In December, Google introduced its next major language model, Gemini, and now its most advanced version, Ultra, has been announced to the public. According to the company, this is the first program to outperform human experts on the MMLU test, which includes questions in mathematics, physics, history and ethics.
Competition between large language models, such as LLMs, continues, and while OpenAI, which developed ChatGPT, used to raise the bar higher and higher, it now appears that Google has taken it a step further. As we reported at the end of last year, the most interesting feature of Google's new model, Gemini, is that it was designed from the ground up to be multimedia, meaning it can handle not only text, but also videos, images, and audio recordings. And program codes, and can provide passage between them. The search giant in one of its promotional videos Extremely exaggerated In terms of system capabilities, but regardless, the previous Pro version of Gemini was also a noticeable step forward.
But now the big brother, Ultra, has arrived, and he is better than the human experts in a complex test.
the According to Google's statement They passed the first MMLU (Massive Multi-Task Language Understanding) test. On his website In fact it says the Gemini Ultra performed better. This comparison test, i.e. standardized test, in addition to natural science questions, also tests models in the social sciences and other fields, from simple problems to professional level.
But this is not the real level. From the words of Google CEO, Sundar Pichai, it is clear that after cloud technologies, artificial intelligence (AI) will be the new guiding thread in the company’s products. They delivered what they promised at their developer conference last spring. Let's start by changing the previous name of their chatbot, Bardot, to Gemini. An AI app with the same name is being created on Android; For some reason, these capabilities carry over to the regular Google app on iOS. Duet AI of Workspace, their office software, is also called Gemini, as is Duet used in cloud software. Those who want to try out Ultra capabilities, such as taking photos, fixing software code or preparing for a job interview, will have to subscribe to Gemini Professional, which will be available in the AI-enhanced version of the premium Google One package for about twenty dollars a month. All this in forty languages, in one hundred and fifty countries, according to what appears to be a mature strategy.
Of course, competitors aren't resting on their laurels either: OpenAI is focusing on “agents,” that is, software that runs, clicks, and swipes on our behalf, and Apple is abandoning not just one AI-based development, but two and a half dozen AI companies. acquired last year. It will become clear in the coming weeks what exactly Gemini Ultra is good for, but it's already clear that it will further accelerate competition in this area.
(Cover image: Rafael Henrique/SOPA Images/LightRocket via Getty Images)
Google's chatbot can now generate images and works better in Hungarian
Thanks to the large language model Gemini, Bard has improved greatly, for example, in summarizing longer texts, and in the English-speaking region, it can also generate images that correspond to the input text.