Earlier this year, GitHub introduced Copilot Chat, a programming-focused chatbot similar to ChatGPT, initially available for organizations subscribed to Copilot for Business. Subsequently, Copilot Chat extended to individual Copilot customers, priced at $10 per month, during its beta phase. Now, GitHub is officially launching Chat for all users.
As of today, Copilot Chat is accessible in the sidebar of Microsoft’s integrated development environments (IDEs), Visual Studio Code and Visual Studio. It is included in GitHub Copilot’s paid tiers and is free for verified teachers, students, and maintainers of specific open-source projects.
Shuyin Zhao, VP of product management at GitHub, emphasized the widespread adoption of Copilot as the most widely used AI developer tool in history. Copilot Chat, powered by GPT-4, OpenAI’s flagship generative AI model fine-tuned for development scenarios, allows developers to seek real-time guidance using natural language prompts. This includes asking Copilot Chat to explain concepts, identify vulnerabilities, or generate unit tests.
Despite its official launch, there have been minimal changes to Copilot Chat since the beta phase. It retains its reliance on GPT-4 and continues to address potential copyright concerns with fair use doctrine. However, legal challenges, such as class action lawsuits alleging open-source licensing and IP violations, have been filed against GitHub, Microsoft (GitHub’s parent company), and OpenAI.
When asked about the possibility for codebase owners to opt out of training, Zhao suggested making repositories private to avoid inclusion in future training sets. This suggestion may face resistance from codebase owners who have reasons, such as crowdsourcing bug hunting, for keeping copyrighted code public. GitHub currently does not provide a specific mechanism for opting out of training data.
Generative AI models like GPT-4 may exhibit hallucinations or confidently generate inaccurate facts, particularly in coding contexts. A Stanford study indicates that developers using AI assistants for coding may produce less secure code due to the introduction of buggy or deprecated code snippets. Zhao highlighted GPT-4’s improved performance against hallucinations compared to the older model and mentioned exploit-mitigating features like filters for insecure code patterns in Copilot Chat.
Despite Copilot’s success with 1 million paying users and ~37,000 enterprise clients as of October, GitHub faces challenges in making Copilot profitable. The underlying AI models reportedly contribute to Copilot’s monthly losses per user, with some customers costing as much as $80 per month. In contrast, Amazon’s CodeWhisperer, a notable competitor, has made strategic moves such as offering a free version, introducing a professional tier with enhanced features, and optimizing suggestions for specific development environments.
Apart from CodeWhisperer, Copilot contends with competition from startups like Magic, Tabnine, Codegen, and Laredo, as well as open-source models like Meta’s Code Llama and Hugging Face’s and ServiceNow’s StarCoder.
