Sundar Pichai, chief executive officer of Alphabet, during the Google I/O Developers Conference in Mountain View, California, U.S., on May 10, 2023. Credit - David Paul Morris—Bloomberg via Getty Images
According to Google, Gemini Ultra, its largest and most capable new model, outperforms OpenAI’s most capable model, GPT-4, at a number of text-based, image-based, coding, and reasoning tasks. Gemini Ultra will be available through a new AI chat feature called Bard Advanced from early next year, the company said. It is currently being refined and is undergoing “trust and safety checks, including red-teaming by trusted external parties,” according to the announcement.
Google DeepMind also announced the launch of Gemini Pro, which is now available to the public through Google’s Bard chat interface, and the smaller Gemini Nano, which will run on Google’s Pixel 8 Pro smartphone. All three models can process text, images, audio, and video and produce text and image outputs.
Google will start to integrate Gemini models into its other products and services, such as internet search and advertisements. From Dec. 13, developers will be able to access Gemini Pro through an API, and Android developers will be able to build with Gemini Nano.
The rollout will put the Gemini suite up against rivals including OpenAI, Anthropic, Inflection, Meta and Elon Musk’s xAI.
DeepMind was founded by Demis Hassabis, Shane Legg, and Mustafa Suleyman in 2010. Google acquired the AI lab for $400 million in 2014, and in April 2023 DeepMind merged with Google’s elite AI research team, Google Brain, to form Google DeepMind, which Hassabis leads.
More From TIME
A year after the acquisition, DeepMind’s founders began negotiating to try and obtain greater independence from their new parent company. In 2017, DeepMind’s founders reportedly tried to break away from Google but failed. In 2020, the founders reportedly pushed for a new legal structure to ensure that powerful AI wasn’t controlled by a single corporate entity, even hiring an outside lawyer to help draft the structure, but according to the Wall Street Journal, the proposed structure didn’t make financial sense for Alphabet.
Google and Google DeepMind have been responsible for a number of the most important AI breakthroughs of the last decade, including AlphaGo, which mastered the complex game Go, inventing the transformer architecture that powers today's chatbots, and solving the protein folding problem with AlphaFold.
But the tech giant has lagged behind competitors such as OpenAI and Anthropic in the era of AI chatbots. A paper from 2021 suggests that DeepMind developed a chatbot, Gopher, as early as December 2020. Google DeepMind chief operating office Lila Ibrahim told TIME that DeepMind decided not to release Gopher because it often gave factually inaccurate responses—a tendency referred to in the industry as ‘hallucinating.’ Before the DeepMind-Google Brain merger, a DeepMind project codenamed Goodall was working to build a ChatGPT competitor, although this was dropped in order to focus on Gemini, The Information reported in August.
Read more: CEO of the Year 2023: Sam Altman
Google announced its own chatbot, Bard, in February 2023, but parent Alphabet’s stock price dropped after analysts judged it to be inferior to competitors. In May it released PaLM 2, an improvement on Bard but judged by commentators to be inferior to GPT-4.
While Google has been slower to bring a consumer AI product out, it has been on the mind of its biggest competitor. Microsoft’s partnership with OpenAI gives it privileged access to OpenAI’s AI models. After the software behemoth announced that it was incorporating OpenAI’s models into Bing search engine, CEO Satya Nadella told The Verge in an interview that he thought AI could help his company challenge Google’s internet search dominance, and that he expects a reaction from Google. “I want people to know that we made them dance,” he said. (Data from analytics firm StatCounter suggests that Google has retained its hegemony in search, although a Microsoft executive disputed this in an interview with the Wall Street Journal in August).
Google DeepMind said in the announcement that it had compared Gemini Ultra to a range of competitor models—OpenAI’s GPT-4, Anthropic’s Claude 2, Inflection’s Inflection-2, Meta’s Llama 2 and xAI’s Grok 1—finding that the large language model outperforms those rivals on tests including professional and academic multiple choice questions and Python coding.
Write to Will Henshall at email@example.com.