Products You May Like
Google has shifted gears after being caught flat-footed by OpenAI and Microsoft. The search giant has integrated its Gemini AI into numerous products, but that’s not the only Google AI. The company also offers a collection of open-source models, and it’s adding some new options today. Google says its latest Gemma models are more efficient and safer, and developers can begin tinkering with them now.
Gemini is Google’s flagship AI, which you can find as a standalone chatbot on virtually every platform. It also powers AI features in products like Gmail, Google Drive, and more. Google doesn’t publish the source code for Gemini, but the company has a history of contributing openly to AI research. It published the first work on the transformer algorithms that underlie all current large language models (LLMs) like OpenAI’s GPT and Meta’s Llama.
The open Gemma models provide tools for developers who want to take a peek at source code, and the latest models offer a bevy of interesting features for them. Gemma 2 2B, ShieldGemma, and Gemma Scope have different intended applications.
Gemma 2 2B is a text analysis and generation model with 2 billion parameters, not to be confused with tokens, which is another metric often used to describe LLMs. Parameters are the internal probability values a model uses to generate outputs, and 2 billion is on the low side. The latest Gemini models have more than 1 trillion parameters, for example. However, 2B serves Gemma well. Google claims this compact model can run locally without specially designed hardware and outperforms all GPT-3.5 models. It’s available for both research and commercial applications.
ShieldGemma is a classifier model that detects and filters undesirable AI outputs to keep users safe. Google designed ShieldGemma to detect hate speech, harassment, sexually explicit content, and other dangerous materials that might emerge from the black box of generative AI models. It’s built on top of Gemma 2 with configurable parameter modes for different applications, from 2 billion for online and up to 27 billion for offline uses where latency isn’t a concern.
Credit: Google
Finally, there’s Gemma Scope, which might be the most important of the bunch. It aims to make Gemma 2’s inner workings easier to understand. That’s a problem for all commercial LLMs, which can produce strange outputs with no way for researchers to dig deeper. Gemma Scope uses sparse autoencoders that allow developers to “zoom in on specific points within the model and make its inner workings more interpretable.”
All three of Google’s new models are available now. Gemma 2 2B and ShieldGemma are both available for download from Google. Gemma Scope is live on “interpretability research” platform Neuronpedia.