Code Llama is official: this is how the new Meta AI works to help you code

It was just a few days ago Target make it official. After some of its features became known, Mark Zuckerberg’s firm launched Code Flame, his new AI model capable of writing and fixing code. The new tool, specifically designed for programmers, works with both natural language instructions and other code snippets.

It is worth noting that Code Llama is not a chatbot like ChatGPT or Bard, which, among its many functions, is able to write code. The new AI Meta was created specifically for this task and includes some very interesting elements.

As the name suggests, Code Llama is based on LLaMA 2, a powerful Mass Language Model (LLM) released some time ago. Menlo Park residents explain that their artificial intelligence is compatible with the most popular programming languages. Among them, Python, Javascript, PHP, C++, C# and Bash.

Code Llama can be used to complete blocks of code, or to find bugs and fix recommendations. But the Californian company says that in addition to programming, AI can generate explanations of code in natural language or from its fragment.

Code Llama will have specialized versions

Another very interesting point to consider is that Meta will be releasing three variants of this model. Code Flame is the underlying model for code generation, while Flame-Python Code This is a specialized version in this particular programming language. Bye Flame-Instruct Code it is optimized for understanding natural language instructions.

It is worth noting that Meta does not recommend using Code Llama or Code Llama-Python for tasks where the input method is natural language instructions. These models are prepared only for specific tasks from the code, the company said. The Code Llama-Instruct, on the other hand, is designed to “better understand what people hope to achieve with their abilities.” hints“.

Californians also understand that not all AI-powered jobs are equally complex or require the same resources. That’s why the three versions of the Lama’s Code will be offered in 3 sizes. The smallest one has 7000 million parameters; intermediate – 13,000 million and the largest – 34,000 million.

Regardless of the differences between them, all versions were trained using 500,000 million tokens of code data and other programming-related information. The big difference lies in the speed of work and the complexity of the proposed results.

This is why the Meta explains that, for example, a version of Code Llama with 7 billion parameters can run on a computer with a single GPU. In addition, it provides a faster response to tasks such as completing lines of code. For its part, the version with 34,000 million parameters is more accurate and offers better results, but is not optimized for real-time performance.

AI malware threat mitigation

With the hype around generative artificial intelligence, tools like ChatGPT and Bard are increasingly being used to create malware. And while they haven’t done a very good job of it so far, it’s a growing concern in the industry. For this reason, Meta confirms that it has analyzed the situation and taken steps to address it. reduce the misuse of your technology.

“Like all advanced technologies, Code Llama carries risks. Responsibly building AI models is critical, and we have taken numerous security measures ahead of the launch of Code Llama. As part of our red team efforts, we have quantified the risk of Code Llama. creating malicious code. We created posts that attempted to request malicious code with explicit intent and compared Code Llama responses with ChatGPT (GPT3.5 Turbo) responses. Our results showed that Code Llama responded with safer responses.”

Target.

The Lama Code is now available and is free to use for both commercial and research purposes.. Those who want to learn in depth (and more technically) about this new AI model can read this detailed description. paper Author: Menlo Park.