Google has introduced three new ‘open’ generative AI models, highlighting their enhanced safety, smaller size, and improved transparency compared to many existing models - a significant claim.
The latest additions to Google's Gemma 2 lineup of generative models, which was first introduced in May, include the new models Gemma 2 2B, ShieldGemma, and Gemma Scope. These models are tailored for various applications and use cases, emphasizing a shared commitment to safety.
Unlike Google's Gemini models, which have closed-source code and are integrated into Google's products and used by developers, the Gemma series strives to promote goodwill among developers, akin to Meta's strategy with Llama.
Gemma 2 2B, a lightweight text-generation model, is designed to work efficiently across a range of hardware, from laptops to edge devices. This model is licensed for exclusive use in specific research and applications, and is accessible through platforms such as Google's Vertex AI model library, Kaggle, and Google's AI Studio toolkit.
ShieldGemma is a set of ‘safety classifiers’ crafted to identify harmful content like hate speech, harassment, and sexually explicit material. Utilizing Gemma 2 as its foundation, ShieldGemma has the capability to filter prompts to a generative model and the resulting content produced by the model for inappropriate material.
Gemma Scope, on the other hand, enables developers to "zoom in" on specific aspects of a Gemma 2 model, making its inner workings more interpretable. According to Google, Gemma Scope consists 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 Gemma 2 models coincides with the recent endorsement by the U.S. Commerce Department of open AI models in a preliminary report. The report emphasizes the benefits of open models, such as increasing accessibility for smaller companies, nonprofits, researchers, and individual developers to generative AI, while also acknowledging the importance of monitoring capabilities to mitigate potential risks.