Updates from GitHub and a consortium comprising Nvidia, HuggingFace, and ServiceNow are set to introduce new options to the wide array of AI coding assistants available to developers. However, experts are cautioning about the adoption of these tools due to ongoing security and copyright concerns.
The new GitHub Copilot Enterprise, a higher tier version of the popular GitHub Copilot AI coding assistant, has been made generally available this week. Priced at $39 per user per month for users of GitHub’s Enterprise Cloud, this version offers customization for organizations using Copilot and generates chat answers, code completion, and pull request difference analysis based on specific codebases. Additionally, an add-on that will provide fine-tuned AI models is in the works, as mentioned in a GitHub blog post.
According to Andy Thurai, an analyst at Constellation Research, coding copilots are proving to be a valuable tool for improving developer efficiency, with many enterprises considering, experimenting with, and implementing them. GitHub Copilot, backed by Microsoft, has the advantage of integration with Visual Studio, giving it an early lead in the market.
A survey of 800 engineering professionals conducted by Sonatype revealed that GitHub Copilot is already one of the most widely used AI coding assistants available, with 70% of respondents using the tool. The survey also highlighted that 97% of DevOps and SecOps leaders are currently utilizing generative AI in their workflows, with a majority using two or more tools daily.
Despite GitHub’s claims of “enterprise-grade security, safety, and privacy” with Copilot Enterprise, concerns around security and copyright issues persist. AI coding assistants, including GitHub Copilot, have been shown to reproduce security vulnerabilities and bad practices from a customer’s existing codebase, as pointed out in recent research by cybersecurity vendor Snyk.
In an effort to address these concerns, a new set of AI models called StarCoder2 has been developed by a consortium comprising Nvidia, HuggingFace, and ServiceNow. These models, which are cheaper to run and can be fine-tuned to provide better quality answers based on specific codebases, aim to alleviate security and legal worries surrounding AI coding assistants.
The StarCoder2 models, available in three sizes ranging from 3 to 17 billion parameters, have been trained using responsibly sourced data under license from the digital commons of Software Heritage. This approach, along with the models being trained in 619 programming languages, is expected to appeal to enterprises looking for alternatives to Copilot due to copyright concerns.
With the rapid advancement of AI technology in the coding assistant space, developers and organizations are urged to proceed cautiously and consider the risks and rewards of implementing these tools. While AI coding assistants hold the promise of enhancing productivity, addressing security vulnerabilities and copyright issues should remain a top priority for those leveraging these technologies.