Faced with GPT, the OpenAI chatbot which has been in the spotlight for more than a year, Google has until now been rather discreet, contenting itself with offering a conversational agent, Bard. Quite anecdotal compared to the competition (ChatGPT, Microsoft Copilot, etc.), Google’s AI has just undergone a significant evolution thanks to the introduction of a new model: Gemini.
What does Google offer with its Gemini AI model?
In a blog post published Wednesday December 6, 2023, Google lifts the veil on Gemini, its artificial intelligence model “the largest and most efficient”.
Developed by Google’s DeepMind division, the first version of this AI model is intended to be multimodal from the outset: “Which means it can generalize and seamlessly understand, leverage and combine different types of information, including text, code, audio, image and video”, specifies the firm.
With Gemini 1.0, Google is pulling out all the stops by announcing a flexible model that can run on both data centers and mobile devices. Three variations are also offered:
- Gemini Ultra – best performing model for very complex tasks;
- Gemini Pro – best model for scaling a wide range of tasks (now available on Google Bard);
- Gemini Nano – most efficient model for tasks on-deviceor those directly embedded in the devices (now available on Pixel 8 Pro).
Regarding the fields of action of this new artificial intelligence model, Google explains that “Gemini 1.0’s sophisticated multimodal reasoning capabilities can help understand complex written and visual information”.
Text, image and audio can thus be taken into account simultaneously in order to offer a detailed and nuanced response to complex subjects, such as physics or mathematics. In addition to this, Gemini can generate and interpret code (Python, Java, C++, Go…).
Regarding the performance of Gemini, Google announces that the Ultra version obtained a score of 90% in the MMLU (Massive Multitask Language Understanding) tests. As a reminder, these allow AI models to be evaluated on 57 different themes. The American company thus declares that its AI model is “the first to surpass human experts”.
Via two comparative tables, Google shows that its model does better than OpenAI’s GPT-4 in almost all areas, whether in text interpretation or multimodal.
Already available on Google Bard and the Pixel 8 Pro, Gemini is preparing to arrive on other Google services in the coming months: Chrome, Search Ads and Duet AI. For developers, Google said that a selection of customers and partners could test the model in its Ultra version from December 13 via Google AI Studio or Google Cloud Vertex AI.