There was news earlierRefers toAfter Google may delay the unveiling of its large-scale natural language model "Gemini", apparently in order to clarify market rumors and compete with competitors, Google announced that it will provide a small model called Gemini Nano, which can run on mobile phones, to Android platform developers starting December 12. It is expected that on December 6, a larger model called Gemini Pro will be available to developers and enterprises through Google Cloud's Vertex AI and AI Studio. The largest model, Gemini Ultra, will be provided to early registered developers and enterprises, with the goal of opening it to the public in early 12.
Gemini is available in three sizes and can be used offline on mobile devices.
Google explained that Gemini boasts higher computational accuracy and content description capabilities, achieving exceptional performance. The model was developed through extensive early testing, emphasizing its responsible approach to building the model. Google also stated that Gemini became the first large-scale natural language model to achieve a score exceeding 90% on the MMLU (Massive Multi-Task Natural Language Understanding) benchmark, highlighting its semantic understanding accuracy.
In addition, "Gemini" can be divided into Gemini Nano, which can be used on mobile devices and supports offline operation, and is optimized for Google Tensor processors, according to different deployment needs. It also includes Gemini Pro, which strikes a balance between execution accuracy and operating speed, and Gemini Ultra with complete functions and execution performance.
Maintain consistent performance across different languages and natively support multi-mode operation
In further explanation, Google stated that Gemini supports not only English but also multiple languages, maintaining consistent performance across different languages. Gemini training is accelerated by Google's TPU 2022, which was released in 4, and the TPU v2023e, which was released in 5. Google also simultaneously unveiled the new, more powerful TPU v5p, emphasizing that it can accelerate the development of artificial intelligence technology without relying on GPU products from NVIDIA and AMD.
From an overall architecture perspective, developers and businesses can use their own AI software frameworks and datasets at the top layer, and use Vertex AI and AI Studio tool resources and development-related API content in sequence. Gemini Nano, Gemini Pro, or Gemini Ultra can be selected as the bottom layer, and the bottom layer is the AI optimization infrastructure.
As for the actual demonstration by Google, "Gemini" can describe the user's drawing of lines and forming images in almost real-time form, and can change its judgment and description content according to different colors and objects, and make relevant responses, even describing them in different languages.
Gemini also boasts extensive computer vision analysis capabilities, such as playing games like guessing the country's location based on tourist attractions or playing coin tricks. It can also distinguish between cookies and oranges and recommend knitting different dolls for different yarns. Other capabilities include the ability to interpret and write more accurate code using AlphaCode 2, and to integrate more technologies and data to meet diverse computing needs.
Google emphasized that "Gemini" can provide answers with higher accuracy and faster response speed, or conduct deeper interactions, with more accurate semantic understanding and visual judgment capabilities, native support for multi-modal operations, and the ability to connect with developers and enterprises to import data sets.
Gemini will be integrated into Google products such as the AI service Bard, with the Pixel 8 Pro being the first to introduce it.
In addition to announcing the upcoming launch of "Gemini", Google also confirmed that "Gemini" will be integrated into Google products such as the artificial intelligence service "Bard" in the form of Gemini Pro. However, it will only support English interactions in the initial stage. More languages will be supported in the future. Subsequently, the integration of Gemini Ultra and the more advanced "Bard Advanced" will be further launched. However, whether it will be provided in a separate billing form is not yet certain.
The Gemini Nano, optimized for Google's Tensor processor, will be first used in the Pixel 8 Pro and will subsequently be compatible with more Android phones.
However, Google did not disclose the specific parameter scale of "Gemini" in this statement, nor did it explain whether the "Gemini" operating mode will be adjusted for different processors and operating system platforms (such as Qualcomm, MediaTek processors, or iPhone models).
As for whether future applications of "Gemini" technology will be similar to OpenAI's software marketplace service that includes various GPT services, Bard and Google Assistant product manager Sissie Hsiao revealed that this may be a future planning direction, but no details were disclosed at this time.
