Google has launched a trio of recent, “open” generative AI fashions that it’s calling “safer,” “smaller” and “extra clear” than most — a daring declare, to make certain.
They’re additions to Google’s Gemma 2 household of generative fashions, which debuted again in Might. The brand new fashions, Gemma 2 2B, ShieldGemma and Gemma Scope, are designed for barely totally different functions and use instances, however share in widespread a security bent.
Google’s Gemma sequence of fashions are totally different from its Gemini fashions in that Google doesn’t make the supply code obtainable for Gemini, which is utilized by Google’s personal merchandise in addition to being obtainable to builders. Moderately, Gemma is Google’s effort to foster goodwill inside the developer neighborhood, very like Meta is trying to do with Llama.
Gemma 2 2B is a light-weight mannequin for producing analyzing textual content that may run on a variety of {hardware}, together with laptops and edge units. It’s licensed for sure analysis and industrial functions and will be downloaded from sources comparable to Google’s Vertex AI mannequin library, the info science platform Kaggle and Google’s AI Studio toolkit.
As for ShieldGemma, it’s a group of “security classifiers” that try and detect toxicity like hate speech, harassment and sexually express content material. Constructed on prime of Gemma 2, ShieldGemma can be utilized to filter prompts to a generative mannequin in addition to content material that the mannequin generates.
Lastly, Gemma Scope permits builders to “zoom in” on particular factors inside a Gemma 2 mannequin and make its internal workings extra interpretable. Right here’s how Google describes it in a weblog put up: “[Gemma Scope is made up of] specialised neural networks that assist us unpack the dense, advanced info processed by Gemma 2, increasing it right into a type that’s simpler to research and perceive. By learning these expanded views, researchers can acquire precious insights into how Gemma 2 identifies patterns, processes info and in the end makes predictions.”
The discharge of the brand new Gemma 2 fashions comes shortly after the U.S. Commerce Division endorsed open AI fashions in a preliminary report. Open fashions broaden generative AI’s availability to smaller firms, researchers, nonprofits and particular person builders, the report stated, whereas additionally highlighting the necessity for capabilities to watch such fashions for potential dangers.