Artificial Intelligence | February 20, 2024
The launch of ChatGPT was a setback for Google, which had been investing in Artificial Intelligence for years only to be overtaken by OpenAI with a very popular product for the general public. Since then, the big G has been trying to catch up, and a year later, we can say that it has caught up with its rival with Google Gemini, at least in some aspects.
The development of AI within Google has accelerated to unprecedented levels, and as evidence, today’s presentation of Gemini 1.5, a new version just days after making Gemini 1.0 available to everyone. In these days, there has even been time for Gemini to reach Android and be available in multiple languages, a brutal release pace that we hadn’t seen even in Google’s best times.
As the “.5” indicates, this new version of Gemini is not a complete overhaul of the generative AI, but it is a substantial improvement in a fundamental aspect: performance and efficiency. In other words, doing more in less time and at a lower cost, which opens the door to other improvements such as the amount of information the AI is capable of processing at once and how quickly it can be trained for different tasks.
Gemini 1.5 is here Google claims that Gemini 1.5 has a window of 1 million tokens consistently; for comparison, Gemini 1.0 Pro ‘only’ was able to process 32,000 tokens, and GPT4 Turbo, the latest version of the model ChatGPT is based on, can process 128,000 tokens. Tokens are blocks used to process information, and they can be parts or subsections of words, images, videos, sound, and code; the larger the ‘context’ window, the more information it can obtain and process after a user request.
All this means that Google’s AI is able to take in more information at once, thanks to a new architecture based on the ‘Mixture-of-experts’ (MoE) technique, which allows selecting the most relevant parts of its neural network depending on the type of data entered; and this is still a multimodal AI, so it supports not only text but also video.
All of this means that Gemini 1.5 Pro is capable of processing one hour of video, 11 hours of audio, more than 30,000 lines of source code, or more than 700,000 words without any problems. However, Google also admits that only 128,000 tokens will be used as standard, and that only a limited group of developers will be able to test it with 1 million tokens in AI Studio and Vertex AI.
Processing a greater amount of information has an obvious advantage in a generative AI, as it allows obtaining a greater context and presenting a more consistent, relevant, and useful result for the end user. This may be what Gemini needed to catch up with ChatGPT and surpass it, something that we will notice in the official Android app as well as in apps and services that use Gemini.
