Google opens doors to safer artificial intelligence. Introducing Gemma 2

Google embarks on a quest for more ethical and transparent artificial intelligence. As part of this effort, it has released a trio of new generative AI models named Gemma 2, which promise higher security, lower hardware requirements, and smoother operation.

Security first

Unlike the Gemini models, which Google uses for its own products, the Gemma series is intended for more open use by developers and researchers. Similar to Meta's Llama project, Google is also striving to build trust and collaboration in the AI field.

The first innovation is the Gemma 2 2B model, designed for generating and analyzing text. Its greatest advantage is the ability to run on less powerful devices. This opens the door to a wide range of users. The model is available through platforms like Vertex AI, Kaggle, and Google AI Studio.

ShieldGemma: Guardian against toxic content

The new ShieldGemma introduced a set of safety classifiers whose task is to identify and block harmful content. This includes manifestations of hate, harassment, and sexually explicit materials. In short, ShieldGemma acts as a filter that checks both the input data for the model and its outputs.

The final addition is Gemma Scope, a tool enabling detailed analysis of the Gemma 2 model's functioning. Google describes it as a set of specialized neural networks that unpack the complex information processed by the model into a more comprehensible form.

Researchers can better understand how Gemma 2 identifies patterns, processes data, and generates results. The release of the Gemma 2 models comes shortly after the U.S. Department of Commerce supported open AI models in its report.

These make generative AI accessible to smaller companies, non-profits, and independent developers. Additionally, the report highlighted the need for tools to monitor and regulate these models to minimize potential risks.