Google presents Gemini 1.5 Flash, what's new: what it can do and how the new AI is trained

Credits: Google.

During the event Google I/O held yesterday in Mountain View, Google announced various interesting news regarding the artificial intelligence sector, opening the curtain on new tools, including Gemini 1.5 Flash, the latest model in the Gemini family that focuses entirely on speed and efficiency. The AI ​​model will be integrated into Android 15 (the next version of Google's mobile operating system) with various additional features. Furthermore, “Big G” announced the availability of Gemini 1.5 Pro in various markets, including the Italian one. Among the new features announced at Google I/O also stands out Yeptechnology text-to-video who “challenges” Sora from OpenAI.

What Gemini 1.5 Flash can do: news and features of Google's AI

Regarding the Features of Gemini 1.5 FlashGoogle has stated that it is the fastest model available through APIs (Application Programming Interface), or the set of “rules” that allows different applications to communicate with each other. According to Google, “it is optimized for large-scale, high-volume, high-frequency activities.”

Gemini 1.5 Flash also comes with multimodal reasoning, so it works via textual, vocal and visual input. Another interesting aspect is that it has an innovative feature context window up to 1 million tokens long, equivalent to 10-11 hours of video or tens of thousands of lines of code. An impressive amount of data and information!

All these features allow Gemini 1.5 Flash to excel in various tasks. In the official Google press release, in fact, we read:

Gemini 1.5 Flash excels at summarizing, chat applications, image and video captioning, extracting data from long documents and tables, and more.

It must be said that the tool is designed for developers who intend to base their products and services using Google's AI, and not for “end” users.

Credits: Google.

How the new Google Gemini Flash works: training by distillation

The “secret” behind the excellent lightness-power ratio of Gemini Flash is the result of the type of training with which the model was trained. Google has in fact opted for the so-called training for “distillation”. How does it work? To put it simply, this process consists of transferring the essential knowledge and skills of a larger model (in this specific case, Gemini 1.5 Pro) to a smaller and more efficient model (in this case, Gemini 1.5 Flash).

Conceptually it is as if the larger model acts as a “teacher” to the smaller model who, therefore, will act as a “student”. The trained and “distilled” model will in fact be able to do the same things as the larger model, but at a lower cost and with reduced latency.

What are the other AI news announced at Google I/O

In addition to the launch of Gemini 1.5 Flash, Google also announced some news regarding the other models that are part of its “package”. In particular, Gemini 1.5 Pro he improved his coding skills as well as analyzing audio tracks and images. Google plans to further improve the model by doubling the context window of Gemini 1.5 Pro (and, by extension, also of 1.5 Flash) to 2 million tokens. This latest innovation will only be accessible by registering on a special waiting list. It was also announced theimmediate availability of Gemini 1.5 Pro in 35 languages, including Italian, in 150 countries for Gemini Advanced subscribers (€21/month after a 2-month free trial).

In conclusion we summarize the other news announced during the Google I/O conference.

  • AI-powered Google Search: through the AI ​​Overviews tool, Google search becomes more effective, as the information and answers to your questions are summarized and grouped intelligently.
  • VEO and Imagen 3: are two generative AI tools which, respectively, allow you to generate ultra-realistic images and videos (lasting more than a minute) starting from textual input.
  • Project Astra: it is a digital assistant capable of processing text, voice and images in real time, capable of understanding context, a bit like GPT-4o.

If you are interested in learning more about the news we have told you about, you can watch the recording of the conference (in English) via the following video.