Meta has revealed its latest addition to the world of large language models with the introduction of Llama 3. This model, which is available to cloud providers and soon to be on model libraries like Hugging Face, boasts superior performance compared to many existing AI models. Llama 3 comes in two sizes - one with 8 billion parameters and another with a massive 70 billion parameters. Parameters are the elements of the model that determine its complexity and its ability to process and understand training data. Despite being text-only at this stage, Meta claims significant improvements over its predecessor, including a wider variety of responses, fewer declined responses, and improved reasoning abilities. More concretely, Llama 3 has outperformed not just older models but its contemporaries too, such as Google's Gemma and Gemini, Mistral 7B, and Anthropic's Claude 3 in some benchmark tests. For example, in the MMLU benchmark which measures general knowledge, Llama 3's versions excelled. Nevertheless, it's worth noting that the validity of benchmark tests is not without concerns, as sometimes the data used in benchmarks can also be present in the model's training set, meaning the AI may already 'know' the answers. Despite that, human evaluators have also marked Llama 3 higher than others, including OpenAI's GPT-3.5, based on a distinct dataset designed to simulate real-world applications. Looking forward, Meta intends to increase the capabilities of Llama 3 by introducing even larger model sizes capable of more complex pattern recognition and multimodal responses. These future models, training at over 400 billion parameters, will hopefully understand lengthier data sequences and perform tasks like image generation and audio transcription. As they are currently being trained, initial tests display promising results on diverse questions, setting the stage for Llama 3's evolution into a multifaceted tool capable of a vast array of AI tasks.