AI in practice

How Google's Gemini AI model got its name

Matthias Bastian
Infographic on Gemini Multimodality

Google

The name Gemini is closely associated with a major shake-up at Google.

In April 2023, Alphabet announced the merger of its two AI units, Google Brain and Deepmind. The resulting Google Deepmind was to focus on developing large multimodal AI models. It was a big move that showed how much pressure Google was under due to the massive success of ChatGPT.

Gemini is Google and Deepmind

Jeff Dean, head of Google Brain until the merger with Deepmind, became the new merger's chief scientist, with a direct line to Alphabet CEO Sundar Pichai. Dean now explains that the name Gemini, Latin for "twin," is directly related to the merger.

"The Gemini effort came about because we had different teams working on language modeling, and we knew we wanted to start to work together," writes Dean.

The twins were the Brain team, which worked on LLMs such as PaLM, and the Deepmind team, which worked on the Chinchilla project.

Together, they tackled a multimodal project that Google unveiled a few days ago under the name Gemini, and which Dean says eventually attracted "many people from all across Google."

The future of Gemini

Although Google has already unveiled Ultra, the most capable model of Gemini, it will not be released until early next year. In many benchmarks, it is just slightly ahead of OpenAI's GPT-4.

However, Google is still playing catch-up with Gemini and has not made any headway against OpenAI and its partner Microsoft. With a similar technology, OpenAI has long since entered the everyday lives of many people and developers.

And what Google has shown of Gemini so far may not be the final vision or version. Demis Hassabis, CEO of Google Deepmind, said over the summer that Gemini would combine AlphaGo's techniques like reinforcement learning with traditional methods like tree search to provide new capabilities like advanced problem-solving, memory, and planning. These capabilities were not discussed in detail at the Gemini presentation.

With Pathways, Google is also working on an AI architecture that aims to generalize better and be more efficient than current systems. Alphabet CEO Sundar Pichai has announced a series of "next-generation" models for 2024 - so perhaps Gemini Ultra is just the beginning of Google's new AI developments.

There is no alternative to Google moving fast anyway: according to rumors, OpenAI could respond to Gemini with GPT-4.5 in December, and Claude chatbot developer Anthropic is set to unveil a multimodal model. The information comes from OpenAI leaker "Jimmy Apples", who has been right several times this year.

Sources: