Gemini
Google’s announced its most anticipated large language model, Gemini, on Wednesday. During a virtual press conference, the leaders from Google DeepMind and Google Cloud showcased the capabilities of the new LLM while also sharing the roadmap.
Initially announced at Google I/O 2023 in the spring by CEO Sundar Pichai, Gemini is the most capable LLM that succeeds PaLM 2, the current foundation model from Google. It is one of the first language models to score over 90% on the Massive Multitask Language Understanding benchmark, which covers 57 subjects across STEM, the humanities, the social sciences and more.
Gemini is one of the most adaptable models, capable of running on both the cloud and mobile devices. It is multimodal in the sense that it can understand input in the form of a text prompt, an image, audio or even video. It has unrivaled reasoning capabilities and is capable of producing sophisticated programs and code. For instance, Gemini can look for errors in a student's written response to a challenging mathematical problem.
Google trained Gemini on its state-of-the-art AI accelerator chip, the Cloud TPU v5e, purpose-built to bring the cost-efficiency and performance required for medium- and large-scale training and inference. It has built-in support for leading AI frameworks such as JAX, PyTorch and TensorFlow, along with popular open-source tools like Hugging Face’s Transformers and Accelerate, PyTorch Lightning and Ray.
Google has also announced the next generation of its AI accelerator, Cloud TPU v5p, which is four times more scalable than its predecessor and 2.8 times faster to train foundation models and LLMs. This infrastructure will be available on the Google Cloud Vertex AI platform for machine learning researchers and developers to train, fine-tune and deploy generative AI models.
While Google didn’t disclose the size of the model, the training dataset or the massive number of parameters it is trained on, Gemini comes in three flavors: Ultra, Pro and Nano.
The Gemini Ultra model is the largest and most sophisticated and will become available early next year. Gemini Pro is already integrated with the Google Cloud machine learning platform, Vertex AI, which will be available to developers on December 13. Gemini Nano, the smaller version of the LLM that can run on Pixel 8 Pro and other Android devices, will be available on December 6 for developers. Gemini will initially support English, while other languages are in the pipeline.
Bard, Google’s answer to ChatGPT, will soon be powered by Gemini Pro. Combined with the ability to use Google Search, which is already available in Bard, the chatbot will be able to provide accurate and reasonable responses. Google is launching Bard Advanced, the next generation of its chatbot, backed by Gemini Ultra. This will be made available to early adopters early next year through the Trusted Tested program.
Gemini marks another important milestone in the evolution of generative AI. Given its remarkable powers, Gemini will undoubtedly serve as the foundation for many other Google services and products, such as Bard, Duet AI and Google Cloud.
