Ten million tokens can be used to power the new Gemini Pro.
The new artificial intelligence (AI) branding that Google has just unveiled is called Gemini, and it is rolling out models in a variety of sizes and levels of sophistication. Already in the process of being replaced is the Gemini 1.0 Pro, which is the model’s middle-tier variant. It has been stated by Google that Gemini 1.5 Pro is ready and has the ability to take in more data than any of its competitors. The result places it on an equal footing with the largest Ultra model that Google offers. But just a select few testers will have the opportunity to experience what the unfettered Pro AI is capable of.
Researchers from Google have just recently published a study that describes a way that can eliminate the memory bottleneck that occurs in GPU clusters. In the past, graphics processing units (GPUs) that were running generative artificial intelligence would process a portion of data and then recompute it when it was time to hand it off to another AI accelerator. With this novel method, every computer unit is able to transmit bits to other parts of the cluster while concurrently receiving blocks of the same kind from other parts of the cluster. Consequently, the focus of the model is refined, which enables it to consume a greater quantity of data simultaneously. This is what caused the version 1.5 update to happen so quickly.
A comparison of the number of tokens that a model is able to process is one method that may be utilized to evaluate the level of complexity and, in theory, the effectiveness of a model. A simple slice of data that has been cut up into easily digestible chunks, such as numbers, words, or parts of words, is referred to as a token. On the other hand, ChatGPT-4 Turbo had 128,000 tokens, whereas Gemini 1.0 Pro had only 32,000. Claude, the chatbot that Anthropic uses, was the previous leader with a frame window that contained 200,000 tokens. Despite the fact that Gemini 1.5 Pro is capable of operating with up to 10 million tokens, you will not be using that product.
Gemini 1.5 Pro is a multi-modal artificial intelligence, which means that you may provide it with image data, text, or even video. It has the capacity to process one hour of video, eleven hours of audio, thirty thousand lines of code, or seven hundred thousand words, all thanks to its enormous context window. In the video that can be found above, Google provided Gemini 1.5 Pro with all 402 pages of the transcript of Apollo 11’s moon landing, which valued about 330,000 tokens. The model is able to comprehend the context and utilize logic in order to provide an explanation for the information. The company claims that it is about equivalent in capabilities to its most powerful Gemini 1.0 Ultra model, but that it will demand fewer resources.
Beginning with a preview for enterprise clients and developers, the new model is being introduced. There are further tiers that can go as high as one million tokens, and the regular version of Gemini 1.5 Pro that you will receive when you go to the Gemini website or download the app will begin with 128,000 tokens. Google is now testing that version with a select group of developers and researchers.
Once the problems in the new Gemini Pro have been worked out, Google has stated that there will be subscription models for the product; however, there is no information regarding pricing at this time. This may indicate a significant change for the product that is still in its infancy. Users with no cost are provided with the Gemini 1.0 Pro model, while upgrading to Ultra costs twenty dollars. If you do not require the Ultra option, the free ride may be coming to an end soon.