AI startup Hugging Face has partnered with ServiceNow to release a free AI coding system called StarCoder.

With StarCoder, Hugging Face and partner ServiceNow are launching a free alternative to artificial intelligence coding systems like Copilot from Microsoft’s GitHub.

Risks and Opportunities of AI in Coding

Systems like Amazon’s CodeWhisperer or OpenAI’s Codex (which runs Copilot) are already lifting part of the veil covering the possibilities of AI in coding and computer programming. For the time being, there are still many issues to consider, both ethical, legal and purely technical.

According to a study from the University of Cambridge, developers spend about half their time fixing bugs instead of programming effectively. This fact costs the software industry more than $300 billion annually. Solving these problems therefore represents a huge step forward: with a decent implementation of artificial intelligence in programming, developers can focus on the more creative side of their work. However, not many systems are currently publicly available.

So now there is one with StarCoder. Anyone can use it royalty-free, even businesses. The system was trained in more than 80 programming languages ​​and with GitHub’s text collection, including documentation and virtual programming notebooks. The system is also complementary to Microsoft’s Visual Studio Code. And just like the well-known ChatGPT, you can give StarCoder instructions like “Create a user interface for an app” or ask questions about the code.

According to Hugging Face, StarCoder is as good or maybe even better than OpenAI. The company relies on the creativity and expertise of the open source community and takes Stable Diffusion as a model. A few weeks after the launch of this AI image generator, dozens of variations and applications were already circulating.

big picture

StarCoder is part of the BigCode project. Hugging Face and ServiceNow want to responsibly develop pioneering AI systems that can write code themselves.

More than six hundred people are now working on the project, divided into different working groups. For example, there is a group that develops models for code, one for evaluation, or a group that deals with the ethical side of this topic. The latter is not unimportant: Under the name Legal, Ethics and Governance, this group deals with everything to do with the licensing of data used, the origin of original code and the risks of malicious code. The group also performs editorial work on potentially identifiable personal information.

BigCode’s intention is to address a number of issues related to the implementation of artificial intelligence in coding. For example, licenses and copyrights are an important point. Both OpenAI and GitHub have been criticized by the nonprofit Software Freedom Conservancy for using public code to train commercial systems. However, both companies deny they are doing anything that isn’t allowed, at least in the US.

StarCoder was trained by The Stack; this is a data collection with more than 19 million unlicensed and compound texts and up to six terabytes of code in more than 350 programming languages. Since The Stack is license-free, the code used from it can not only be copied, but also modified and redistributed. BigCode even helps developers remove their work from The Stack.

The project goes one step further and is actively working to remove potentially identifiable personal data from The Stack. Think of names, IP addresses, usernames or even passwords. Additionally, BigCode uses Hugging Face’s malware detection system to extract dangerous files from The Stack.

Important issues addressed

For now, privacy will remain a key focus when using generative AI. After all, such systems train with immense amounts of data that are accessed almost unfiltered from the Internet. Some experts even warn that AI coding can pose a threat to companies by using sensitive information or copyrighted text without the companies’ knowledge.

Because some systems remove licenses from code in use, it’s often difficult to tell which code is free to use and which code is restricted. In response to this criticism, GitHub installed a virtual switch that allows users to avoid potentially copyrighted code. Amazon has also built something similar into its CodeWhisperer.

caveats

Given these concerns, what exactly does ServiceNow get out of StarCoder? ultimately it is also a commercial company (mainly focused on automation software). The answer, according to ServiceNow, is “a strong responsible AI model that enables commercial opportunity.” It is therefore expected that ServiceNow will integrate StarCoder into its commercial products in the near future.

Exactly how much the company has invested in BigCode is not known, but ServiceNow has already indicated that it is a significant sum. The company also emphasized the many benefits of participating in the project anyway.

StarCoder may be advertised as open source, but that’s not entirely true. The system is released under a license scheme called OpenRAIL-M. This means that there are actually things that users should consider. For example, StarCoder may not be used to write malicious code. After all, specialists have already shown how AI can be used to make malware invisible to detection.

StarCoder is available on GitHub this week; initially with fewer features than Copilot, but ServiceNow is counting on the community to make improvements and even provide their own models. Meanwhile, BigCode continues to search for new models while improving the current systems. This is necessary because when Hugging Face and ServiceNow launched StarCoder, they said there might still be a risk of malicious code, along with potential privacy and deceptive content issues.