So check this out, Meta has dropped another one of their machine learning models on us. This time it’s all about generating source code for software. They call it Code Llama, and it’s part of their LLaMA family of language models. Now, LLaMA stands for Llama 2 model, which was released a while back. But this new version, Code Llama, is all about generating and discussing source code in response to text prompts. It’s a pretty cool tool for programmers looking for some help writing better code.
But of course, there are risks with this kind of cutting-edge technology.
According to Meta, Code Llama can generate code for various programming languages like Python, C++, Java, and more. So if you need a function that spits out the Fibonacci sequence, just ask Code Llama and it’ll give you the code and a nice explanation to go with it. But here’s the thing, you gotta talk to Code Llama in English, because it hasn’t been tested for other languages yet. Who knows what kind of crazy stuff it might say if you try to chat with it in another language?
Now here’s what Meta has to say about the risks of using Code Llama. They did some testing to see if it could create malicious code, and turns out Code Llama is actually safer than ChatGPT, another AI model from Meta. So that’s a good sign.
According to Meta, Code Llama performs better than other open-source code models and even matches the performance of OpenAI’s ChatGPT. It comes in three different sizes, with the largest one having 34B parameters. To give you an idea, that’s way more parameters than the previous version of OpenAI’s Codex.
Now here’s something interesting. The smaller versions of Code Llama can actually fill in missing source code without any additional fine-tuning. They’re perfect for code completion tasks where speed is important. And there are two different variants of Code Llama: one focused on Python code and the other fine-tuned for code generation based on input and output patterns.
So, is it reliable?
Well, here’s the thing with language models like Code Llama. Sometimes they give incorrect answers, especially when it comes to programming prompts. But despite that, many developers still use them to quickly recall patterns and API parameters without having to search through documentation. And what’s cool about Code Llama is that it can handle really long code sequences, up to 100,000 tokens. So you can throw a bunch of code at it and get a detailed response.
Meta says that having longer input sequences opens up new possibilities for Code Llama. You can provide more context from your codebase to make the generated code even more relevant. It’s also useful for debugging larger codebases where keeping track of all the code related to a specific issue can be a challenge for developers.
So you can give Code Llama more context to get better results.
Code Llama is part of a growing trend of AI models designed to understand and generate code. It all started with OpenAI’s Codex, and then there’s GitHub’s Copilot, DeepMind’s AlphaCode, OpenAI’s GPT-4, Amazon Code Whisperer, and Google’s Bard. There have also been some open-source models like StarCoder and XGen. So there’s definitely a lot going on in this space.
Meta has released Code Llama under a community license, which is cool and all, but it’s not exactly open source. They’re all about this “open approach to AI,” but it’s limited to their own models. You can’t use Code Llama to improve any other large language model. And if you’re a mega-service with more than 700 million monthly active users, you gotta get a special license from Meta to use their stuff. They gotta keep things exclusive, you know?