Japan Firms Jumping on Generative AI Bandwagon

By Moe Tajima / Yomiuri Shimbun Staff Writer

15:47 JST, June 18, 2023

ChatGPT, a conversational AI model created by a U.S. startup, has recently stormed onto the scene, prompting Japanese tech companies and research institutes to embark on developing the basic technology for homegrown generative AI tools.

However, generating profit will be difficult if information only in Japanese is fed into their systems. Companies are also beset with issues such as copyright infringement.

CyberAgent, Inc. released last month its Large Language Model (LLM), the basic technology for generative AI capable of learning Japanese. The company has been developing the system since last year.

“There are few LLM’s that are strong on learning Japanese language and culture,” a company official said.

LLM is a technology that serves as the core of generative AI. It can read vast amounts of text and make predictions of the probability of the words to follow and process them to create sentences, summarize them and answer questions. A gauge to show the scale of the model’s learning is called a parameter.

Combining this technology with ChatGPT, CyberAgent has begun working on automatically creating advertising copy. For example, after typing “cosmetics for working women in their 20s who are busy in the morning,” sentences such as “We present working women with time-saving, antiaging cosmetics” are created.

Another startup, rinna Co., which is headquartered in Shibuya Ward, Tokyo, known as Bit Valley, has released a large-scale language model specializing in Japanese. When asked to recommend a local sightseeing spot, the system responded: “The Hachiko statue.” The company had its model learn information mainly about Japan.

Fujitsu Ltd. and the Tokyo Institute of Technology, among others, have begun developing LLMs specializing in Japanese with the help of the supercomputer Fugaku. Nippon Telegraph and Telephone Corp. (NTT) has also announced its entry into the generative AI field.

Profit, copyright protection

Developing generative AI models takes an enormous sum of money, and Japanese companies have lagged behind foreign rivals. Building them requires high-performance computers that can process large amounts of data, and electric bills are high because they use much electricity.

“We will have to monetize it somehow at some point; the computing costs are eye-watering,” Sam Altman, CEO of OpenAI, the U.S. company that developed ChatGPT, posted on Twitter last year.

Another weakness in using information only in Japanese is that the data that serves as the basis for learning is limited.

“It’s mainly Japanese companies that use LLMs that are fed with information in Japanese. The small size of that market has also left Japan lagging behind in development,” said consultant Otoya Shirotsuka of NTT Data Intellilink Corp.

Generative AI tools — which produce sentences and images that look as if they were created by humans — are beset with many challenges.

According to Bloomberg, Dow Jones & Company Inc., the publisher of The Wall Street Journal, criticized OpenAI for unauthorized use of its articles. CNN of the United States is also said to have expressed concern.

A company in Japan that operates a website for illustration submissions at the end of May banned the use of AI in learning the work of others and submitting very similar work without the original submitter’s permission. The move was made in response to widespread discontent by creators who feel they are at a disadvantage.

“There is no clear conclusion about in what case a text or image created by generative AI is blamed for infringing on a copyright,” said Shino Uenuma, a lawyer specializing in intellectual property. “We should advance discussions on this issue promptly to protect and foster creators in Japan.”