Meta: Llama 3's artificial intelligence is even better than Google Jamnai

Meta claims that Llama 3's AI is better than its competitors, including Gemnai and Gemma.

The Llama 3 large language model, soon to be released by cloud service providers such as AWS, outperforms most current artificial intelligence models, Meta said in a press release.Llama 3 currently has two versions with eight billion parameters and 70 billion parameters, which currently only provide text-based answers. Llama 3 shows more variation in response to requests than competing large language models.Llama 3 has fewer false refusals when it refuses to answer questions and is able to reason better. Llama 3 understands more instructions and writes better code than ever before, Meta says.

Meta claims that the first two versions of the Llama 3 outperform comparable models such as the Gemma and Gemnai, Mistral and Claude 3 in certain benchmarks. On the MMLU benchmark, which typically measures general knowledge, the eight-billion-parameter version of Llama 3 outperformed the seven-billion-parameter version of Gemma and Mistral.

The 70 billion parameter version of Llama 3 was in close competition with Jamna Pro 1.5.

Interestingly, the GPT-4 large language model does not appear in meta comparisons. It should be noted that the benchmarking of AI models, although useful for understanding how powerful they are, is incomplete. The datasets used to evaluate the models are part of their training; This means that a big language model already knows the answers to the questions.

Human evaluators rated the Llama 3 higher than other models, including the GPT-3.5, Meta says. The company created new datasets for human evaluators to simulate real-world scenarios in which Llama 3 might be used.

The aforementioned datasets include things like consulting, summarizing, and creative writing. Meta says the team working on the model did not have access to the new evaluation data, so the model's performance was not affected.Llama 3 will also have a 400 billion parameter version that can handle longer strings of instructions and data, and a multi-mode model that can be used to generate an image or transcribe an audio file.The 400 billion-parameter version of Llama 3 can learn more complex patterns than the smaller versions and is currently being trained.