When one thinks about the development of artificial intelligence models, the two names that initially come to mind are OpenAI and, perhaps, Google. However, they are not the only ones. Meta, formerly known as Facebook, is also dedicating a lot of resources to this work. And proof of this is the new Code Llama 70B, a completely free artificial intelligence model that surpasses the well-known GPT-3.5 and, in turn, reduces the distance with GPT-4, the most advanced OpenAI model.
Code Llama 70B, as its name indicates, is an artificial intelligence model focused on programming and composed of 70 billion parameters (or 70 trillion, according to American notation). A higher number of parameters translates into greater complexity and, generally, better results.
Code Llama 70B can be used 100% free for both commercial and research uses. The company has also released three different models: the generic one (Code Llama 70B) and, in addition, one focused on Python and another on Instruct.
According to the tests published by Meta, the generic version of Code Llama 70B achieves 53% accuracy in the HumanEval test, outperforming the GPT-3.5 model (48.1%) and reducing the distance with GPT-4 (67%) and Gemini Ultra (74.4%), according to data published by both companies.
However, if we take the Instruct and Python versions as a reference, the comparison becomes more favorable for Code Llama. Instruct achieves 67.8% accuracy, narrowly beating GPT-4 In this stage; while Code Llama Python reaches 57.3%.
Model | HumanEval |
---|---|
Code Llama 70B | 53% |
Code Llama 70B Python | 57.3% |
Code Llama 70B Instruction | 67.8% |
OpenAI GPT-3.5 | 48.1% |
OpenAI GPT-4 | 67% |
Google Gemini Ultra | 74.4% |
“Writing and editing code has become one of the most important applications of artificial intelligence models today. It has also been shown that the ability to program is important for artificial intelligence models to be able to process information in other domains in a way more rigorous and logical,” Mark Zuckerberg detailed in your Facebook profile. “I am proud of this progress,” she said.
The launch of Code Llama 70B is just a sample of the efforts that the company led by Mark Zuckerberg is making in the field of artificial intelligence. The manager, in a recent interview, marked the development of AGI (artificial general intelligence) as one of the company’s goals. And, at least for the moment, everything is being done in a way open sourcean approach the developer community praises considering that it is something as influential in society as artificial intelligence.
The emphasis on artificial intelligence has, however, made it his commitment to the metaverse has less prominence. At least publicly. The company maintains its investment in AR, VR and MR through the Reality Labs division and continues to bring products to market. However, the impact and interest around the concept of metaverse has fallen compared to 2021, when the company even changed its name from Facebook to Meta.