Google launches Gemini – AI model 'more powerful than GPT-4'

Gemini launched on the evening of December 6, is Google’s most advanced and general AI model to date, competing with OpenAI’s GPT-4.

Unlike other popular large language models in recent times, Gemini is built in a multimodal direction, meaning it can generalize, operate and combine on many different types of information including text, code, and sounds. audio, images and video.

To meet flexible usage needs, from data centers to mobile devices, Google said Gemini 1.0 will be offered in three different sizes, including: Gemini Ultra, Gemini Pro and Gemini Nano . Of these, Gemini Ultra is the largest size and most powerful model.

According to test results published by Google, Gemini Ultra scored 90% in the Massive Multitask Language Understanding (MMLU – Massive Multitask Language Understanding) test. This model uses a combination of 57 subjects such as math, physics, history, law, medicine and ethics to test both world knowledge and problem-solving ability, and can “use the ability to my ability to think more carefully before answering difficult questions.”

With this result, Gemini is the first AI to surpass humans at the expert level, which scored 89.8% on the same test. The result of GPT-4 was 87%, LLAMA-2 reached 68% and Anthropic’s Claude 2 reached 78.5%.

In addition, this strongest version of Gemini also exceeded 30 out of 32 standards in large language model research and development, scoring 59.4% in MMMU (massive multimodal understanding on multimodal understanding) capabilities. industry), including multimodal tasks spanning different domains that require intentional reasoning.

Demis Hassabis, CEO of Google DeepMind, representative of Gemini Team, said the company wants to build a new generation of AI models inspired by the human way of recognizing and interacting with the world. Thanks to that, AI will not just stop as smart software, but can become more useful and intuitive, similar to a partner for users.

“Today, we’re one step closer to this vision by introducing Gemini, the most advanced and general AI model ever developed by Google,” Hassabis said.

In addition to strong performance, Google says Gemini 1.0 is trained to recognize text, images, sounds and more at the same time, helping it better understand nuanced information and respond. Questions related to complex topics. This model can also be interpreted and coded in today’s popular programming languages such as Python, Java, C++ and Golang.

According to Google, these features help Gemini read and extract information from hundreds of thousands of documents, thereby opening up the possibility of creating new breakthroughs in many fields, from science to finance in a short time. .

During the launch, Google said that Gemini Ultra version is the version for the most complex tasks and is in the process of completing safety testing before officially launching . Gemini Nano is a version for tasks performed on mobile devices, which will be equipped on Pixel 8 Pro. At that time, the phone will have some additional capabilities such as summarizing recording content and smart replies on the Gboard keyboard. These two Gemini versions will hit the market next year.

Meanwhile, the Pro version is now used in the Bard chatbot. Users can experience changes through a number of reading comprehension, summarizing, reasoning, programming, and planning requirements.

This is also the biggest upgrade for Bard since its launch. However, Bard uses Gemini Pro which currently only supports English and can be used in 180 countries and territories. Google said it will expand languages for Bard in the near future. Next year, Bard will be upgraded and use the most powerful version of Gemini Ultra.

Luu Quy