Curated on
April 25, 2024
Meta has announced the release of its latest large language model, Llama 3, which boasts superior performance over many of its contemporaries. The new generation AI, which will soon be available on cloud services like AWS and in model libraries such as Hugging Face, introduces innovative capabilities with two variants: an 8 billion parameter and a 70 billion parameter model. The 'billions of parameters' descriptor signifies the complexity and understanding capacity of the AI model. Unlike its predecessors, Llama 3 offers more diverse responses, shows improved reasoning, has fewer instances of declining to respond to queries, and writes better code.
Head-to-head against similar-sized models, including Google’s Gemma and Gemini, Mistral 7B, and Anthropic’s Claude 3, Llama 3 shines in benchmark tests. In particular, Llama 3's 8 billion parameter version outclassed the competition in the MMLU benchmark, famous for assessing general knowledge. The 70 billion parameter iteration was also noted to have a slight edge over Gemini Pro 1.5. However, it's important to note the imperfections of benchmark testing, as the datasets used can overlap with a model's training data, skewing results. Despite the potential flaws in benchmarking, Llama 3 was favored by human evaluators who tested it in scenarios mimicking real-world applications.
The real-world applications ranged across 12 key areas, including creative writing, advice provision, and coding, which gave Llama 3 a more authentic environment for assessment. Moreover, Meta is planning for Llama 3's future iterations to handle larger data inputs and to offer multimodal responses, like image generation or audio transcription, underlining the model's scalability and adaptability. The forthcoming versions, boasting over 400 billion parameters, suggest Meta's ambition to push AI capabilities even further, potentially redefining the limits of what language models can achieve.