Alphabet Inc’s Google is introducing new open large language models that it’s calling Gemma, reversing its general strategy of keeping the company’s proprietary artificial intelligence technology out of public view.
Gemma, which will handle text only, has been built from the same research and technology used to create the company’s flagship AI model, Gemini, Google said Wednesday in a blog post. It will be released in two sizes, one targeted at customers who plan to develop artificial intelligence software using high-capacity AI chips and data centers, and a smaller model for more cost-efficient app building.
Google developed the technology that underpins generative AI, but has lagged behind OpenAI and the startup’s partner, Microsoft Corp, in creating products that can automate tasks like summarising reports, writing software code or creating ad campaigns. While the release of the more open Gemma models might seem in contrast with the company’s earlier goals in AI, Google said it’s simply a continuation of its history of innovation in the open-source world.
Company leaders pointed toward earlier moves to make its internal AI engine TensorFlow free for developers and introducing so-called transformers to the world – the key building blocks that make up today’s most widely used large language models, including OpenAI’s popular AI chatbot, ChatGPT.
“It’s a new opportunity for us to build something that we can work on with the community to create new opportunities in AI research and development, in partnership with people both inside and outside of Google,” Tris Warkentin, a Google DeepMind director of product management, said in an interview.
Some observers – including a few of Google’s employees – have criticised the Internet search giant for losing its edge in AI to the open-source community. Last year, Google senior software engineer Luke Sernau penned a widely-shared critique of the company, initially published on an internal system, arguing that many independent researchers have used open-source tools to make rapid and unexpected advances in AI. “We have no secret sauce,” Sernau wrote at the time. “Our best hope is to learn from and collaborate with what others are doing outside Google.”
In the months since Sernau’s missive went viral, Google’s focus on proprietary AI has given way to a broader approach – including moving deeper into the world of open-source tools. In January, the company forged a deal to host AI software from startup Hugging Face, which makes AI software and provides a platform for other companies, on its cloud computing network, giving open-source developers greater access to the technology. Google has also worked with the startup to have the Gemma models fully supported with comprehensive integration in Hugging Face at launch.
Gemma’s release will supplement Google’s flagship AI model Gemini, as the company solicited feedback from developers and “found that there are certain sets of developers that are going to not only use APIs, but they’re going to use open models as well,” said Jeanine Banks, Google’s head of developer relations. Banks said it’s common for developers to get started on building a piece of software using APIs like Gemini’s, which allow them to easily prototype, iterate and test their ideas out. But open models are useful for when a developer might want to build on their apps using their own data, or see how their software performs using their hardware or infrastructure. “It’s really this cross-platform story that we think Gemma’s open model family really supports,” Banks said.
The Gemma models integrate with Google Cloud’s enterprise platform, Vertex AI, allowing developers to stay within Google’s ecosystem for every phase of building their AI tools. Meanwhile, the apps that developers create using the model will be “commercially permissive,” according to Banks, meaning developers will take ownership of the software products and be able to apply business models that make sense to them.
Open-source AI models, however, carry a higher risk of abuse by those who may want to train the model on unsavory content and use it to generate explicit or hateful output. In order to mitigate this risk, Google said it developed Gemma in line with the company’s responsible AI principles.
When developers accept the company’s license terms for Gemma, they agree to abide by Google’s expectations for how the AI model should be utilised, and must use platforms like Kaggle and Hugging Face that require people to develop software in the open, Google’s Banks said. “While it’s important for AI to be widely available for good, there’s also potential for harm,” she said. – Bloomberg