Stability AI has announced StableCode, it’s very first LLM generative AI product for coding. StableCode provides AI-assisted coding capabilities that aim to boost programmer productivity and lower barriers to entry for aspiring developers.
While conversational AI assistants like Llama, ChatGPT, and Bard are already capable of writing code, they are not optimized for the developer experience. StableCode joins tools like Microsoft's GitHub Copilot and other open-source models that can be directly integrated into an IDE such as VSCode.
StableCode incorporates three specialized models to offer intelligent autocomplete, respond to instructions, and manage long spans of code. Trained on an extensive dataset of over 500 billion tokens from diverse programming languages, the system can understand and generate code in languages like Python, Java, JavaScript, and more.
The standard autocomplete model, StableCode-Completion-Alpha-3B-4K, draws on this understanding of code structure and dependencies to offer single and multi-line recommendations as developers type. It is intended to do single/multiline code completion from a long context window upto 4k tokens.
The instruction model, StableCode-Instruct-Alpha-3B, allows users to leverage natural language prompts to perform coding tasks. While the model can independently generate new code completions, Stability AI recommends that developers use it with HuggingFace's VSCode extension.
Finally, StableCode-Completion-Alpha-3B has a long context window (upto 16k tokens) and can handle a lot more code at one. Stability AI says this can allow a user to "review or edit the equivalent of up to five average-sized Python files at the same time".
StableCode joins a proliferation of open source AI coding models like StarCoder, CodeGen, Replit-code-v1-3b and Pangu-coder2. However, with twice the cotext length of most existing models, ample training data, and carefully tuned models, Stability AI believes StableCode can significantly enhance developer workflows. The system not only promises greater efficiency but also has the potential to lower the barrier to entry for new programmers.