Rohto Pharmaceutical's Generative AI Initiative

Impress Taguchi (Jun): This time, I would like to ask Mr. Itabashi about the inside story of Rohto Pharmaceutical, and ask various questions together with Mr. Taguchi (Kei) and Mr. Onishi. First of all, could you tell us about Rohto Pharmaceutical's generative AI journey?

Rohto Pharmaceutical Itabashi: Rohto Pharmaceutical first encountered generative AI two years ago. It was around the time that Midjourney and GPT3.5 were coming out, and I think many people felt that it was interesting but not usable for work. GPT-4, which appeared a year ago, made it much more practical, but it was still difficult to use for a company. After that, various generative AIs like GPT using OpenAI's engine appeared, and we started using ExaWizards's from around June 2023.
About half a year has passed since then, and although it was good for tasks such as translation and summarization, it was difficult to get the answers I wanted. Therefore, from around the end of 2023, I decided to have employees learn how to use Prompts so that they could use them properly.

This kind of prompt engineering only returned general answers, which was not yet sufficient for work use, but we found that it could be useful to use it as an assistant equipped with a little more specialized knowledge by using RAG (additive learning) to train it on in-house specific information. However, some of the information was highly confidential, and we were worried about using only tools on the in-house network, so we developed our own generative AI called "ROHTOPilot" using OpenAI's API.

We are currently working on this idea, thinking that by building something like this within our own company's network, we can broaden the scope of use not only for RAG but also for various business databases and core systems within the company. When we gave it one file for additional learning, it learned a great deal about the company just from that content, so everyone was excited, saying "this is great." However, after everyone gathered together to train it on various files containing text, graphs, figures, etc., we found that the accuracy rate differed greatly depending on the format.

Macnica Onishi: Is there any point in choosing from the perspective of "what to apply it to"? For example, I think it's good to understand call centers, but when applying it to your own work or business, you might think, "It seems like it will be easier to get results in this kind of place," or "I don't know about that, so I'll think about it like this."

Rohto Pharmaceutical's Itabashi: Actually, that's not the case, but since it's artificial intelligence, it should be usable for most jobs that require human thought. There was a time when I thought that generative AI might replace white-collar jobs, but when I went to Hannover Messe in April 2024, I realized that it could also be used in the field of production. In a way, it's like having a knowledgeable veteran employee next to you. Our company holds prompt engineering workshops, and people from all departments, including sales, marketing, R&D, production, and indirect departments, participate.

Impress Taguchi (Jun): Mr. Itabashi took the lead in starting this initiative in June 2023. How many people were in the team?

Rohto Pharmaceutical Co., Ltd. Itabashi: There were only a few of us, but there were many people who wanted to try it. When I tried using ChatGPT, I thought I could do it, but at the time, there was a concern that OpenAI would learn the prompts and leak confidential information. We didn't want that to happen, so we introduced SaaS for businesses. At first, we reached out to people who seemed to have literacy, and expanded the number of users through trial and error.

Impress Taguchi (Jun): Mr. Taguchi and Mr. Onishi, could you tell us your opinions on whether the journey you introduced earlier is an advanced one, or whether there are still many companies that are sticking to 4.0?

Digil Taguchi (Kei): There are some parts that are influenced by the internal environment of the company. For example, I think that there is a trend for companies that have high literacy when taking the first step and have people who can create a usage environment from 0 to 1 to switch to new things. On the other hand, many companies understand that generative AI is very important, but there are few people who actually put it into practice. It has been like this for a long time, but Japanese companies have not been good at manufacturing things themselves or trying new things, so I think that a huge gap is emerging.

Impress's Taguchi (Jun): If someone in your position like Mr. Itabashi says, "There are risks, so we should take things slowly while watching how other companies do," it puts the brakes on. Mr. Itabashi said, "I'm not in the IT field, I'm in the chemical field." How did you understand that?

Rohto Pharmaceutical's Itabashi: Generative AI is still a new field, so if you've been working on it for 2-3 years, you can call yourself a veteran, and we're not the first to do it. There are many companies in Japan that are more advanced than us, but by chance I knew someone among them, so I was able to learn a lot from them.
Another feature of our company is that our chairman, Yamada, has a good understanding of AI and IT. Before becoming president, he was the head of the IT department. When we first started to pay attention to generative AI, he was also paying attention to it, and he was disseminating information about it within the company himself. When starting something new, I think it is very easy to do if the management understands it.

Impress Taguchi (Jun): I have the impression that compared to humans, AI is still not very good at reading lines and graphs, but Mr. Taguchi and Mr. Onishi, what common points do you see among companies that are successful with RAG?

Digil Taguchi (Kei): I haven't come across a place that's really working well yet, but I think that there will be candidates that can do it if they focus on a very single-function task. The combination of RAG performance and LLM, for example, has its pros and cons, and I get the impression that there are more companies that are experimenting.

Macnica Onishi: I get the feeling that people have high expectations for RAG, that it will be able to search for anything, but I don't think we've reached that point yet.

Digil Taguchi (Kei): When considering using RAG in a company, the positive aspects such as scalability tend to be focused on, but after inputting the data, you may find that the input information was ridiculous to begin with, or that the company's documents have not been properly updated, meaning that the quality is poor. This is because incorrect information is being output as a result of processing incorrect information, so RAG is not to blame, and improving it is completely unrelated to the system. If the documents themselves are not structured, it's not like you can use RAG to make them look nice. I'm sure you all probably struggle with these points.

Impress Taguchi (Jun): In the case of Rohto Pharmaceutical, isn't it the case that all document data is collected in an ideal form and someone is maintaining it so that there are no inconsistencies?

Rohto Pharmaceutical's Itabashi: That's right. Entering incorrect information is a different matter, but even if you enter correct information, you would still need to update it. However, RAG will not update if you leave it alone, so it will need to be replaced after updating the data. So is it enough to simply replace it? Old information can be necessary even if it is old. However, if generative AI is used to break down the components and vectorize them, time information is lost, and it becomes impossible to tell whether it was last year or the year before. It is not enough to just have the latest information, so it is also necessary to manage updates, such as how to manage it on a time axis.

Impress Taguchi (Jun): If you don't take the timeline into account, it could appear that there is some contradictory information.

Rohto Pharmaceutical's Itabashi: That's right. That's why various ideas have been proposed to give generative AI a concept of time.

Impress Taguchi (Jun): Mr. Onishi, are the technical means to solve these issues gradually emerging?

Macnica Onishi: Yes. However, the characteristics of each data are different, so you have to think about each one individually, which I think will cause problems in many places. For example, if Japanese and English are mixed, you need to change the database, and there are also differences in formats and versions, so it poses many challenges. In such a situation, everyone says "RAG is convenient," so I think there is a gap between the image and reality.

Impress Taguchi (Jun): Although there are various challenges, the important thing is to have the attitude of solving them one by one, and when you hit a plateau, it's not right to say, "Let's quit RAG."

Digil Taguchi (Kei): I think that new technologies other than RAG will emerge in the future, but it is very important that we continue to use them proactively. It's not a bad idea to use something that other companies have done well in our own company, but after all, only we know the circumstances of our own company.

Expectations for AI-related companies

Impress's Taguchi (Jun): I think it will be important to utilize open source-based language models that can be controlled in-house, such as RAG and SLM, in the future. What kind of services does Macnica provide in this field?

Macnica Onishi: We provide a variety of data and AI infrastructure services, primarily Databricks. We also provide consulting and implementation support for creating in-house AI and resolving operational issues, and we offer a range of products to make it easier for customers to use generative AI while meeting a wide range of needs, including privacy and security.

Rohto Pharmaceutical Itabashi: As is well known, generative AI technology is evolving quickly, and I don't think it's an area where you can say, "Let's do this now." The general trend that people who buy more GPUs will be able to create better-performing LLMs is likely to continue for a few more years, but we need to think about what will happen beyond that. At that point, private AI that combines SLM and RAG will likely emerge.
ChatGPT is a very general and public AI, but while the scope of use for corporate AI is gradually narrowing as it is used in the form of group AI, the trend is that it is very useful. As AI evolves, the necessary tools and modules also change rapidly, so we are grateful that Macnica is able to introduce such things to us in a timely manner.

Digil Taguchi (Kei): Recently, Microsoft has also been working on this, but in the future, we will move in the direction of having an SLM model on each individual's computer, with a company-specific RAG and an agent specialized for the individual on their own device. With technology advancing at an incredible speed, I feel that companies themselves must think about how to use these things.

Rohto Pharmaceutical Co., Ltd.
Mr. Yuichi Itabashi
Joined Fujifilm as a chemical engineer in 1985. In R&D, he developed a groundbreaking digital color printing technology using microcapsules. To commercialize the technology, he moved from R&D to the Imaging Division, and in the midst of the crisis of losing his main business due to the digitization of photography, he worked on business transformation through commercialization and marketing of digital cameras and digital printing systems, and also led the revival of the Instax business. He then contributed to the company's management transformation using digital technology as head of the Digital Marketing Department and ICT Strategy Department. In 2021, he joined Rohto Pharmaceutical and promoted the company's digital transformation and use of AI.

Digil Co., Ltd.
Keiji Taguchi
He has led strategic IT/DX initiatives at a major telecommunications company, a foreign information security company, a major retail company, and a real estate group company. He is currently engaged in support consulting for companies that continue to take on challenges, focusing on corporate transformation, DX promotion, and in-house organization construction.

Impress Corporation.
Jun Taguchi
Joined Nikkei Business Publications in 1984. Worked as a Nikkei Computer journalist covering the field of corporate information systems. Served as editor-in-chief of Nikkei AI, Nikkei IT Professional, and Nikkei Computer. In 2008, left Nikkei Business Publications and moved to the Impress Group, where he launched IT Leaders, a media outlet specializing in IT. Currently works as editor-in-chief and producer of IT Leaders at Impress. Other roles include chairman of the IT Skills Research Forum, director of the Japan Data Management Consortium, director of the Business System Initiative Association, director of the IT Education Business Association, member of the "DX Brands" committee, and hometown ambassador for Himeji City (born in Himeji City, Hyogo Prefecture).

Macnica
Kosuke Onishi
Joined the company as a new graduate in 2005. Works in the Networks Data & Applications Division. Has been in charge of software products since joining the company. Has been involved with data since 2013 and AI since 2020, and currently, as deputy general manager, promotes business such as planning and market expansion of new products and services in the data and AI fields. Also discovers and investigates overseas venture companies that handle cutting-edge products that meet market needs.