The Tokyo Institute of Technology and other organizations have announced the start of development of highly performing generative artificial intelligence using the Fugaku supercomputer, one of the world’s fastest in terms of calculation speed.
“We aim to accelerate the development of generative AI using the highly performing Fugaku and create an environment where we can produce the world’s best AI,” said Rio Yokota, a professor of advanced applications of high-performance computing at the Tokyo Institute of Technology, who leads the project, on May 22.
Generative AI has become widely known thanks to ChatGPT, an AI chatbot developed by the U.S. startup OpenAI.
It is rapidly becoming widespread mainly among U.S. companies as computers and generative AI tools’ ability to learn vast amounts of texts has grown more sophisticated.
However, major generative AI tools are being fed with information that is mainly in English, meaning they are prone to giving incorrect information when asked questions about Japanese history or Japanese people.
The announced project will input information that is in Japanese into AI.
Researchers involved in the project will make public how they develop its generative AI, thereby helping to expedite the development of homegrown AI.
The Riken research institute and Fujitsu Ltd., both of which jointly developed Fugaku, and Tohoku University, which researches language processing AI, have all also joined the project.
They will use Fugaku to develop a method to enhance the performance of the generative AI that has been fed with texts from the Japanese version of Wikipedia, the free online encyclopedia, and other Japanese sources.
Researchers in the project will make public the generative AI they will develop and how they develop it by the end of this fiscal year so that companies and research institutions in Japan can utilize the information.
After that, the researchers will continue their research to produce AI that can compete with the AI produced by Google and other organizations.
The project also aims to find solutions to issues that ChatGPT has faced, including the occasional tendency to give completely incorrect answers.
Source: Asai Shimbun / TAKAHIRO TAKENOUCHI
