Google's announcement on its new linguistic model, Gemini, raised a veil of curiosity and skepticism. Although the promotional video promised extraordinary performance, some investigations and observations have shed light significant discrepancies between what is shown and actual capabilities of the system. According to eminent sources, very careful, its potential would be lower than what is described.
The truth behind the Gemini promotional video: all smoke and fire?
During the recent presentation of Gemini, Google hit the public with a video showcasing the supposedly advanced capabilities of its new AI language model. However, it wasn't long before some media outlets, including Bloomberg, they began to doubt the authenticity of these performances.
In the video, Gemini appeared to be able to instantly interpret a series of gestures, including rock, paper and scissors. However, as later emerged, to obtain accurate answers from Gemini, you need to show all three gestures and provide specific guiding questions. This suggests that the actual process is more complicated and less intuitive than was initially portrayed.
Accusations of having "embellished" the real capabilities of the neural network have raised doubts about the Google's transparency and honesty in presenting their innovations. According to some Google employees, who preferred to remain anonymous, the demonstration video did not accurately reflect Gemini's real capabilities. This view was further confirmed by Eli Collins from Google DeepMind, which admitted that some of the demonstrated features, such as the duck drawing, they are still at research level and not fully integrated into the product.
Despite these revelations, Google maintains that the video was intended to be a simplified representation of Gemini's capabilities for the sake of entertainment and clarity. The description of the video on YouTube, in fact, admits that the presentation had been shortened for these purposes.
During a test conducted with Gemini through the Google Bard chatbot, journalists noticed that the capabilities of the new AI model did not differ significantly from those of GPT-3.5. This raises questions about the actual technological progress represented by Gemini compared to existing models.
The three variants of Gemini
There are three versions of the neural network: Incredibly , Pro e Nano. The former is designed to work in data centers and enterprise applications, the latter is focused on Bard and other Google services, and Nano is designed for offline work on Android. The latter has already been announced for Pixel 8Pro. As an example of its use, Google showed the summary of information from voice recordings.