Ugly move by Google. Revolutionary AI does not work as shown in the film

Gemini – because we are talking about it – makes a huge impression, taking into account only a few video materials that Google has published so far.

Google showed amazing artificial intelligence. There is only one “but”

The video that aroused the most emotions, lasting over six minutes, showed the potential use of the Gemini model by an average user. In one day, the material was viewed 1.6 million times, and the possibilities of artificial intelligence presented in it are both admiring and a little terrifying. Gemini analyzes sound, images and even videos displayed on your phone’s screen in real time and gives you sensible and concise tips.

Google Gemini in action photo: screenshot from a Google video on YouTube

The program recognizes handwritten sketches, decides what is the best choice in the situation described in the picture, advises how to use the materials shown in front of the camera or instantly recognizes the final image in a connect-the-dots task. What’s more, it recognizes the three-cup game on its own and has no problem guessing which cup the ball is under. It also easily recognizes the location of objects on a paper map and much more. The entire video, lasting over six minutes, presents dozens of tasks that Gemini can handle without any problems, and in just a few seconds at most.

The problem is that – as it turns out – the film does not precisely show the actual AI reactions to the given command. Parmy Olson noticed that the Google program copes with tasks extremely well and quickly. Indeed, in the video description, Google notes that “latency has been reduced for this demo and Gemini output has been condensed for brevity.”

However, Google’s artificial intelligence does not work like in the movie

Google admitted in response to Bloomberg’s questions that the film does not show a recording of an actual conversation between a human and AI, but combines recorded fragments of raw material with Gemini’s statements superimposed on them, but added at the editing stage in voice and written form. Therefore, one may suspect that in reality one has to wait much longer for the correct AI answers from Google, and the AI ​​responds in a less factual and concise way. Therefore, it is probably impossible to have such a fluent conversation with Gemini as shown in the film.

notes that such modification of videos showing the programs’ capabilities is a standard trick of their producers. Even supposedly live demonstrations are often pre-recorded to avoid any unfortunate mishaps or technical errors during the presentation. However, Google often distorts reality in such demonstration videos. In response to a request for additional explanations, Google also referred the website’s editors to a post by Oriol Vinyals from Google DeepMind, who explained the process of creating the video.

Vinyals emphasizes that all user commands and program responses presented in the video are real. However, their content has been slightly shortened. He added that AI answers were obtained by sending questions in text form with attached photos of the elements to be recognized. Therefore, the program did not recognize the image and describe the captured elements from the recording in real time. Vinyals added that we will be able to test the model’s capabilities ourselves on December 13, when Google will release the Gemini Pro version, but he also admitted that the Gemini Ultra version, a more advanced variant of this model, was used for the presentation.