Google Launches Gemma, A New Family of Open AI Models

Google has unveiled Gemma, its latest lineup of generative artificial intelligence (AI) models for natural language processing. Building on its existing Gemini models revealed just last week, Gemma launches with two configurations, Gemma 2B and Gemma 7B.

While specifics around benchmark performance compared to competing models from Meta, Anthropic and others remain scarce for now, Google states Gemma represents “state-of-the-art” AI capable of assisting developers with a wide range of text generation and comprehension tasks.

Decoding Google’s New Open AI Approach

Google is quick to describe Gemma as “open” models. However, this does not mean the same thing as open source. Jeanine Banks, senior product manager at Google, clarified in a press briefing that Gemma is better referred to as an open access model.

Essentially this allows developers to leverage the pre-trained models for inference, personalization and experimentation permissive terms of use. But it stops short of enabling full redistribution or ownership of derivative works under legal licensing associated with open source.

While this may limit certain applications, Banks asserts that positioning Gemma models as open access strikes the right balance of control and customization for Google Cloud customers. It unlocks the advanced AI capabilities to enhance services, while protecting commercial interests and addressing responsible development concerns.

Compact Yet Powerful Foundation for AI Apps

The initial Gemma configurations weigh in at 2 billion and 7 billion parameters respectively. This puts them in the lightweight class of generative AI models, compared to behemoths like Anthropic’s Constitutional AI topping out at 20 billion parameters.

However, thanks to efficiency gains in natural language AI over the past year, Google contends compact models like Gemma now unlock state-of-the-art performance across text generation and comprehension capabilities:

• Summarization
• Sentiment Analysis
• Search Relevance
• Dialog Systems
• Content Moderation

Tris Warkentin, Google DeepMind’s product management director, suggests that the moderate model sizes open exciting new avenues for developers. Specifically, inference and fine-tuning can run locally on basic hardware like RTX desktop GPUs or individual Cloud TPU processing cores. This enables AI augmentation of apps and workflows without reliance on based large batch parallel processing.

Warkentin said, “The generation quality has gone significantly up in the last year. Things that previously would have been the remit of extremely large models are now possible with state-of-the-art smaller models.”

Seamless Access & Integration Across Google’s AI Stack

To further reduce barriers to leveraging Gemma’s generative power, Google has built turnkey support into its existing cloud services and developer frameworks spanning:

● Colab Notebook Environment
● Kaggle Machine Learning Platform
● Hugging Face Model Hub
● Max Compute AI Engine
● Nvidia NeMo Toolkit

This means developers can simply import Gemma for inference and fine-tuning without extensive configuration. It also enables broad deployment flexibility from desktops to cloud backends.

Responsible AI Development Kit & Model Debugger

As with any leading-edge generative models, there exists potential for misuse or unintended bias. In conjunction with releasing Gemma access, Google also published two tools to encourage accountability:

● Responsible Generative AI Toolkit – Provides guidance and tools for auditing model outputs and mitigating risks like privacy violations.
● What-If Tool – Enables interactive probing, testing and debugging of AI model behavior to quickly identify and correct problems.

Both resources reflect Google’s expanded investment in responsible AI principles and development tooling to promote trust and safety.

But Questions Around Performance & Openness Remain

Reactions among insiders to the Gemma launch are mixed. While broader access to Google’s generative AI capabilities furthers ecosystem innovation, lack of comparative test data makes it difficult to evaluate strengths against similar models.

And Google avoiding full open-sourcing aligns with its cloud services business model but could limit creative exploration versus AI systems like Anthropic’s Claude which permits redistribution and modifications.

As Nenad Markus, founder of AI startup blink.bi stated following the news, “It’s great that Google keeps pushing state-of-the-art AI innovations and opening access. However both researchers and developers would really benefit from more performance insights relative to tools emerging from openAI, Anthropic and others. That data is critical in discerning the right platform for any given problem.”

As models and supporting infrastructure continue evolving at a breakneck pace, comparative testing will help guide adoption. But responsibility also dictates evaluating AI not just by metrics alone but also based on principles of ethics, transparency and trust aligned with use case impact. Google’s latest toolbox investments reflect this expanding scope for AI success.

---

If you need assistance understanding how to leverage Generative AI in your marketing, advertising, or public relations campaigns, contact us today. Custom training workshops are available. Or, schedule a session for a comprehensive AI Transformation strategic roadmap to ensure your marketing team utilizes the right GAI tech stack for your needs.

Read more: Google Launches Gemma, A New Family of Open AI Models