Google shows off Gemini, its latest bid to keep pace with OpenAI
Add Axios as your preferred source to
see more of our stories on Google.
A row of Cloud TPU v5p AI accelerator supercomputers in a Google data center. Image: Google
Google is announcing details around Gemini, the next version of its large language model that will power Bard and other products.
Why it matters: Google is eager to show it can keep up with rivals, especially OpenAI and Microsoft.
Driving the news: Gemini will come in three flavors: an Ultra version for the most demanding tasks, a Pro version it says is suited to a wide range of tasks and a Nano version that can run directly on mobile devices.
- It's designed to be multimodal, recognizing video, images, text and voice at the same time. However, for now Gemini will only return results in either text or code.
- Google is rolling out Gemini in stages. Starting today, Bard will use a fine-tuned version of Gemini Pro.
- Gemini Nano will power the Google Pixel 8 Pro's new generative AI features, starting with a summarize feature in the voice recording app and a smart reply option in Gboard.
Beginning December 13, developers and enterprise customers will get access to Gemini Pro through either the Gemini API in Google AI Studio or Google Cloud Vertex AI.
- Google also announced it is using a new, higher performance version of its Tensor chips — the Cloud TPU v5p — to train and serve its AI products.
Yes, but: Gemini Ultra is not being made available while the company completes additional safety and testing work.
- "As part of this process, we'll make Gemini Ultra available to select customers, developers, partners and safety and responsibility experts for early experimentation and feedback before rolling it out to developers and enterprise customers early next year," Google said.
- It also plans a higher-end version of Bard, Bard Advanced, that will offer access to Google's "best models and capabilities, starting with Gemini Ultra."
- Google declined to discuss whether or how much it plans to charge for Bard Advanced.
Of note: Google showed reporters a video of Gemini recognizing and interacting with drawings and objects and offering suggestions without text prompts.
- In one of the most compelling examples, Gemini was given images of two spools of yarn in different colors and offered suggestions — with pictures — of what the user could knit.
- That demo relied on Gemini Ultra, Google said, while most of the other examples were done using Gemini Pro.
The big picture: Google has been under pressure to detail Gemini with its rivals continuing to make improvements to their models.
- OpenAI announced its GPT-4 Turbo model at its DevDay event in November, while Microsoft announced some improvements to Copilot yesterday.
- Anthropic announced version 2.1 of its Claude chatbot last month.
Between the lines: Google touted the improved performance of Gemini, saying Gemini is the first AI model it knows of to score better than human experts at the massive multitask language understanding (MMLU) test, which incorporates both knowledge and reasoning tasks across 57 subjects including math, physics, history, law, medicine and ethics.
- Google released a technical paper showing Gemini Ultra outperforming GPT-4 on a range of benchmarks.
- Google declined to say how many parameters are used in the Pro and Ultra versions of Gemini, however the technical paper said that Gemini Nano incorporates up to 3.2 billion parameters.
Editor's note: This story has been updated with details from a technical paper published by Google.
