Google introduces its ‘most capable’ multimodal AI model Gemini By Elets News Network - 07 December 2023

Gemini

Google unveiled Gemini, their self-described “most capable” artificial intelligence model. For the time being, Google has optimized the original Gemini 1.0 model, which is available in three sizes: Ultra for extremely complicated activities, Pro for scalability across a variety of workloads, and Nano for on-device tasks. These three variations are arranged in decreasing order of size, as the name implies. Because of its adaptability, Gemini can operate effectively on a variety of infrastructures, including mobile devices and data centers.

One of the characteristics of this AI model is that it is multimodal by design, meaning it can operate and comprehend context across and combine several input formats, including text, code, voice, images, and videos. “Gemini was pre-trained on many modalities from the beginning and is inherently multimodal. Then, to increase its efficacy even more, we adjusted it using more multimodal data. This is considerably superior to current multimodal models in helping Gemini comprehend and reason about a wide range of inputs from the ground up, according to the company blog.


Google reports that on 30 of the 32 commonly used academic benchmarks used in large language model (LLM) research and development, Gemini Ultra’s performance was shown to surpass state-of-the-art findings. Additionally, the Ultra model achieved a score of 90% on MMLU, or massive multitask language understanding, making it the first model to surpass humans in terms of problem-solving and domain knowledge. This suggests that Gemini Ultra considers a question “more carefully” before responding to it, as opposed to relying solely on gut feeling, according to Google.

Google’s v4 and v5 tensor processing units (TPUs) were used to train Gemini 1.0. Google also launched Cloud TPU v5p, the “most powerful, efficient, and scalable TPU system to date,” in addition to Gemini. It is anticipated to make large-scale generative AI model training faster for developers and businesses.
Finally, AlphaCode 2, a code-generating tool developed by Google utilizing a customized version of Gemini, has superior competitive problem-solving capabilities compared to its predecessor. It involves more than just coding; it also involves tackling theoretical computer science puzzles and difficult math.

Also Read | Google explores AI Tools for Enhanced Protection


Related News