Surprise, Google is already launching the Gemini 1.5 and 1.5 Pro versions of its conversational robot, so what’s new?

Google announced on Thursday the launch of Gemini 1.5, the new generation of its AI model, with improved performance. Gemini 1.5 Pro also appears, with solid long context understanding capabilities.

Barely a few days after changing Bard to Google Gemini, the search engine giant unveiled its latest advances in artificial intelligence, the Gemini 1.5 model, on Thursday. By emphasizing the understanding of the long context, it broadens its capacity to process information. Gemini 1.5 Pro, a mid-size model that Google is unveiling in parallel, matches the performance of Gemini 1.0 Ultra, but consumes less computing. Let’s look at all this in more detail.

Spectacular performance for Gemini 1.5

Let’s start by talking about Gemini 1.5. Google’s ambition is not meager: the company wants ” revolutionize its approach to AI », with a new generation which displays improved performance, supported by a Mixture of Experts (MoE) architecture, which makes it possible to increase the precision of the model and its number of parameters.

More precisely, this MoE model divides the neural network into “experts”, which will optimize the efficiency of the model by selectively activating the most relevant expert pathways. This is a pioneering approach at Google, which will therefore allow Gemini 1.5 to learn complex tasks more quickly, while maintaining quality.

The Pro version, comparable to the Ultra 1.0 model, must shine in its understanding of the long context, processing up to 1 million tokens. Efficiency gains enable faster iteration and easier deployment. Concerning Gemini 1.5, we can speak of a major breakthrough and a possibility for developers and companies to extend their scope in the field of AI.

Huge amounts of data processed in one go for Gemini 1.5 Pro

Google is launching its first multimodal model, with preliminary tests around Gemini 1.5 Pro. This one offers a standard pop-up of 128,000 tokens, comparable to the Ultra 1.0 model. We were talking about a pop-up of up to 1 million tokens. This capability allows the model to process 1 hour of video, 11 hours of audio, code bases of 30,000+ lines or 700,000+ words, all in one go. The prospects are frankly promising.

Google also tries to reassure security and ethics in the development of its AI models. Rigorous testing ensures thorough risk assessment, integrating learnings into governance and development processes.

Since version 1.0 Ultra, the company has been committed to improving security, testing Red Teaming techniques, to assess potential hacking risks. Note that Gemini 1.5 Pro is available this Thursday, February 15 in limited preview via AI Studio and Vertex AI, for developers and businesses.