GitHub is currently facing a lawsuit for using open-source code in its training data, and Masad says Meta may have restricted the training data to avoid such complications. Copilot costs $10 per month for consumers and $19 per user per month for businesses.
It seems that Copilot has been a hit with the developers. According to data from GitHub in June, it is used by more than a million developers and more than 200,000 organizations. The company’s own studies also suggest that Copilot accelerates the speed at which individuals complete tasks, resulting in a 30% increase in productivity.
Meta will release two versions of Code Llama, one geared towards generating Python code and the other optimized for converting natural language commands to code. In addition, three model sizes are available. The smallest runs on a single graphics processing unit, or GPU.
Meta claims that Code Llama is trained on public-domain code. In two popular coding benchmarks: HumanEval has mostly basic Python issues; its performance is much better than existing open coding models and is “on par with ChatGPT,” the company emphasizes.
Amjad Masad, CEO of Replit, a coding platform online that offers various generative AI tools, doesn’t expect Code to replace Llama Copilot as its training data will likely be more limited. However, he believes the release would allow developers to experiment with agents that perform useful tasks, such as searching for information on the Internet or resorting to an application programming interface to book a flight or order food. “It’s a really exciting field,” says Masad. “The interactions where you write instructions in natural language and the model “chews” the information allow you to do interesting things in the world.”
The release of Code Llama could also bring Meta profits. The company may not have ChatGPT or an AI-powered search engine, but if it establishes itself as a free AI provider for many developers, companies, and researchers, it would be in the race to leverage generative AI. Meta opted for an open approach after an early version of Llama was leaked online in May.
Neither Llama 2 nor Code Llama are released under conventional licenses from open source software, which would allow unrestricted commercial use. For example, under the Meta license, users are prohibited from using the models in any application or service that has more than 700 million monthly users.
A research report published online this month highlights that the adoption of artificial intelligence tools could bring significant indirect benefits to the companies behind them, for example by keeping researchers in their systems and providing them with new insights that they could exploit at scale.