Google makes world’s most powerful neural network available to public
Google announced the launch of open beta testing of the advanced neural network Gemini 1.5 Pro, whose main feature is the ability to process up to 1 million tokens.
The Gemini 1.5 Pro neural network operates with a context window of 128k tokens, expandable to 1 million. It has the capacity to process 700 thousand words, 30 thousand lines of code, 1 hour of video or up to 11 hours of audio material, Gazeta.ru reports.
Under the research, Google also successfully tested the processing of up to 10 million tokens.
Using the Transformer architecture and MoE, Gemini 1.5 Pro demonstrates high performance in a variety of tasks, including historical data analysis such as transcripts of the Apollo 11 mission. In addition, the neural network not only analyzes large volumes of information, but also quickly finds specific fragments of text, and also works effectively with large blocks of code.
In Needle in A Haystack (NIAH) testing, the neural network showed 99 per cent accuracy in identifying specific data in extensive texts. The adaptive learning capabilities demonstrated in the Single Book Machine Translation (MTOB) benchmark make Gemini 1.5 Pro one of the best today.
Right now anyone can try Gemini 1.5 Pro in the interface Artificial Intelligence Studio From Google. The neural network understands and can respond to Russian, but it does not work in Russia. You can find out in which regions the new service operates Online companies.