Japanese companies jumping on the bandwagon of generative AI

The recent emergence of ChatGPT, a conversational AI model created by a US start-up, has prompted Japanese technology companies and research institutes to develop the underlying technology for domestic generative AI tools.

However, it becomes difficult to make a profit if you enter only Japanese information into the system. Companies are also plagued with issues such as piracy.

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

A company official said, “There are few LLMs who are good at learning Japanese and Japanese culture.”

LLM is the core technology of generative AI. It can read vast amounts of text, predict the probabilities of subsequent words, process them to form sentences, summarize, and answer questions. Gauges that indicate the extent of model learning are called parameters.

CyberAgent combined this technology with ChatGPT and started working on the automatic creation of ad copy. For example, if you enter “cosmetics for working women in their twenties who are busy in the morning”, a sentence such as “I present working women with time-saving anti-aging cosmetics” is created.

Another startup, rinna Inc., headquartered in Shibuya Ward, Tokyo, known as Bit Valley, has also released a large-scale Japanese-specific language model. When asked for a local sightseeing spot recommendation, the system answered “Hachiko statue”. The company trained the model primarily with information about Japan.

Fujitsu, Tokyo Institute of Technology, and others have started developing an LLM that specializes in Japanese using the supercomputer “Fugaku.” Nippon Telegraph and Telephone Corporation (NTT) also announced its entry into the generative AI field.

profit, copyright protection

Developing generative models for AI costs huge sums of money, and Japanese companies are lagging behind their overseas rivals. Building these requires high-performance computers that can process large amounts of data, and consumes a lot of electricity, resulting in high electricity bills.

“Someday we will have to monetize it in some way.

Also, when using information only in Japanese, there is also a weakness that the data that serves as the basis for learning is limited.

“It is mainly Japanese companies that use LLM, which provides information in Japanese. said Otoya Shirozuka, a consultant at Link Corporation.

Generative AI tools that produce sentences and images that look as if they were written by humans present many challenges.

According to Bloomberg, The Wall Street Journal’s publisher, Dow Jones & Co., has criticized OpenAI for misappropriating articles. CNN also expressed concern.

At the end of May, a Japanese company that operates an illustration posting site banned the use of AI to learn about other people’s works and post works that are very similar without the author’s permission. The move comes in response to widespread frustration from creators who feel they are being disadvantaged.

Shino Uenuma, a lawyer who specializes in intellectual property, said, “There is no clear conclusion as to when texts and images created by generative AI constitute copyright infringement.” “In order to protect and nurture Japanese creators, we should proceed with discussions as soon as possible.”