StarCoder 2 is reportedly more efficient than some other code generators, including Code Llama 33B, and can be deployed locally, allowing it to learn a developer's source code. Unlike some competitors, it was trained only on data under license from the Software Heritage, avoiding potential copyright issues. However, its license, the BigCode Open RAIL-M 1.0, may still pose challenges for some developers, as it imposes certain restrictions and may conflict with AI-related regulations such as the EU AI Act.
Key takeaways:
- AI startup Hugging Face, in collaboration with ServiceNow and Nvidia, has released StarCoder 2, an open-source code generator that comes in three variants with different parameter sizes. The models can be used to complete unfinished lines of code and retrieve code snippets when asked in natural language.
- StarCoder 2 has been trained with significantly more data than the original StarCoder, leading to improved performance at lower operating costs. It can be fine-tuned quickly using a GPU to create applications like chatbots and personal coding assistants.
- Compared to other code generators, StarCoder 2 is more efficient and can be deployed locally to learn a developer's source code. It is also considered more ethical as it was trained only on data under license from the Software Heritage, giving code owners the option to opt out of the training set.
- Despite its advantages, StarCoder 2's license, the BigCode Open RAIL-M 1.0, may pose a challenge for some developers as it imposes certain restrictions and may conflict with AI-related regulations such as the EU AI Act.