Google has released a trio of new “open” generative AI models that it calls “safer,” “smaller,” and “more transparent” than most — a bold claim, to be sure.
These are additions to Google Gemma 2 family of generative models, which debuted in May. The new models, Gemma 2 2B, ShieldGemma, and Gemma Scope, are designed for slightly different applications and use cases, but share a common focus on security.
Google’s Gemma model series differs from its Gemini models in that Google does not make available the source code for Gemini, which is used by Google’s own products and is also available to developers. Rather, Gemma is Google’s effort to foster goodwill within the developer community, much like Meta tries to make it with Llama.
Gemma 2 2B is a lightweight model for generating analytical text that can be run on a range of hardware, including laptops and edge devices. It is licensed for select research and commercial applications and can be downloaded from sources including Google’s Vertex AI model library, the Kaggle data science platform, and Google’s AI Studio toolkit.
ShieldGemma is a set of “safety classifiers” that attempt to detect toxic content such as hate speech, harassment, and sexually explicit content. Built on top of Gemma 2, ShieldGemma can be used to filter prompts to a generative model as well as the content generated by the model.
Finally, Gemma Scope allows developers to “zoom in” on specific points in a Gemma 2 model and make its inner workings more interpretable. Here’s how Google describes it in a blog post : “[Gemma Scope is made up of]specialized neural networks that help us unpack the dense, complex information processed by Gemma 2, expanding it into a form that’s easier to analyze and understand. By studying these expanded views, researchers can gain valuable insights into how Gemma 2 identifies patterns, processes information, and ultimately makes predictions.”
The release of the new Gemma 2 models comes shortly after the U.S. Department of Commerce endorsed open AI models in a preliminary report. Open models expand the availability of generative AI to small businesses, researchers, nonprofits, and individual developers, the report said, while also emphasizing the need for monitoring capabilities for these models to detect potential risks.
