As usual to properly launch its product, Google subtly changed the name from Bard to Gemini (as it had gone from Google Apps to GSuite then Workspace, or from Google Local to Places to MyBusiness to Business Profile, or Google Hangouts to Chat, etc.) while giving it new features.

It must be recognized that Google's artificial intelligence engine is powerful.

The next few weeks will undoubtedly allow us to explore in detail what it has under the hood, but we already know, for example, that Gemini now has a pop-up of one million tokens. This pop-up window is what allows the agent to remember the exchanges from one question to the next. By comparison, ChatGPT 3.5 (the free version) has a window of around 16.000 tokens and version 4 allows up to 128.000 tokens and to give an order of magnitude, this last number represents the equivalent of a book of pocket roughly, where Gemini can remember a complete work of 1.500 pages…

Or a film of around an hour and a half: Gemini allows the user to provide images or videos as contextual input on which to base their responses; to quickly summarize a video, this may prove particularly interesting in the near future.

But alongside these undeniable technical advances and truly interesting power, Gemini stood out above all for the very rapid identification of a problem that was quite embarrassing for the Mountain View firm in California: undoubtedly wanting to catch up in in terms of image production from a textual description – OpenAI has indeed been allowing its clients to produce images for several months directly from ChatGPT – Gemini was equipped with this possibility but in use, it quickly became apparent that certain requests were simply not accepted or the gap between requests and results was so great that, very quickly, social networks took over the matter.

No doubt: when Gemini is asked to produce images with historical content or representing certain ethnic groups, the latter interprets the request a little too specifically.

Thus, obtaining the image of a medieval knight or a pope results in the production of images systematically in gross disagreement with reality: Google's artificial intelligence enjoys creating, with suspicious enthusiasm, knights medievals of all possible ethnicities but the blond knight with blue eyes is strangely absent; the images of popes produced happily draw on women, possibly Indian or black; as for the Roman emperors, they are all surprisingly very African.

The milestone was reached when faced with the request to represent German soldiers in 1943, Gemini saw fit to produce a series of resolutely inclusive images including proud black representatives of the Wehrmacht... Who still doubted that historical reality could bend as easily to the most modern constraints?

read the article