Recently, Google introduced three new models of artificial intelligence: Gemma 2 2B, ShieldGemma, and Gemma Scope. They were officially added to the Gemma 2 series, launched in June this year. These models are considered safer and easier to use compared to similar products from other companies. The special feature of the updated Gemma series is its availability for programmers. In this way, one of the company's strategies is realized – to build strong relations with the development community. By the way, we can see the same thing at Meta with their Llama model.
Gemma 2 2B is a model designed for the creation of texts that can work on a variety of equipment, including laptops and peripherals. It is available for scientific and commercial projects and can be downloaded from resources such as Vertex AI Model Library, Kaggle, and Google AI Studio. Despite its compact size of 2.6 billion parameters, Gemma 2 2B competes effectively with larger models. It successfully demonstrates that size does not necessarily determine performance. The model is accessed through the Hugging Face platform. In addition, it supports various software frameworks, including PyTorch and TensorFlow.
ShieldGemma is a set of tools for detecting unwanted content, including images, stalking, or sexual content. Using technologies from the Gemma 2 base model, ShieldGemma can filter both incoming data and content generated by the model itself.
Gemma Scope gives developers a deeper understanding of how the Gemma 2 model works. This tool uses special neural networks to break down the complex data being processed by the model into more understandable elements.
According to Google, its new open models of artificial intelligence will provide wider access to generative AI for small companies, scientific researchers, non-profit organizations, and individual developers. At the same time, the company emphasizes the need to monitor these models to identify and minimize potential risks.