Google has officially launched Gemma, an open-source large language model that claims to surpass Meta's Llama-2 competitors
Tencent Technology News reported on February 21 that according to foreign media reports, Google is launching a new open-source large language model and named it Gemma, which subverts the company's long-standing strategy of keeping proprietary artificial intelligence technology secret.
In a blog post published Wednesday, Google revealed that Gemma focuses on text processing and is built on the same foundation as the research and technology of the company's flagship artificial intelligence model, Gemini. Gemma will be available in two versions, with 7 billion parameters and 2 billion parameters. The former is aimed at customers who plan to develop AI software using high-performance AI chips and data centers, while the latter is more suitable for customers who want to build more cost-effective applications.
While these models are relatively simple to build, they offer advantages in terms of speed and cost. Google claims that Gemma "significantly outperforms larger models" on key benchmarks, including Meta Llama-2, and is "able to run directly on developers' laptops or desktops."
Although Google has developed technology to support generative AI, it has lagged behind OpenAI and its partner Microsoft when it comes to building products that can automate tasks such as summarizing reports, writing software code, or creating ad campaigns. However, the release of Gemma appears to be in stark contrast to Google's early goals in the field of artificial intelligence. In response, Google explains that this is just a continuation of its open-source world innovation journey.
Google's leadership highlighted that the company's earlier moves included opening up its in-house AI engine, TensorFlow, to developers for free and rolling out to the world so-called "transformers," a key component that makes up today's most popular large language models, including OpenAI's popular AI chatbot ChatGPT.
"This is a new opportunity for us to work with the community to co-develop new products and to work with people inside and outside Google to create new opportunities in AI R&D," Tris Warkentin, director of product management at Google's DeepMind, said in an interview. ”
Over the past year, Google's attitude towards open-source AI technology has changed significantly. Last year, a critical article written by Luke Sernau, a senior software engineer at Google, circulated widely internally. He bluntly noted that many independent researchers have made amazing strides in the field of artificial intelligence using open-source tools. In the article, Serno wrote: "We don't have a secret recipe, and our best hope is to learn from and work with companies other than Google. ”
Over time, Google's focus on "proprietary AI" has diminished in favor of a broader approach that includes digging deeper into open-source tools.
In January, Google partnered with startup Hugging Face to host its AI software on Google's cloud computing network, making the technology more accessible to open source developers. This time, Google is also working with Hugging Face to ensure that the Gemma model is fully supported by it at launch.
Jeanine Banks, Google's head of developer relations, explained that the launch of Gemma complements Google's flagship AI model, Gemini. She said that after soliciting developer feedback, the company found that many developers use not only APIs, but also open source models. The open-source model is especially popular when developers want to build applications with their own data, or test software on their own hardware or infrastructure. She further noted, "What we think Gemma's open source model really supports is this cross-platform model. ”
It is worth mentioning that the Gemma model is integrated with Google Cloud's enterprise platform, Vertex AI, allowing developers to stay within Google's ecosystem throughout the AI tool building process. In addition, applications developed by developers using the model will be "commercially licensed," which means that developers will own the software product and be able to apply a business model that makes sense to them.
However, open-source AI models are also at risk of being misused. To mitigate this risk, Google emphasized that the development of Gemma is in line with its responsible AI principles. Banks said that when developers accept Gemma's license terms, they also agree to abide by Google's expectations for how AI models can be used, and must use platforms such as Kaggle and Hugging Face that require software to be developed publicly. "While the widespread application of AI is crucial, it also has potential harms," she warned. (Compiled by Golden Deer)