This New UNRESTRICTED AI BEATS Both GPT-4o and Gemini Pro

All right, let's talk about something pretty exciting happening in the AI world, particularly for those interested in having more control and flexibility with AI. Instead of going with the big players like OpenAI, Google or Meta, NA Research has their own thing going, a set of unrestricted AI models. Think of these as a fresh new spin on AI tools with the power to compete with GPT4O or Gemini, but without the usual limits. And they've recently dropped two very interesting new products, Noosechat and Forge Reasoning API Beta. So let's unpack what that's all about. So to kick things off, Naus Research started with their Hermes models. And for a while they kept these pretty technical, meaning you had to download and run them on your own machine or access them through partner sites. But that's changed with the release of Noosechat, their first chatbot interface. And this is where things get interesting. Mouse Chat is pretty much like the ChatGPT or Claude interfaces we're all familiar with. A big text box at the bottom to type in your prompts, and a spacious area at the top where the AI responds. But here's the twist. Mouse Chat uses their Hermes 370B model, a fine tuned variant of Meta's Llama 3.1 model. This isn't your average chatbot. It's a powerful open source model you can access without needing to set it up yourself, which is super convenient. Now with Noose Chat, you're getting this AI model in a user friendly way without deal with the hassle of running code. The team designed News Chat to feel accessible and cool using this retro style look with vintage fonts that resemble early computer terminals. So if you're into that aesthetic, it's a nice touch. They've even added light and dark modes to switch between, so that's a bonus. But let's get a bit technical for a sec, because that's where this chatbot really shines. One of the key features NUS Research has baked in is this array of pre written prompts. If you're into different tasks like knowledge and analysis, creative writing, problem solving, or research and synthesis, you can click on these categories and the model will automatically load a related prompt to kick things off. For instance, if you choose knowledge and analysis and ask for a summary on something like intermittent fasting, the model pulls up a nicely structured answer, adding some serious value. One of the questions some people have is whether Noschat is truly unrestricted. Now, Naus Research promised to deliver something that gives users full control over what they can ask the model but they still have some guardrails in place. For example, if you try to ask about illegal narcotics, the model won't respond in full detail. The team explained this as a common sense filter, not a strict block, meaning it's more of a soft warning. Still, some AI jailbreakers, people who specialize in testing limits, have found ways to push these boundaries, like using creative prompts to get past the filters. Now, what sets News Chat apart, even though it might not be as advanced as some corporate tools, is its speed. Responses pop up within seconds, making the interaction feel snappy. Yet it does have its limits. For one, it lacks access to the latest web data since its knowledge cutoff is April 2023. So while it can't give you up to the minute information like OpenAI's latest models with Internet access, it does well within its limits and even provides web links for further reading. Though some of these can be hallucinated, meaning they're not real, you might get a link that looks legitimate but leads nowhere. If we zoom out for a second, Noose Chat is part of a broader push by Naus Research to build tools that aren't tied down by the typical corporate restrictions. They want AI to be more of a toolkit that researchers and developers can use as they like without the rigid constraints we see elsewhere. That's where their other new tool, the Forge Reasoning API Beta, comes into play. This API takes reasoning up a notch by adding a code interpreter and other advanced reasoning capabilities to AI models. What makes Forge unique is that it boosts the Hermes model so much that it's competing with some of the biggest names OpenAI, Anthropic, and Google in reasoning tasks. One of the cool things they did to showcase Forge's abilities was to test it with math problems with from the AME competition if you're not familiar, AIME is a math competition that's one of the qualifiers for the US Math Olympiad, so it's intense. Forge powered up the Hermes 70B model to the point where it performed on par with models that are way bigger in size. That's a big deal because it shows that Hermes, with the help of Forge, can handle reasoning tasks without needing to be as massive or resource hungry as some of the industry's larger models. Now, Forge doesn't just work with Hermes, it's designed to be flexible and supports multiple models, including Claudsonnet 3.5, Gemini, and GPT4. You can either use a single model to handle tasks or combine several models, allowing for something called a mixture of agents approach. With this, different models Work together almost like a team of specialists, each bringing their own strengths to answer a question or solve a problem. This collaboration helps provide more well rounded answers, which is huge for users who need reliable, nuanced information. So how does Forge pull this off? It's built on three main architectures. MCTS, Monte Carlo Tree Search, CoC, Chain of Code, and MOA. Mixture of agents let's break these down quickly. MCTS is great for making decisions in complex scenarios, like planning moves in a chess game. It explores possible outcomes, then picks the most promising path. CoC, or chain of Code, connects logical reasoning with a code interpreter, allowing the API to handle math heavy or code based tasks much better. And finally, MOA allows multiple models to work together to analyze and debate a question, then arrive at a consensus with Forge. Noose is setting the stage for some powerful tools. This isn't just AI giving you one response. It's about synthesizing input from multiple perspectives and adapting to the task at hand. Right now Forge is in a closed beta, which means only select users are testing it. They're gathering feedback from these power users to fine tune the system. If everything goes as planned, we could see Forge opening up to a broader audience soon. But Forge isn't just for developers who need raw computational power. It's actually adding a layer of intelligence that brings these models closer to what Noos envisions as a telecommunications brain. This concept is still a bit futuristic, but basically the goal is to create a system that can understand the structure of cellular networks, predict changes and adapt without constant human oversight. Large language models or LLMs like Hermes are starting to fill in these gaps, but we're not fully there yet. This vision of a telecommunications brain is part of a longer journey. Right now, most AI and telecoms helps out as an add on or support. Think chatbots that understand telecom specific language or tools that help set network configurations. Full autonomy is still a stretch because of the limitations in LLMs today, but with tools like Forge, we're seeing steps towards more autonomous AI that can make complex decisions and adapt to different situations. Naos Research has designed Hermes as a multi agent system that tackles network modeling with blueprints which are like detailed guides for setting up network digital twins ndts. These blueprints walk the AI through the steps needed to model things like network performance, signal quality or energy use. Hermes uses a modular approach with a designer agent that drafts blueprints and a coder that translates those into executable Python code. This structured process helps the AI avoid common pitfalls like errors in calculations or logical missteps that you'd expect if a model was winging it alone. And if you're wondering how all this holds up in real world scenarios, NA Research has done some testing. They ran Hermes through tasks like power control, energy saving measures and deploying new base stations and the success rates were impressive. For instance, Hermes hit 85% accuracy in power control tasks, way ahead of other methods. As the complexity of the tasks increased, the Hermes pipeline still maintains solid performance, which shows how the multi agent blueprint based approach really pays off. Lastly, if you're someone who's curious about open source alternatives, KNA has been testing Hermes with models like Llama 3.1. They found that with smaller open source models the success rate is lower than with big proprietary models. But there's still hope. When they gave these open source models access to a library of expert designed modules, the performance improved significantly. This could make open source options more viable for specialized tasks, especially if they're set up with ready made components. If you're interested in checking out News Chat, they've made it free for now on their site and Forge is available in beta for a select few. And that's the scoop on Noose Research's latest AI tools. Would you give them a try or see them fitting into your workflow? Drop your thoughts below and if you enjoyed this, hit like and subscribe for more AI updates. Thanks for watching and see you in the next one.