Geminis are undoubtedly the result of la lot of red lights that went on at Google about a year ago, when ChatGPT and its cousin Bing were getting almost everyone’s attention when it came to chatbots and artificial intelligence. The first reaction was, as we told you at the time, to report the Bard, but it was clear from the first that this was only an emergency plan, but in fact his plans went much further.

We had the first confirmation when we first heard about Gemini, at Google I/O 2023, and the scope of this new model clarified for us with its introduction only two months and a few days ago. The Pro version of this model has started to be integrated into the Bard, while the debut of the Ultra was announced earlier this year and the Nano features will be coming to some Android devices.

So both the arrival of the most advanced model and the replacement of Bard with Gemini (which is both the name of the model and the service) happened exactly a week ago today. So we expected that further news about it would be related to its international expansion (territory and languages) and the development of new features. However, Google surprised us today with the introduction of Gemini 1.5a remarkable model update that, yes, is not yet available.

And is the evolution from version 1.0 to version 1.5 so surprising? The truth is, yes, and it’s mainly for two reasons. The first is that performance and efficiency have improved simultaneously. So as we can read in the official announcement, Gemini Pro 1.5 is on par with Ultra 1.0 in resultsbut it also does with less demands on computing technologysomething that has resorted to an increasingly common technique known as “Mixture of Experts” (MoE), which allows only specific parts of the model to be used instead of using the whole for each query.

This type of optimization plays a key role in Gemini 1.5’s surprise novelty, which is Google managed to expand the popup to one million tokens. For those who don’t remember, a popup is the maximum amount of information, measured in tokens, that a model can handle in a work session, and to give some context, last May Antrophic surprised us with a VdC of 100,000 tokens. And in case you were wondering, GPT-4 currently offers a max popup of 128,000 tokens. And if that seems like an excessive number to you, I wonder what you’ll think when you know that Oriol Vinyals, VP of Google Deepmind Research, calculated on his twitter account who tested it with VoC up to 10 million text tokens.