New Llama 3.3 Shocks the AI World - Crushes GPT-4 and Costs Almost Nothing

So Meta just unleashed llama 3.3 and get this, it uses only about one seventh the parameters of their previous 405 billion parameter colossus, just 70 billion, but it still hits almost the same performance. With that kind of efficiency, we're talking lower costs, smaller GPU demands, and the power to turbocharge everything from everyday AI tools to immersive VR worlds. So let's talk about it. So first off, Llama 3.3 is Meta's new multilingual large language model. And the headline here is that it has 70 billion parameters. That might sound huge, and it is, but what's really wild is that it's performing close to Meta's previous giant Llama 3.1 model that had a staggering 405 billion parameters. In other words, Llama 3.3 gets you near top tier performance while being way more efficient. It's like having a sports car that's nearly as fast as the ultra expensive model, but consumes way less fuel and space. Mark Zuckerberg himself announced that Llama is now the most adopted AI model in the world. Meta claims it's had over 650 million downloads. That's massive uptake, especially considering how many developers are now building on these open source AI protocols. Open source is key here. Meta wants to be the backbone for tons of AI projects. Although open source might sound like they're just giving it all away, Remember that if everyone is building on Meta's foundation, Meta becomes sort of the underlying infrastructure that could mean a whole lot of market influence over time. Now, alongside this AI push, they're also expanding in VR. They're working on making their VR tools the industry standard. Think about it, if they control the leading AI tools and the leading VR tools, they're kind of setting the stage for what the future of digital connectivity looks like. Maybe even the Metaverse they keep talking about. They're teaming up with third parties to do this, because the more people rely on Meta's tools, whether that's for AI or VR, the more Meta becomes embedded in the future of digital interaction. On the practical side, Zuckerberg also mentioned plans for a new AI data center in Louisiana and a new undersea cabling project to support all this. Meta AI apparently has around 600 million monthly active users. But let's be real, Meta has over 3 billion users across its apps. Facebook, Instagram, Messenger, WhatsApp, and they've basically stuck their AI assistant right into all of them. They're also nudging people to try AI image generation features and other AI powered experiences. Sure, that might drive up the usage stats, but how meaningful is that usage if people aren't exactly flocking to it voluntarily? This is a fair point, but maybe AI assistants inside social apps don't feel supernatural. Yes, you can generate neat images or ask questions, but for many everyday users, what's the compelling reason? Still, it's early days and maybe the real value is going to be down the road, especially as VR becomes more mainstream. All these puzzle pieces, AI, VR, wearables, the undersea cables are building towards something bigger. Okay, let's move into more of the technical weeds. Llama 3.3 is open source and multilingual, supporting English, German, French, Italian, Portuguese, Hindi, Spanish and Thai. It's been trained on a whopping 15 trillion tokens compared to llama2's 2 trillion token training set. That's a huge jump in training data, leading to better performance on reasoning tasks, coding, benchmarks, stem problem solving, and even trivia. Another neat trick is that this new model supports super long context Windows 128,000 tokens. That's about the length of a decent sized book, meaning the model can handle really long documents and keep track of what's going on over many pages. Meta's also done a lot under the hood to make sure llama 3.3 can run efficiently. For one, the new model uses something called grouped query attention, which helps with memory efficiency and speeds things up during inference. That's the process of generating answers or predictions when you're actually using the model rather than training it. GQA improves scalability and makes it cheaper to run. According to meta, llama 3.3 can be super cost offensive for developers, generating text at as little as $0.01 per million tokens, way cheaper than the likes of GPT4 or Claude 3.5 in many scenarios. There are some incredible hardware savings too. The older 405B parameter model required a massive amount of GPU memory, up to nearly 2 terabytes of GPU memory for inference. In contrast, the 7db llama 3.3 might only need as little as tens of gigabytes. We're talking about potential GPU load reductions that could save developers hundreds of thousands of dollars in upfront GPU costs, plus huge ongoing power savings. Imagine trimming down from something like $600,000 in GPU costs to something more manageable because of the reduced memory footprint for developers and researchers. Llama 3.3 also comes under a specific community license agreement. It's mostly free and open source, but there's a catch if your organization has over 700 million monthly active users. You need a commercial license directly from Meta. Also, if you use it, you must credit Meta with something like Built with Llama and follow their acceptable use policy. This policy aims to prevent harmful content generation, cyberattacks or any activity that breaks the law. So it's open but with certain guardrails. Meta's not just throwing this model out there without any safety measures. They've really leaned into safety and trust. The model uses things like supervised fine tuning and reinforcement learning with human feedback to ensure it's aligned with helpfulness and safety standards. They've built a bunch of safeguards like Llama Guard 3 and Prompt Guard to keep the model from doing harmful things or spitting out unsafe content. They've also done extensive red teaming where security experts try to trick the model into doing bad stuff to find weaknesses and fix them. They've looked at risks from child safety issues to cyber attack enablement and tried to mitigate them. They've even factored in environmental considerations training llama 3.3 took a lot of GPU hours, about 39.3 million GPU hours on H180GB hardware. This generated around 11,390 tons of CO2 equivalent emissions, but Meta claims they used renewable energy to achieve net zero emissions for the training phase. They've also publicly shared how much energy it took, so others can understand the environmental cost of these massive models. Performance wise, Llama 3.3 kicks butt on a range of benchmarks, often beating similarly sized models. On MMLU, which tests knowledge across various subjects, it hits around 86% accuracy on math benchmarks like math, it scores around 77% for coding tasks such as Human Eval. It gets a pass at one of 88.4%, which is pretty darn good. It even handles multilingual reasoning tasks well, scoring about 91.1% on MGSM. While it might not beat the biggest models on every single test like GPT4 encoding tasks, it gets impressively close in terms of ecosystem. You can grab llama 3.3 from Meta's site hugging face, GitHub or other platforms. Developers can integrate it with various tools like LangChain or weights and biases, and it's cloud friendly on aws, GCP or Azure. You can also fine tune it for your own applications with Meta's torchtune library. It's flexible enough for a wide range of scenarios, everything from natural language understanding to coding assistants to maybe even VR experiences down the road. Safety remains a big talking point. Meta wants developers to take responsibility for how they use the model. For example, if you integrate llama 3.3 with a tool that can access external services, it's your job to make sure nothing malicious happens. The model is like a building block. You can create something helpful and safe, or you can create something harmful. Meta provides guidelines, best practices and extra tools like Prompt Guard and Code Shield to help you avoid dangerous outputs or insecure code. Meta's ultimately betting that open source, powerful and efficient AI models will form the infrastructure of tomorrow's tech stack. Combine that with their push in VR, like their risk based electromyography devices that measure muscle signals to create control virtual objects and you can see where they're headed. They're not just making an AI model, they're building out the foundation for the next stage of computing where VR and AI blend seamlessly. Will it become the standard for AI assistance or VR interactions? Meta sure hopes so. Right now it's definitely cheaper and more efficient than some of the top tier closed source models out there. Let's keep an eye on how developers actually use it, what kind of apps come out, and how users respond. All right, let me know what you think in the comments and if you enjoyed this. Make sure to like and subscribe for more AI updates. Thanks for watching and see you in the next one.