To promote its new Gemini AI, Google has come to terms with reality a little

Google just unveiled Geminihis response to ChatGPTand the least we can say is that the demos made by Google are stunning… A little too stunning even.

The AI ​​war is declared and in this conflict Google has come armed with Gemini, its great language model.clearly ahead» on competition. The search giant notably highlighted Gemini’s in-depth understanding of audio and video content. And if Google’s digital brain is not yet available in French at the moment, we must admit that the demonstration made in English by Google was enough to take your breath away.

A demo too good to be true

For 6 short minutes, we see the AI ​​analyze and react almost in real time to what the camera shows. The machine seems to understand immediately when a hand starts playing rock-paper-scissors, recognize failed imitations of The Matrix and even play pieces of music adapted to the instruments scribbled on a post-it. Unfortunately, everything is not as smooth and efficient as Google has led us to believe.

As detailed in a Google blog post, the responses given by Gemini in the video are in fact much more fragmented than that and the “prompts” given to the machine are much more precise than the voice-over might lead you to believe. So the video gives the impression of an almost casual discussion with the AI ​​when the reality is much more tedious than that.

A gifted AI, but not autonomous

None of the answers given by Gemini were invented, but some were merged or shortened to give the impression that the AI ​​knows how to hold a discussion and string together related elements of answers without having to be restarted. Which is not the case. For example on the recognition of instrument drawings, the video makes you believe that Gemini is capable, without any intervention, of recognizing the drawing and automatically playing an adapted piece, whereas the prompt was in fact separated into two stages and precisely details all the actions that Gemini must perform.

In its defense, Google made it clear in the description of the video that “For the purposes of this demo, latency has been reduced and messages from Gemini have been shortened“. Oriol Vinyals, head of AI research at Google, even explained that “video illustrates what multimodal experiences could look like […] with Gemini“.

But between a video named “Getting started with Gemini» and an illustration, partially true, of what Google’s AI would potentially be capable of doing, there is still a world ago.

Source : Google for Developers