JetBrains, renowned for its suite of app development tools, has recently unveiled its first open AI model, Mellum, designed for coding tasks. On Wednesday, the company released this code-generating model on Hugging Face, an AI development platform, after initially launching it for its software development tools last year.
Mellum is a sophisticated code completion model, trained on over 4 trillion tokens and comprising 4 billion parameters, making it adept at completing snippets of code based on context. The training process utilised a collection of permissively licensed datasets, such as code from GitHub and content from English-language Wikipedia, and was executed over approximately 20 days on a cluster of 256 Nvidia H200 GPUs.
The parameters of an AI model signify its problem-solving ability, while tokens are the data inputs processed by the model; notably, one million tokens equate to around 30,000 lines of code. Mellum is designed for seamless integration with professional development tools, aiding in intelligent code suggestions within integrated developer environments (IDEs), supporting AI coding assistants, as well as contributing to research in code understanding and generation. Furthermore, its capabilities extend to educational applications and experimentation with fine-tuning, although JetBrains advises that the base model requires fine-tuning before deployment.
While JetBrains has provided some Mellum models specifically fine-tuned for Python, the company notes these versions should only be regarded as estimates of potential capabilities, rather than for use in production. This reality underscores the importance of ensuring a model is properly honed before implementation.
Despite the advantages of AI-generated code in transforming software development practices, it brings forth security challenges as well. A recent survey by developer security platform Synk revealed that over 50% of organisations encounter security issues related to AI-generated code frequently or occasionally.
JetBrains has acknowledged that Mellum may inherit biases from public codebases, which could influence the style and security of the code it generates. Its suggestions may not always guarantee secure or vulnerability-free outcomes.
In a reflective statement, JetBrains conveyed that this initiative represents the onset of a broader endeavour, emphasising their commitment to focusing on specific use cases instead of pursuing general solutions. They expressed hope that Mellum could inspire significant trials, contributions, or collaborations within the developer community.
This move marks an important step for JetBrains, as they strive to support developers by leveraging open-source AI technology, enhancing productivity while cautiously navigating the associated challenges in security and bias.
Fanpage:Â TechArena.au
Watch more about AI – Artificial Intelligence
