Just yesterday, we told you that the arrival of Gemini, Google’s new AI model, is expected to be imminent and likely to happen this week. And we weren’t wrong, at least you could say we were overly conservative in determining the time left until that happened, because what was intended to be a few days, It ended up taking a few hours.. And in addition, with a faster disembarkation than we expected.

Before I talk in depth about the model, I think it is interesting to address this point. As you probably remember, it was leaked less than a week ago that Google decided to delay the launch of Gemini due to some issues with its releases in languages ​​other than English. That explains it for us, of course His debut will be produced in Shakespeare’s language, with other languages ​​to be added later. Personally, I think this decision is the best.

But let’s see, what is a Gemini? The first impulse calls us to classify it as LLM, and this explains that the comparisons we see, including some from Google itself, put it face to face with GPT-4. However, this misses an important part of the picture, which is better understood if we compare it with other multimodal models. And the latter are, so to speak, LLMs, to which other models capable of analyzing image and video, sound, etc. were later added. However, From its inception, Gemini was designed to support various types of input and output formatswhich translates into native and therefore full integration of the mentioned media types.

Although we are talking about Gemini, there are two important points to clarify in general. It’s the first one Google specifically refers to Gemini 1.0, which leaves us in no doubt that we are talking about a model that we will see develop. And this may seem silly, but it is not at all, because it shows us a more than likely competition, the toughest (and therefore the most interesting) between Gemini and GPT.

Another key aspect is that there are actually three implementations of the model, and Google defines them as follows:

• Gemini Ultra – our most powerful and largest model for highly complex tasks.
• Gemini Pro – our best model for scaling across a wide range of tasks.
• Gemini Nano – our most efficient model for performing tasks directly on the device.

They are on first reading Many people draw the parallel that Ultra would be an equivalent/rival to GPT-4, while Pro would arrive to compete with GPT-3.5. However, many of these comparisons do not take into account what I mentioned above, which is native multimodal support, which is intended to translate into more coherent and faster responses to problems. prompts users.

Clearly, Ultra is aiming very, very high. As stated by Google «With a score of 90.0%, Gemini Ultra is the first model to outperform human experts in MMLU (Massive Multitasking Language Understanding), which uses a combination of 57 subjects including mathematics, physics, history, law, medicine and ethics, both of which are tested for proficiency of the world and the ability to solve problems«.

yes indeed This is the Gemini mod that will take the longest to market, and it may also be the one where we see the debut of the Bard payment method, a model similar to that adopted by OpenAI with ChatGPT+. Whether it is paid or not is still unclear, but the company has announced that one of the ways to use Gemini Ultra will be Bard Advanced, «the new cutting-edge AI at Bard, giving you access to our best models and capabilities«,

Gemini Pro has already started to be implemented in Bard and will soon be implemented in other Google products and services. Of course, as indicated earlier, for now it will only take place in English as quite broad but not global reach, which the company estimates in more than 170 countries worldwide. So yes, it seems quite likely that, as happened in May, we will still have to wait to enjoy its benefits within the common European space.

As for what other products and services Gemini Pro will receive, Google is not specific search engine, ads, Chrome and Duet AI, among others, although he has not yet specified what he will contribute to each of them. However, it can be expected to be a fairly significant contribution, as this is undoubtedly the moment and model that the company has been waiting for to reach cruising speed in terms of AI implementation.

Now, I personally was most interested in undoubtedly the Gemini Nano advertisement, the smallest of them, as its name suggests, but which, as we can see in its description, has been optimized for direct use in the device, that is, those devices in which this model is implemented will be able perform more and more AI-based tasks autonomously, without having to rely on your connectivity and server infrastructure that provide the specified service.

And where will the Gemini Nano debut? Surely you already imagined it, so I can confirm, this model will soon land on the Google Pixel 8 Pro, the top smartphone of the company, in which we can already find advanced use of artificial intelligence and that with this qualitative leap it can develop substantially in the coming months.

This is very important because after the boom of artificial intelligence based on remotely launched services artificial intelligence in the client gives the user more control, except that the use of these functions does not require an Internet connection. Of course, the deployment of locally executed AI functions will not be quick, but it points to a clear trend direction that will undoubtedly expand in the short to medium term.

More information