Within BIGDATA AI Bulletin Board for News Briefs

Introducing the inside BIGDATA AI News Briefs Bulletin Board, a timely new service that offers you the most recent viewpoints and industry insights on all things related to AI, including generative AI, deep learning, huge language models, and transformers. We’re working nonstop to unearth the most intriguing and topical nuggets beneath the most widely used technology of the day. Since this sector is developing quickly, we want to provide you with a frequent resource that will keep you up to date and informed. The news snippets are uploaded all the time, with the most current ones at the top and the dates reversed. You can use our bulletin board to frequently check in on developments in our industry, which is growing at a rapid pace.

[December 8, 2023] Google recently released Gemini 1.0, their biggest and most powerful AI model. The Cloud TPU v4 and v5e, the company’s specially created AI accelerators, were used to train Gemini. It’s the first model of the Gemini period, designed from the ground up to be multimodal. Ultra, Pro, and Nano are the three sizes in which Gemini is optimizer. In benchmark testing, 30 out of 32 tests show that Gemini performs better than OpenAI’s GPT-4, especially in multimodal understanding and Python code generation.

Every model is intended for a certain use. Using all of Google’s AI capability, Gemini Ultra can handle challenging jobs in data center’s and enterprise applications. Gemini Pro works well with Google’s own AI service, Bard, and offers a broader range of AI services. Finally, there are two iterations of Gemini Nano: Nano-1, which has 1.8 billion parameters, and Nano-2, which has 3.25 billion. With an emphasis on speed optimization’s in Android contexts, these models are specifically designed for on-device activities. Gemini employs Alpha Code 2, a code-generating system, to demonstrate its ability to comprehend and produce high-quality code in multiple languages.

The core of the Gemini models is an architecture based on improved Transformer decoders designed especially for Tensor Processing Units (TPUs) from Google. The models are faster and more economical than earlier iterations like PaLM thanks to this hardware-software coupling that allows for better training and inference procedures.

The multimodal nature of the Gemini suite, which was trained on a wide range of datasets including text, graphics, audio, and code, is one of its main features. According to reports, Geminis outperform OpenAI’s GPT-4 in a number of performance metrics, particularly multimodal comprehension and Python code production. The more sophisticated Gemini Ultra model, which is anticipated for release next year, is lighter than the recently introduced Gemini Pro model. Now powering Google’s rival ChatGPT, Bard, Gemini Pro claims enhanced comprehension and reasoning skills.

It is claimed that Gemini Ultra is “natively multimodal,” capable of handling a wide variety of data types, such as text, photos, audio, and video. In visual problem domains, this capability outperforms OpenAI’s GPT-4; yet, the gains are negligible in many ways. For instance, Gemini Ultra barely beats GPT-4 in a few benchmarks.

Google’s lack of transparency regarding the model’s training data is a worrying feature of Gemini. The origins of the data and the rights of the creators were not addressed. This is crucial since the AI sector is increasingly being sued for allegedly utilizing copyrighted material without giving credit or payment.

After making its major premiere on December 6, 2023, Gemini is receiving mixed reviews. However, people may lose faith in the company’s multimodal technology and/or honesty after learning that the most spectacular demo of the product was essentially staged. It was Parmy Olson of Bloomberg who first brought attention to the disparity. TechCrunch does a fantastic job of enumerating the problems in the following video.