Episode 8: This week's AI news is insane...

Meta released Llama 4. What is different compared to their previous models? From Meta’s website:Meta's Llama 4 models are cutting-edge multimodal AI systems designed to handle both text and image inputs seamlessly. They employ a mixture-of-experts (MoE) architecture, which activates only a subset of parameters for each input, enhancing efficiency and scalability. Basically, what this means is that only a smaller set of the model’s parameters are activated at a single request. The activated parameters represent one or more experts. These experts to be activated are chosen from an internal router. This is something really interesting, especially to that extent. You can request access to the models from https://www.llama.com/llama-downloads/. Leave a link in the description. Also, something that we have not seen before, Llama 4 Behemoth pushes the parameters to 2 trillion. This is an outstanding number for the industry. Source: https://ai.meta.com/blog/llama-4-multimodal-intelligence/ Let’s discuss the key differences between the models. Key Features of Llama 4 Models: