In the “gold rush” relating toArtificial intelligencethe latest update of Meta could mark an important step in its approach to GPT-4.
We are talking about Code Llama 70Bdefined bluntly as “The largest and best performing model” in the context of AI. According to what is claimed on the official Meta blog, this tool is capable of handling more queries than previous versions. This translates into the possibility of giving more instructions during programming, therefore more accurate results.
Tests in hand, the platform’s claims are anything but far-fetched. Taking into consideration the benchmark HumanEvalIn fact, Code Llama 70B obtains an accuracy result of 53%, exceeding GPT-3.5 (48,1%) and approaching GPT-4 (67%).
Code Llama 70B surpasses GPT-3.5 and begins to put some pressure on OpenAI
On the other hand we are talking about the evolution of Llama 2, capable of creating code strings from prompts and debugging code created by developers. Meta stated that models bigger than him “They return the best results and allow for better coding assistance“.
Code Llama 70B is still available, at least at the moment, free of charge for research and commercial use. The large model was trained with 1TB of code and is hosted on the repository Hugging Facewhich offers access to GPU to run Artificial Intelligence models.
If Meta is desperately trying to get closer to what is proposed by OpenAI, it should not be forgotten that there are also other competitors. Amazon, with its CodeWhispereris a name not to be underestimated. Microsoftfor its part, has relied on OpenAI to significantly improve its Copilot.
Simply put, in the coming months and years we will most likely experience a compelling challenge between the large technological giants in the field of AI.