• By quade
• 1 March 2024

New AI Code Generator Set to Release With 619 Programming Languages

With AI technology expanding, many people are working to create an AI model that can take over repetitive tasks to free up workers to handle other projects. Among the tasks that many businesses hope to outsource to AI is programming.

Coding is an essential task for all online businesses as it is how you can set up your online presence and maintain your website. The problem is, that placing that code requires a lot of time and patience, something many people do not have. For that reason, ServiceNow Inc., Hugging Face Inc., and Nvidia Corp are working on a project that can fix that problem.

They believe their new AI model, StarCoder2 will be able to handle the heavy lifting of coding, allowing businesses to automate their services. This model is a continuation of the existing StarCoder series, a model used by programmers for coding. However, the creators at Nvidia Corp are confident that this latest AI can easily surpass their predecessors with its robustness and capabilities.

What Can StarCoder2 Do?

Starcoder2 is the latest AI model in the StarCoder2 business, providing much-needed updates for predecessors. Nvidia Corp claims that this version is faster and more flexible than the previous models, giving you more options with your coding requests.

More importantly, there have been additional guardrails and security measures that prevent AI from committing intellectual property infringement. This ensures that the code it creates for you will be safe to use without the risk of being sued.

More importantly, StarCorder2 is also more flexible as it can generate code in 619 programming languages. This is possible thanks to their partnership with BigCode Community. This research project is managed by ServiceNow and Hugging Face and was used last year to create the original StarCoder. Now their continuous research has been used to update the AI model.

The model is built around a larger dataset called Stack v2, which is seven times larger than the previous Stack v1. This new dataset gives the AI more information, including those about less popular programming languages like Cobol. The dataset also explains the mathematics and programming code to help the AI generate more accurate code.

You also can implement StarCoder into existing business software so it can create source code for you based on existing data. This can be used to create workflows and summarize text which can speed up coding by summarizing code and finding code snippets.

StarCoder2 Flexibility

The biggest selling point of StarCoder2 is its flexibility. The model can be programmed to work on countless different programming languages and comes in different model sizes, all trained by different groups. There is a 3-billion-parameter model trained by ServiceNow, a 7-billion-parameter model trained by Hugging Face, and a 15-billion-parameter model built by Nvidia. The latter uses the NeMo generative AI and is trained on Nvidia infrastructure.

There are different advantages to picking different-sized models. The smaller ones are more energy efficient as they need fewer parameters when they process requests in the inferencing stage. This makes them compatible with consumer-level GPUs.

Despite that, the StarCoder2 creators boast that the 3 billion-parameter model performs just as well as the 15 billion one in processing, providing accurate predictions. However, the 15-billion model will provide more context-aware predictions because of the increased data.

AI has been a popular tool in software development, especially since GitHub’s Copilot and Amazon’s CodeWhisperer have become available to the public. A survey by CoderPad Inc. showed that 91% of U.S. developers made use of AI coding tools in some way, showing many people are open to the technology. However, not everyone is on board as 25% of developers are unsure whether the AI is beneficial and another 28% refuse to use it at all.

Issues with AI Coding

One reason for the hesitation in this AI adoption is the fear of its code quality. Without human developers constantly double-checking the code, many fear the AI could create inefficient code and leave the company open to legal troubles with intellectual property infringement. This fear is far from unfounded either as recent studies conducted by Stanford University have shown that some AI models did exactly that and created insecure codes.

This is where the companies who manage StarCoder2 have worked to address these issues. They insist that the model was built with data from Software Heritage. All of the code there is not copyright protected which reduces the risk of the AI using stolen code. Most of the model code also resides on BigCode’s GitHub page, under the OpenRAIL-M license, which is open to royalty-free access and use.

However, while it is open-source, there are still limitations to what the Rail-M License can provide. Things like medical or legal advice are strictly prohibited.

Here at geniusOS, we have followed closely with this story and have eagerly implemented this model into our list of tools. However, we also understand its limitations and work to ensure that all code we generate is kept secure and doesn’t violate any copyright laws.