A discreet launch. Since the end of May, Google has been offering an application to use artificial intelligence locally and without internet connection. His name: Google AI Edge Gallery. Available on Android smartphones, and soon on the iPhone, it is presented as “an experimental application that places the power of maximum generative models in their hands, turning completely on their devices.”
With Google ai Edge Gallery, it is possible to discuss with an AI, send an image and question a model or get a summary of a text. But what is this new application worth? Tech & co tried it.
How to download and use Google ai Edge Gallery
To start, Google AI Edge Gallery is not available in Play Store. To download it, it is necessary to go through the github platform and download an APK file. An installation that is performed quickly and without difficulty. Google’s application has no danger. However, it may include errors, which is still in the experimental stage.
Once Google AI Edge Gallery installed on your smartphone, three functions are offered: “Ask Image”, “Fast Laboratory” and “AI Chat”. For each of them, Google specifies the number of available models. For the first, which makes it possible to question an AI in an image, two Gemma models are offered, which come from the company.
But before I can use them, it is necessary to download them. Therefore, an Internet connection is necessary. It is only once it will be possible to do without the Internet.
Identification, translation …
To “ask image”, it will be enough to send an image and then question the model, which responds in seconds.
At the request of the user, you can indicate the objects present in the image, translate the present text or even explain how to get to a place. But as Tech & Co pointed out, the model’s responses are not always precise.
When questioning what he sees in the photo above, the model, for example, replied that the two figures were lego, while this is not the case. The one on the left is a playmobil, while the one on the right is the real -brand pet. He also assured that the notebook in which the two figures were placed was a digital tablet.
AI can also find difficulties in translation. When an image was presented asking him to translate the text that appeared there, she replied leaving the text in English.
She left much better with another image, including the instructions for a connected coffee maker, but there were still some mistakes. For example, he translated “Connect the cap from the hands of the coffee machine to a wall socket” to “connect your coffee machine to a wall” instead of “wall socket”.
Finally, the AI model can also be used to discover how to get to a place. He managed to identify the triumph arc and indicate how to get there from the defense, either per meter, by car or by bicycle.
Options for which the estimated duration specifies, the cost, as well as the advantages and disadvantages.
To conclude, the AI has summarized everything and has advised the use of certain applications, such as Google Maps, to “obtain real -time routes.”
Difficult discussions
The new Google application also allows you to chat with an AI as we do with Chatgpt or Gemini. But only by text. There are four models available, including three from Google and one of the Chinese company Alibaba. “Hello (…) I am a virtual assistant who can help you in many areas,” replies the last one when it is said to be “Hello.”
But the problems quickly manifested in the conversation. The Alibaba model cannot write the letter “É”, systematically replacing it with two special characters, as shown in the image above.
And worsened when asked why, with an answer full of absurdities. If the experience is better with Google models, it is far from being perfect. For one of them, for example, each response is interrupted by the character chain “<0xC2>“
It can also be an inconsistency in their words, integrating scripts at all costs and prayers that make no sense.
Very useful functions but …
Finally, the laboratory notice, composed of the same four models as “AI Chat”, offers four functions. Among them, the summary of a text and the change in the tone of a message. For the first, several options are proposed, including “Conci summary” (around 50 words), key points (3 to 5) or “short paragraph”.
With Alibaba’s Qwen model, the results were far from agreed. When sending a Tech & Co article to summarize in 50 words, the AI responded with a series of hashtags, with words that had nothing to do with the subject. The article in question was to talk about Samsung and a future “ultra” version of its folding smartphone and in the hashtags mentioned by AI are the words Apple and Ipad.
She improved when they asked for a small paragraph, but he was in English. A problem that is also recurrent with this function and that is not only reserved for the Alibaba model. Another problem: AI can be repeated in summary, mentioning, for example, twice that this future version “ultra” would have a “powerful camera”.
With Google Models, AI has generated good summaries in general. However, he was wrong, in the 50 words summary, saying that Samsung had presented an “ultra” version of his folding smartphone, which is not the case. The summary in the keywords was much better, with all the correct information that was summed up well.
Finally, either with Google or Alibaba models, the only problem with the function that allows you to change the tone of a text was that the result was systematically in English. Several options are offered (formal, friendly, educated, enthusiastic …), with convincing results. And in some cases, the AI is not only content to transform the text, but also explain its choices.
For the “enthusiastic” tone, the AI integrates, for example, emojis in its message and explains that this “adds enthusiasm and a pleasant atmosphere”, as well as the choice of certain terms.
If the Google application is still far from being perfect, we must not forget that it remains an experimental version. Therefore, it leads to improve over time. But in the meantime, the results are not very convincing for most of the functions offered.
Source: BFM TV
