Despite concerns about potential security vulnerabilities and "code sprawl", StarCoder 2 is seen as more efficient than some other code generators. It can be deployed locally and "learn" a developer's source code, which is attractive to those wary of exposing code to a cloud-hosted AI. Unlike some competitors, StarCoder 2 was trained only on data under license from the Software Heritage, reducing the risk of copyright infringement. However, it may generate code reflecting stereotypes about gender and race, and performs weaker on languages other than English and "lower-resource" code.
Key takeaways:
- AI startup Hugging Face, in collaboration with ServiceNow and Nvidia, has released StarCoder 2, an open source code generator that offers improved performance at lower operating costs. The tool can be fine-tuned to create apps such as chatbots and personal coding assistants.
- StarCoder 2 is licensed under Hugging Face’s RAIL-M, which imposes "light touch" restrictions on both model licensees and downstream users. However, it does not permit developers to use StarCoder 2 for every conceivable application.
- Despite its advantages, StarCoder 2 has its limitations. It is susceptible to bias and performs weaker on languages other than English and "lower-resource" code like Fortran and Haksell.
- ServiceNow, Hugging Face, and Nvidia are offering paid services on top of the open source releases. ServiceNow has used StarCoder to create Now LLM, a product for code generation fine-tuned for ServiceNow workflow patterns, use cases, and processes. Hugging Face and Nvidia are providing hosted versions of the StarCoder 2 models on their platforms.