version of the small Llama model, which means this is a LLM that you could run on most laptops and it can and will do anything for you. Now, it's not just based on llama 3.1, it's kind of pieced together from Meta's models and Quen models, and apparently they even use Deep SEQ to augment the data this was trained on. But the point here is this. The consensus seems to be that this is the new go to uncensored model that you want to use on your local machine. So what I did is I quickly downloaded this into my LM Studio, which is a free app. You can download some of these models and if you look for dolphin 3.0, llama 3.1, this one from cognitive Computations, then you can download the model. If you have a really small machine, you can get the smallest one, the Q2K. But the consensus seems to be that this Q6K gives optimal performance. So. So if you can run it on your computer, try this one out. And now if I go to a new chat, I can pick this Q6 model of the Dolphin 3.0, and I have a fully uncensored LLM running locally. Now, mind you, this is only an 8B model. This is not going to be as competent as something like a GPT4.0, not even close to something like a 01, but you could ask this anything. And I came prepared with one prompt that I would like to share with you today that I personally found really useful over the past few weeks. And it's something that surfaced in our community concretely. Stephen, from our community races. Thank you so much, Steven, for bringing up a lot of very interesting prompts. Thank you. I actually built a ChatGPT project out of this with more of an intricate prompt right here. But to demo this uncensored model, we'll go with something a little simpler. This prompt basically looks at an article or a piece of text and highlights all underlying assumptions. If you're looking to improve your thinking or you're not sure what to think about a certain topic, this is a fantastic place to start because it highlights all the assumptions that are being made in the text. And this can be a really useful way to use some of these uncensored models because now you can literally input any, any sort of text, any conversation you might have had, any thoughts you might have had, any article. No matter what the assumptions in it are, no matter what the content are, ChatGPT or Claude are often going to stop you because the content might not align with their content policies. So if I simply take this article right here, I can feed it to the model, hit enter, and get analysis of all the assumptions that are being made there. And even though ChatGPT can run some of these prompts, a lot of the times the answers do differ as an uncensored model like this just doesn't care. It will tell you about every assumption because there's no content moderation policy here. So if you want to improve your thinking, I would recommend assumption tracking prompt like this. And if you feel like the LLM is not being fully honest with you, you can Download this Dolphin 3.0 model for free and run the LLM locally like so. Again, just keep in mind this is an 8B model. The most capable Alma model has 400 billion parameters and something like GPT4 is probably around that size or above. All right, next up, as we're talking about models just very Quickly, Microsoft released 5, 4 and when we look at the benchmarks, this seems to be the new best perform performing tiny model. This one is 14 billion parameters right here you can see the numbers for GPT4O mini and while the size is the same, this Phi 4 just performs better on several of these benchmarks. And Microsoft put this out under MIT license, which means it's fully open source and free for people to use. So yeah, these models are still progressing on all fronts. It's always good to see the open source movement progress. On the close side, GROK seems to be preparing for something major. There's a lot of rumors on Twitter, so this is not something you can use. But the rumor mill has been buzzing with the announcement of Grok 3, which is supposed to come late January. We'll see we'll have a look at this as soon as it comes out. But for now they're shipping minor features and now they even have iOS app. So if you enjoy using Grok or want to gear up for Grok3 coming up, you can download the iOS app for free from the link in the description. And I personally am very excited for the future of this because they always had this very open ethos about them. I'm a big believer in freedom of speech and Grok seems to be aligning with that the most out of all the models. But they've just been behind on model quality, which is supposed to change, change soon. So we'll see as soon as more comes out and we'll judge it by the outcomes and results it's able to produce. So most of you know me as the generative AI YouTuber, but honestly, I'm more of an entrepreneur than anything else. I mean, even at the age of nine, I was independently going into the forest, collecting chestnuts and reselling them next to the main street. And that is a part of the reason why we do so much of the AI advantage between the YouTube channel, newsletter products and the community. There's a lot that happens in this entire ecosystem. And one thing that I learned over the years is that in a business like this, customer communication is essential. It's really all we have. And that's because this entire business is built around people. But as more and more people come in, this can get increasingly tedious and scaling support efforts can be a struggle. And that's why I'm happy to introduce you to RunBear, the sponsor of today's video, which solves this. So RunBear uses AI to optimize your customer communications, and it does it in the simplest way possible, which is perfect for a business like mine, where we're trying to juggle a million things at the same time. So concretely, it allows you to create really simple agents for different roles and department. And just because the creation is simple doesn't mean the functionality is. So you can have one agent for support, one for sales, and it all easily integrates into whatever software you might be using to communicate. And because each agent is unique, you can expand your team of experts that are available 247 with Run Bear. Let me quickly show you how simple this is. Okay, so I'm on a brand new account here that you can create by clicking the link in the description. And then here I'm just going to proceed to creating an assistant. I'm going to go with an OpenAI1 I'm going to call him Support Agent 1 and I'm not even going to bother with prompt engineering here as they have fantastic templates over here. I'm just going to go with a basic Q and a bot right here and GPT4O. And you could and probably should obviously customize this, especially this handbook parts. That gives it a little bit of context. All of this down here is just a placeholder for this demo. It will I will do though down here I will upload files and at this point in time, I know a lot of you might not even have some files to put into here, so let me show you how to do that too because I can just head on over to my website, take all of my terms of use. I'll throw these into a brand new Word document which I'll save as a.txt file. This is the format that usually works the best. And then it will simply upload this to the knowledge source of my RunBear agent. And here under actions, I can give it additional capabilities. Concretely, I'm going to give it access to the current date and the ability to fetch fetch a concrete URL. So if they reference a website, the agent can look at it as I ask it something. There's even more advanced capabilities, but I like the default settings here, so I will leave them as so and I can click Create. And with that being said, I have my default agent that is trained on the context of this fictional company. And all I need to do now is connect my Slack channel. And here I'll create a new token which I'll copy into this refresh token area. And as I allow this, that's it. I can access my custom agent that I could extend sent with extra knowledge at any time inside of my company Slack. And then over here in my Slack I could go ahead and mention AI support port at any times and I could ask anything about these terms of use. On first use, you need to invite the bot. And there you go. The knowledge from the custom knowledge base appears right here in the conversation. So as we're in the middle of a conversation, we can simply create these different agents that we can talk to without interruption. And it's really that simple. So if you're looking to upgrade or streamline your business, just check out RunBear. You can sign up for the link in the video's description. Thank you so much to RunBear for sponsoring this video. And now let's move on to the next piece of AI news that you can use. All right, Next up, we have a whole host of announcements coming out of Nvidia. They announced so much this week during CES that I don't think I will be able to give you a comprehensive summary of absolutely everything. There's a lot of other videos on YouTube that did that really well. Matter of fact, even if you're slightly interested this, I do recommend watching the entire keynote from Jensen Huang. They actually designed this in such a captivating manner that even though you might not understand a lot of the hard hardware or robotics talk, there to me has become the new Apple Keynotes that I've always been looking forward to, back when Apple was really pushing the envelope and innovating every single keynote. That's Nvidia now, and this was a lot of fun to watch. With that being said, let me highlight a few things that stood out to me and let's also talk about availability of these, because that's one of the core things we care about on this show. Starting out with Project digits, this one was to me personally the most interesting because this is essentially a small supercomputer that only focuses on AI, whether it's data science, workflow, fine tuning models, or just running inference. This little computer, or add on to your computer as you could use it comes at a $3,000 price tag and can run 200 billion parameter models. And you can daisy chain these. So if you buy two of these and put them on your desk, you could run a llama 405B locally, something that right now is pretty much impossible with any common hardware. You would really need to build a super rig that would probably cost 15000 or more, plus it wouldn't run as effectively. I mean there's this one hack where you can stack mult multiple Mac minis and daisy chain them together and run custom software. I suppose that could do it a little cheaper, but not as cheap as $6,000 to run something at the strength of GPT4O in your bedroom. That is quite insane to me and a clear leap in what's possible for the individual. Just to be clear, if you want to do that right now, it's absolutely doable, but you will have to pay for GPUs from some cloud service provider. And this will allow everything that we talk about here, whether it's running LLM like dolphin or llama 405B, or running text to IM videos, or even running image to video models locally, very affordable and very efficient. So this thing will be coming out in May and I think this is going to hit the sweet spot for a lot of individuals or small organizations that really want to lean into generative AI locally. Other than that, as I said, they had a lot of announcements, but I want to highlight two more things. One of them is they announced a bunch of new AI foundational models that will be running on Nvidia hardware. And along with them they talked about a lot of these Nvidia blueprints. Okay. And there will also be shipping more and more of these. One of them they introduced was the PDF to podcast, something that we'll talk about later in this episode that various open source developers are doing. But they're creating an entire ecosystem or operating system for all these applications where you have all in one place. So if you have Nvidia graphics cards, you can run these AI blueprints and just pick from their catalog. And sure, many of these are focused on an enterprise or industry like industrial robot digital twin simulations, but a lot of them will make things where you currently have to pay a subscription fee to some SaaS run locally. And this has not been as intuitive up until now. But with things like this hardware becoming more accessible and then these blueprints being really simple, this will be the year where a lot of these LLM powered workflows and then eventually agentic workflows will happen locally on your computer. They talked about a bunch of these, but I personally really enjoyed this one where you can build a video search and summarization agent and you can pretty much throw any video footage into this. You even get all the details on the architecture of this thing right here if you want to have a look at how it works under the hood. But essentially for me as a content creator, I could run video summarization and visual question answering locally from all the files that I have on the computer currently. There's a lot of steps in between that and I use external tools. I'm looking forward to bringing all this closer to me. And then one of the many other projects they also talked about is something they called Project R2.x and it's basically agentic framework that runs on Nvidia hardware, but that also supports supports these digital humans or AI avatars. So it's not just an LLM, but it's a full stack where you can actually work on your computer and you have this AI avatar sitting there assisting you and you can let it participate in meetings and more. So the conclusion here is really that a lot of this is finally coming together and will become available throughout the next one to three months. The new GPUs they introduced are available now. Also, their Nvidia Cosmos model, which is basically a video model that is used to train autonomous vehicles and robots. This thing is available right now, but all the people working on this will probably know that already. I just wanted to highlight it. Here's the entire GitHub with all the different models. You can try them out and see little examples of what it generates. And this helps robots learn about the world. So, yeah, a lot has been announced from Nvidia, most of it coming up, but I'm looking forward to taking a closer look at some of these as they become available. Generally speaking, in this show, we do only filter for things that are available today, but I thought it was important to give you a bit of a roadmap for what's coming this year so you have a bit more context on where the industry is going and how currently currently available apps match up against some of the stuff Nvidia has been working on here. All right, next up, it's a quick update out of hey Gen and they added something that they call hey Jin. Community avatars. And these are basically avatars that have been prompted into existence by the community. They look very real, though. And with the recent looks feature that we also covered, you can get all of these different camera angles from just one of these avatars. So if you want Oksana over here to teach your next yoga class, well, there's all these camera angles that could help you produce something proper. So the reason I wanted to share this is because obviously this is just one step in the direction that we're heading here. I think the next step here will be where people can actually submit their own likeliness. I mean, we're seeing this with other apps, with example, selling a clone of your voice. I just speak a word of warning at this point. It's so early in this space, and I would just be careful when approaching that. I love to be the early adopter on a lot of things, but selling your looks or your voice doesn't seem like a smart move to me, especially if it's just for pennies. Nevertheless, hey Jen has gotten so good over the course of last year that if you need to do a digital avatar production, I'd probably recommend them. And now they have even more avatars and even more looks available on their site. All right, next up, I have two super quick little tools that I just wanted to show you. These are basic LLM wrappers, and I think that will blow your mind, but nevertheless, two tools that you might want to bookmark as they are free and I found them to be useful in specific situation. This first one we actually featured in our newsletter yesterday. It's called Git Podcast and basically you can put in a GitHub repository and it will summarize it for you. And this is especially useful to all the non technical people because it will take all of the documentation and turn it into a listenable podcast for you where you have a better chance at understanding it. So you can just take one of these example repos here with Streamlit and at this point you don't even need to log in. You just create a five minute podcast that will help you understand the repository. And as on this show, here and there feature some GitHub repositories. Like last week we talked about this one. You could simply take the link, throw it in here and put on a podcast about that repo and learn about the new tools while on the go. Welcome to the Streamlit Innovators podcast. I'm your host, Ava. So much more digestible than technical documentation like this. And it's kind of funny to see the evolution of the fact that Notebook did really well last year and then all of these different apps are being built with similar functionality in different niches. But yeah, what can I say except that consuming something for this podcast format to me and many others seem to seems to be preferable to just reading walls of text. And then the second tool I wanted to show you is actually even simpler. It provides something like the Spotify Wrapped where you review the entire year and look at some trends for your Twitter handle or X account. I suppose. So if I just go ahead and paste my X account on here, then it will run analysis on the account and then it shows me most popular tweets, some milestones. These are relatively accurate considering the fact that I don't use Twitter Twitter that much and it will even roast you. Look at this one. You call yourself an AI educator, but really you're just a ChatGPT abuser. We all know it. I would say that's pretty fair. And also this one. You're like that friend who always talks about the newest tech. We get it, you're cool. Can we talk about something else now that is so accurate. For anybody visiting family over dinner and opening up the topic of hey, this brand new thing released, you guys should check it out. Just to be met with blank faces that can't wait until you wrap up your little rant and they can go back to talking about the newest show on Netflix or whatever. Hey, on this channel, you're in the right place. This is where people of interest similar to ours gather and we can all nerd out together. Anyway, if you want to run this on your Twitter account, this also is completely free and you can try out the tool now links below. On to the next one. All right, and to round out this episode, I just want to highlight Sam Altman's blog post, which has been looking back at the past two years at OpenAI. There's one sentence here that you should really know about if you haven't heard of this this yet, and that's him basically saying that we, OpenAI are now confident we know how to build AGI as we have traditionally understood it. This is the very first time he openly says that. He follows up by highlighting that OpenAI believes that in 2025 we may see the first AI agents join the workforce and materially change the output of companies. If you're following the channel, this should be no surprise, because already AI agents are exist in so many shapes and forms. But what he's saying here is that in 2025, you can expect the first ones to be so capable and so deep that they will really start moving the needle at scale in a way that we haven't seen before. That's at least what I'm reading into this now. I'm super glad to be on this journey towards discovering all this technology together with you. My name is Igor and I'm here every Friday discussing all the latest AI releases that you can put to work. I like to throw in a look into the future here and there. And with that being said, I hope you enjoyed today's episode and I hope you have a wonderful day. See you soon.