Google launches Gemini AI: it is the "world's best" artificial intelligence that will make GPT-4 lag behind
While the debate continues on whether massive experiments with artificial intelligence should be continued and whether they pose a threat to humanity, Google has made its move. It has launched a new AI model called Gemini. According to the developers, it can outperform both the OpenAI GPT-4 model and "expert-level" humans in a number of intelligence tests.
According to the New Scientist, the CEO of Alphabet Inc. and its subsidiary Google, Sundar Pichai, announced the existence of Gemini in May. At that time, the model was in the training stage. Now it has been presented to the public.
Gemini is currently available in three versions: Nano, Pro, and Ultra, which differ in size and capabilities. The smallest of the three, Nano, is designed to run locally on smartphones. It has two versions. One will run on slower phones and has 1.8 billion parameters, and the other runs on more powerful devices and has 3.25 billion parameters. Google refused to disclose the number of Pro and Ultra parameters, their size, as well as the scale or source of the training data. For comparison, it is rumored that GPT-4 includes up to 1.7 trillion parameters, and Meta's LLAMA-2 has 70 billion.
According to Google, the mid-range version of Gemini, called Pro, is superior to some other models, such as OpenAI GPT3.5. However, the "oldest" in the Ultra line surpasses the capabilities of all existing AI models. It managed to beat a human in the standard MMLU test for the first time, scoring 90%. It is known that an "expert level" human is expected to score 89.8% in this test. The test includes a wide range of challenging questions on topics including logical fallacies, moral issues in everyday situations, medical issues, economics, and geography.
In the same test, the GPT-4 scored 87%, the LLAMA-2 scored 68%, and the Claude 2 from Anthropic scored 78.5%. Gemini outperformed all of these models in eight of nine other common performance tests.
The company reported that the Pro model will be integrated into Google's Bard, an online chatbot launched in March this year. And in early 2024, Bard Advanced will be released, which will include an even larger model - Gemini Ultra.
As of today, the new version of Bard will be available in English in more than 170 countries. However, it will not be available in the UK and Europe. According to Google's Cissy Xiao, the delay is related to regulation, not development: "We're working with local politicians and regulators to make sure we comply with local laws and things like that before we launch in other regions," she said.
In turn, Ali Collins from Google DeepMind called Gemini the company's most powerful model and the most versatile at the same time. According to the developers, it can be adapted to different tasks. It is trained to perceive both text and sound and provide output in different formats. However, at this stage, the launch of Bard will only allow the use of text prompts. It is expected that sound interaction and work with images will become available in the coming months.
At the presentation, no Gemini version was presented for testing, but Google demonstrated a recording of their AI helping with homework and working with real-time video input. The company also assured that the new model is better at software development than previous versions. It is known that last year Google's DeepMind division released a code generator based on artificial intelligence called AlphaCode. The developers claim that it can outperform half of human developers. The updated version based on Gemini, according to the company, can outperform 85% of human programmers in coding.