They sound like the names of the latest generation smartphones: Gemini Nano, Gemini Pro and Gemini Ultra. In reality they are three different versions of Google’s new artificial intelligence, which seemed to be postponed to 2024 and which, instead, is available – in the Pro version – already today in 170 countries. Italy is not among these. In Europa, dice Sissy Hsiaovice president at Google and general manager of Google Assistant and Bard, Gemini will arrive “very soon”.
Sundar PichaiCEO of Google and Alphabet who a few months ago compared the impact of generative AI on human life to that determined by the “discovery of fire and electricity”, launched Gemini stating that “the transition we are currently experiencing with artificial intelligence it will be the most profound of our lifetime, much greater than the transition to mobile devices or the web that preceded it.”
“We’re just scratching the surface of what will be possible,” Pichai said.
“Gemini is also our most flexible model created to date,” he said Demis Hassabis, CEO at Google DeepMind, the team born from the merger of Google Brain and DeepMind (AI laboratory acquired by Google in 2014) which develops the Mountain View company’s most advanced artificial intelligence. “Gemini can run efficiently on everything from data centers to mobile devices,” added Hassabis.
This, then, is why they exist three “versions” of the same AI: Gemini Ultra it is the most powerful model, designed for complicated tasks; Gemini Pro it is the most scalable model, suitable for different tasks; Gemini Nano is the model designed to run AI directly on mobile devices: the first to receive it will be the Google Pixel 8 Proand in particular its “Recorder” app which already today, thanks to AI, does an extraordinary job in terms of transcribing content.
We imagine that generative AI, on Pixels (“and later on other Android devices” says Google), will also give a significant hand to put the transcribed notes in order: proposing summaries or bullet points, for example.
Certainly Gemini will not take Bard’s placethat means Google’s free chatbot – similar to ChatGpt – which answers users’ questions with natural language and, when required, with a good dose of “creativity”. Gemini, in fact, will be Bard’s “engine”.. The Pro and Ultra versions of the new AI model, trained on a huge amount of data [Google non ha specificato il numero di parametri, nda] they will allow Bard to solve increasingly complex problems and provide increasingly accurate answers to its users.
Before making it available to the public, Google subjected Gemini Pro to a series of industry benchmarks [vale a dire misure di riferimento comunemente accettate e utilizzate per valutare le prestazioni o le caratteristiche di un sistema, nda]. In six of the eight benchmarks Gemini Pro surpassed GPT-3.5. This includes the MMLU (Massive Multitask Language Understanding) benchmark, one of the leading standards for measuring large AI models, and the GSM8K benchmark, which measures elementary school-level mathematical reasoning.
Furthermore, with a score of 90%, Gemini Ultra – again according to what Google reports – it is the first model to outperform human experts in the MMLU (massive multitasking language comprehension), which uses a combination of 57 subjects such as mathematics, physics, history, law, medicine and ethics to test both world knowledge and problem-solving skills.
On paper, Google’s new AI looks promising. Eli Collinsvice president of Google DeepMind, explained that Gemini is “natively multimodal.”
“So far the standard approach to creating multimodal models [IA capaci di interagire con diverse modalità di input e output, dai testi alle immagini] was to develop separate components and then put them together,” Collins said. These models are very efficient when it comes to doing a certain operation, such as describing an image for example, but they struggle when dealing with difficult concepts or complicated reasoning. Gemini, on the other hand, was trained from the beginning on different types of data such as text, images, audio and so on. In this way Gemini can intuit the nuances of certain information contained in images or audio, for example, and can reason about mathematics or physics problems.”
Collins’ words were followed by a practical demonstration during the meeting organized by Google to reveal Gemini. In a pre-recorded video, Sam Cheung – Interaction designer at Google – shows how the new AI is capable not only of “reading”, analyzing and solving a mathematics problem written on a sheet of paper, but also of checking the answers given by the user to that problem and to explain to the human being where he went wrong and why.
Just two weeks ago, the news of the progress of OpenAI – the company he created ChatGpt – in solving elementary mathematics problems he had even raised fears for the future of humanity. There was talk of a project Q* – whose existence was confirmed by CEO Sam Altman – which would bring OpenAI’s AI closer to breaking latest news, general artificial intelligence that could one day match human cognitive abilities.
“I don’t know the details of OpenAI’s work – Collins replied to those who asked him if Gemini’s performance also suggests progress towards breaking latest news – so I can’t say anything about it. However, I can say that with Gemini, progress has been made regarding multimodal reasoning and reasoning regarding mathematics and physics”.
“With Gemini we have also made enormous progress in terms of factuality,” Collins said, referring to the AI’s ability to base its responses on concrete facts and objective realityso as to avoid the “hallucinations” typical of generative artificial intelligence, i.e. the tendency to produce plausible and coherent answers but with invented content.
How to recognize fake news and how to defend yourself in the era of Emanuele Capone’s artificial intelligence November 27, 2023
“Gemini is our best model from this point of view – added Collins – but the issue of possible errors is an AI problem that is still unsolved. This is why on Bard we have an integrated tool that allows you to verify the information generated [si tratta della “G” simbolo di Google che permette di accedere alla ricerca tradizionale sul web sull’argomento chiesto a Bard, nda]”.
Of all the companies involved in the AI race, Google is the one that needs to be most careful, since Bard’s wrong answers – despite the “Experimental” label and the warning “may show inaccurate information” – could undermine the credibility of a company that makes research accurate a business worth 162 billion dollars a year. So much money comes from advertising on Google, equal to approximately 58% of the company’s total revenues.
by Green&Blue editorial team 27 November 2023
In Mountain View, the California city that hosts Google’s headquarters, there is talk of a new era. Or rather one “Gemini was”, as CEO Sundar Pichai stated when presenting the new AI, not surprisingly branded “Gemini 1.0”. For Google it is the first version of a technology that could have a huge impact not only in the scientific field, but also in everyday life.
We saw it with our own eyes, during the meeting organized by Google to present its new artificial intelligence. We saw it with our own eyes, during the meeting organized by Google to present Gemini. A very advanced version of AI, not available to the public, “saw” through a video recording the various actions taken by a human being and commented on them in real time – with voice and text – dispensing information and jokes.