Google unveiled its newest and most powerful AI model, Gemini, with three sizes. Executives remarked on Tuesday that Gemini Pro excelled OpenAI's GPT-3.5 but declined to make comparisons to GPT-4. Google plans to permit customers access to Gemini through Google Cloud to employ in their own programs. Moreover, this AI can also be used to power Google consumer-facing applications, such as the Bard chatbot and the Search Generative Experience.
On Wednesday, Google unveiled its largest and most sophisticated artificial intelligence model, Gemini, as its efforts to monetize AI continue. The model consists of three sizes: Gemini Ultra, the most powerful; Gemini Pro, which can tackle various tasks; and Gemini Nano, designed for mobile applications. Gemini will be available for customers to license through Google Cloud for their own applications starting Dec. 13. It can be accessed through Google AI Studio or Google Cloud Vertex as well as by Android developers. Additionally, it will power Google products like the Bard chatbot and its forthcoming Search Generative Experience.
Gemini Ultra has been found to surpass human experts in the MMLU (massive multitask language understanding) test, which incorporates 57 different subjects such as mathematics, science, history, medicine and law. It is also able to distinguish subtle nuances and comprehend complex topics.
CEO Sundar Pichai noted in a blog post Wednesday that Gemini was created with large-scale collaborative efforts by teams across Google, including those at Google Research. He pointed out that Gemini was built from the ground up to be able to handle multiple forms of information, including text, code, audio, image and video.
Starting today, Google's chatbot Bard is utilizing Gemini Pro for advanced reasoning, planning, understanding and other capabilities. Early next year, an updated version known as "Bard Advanced" with the addition of Gemini Ultra will be released according to company executives on a call with reporters Tuesday. This is the most significant upgrade to Bard since the launch of its ChatGPT-like chatbot eight months ago and the debut of GPT-3.5 by OpenAI one year ago. When asked how Gemini Pro compares to GPT-4, executives avoided giving an answer.
Addressing the potential for charging for access to "Bard Advanced," Google's general manager for Bard, Sissie Hsiao, said the main focus was on providing a good experience and no monetization details had been decided yet. On being asked by reporters if Gemini has any distinct capabilities compared to other language models, Eli Collins, Vice President of Product at Google DeepMind, replied that he suspected it did but the company was still assessing the novel features of Gemini Ultra.
Initially planned for release earlier, Google postponed the launch of Gemini, reminding people of the company’s rocky start with AI tools earlier this year. When asked about the delay on the press briefing, Collins responded that testing more advanced models takes time, indicating that Gemini has been the most tested Google model thus far with the most detailed safety evaluations. He also mentioned that despite being their largest model, it is also highly cost-effective.
Google announced the release of a technical white paper with more information on the model Wednesday, although the perimeter count details will not be revealed. Last August, CNBC reported that PaLM 2, Google's then-latest AI model, used five times the amount of text data for training in comparison to its predecessor language model.
Also on Wednesday, Google unveiled their next-generation tensor processing unit for training AI models. Called the TPU v5p chip, Salesforce and startup Lightricks have already started using it, claiming it offers better performance at a lower cost than the TPU v4 introduced in 2021. Google, however, did not share any performance information in comparison to the current market leader Nvidia.
At the third-quarter earnings conference call in October, investors put pressure on executives to explain how they will convert AI into actual profits. This August, the company launched a “trial” called Search Generative Experience (SGE) which allows people to experience what using a generative AI-based search engine is like, yet this project has not yet been made publically available. Responding to investors' questions about the timeline for SGE since its first announcement at the Google I/O conference in May, executives gave vague answers on Wednesday's Gemini announcement.
Pichai expressed his enthusiasm for what Gemini can make possible and what lies ahead in his blog post on Wednesday, concluding with: "I'm genuinely excited for what's ahead, and for the opportunities Gemini will unlock for people everywhere."
Comments