Google is stepping up its commitment to artificial intelligence by highlighting a highly promising tool it has been working on for years: Gemini. Google Gemini aims to position itself as the best artificial intelligence model that has ever existed.
It is a clear demonstration of the high level of competitiveness and thirst for innovation amongst companies dedicated to the development of AI; in particular, it highlights the struggle between OpenAI and Google for taking the crown in artificial intelligence.
Although it could be argued that full credit for Gemini cannot be attributed entirely to Google, but rather to DeepMind, a company specialising in artificial intelligence that was acquired by Google (or Alphabet, its parent company) in 2014.
In this Euroinnova article, we’ll tell you everything we know so far about Google Gemini. We’ll explain the differences between each of its variants, how it works and what sets it apart from its main competitor, GPT-4. We’ll also tell you How to try out Gemini on Google’s chatbot, Bard.
Gemini Pro vs. Gemini Ultra vs. Gemini Nano
Google has announced the launch of Gemini in three different formats, which it has named Gemini Nano, Gemini Ultra and Gemini Pro. Each has its own specific features and unique operating modes, which we could discuss at length. However, for the sake of practicality, we shall limit ourselves to comparing the approaches taken by each product in the Gemini family.
Nano
Nano is Gemini’s lightest and least powerful range, designed primarily for use on mobile devices. There are two versions of Nano:
- Nano-1: It can handle up to almost two billion parameters. It is designed for devices with limited memory.
- Nano-2: It is capable of processing three billion parameters. It is designed for devices with a high storage capacity.
Gemini Nano will be the member of the family that is to be implemented on the mobile phone Google Pixel 8 Pro. The integration of this advanced AI model will be particularly noticeable in the camera. Therefore, significantly improves the camera’s performance in low-light conditions and autofocus, as well as automating settings to capture the best photos in every situation. It will also introduce new features for photo and video editing.
Pro
Gemini’s Pro mode is the mid-range model that sits between its younger sibling, the Nano, and its older sibling, the Ultra. We can currently see what this model is capable of using the Bard tool. In December 2023, Google integrated Gemini Pro into Bard, boosting its performance compared to ChatGPT.
Ever since Google launched Bard to compete with OpenAI, Bard has been completely overshadowed by the popularity and positive reviews of the renowned ChatGPT, particularly following the launch of GPT 3.5. However, this could change radically thanks to Gemini’s Pro mode.
Would you like to put Gemini Pro’s potential to the test? All you have to do is access the Bard chatbot and sign up to find out.
Ultra
It is the most powerful version yet, which is set to outperform GPT-4 on many fronts, according to Google’s estimates and experimental tests. Whilst GPT-4 already does a superb job of processing the context of words and images, Gemini goes a huge step further by being able to to distinguish very subtle nuances in data provided in a wide variety of formats, such as text, images, audio and video.
Furthermore, it demonstrates a better understanding of very complex issues, such as maths or physics. It is also predicted that it will be able to answer any questions we ask in real time, with minimal latency.
Although Gemini Ultra is not currently available for use, Google has announced that it will definitely be implementing it in its Bard chatbot, which will be renamed Bard Advanced.
How do Gemini’s artificial intelligence models work?
Gemini’s models are based on transformers that the Google team has optimised in order to train them reliably on a large scale and carry out inferences – that is, process new data – using Google’s Tensor Processing Units (TPUs).
Gemini’s multimodal models are capable of processing long texts of up to 32,000 words and use mechanisms and AI modelling techniques such as MQA (Multiquery Attention).
As well as handling text with ease – something we already see in AI text generators such as ChatGPT and Bing – Gemini can work with a variety of inputs in audio and graphic format such as images, graphics, screenshots, PDFs and videos. In turn, they produce outputs in both text and image formats. This visual encoding is made possible by a number of advanced modelling techniques also developed within Google, such as Flamingo and PaLI.
When it comes to video processing, Gemini treats videos as sequences of frames in a wide range of resolutions that can be interwoven with text and audio.
The Gemini product family has been developed on the basis of machine learning algorithms enhanced by infrastructure and vast sets of proprietary data.
Differences between Gemini and GPT-4
If there’s one thing DeepMind and Google take pride in, it’s having surpassed OpenAI’s most advanced multimodal language model, GPT-4, in terms of general capabilities. In fact, the first thing you’ll see when you visit the Gemini website is a statistical showcase of Gemini’s outstanding performance results compared to those of the GPT model.
Thus, Gemini Ultra outperforms GPT-4 in many respects:
- Answers to questions on various topics.
- Reading comprehension.
- Solving arithmetic operations and mathematical problems.
- Programming in Python.
- Reasoning in multidisciplinary tasks.
- Image processing.
- Document processing.
- Mathematical reasoning in graphical contexts.
- Subtitled.
- Answers to frequently asked questions in video format.
- Automatic audio translation.
- Speech recognition and processing.
Among the significant improvements that Gemini Ultra offers over its competitor, one that stands out is image recognition and the processing of audiovisual formats. So far, we have mainly experimented with text-based formats, rather than audiovisual ones. This innovation would go a long way towards enabling AI models to better understand and interpret data in visual and auditory formats.
Another factor in Google’s favour is that Gemini, as a Google-branded product, would have access to the entire database of information provided by the world’s leading search engine in much of the world. It is estimated that Google stores up to 15 exabytes in its databases (one exabyte is equivalent to one trillion bytes) of storage, and this figure is rising every second. In other words, Gemini’s models are trained using an almost infinite and unrivalled sea of data.
Conclusions: Will Gemini’s expectations be realised?
Whilst the launch of Google Gemini represents a significant milestone in the field of artificial intelligence, we cannot yet confirm that Gemini Ultra will live up to the high expectations it is generating. With its advanced multimodal capabilities – should it live up to them – Gemini would mark a major step forward by integrating visual and auditory data, and Google’s vast database would enhance its training with a virtually unrivalled amount of information.
Español 
English (UK)