As part of its long-term plan to build next-generation social infrastructure, SoftBank Corp. (TOKYO: 9434) has been making good progress on its initiatives to build a generative AI computing platform and develop Large Language Models (LLMs). President & CEO Junichi Miyakawa unveiled new milestones, future plans and more at SoftBank's earnings results briefing for the nine months ended December 31, 2023, of the fiscal year ending March 31, 2024 (FY2023), held on February 7, 2024.
At the conference, Miyakawa noted that SoftBank’s generative AI computing platform became operational in 2023, as previously announced, and that the full-fledged development of homegrown LLMs specialized for the Japanese language is well underway. “At the end of December last year, we reached the 130 billion parameter mark for our LLMs,” he said.
Miyakawa noted that SoftBank is now developing multimodal LLMs, which refers to two or more different types of data, such as text, audio, images, videos, sensor information and others. “We’re now working to complete multimodal LLMs with 390 billion parameters by the end of the next fiscal year*,” he said.
- *
The fiscal year ending March 31, 2025 (FY2024)
Miyakawa noted that SoftBank’s further aim is to exceed 1 trillion parameters, and to achieve this, it will be further expanding its generative AI computing infrastructure. Miyakawa explained that the recent practice in the industry is to build large models at once and then optimize them to sizes that are fit for purpose. This method, commonly referred to as “knowledge distillation” will be the one that SoftBank adopts for generative AI in its goal to achieve 1 trillion parameters, he added.
Related News
(Posted on February 14, 2024)
by SoftBank News Editors
