DUBAI: Google announced the launch on Wednesday of its multimodal AI model, Gemini, which will power its chatbot Bard.
Gemini has been optimized for three sizes — Ultra, Pro and Nano — “which means it’s able to run on everything from mobile devices to large-scale data centers,” said Eli Collins, vice president of product at Google DeepMind, during a press briefing.
The most advanced version, Ultra, outperforms on 30 of the 32 academic benchmarks used in large language model research and development, Collins said.
He explained that Gemini was designed to be “natively multimodal” unlike some AI models, which means that it was trained on different formats from the beginning, enabling it to “understand nuanced information (such as) text, images, audio and code,” and “answer questions relating to complicated topics and reason in math and physics.”
“With Gemini, Bard is getting its biggest and best upgrade yet,” said El-Sisie Hsiao, vice president and general manager of Bard and Assistant.
“A specifically tuned version of Gemini Pro” that has “more advanced reasoning, planning, understanding and other capabilities” is now integrated into Google’s chatbot Bard, she said.
Google will integrate the AI model into other Google products such as search and adverts in the future, and next year launch Bard Advanced, “which is our largest and most capable model, and it’s designed for highly complex tasks,” Hsiao said.
Google is strengthening its foothold in the field of AI nearly eight years into its journey as an “AI-first company,” wrote Google and Alphabet CEO Sundar Pichai in a blog post.
He wrote: “We’re approaching this work boldly and responsibly. That means being ambitious in our research and pursuing the capabilities that will bring enormous benefits to people and society, while building in safeguards and working collaboratively with governments and experts to address risks as AI becomes more capable.”
Gemini has a score of 90 percent on the MMLU (massive multitask language understanding) test and is the first model to outperform human experts (89.8 percent), as well as GPT-4 (86.4 percent) in various tasks across 57 subjects including maths, physics, history, law, medicine and ethics.
Gemini Nano is currently available to developers, while Gemini Pro will be available to enterprise and Vertex AI customers as well as developers in AI Studio from Dec. 13. Gemini Ultra will be rolled out in 2024.