On December 7th, Google announced at the press conference that it had launched the largest and most powerful artificial intelligence model Gemini, Google said that it is still one of the greatest companies in the world, saying that Gemini performed better than OpenAI's GPT-4 model in a series of tests, especially in multimodal **, voice tasks, Gemini test performance is better than that of professional humans in various fields!
Yesterday, Google CEO Sundar Pichai and Head of R&D Demis Hassabis, as representatives of Google's large model team, officially launched the large model Gemini!
The name Hassabis is very familiar, it is the CEO of Deepmind, who previously led the team to develop AlphaGo, defeated human chess players Lee Sedol and Ke Jie, and further pushed the deep learning Xi represented by neural networks to a climax!
For a long time, Google has been regarded as a global leader in technological innovation, but since Microsoft released the GPT model, especially the birth of ChatGPT, Google has been caught off guard, and DeepMind and Google Brain have now completed the integration.
Less than two weeks after the release of ChatGPT last year, Google hurriedly took out Bard, but it was wrong in the demonstration, causing Google's stock price to evaporate more than $100 billion overnight, and then Google also injected capital into Anthropic and launched Claude 2 to deal with ChatGPT.
Within Google, Gemini has always been expected to surpass ChatGPT, and Eli Collins of Google's "Deep Thinking" said that Gemini is the company's largest and most capable model, but it is also the most versatile multimodal model.
Gemini can be used to process multiple forms of information such as **, audio and text, as you can see in the demonstration, when a human draws a duck, Gemini quickly recognizes it, and after adding wavy lines, Gemini can also understand that the duck is swimming in real time.
Hassabis launches Gemini 10, which is divided into three versions with different parameters, namely Gemini Nano, Pro and Ultra, the smallest of which is the Nano, reminiscent of Apple's discontinued iPod line, this version is specifically designed for mobile and can run natively on smartphones.
And the Pro version has been able to beat OpenAI's GPT35. Ultra is the most powerful multi-modal model today, benchmarking GPT-4, which can crush the existing AI large model in all aspects, surpassing 90% of human experts in MMLU (Large-scale Multi-task Language Understanding), and is also the large model with the highest accuracy rate!
Hassabis emphasized that Gemini Ultra is superior to GPT-4 mainly because of the understanding and interaction ability of ** and audio, OpenAI adopts the method of GPT+DALL·E+Whisper to build multimodality, while Gemini focuses on multi-modal mixing from the beginning, and it is expected that Gemini Nano will be launched in Pixel 8 Pro, Gemini Pro will open the Gemini API interface to enterprise users and developers on December 13th.