Tuesday, May 28, 2024

How Google’s AI model Gemini got its name

Must read

Gemini got its first mention at I/O in 2023, when we briefly shared our first natively multimodal model was in training and already showing impressive capabilities. One year later, we’ve brought Gemini models to Search, Ads, Workspace, Pixel and more, including some of the biggest innovations we shared on the I/O stage yesterday.

As Gemini continues to make a name for itself, we wondered, where did that name come from? To find out, we asked the Google DeepMind team to share the origin story behind why they called it “Gemini.”

The story behind the name

Early on, a placeholder title for the project was “Titan,” the name of Saturn’s largest moon. “I wasn’t a huge fan of that name,” says Jeff Dean, Gemini’s co-technical lead. But it gave him an idea — or perhaps a sign — for a name grounded in space.

Gemini is Latin for “twins.” In astronomy, it’s the name of a constellation associated with Greek mythological twins Castor and Pollux, for which its two brightest stars are named. Naturally, then, the meaning behind our AI model’s name is two-fold.

For one, a key characteristic of the Gemini zodiac sign is a dual-natured personality, capable of adapting quickly, connecting to a wide range of people, and seeing things from multiple perspectives — themes well suited for what was happening at Google at the time.

For nearly a decade, DeepMind and the Brain team from Google Research were responsible for some of the world’s biggest research breakthroughs in AI, including deep learning at scale, deep reinforcement learning and AlphaGo, the Transformer architecture that underpins nearly all large language models (LLMs) today and much more. To further accelerate our progress, in April 2023, the teams joined forces to form Google DeepMind, bringing their talent in AI, computing power, infrastructure and resources together all under one team.

The first mission that that crew set out to launch? A series of powerful, multimodal AI models.

“The Gemini effort came about because we wanted to bring our teams working on language modeling closer together,” Jeff says. “I felt the twins aspect of the name ‘Gemini’ was a great fit. The twins here are the folks in the legacy Brain team and the legacy DeepMind team, who started to work together on this ambitious multimodal model project.”

The other inspiration for the name is also space-related: NASA’s early moonshot program, Project Gemini, which lasted from 1965 to 1968.

The pivotal space program was the bridge between the Mercury missions, which determined humans could survive in space, and Apollo, which put the first person on the moon. Project Gemini, so-named for its two-person spacecraft (which was, coincidentally, powered by a Titan rocket), was designed to test equipment and techniques for keeping astronauts in space for extended time ahead of Apollo. It flew 10 crews to space and led to significant feats, such as the first U.S. spacewalk and the first-ever linking of two spacecraft together in Earth orbit.

Project Gemini’s significance to the success of the Apollo program resonated with the team. At one point, Jeff proposed the name “Gemini” in a comment in a Doc the team was working on, and it stuck. “I was immediately sold on the name, because the monumental effort of training LLMs resonated with the spirit of launching rockets,” says Oriol Vinyals, Gemini’s co-technical lead. “It was very fitting to name the most ambitious project we have ever embarked on as Gemini.”

One giant leap for AI models

In many ways, AI has the potential to solve some of the world’s biggest challenges, with breakthroughs as important to humanity as landing on the moon. Our Gemini models are a critical step forward, and represent how we’ll continue reaching for those stars safely and responsibly. “Now the question is, will there be a follow-up to Gemini named Apollo?” Oriol adds, with a smile.

Whatever the future may hold, Google’s Gemini era is just beginning. In December, we started with Gemini 1.0 in three sizes — Ultra, Pro, and Nano. A few months later, we introduced 1.5 Pro, and yesterday, we announced a new 1.5 model, Flash. “Other ideas discussed in relation to how to name models were with names of stars in the universe, as there are quite a few of different sizes, like our models,” Oriol says. “One could say that we are fascinated by the universe in Gemini!”

Since its launch in December, we’ve brought Gemini to billions of people through Google products and hundreds of thousands of developers and businesses have been building with the Gemini API.

“Gemini was born as the research program to create the most capable models in the world,” Oriol says. “By making it the name of the product too, I’m hoping our users feel the continuous improvement, creativity and innovation that our research teams will be bringing directly to them.”

We’re no astrologists, but it’s safe to say Gemini’s future is looking bright.

Latest article