Replit is a company that develops an IDE accessed through a browser and has recently started using AI to assist in code writing. They have introduced a feature called Code Repair which utilizes LLM customized to focus on bug fixing specifically.
Replit mentions that the use of LLM in everyday code writing is a result of training with source code data combined with human conversations such as threads in StackExchange or GitHub issues. This approach has limitations due to the lack of context in the development environment each programmer uses.
As Replit is a company that directly creates IDEs, they already have real development environment data available. This data can be leveraged to help address a significant issue for programmers, which is bug fixing in existing code.
The Replit IDE is based on VS Code technology, using the Monaco editor as a base and implementing the Language Server Protocol (LSP) programming language protocol. This allows Replit to gather a vast amount of error data from LSP and utilize it to train their models effectively.
Replit uses this error data to train an open-source code writing model with around 7B parameters, selecting the DeepSeek-Coder-Instruct-v1.5 for efficiency reasons. The results show that the Replit Code Repair model excels in bug fixing compared to larger models like GPT-4 Turbo and Claude 3 Opus.
Currently, the Replit Code Repair model is still in an internal experimental phase within the company and will continue to be developed in various aspects, such as further language support and improving stability, before being used at a production level.
TLDR: Replit is using AI to enhance code writing with their Code Repair feature, leveraging LLM trained on source code and human conversations. The model shows promise in bug fixing and is being further developed internally for future improvements and implementation.
Leave a Comment