In a groundbreaking move, Google has entered the arena of generative AI with the introduction of Gemini, its latest and most advanced model, boasting state-of-the-art performance across various benchmarks. This innovative AI marvel comes in three iterations, with Gemini 1.0 leading the pack.
Gemini 1.0 offers optimization for different sizes: Ultra, Pro, and Nano, each tailored for specific applications. The Pro version, fine-tuned for advanced reasoning, planning, and understanding, will be utilized by Google's AI chatbot, Bard. Notably, Gemini 1.0 will be accessible in English across more than 170 countries and territories, with plans for expansion into different modalities and support for additional languages and locations in the near future.
The integration of Gemini extends beyond standalone models. Pixel 8 Pro will harness the power of Gemini Nano, introducing new features like Summarize in the Recorder app and Smart Reply in Gboard, initially available on WhatsApp and extending to more messaging apps in the upcoming year.
- Bard will utilize Gemini Pro for advanced functions, marking the most significant upgrade to Bard since its launch.
- Available in English across 170+ countries and territories.
- Future plans include expanding modalities and language support.
- Pixel 8 Pro integrates Gemini Nano for new features in the Recorder app and Smart Reply in Gboard.
- Upcoming availability in Google products and services such as Search, Ads, Chrome, and Duet AI.
- Early experiments in Search show a 40% reduction in latency in the U.S., along with quality improvements.
Google's ambitious plan involves integrating Gemini into various products and services, including Search, Ads, Chrome, and Duet AI in the coming months. This comprehensive approach aims to leverage the model's advanced capabilities across diverse applications.
Gemini's groundbreaking performance is evident in its multimodal design, seamlessly understanding and operating across different types of information, such as text, code, audio, image, and video. Gemini Ultra, the largest model, outperforms human experts on massive multitask language understanding (MMLU) with a score of 90%, covering 57 subjects ranging from math and physics to history, law, medicine, and ethics.
The model's sophisticated reasoning capabilities extend to natural image, audio, and video understanding, as well as mathematical reasoning. Gemini 1.0 can understand, explain, and generate high-quality code in popular programming languages like Python, Java, C++, and Go, making it a versatile tool for various applications, including advanced coding systems.
Google trained Gemini 1.0 at scale on its AI-optimized infrastructure, utilizing Tensor Processing Units (TPUs) v4 and v5e. The announcement also coincides with the introduction of Cloud TPU v5p, the most powerful, efficient, and scalable TPU system to date, designed for training cutting-edge AI models.