The First Step Towards Social Implementation of Generative AI!
Matching Event Held Between GENIAC Developers and User Companies

Promoting Competitive Foundation Model Development and Social Implementation

Takuya Watanabe, Director of the Software and Information Service Strategy Office, Ministry of Economy, Trade, and Industry, explained the event's purpose, the activities of GENIAC since its inception, and future initiatives.

"Today is a matching event between foundation model development companies and potential application development/user companies. We encourage developers to showcase their strengths to potential customers and use this opportunity to gather feedback for future improvements. We also expect user companies to explore new solutions and integrate them into their operations. Through opinion exchanges, we aim to promote the development and social implementation of competitive foundation models," said Watanabe.

GENIAC's Next Actions

Watanabe announced the following three new initiatives for GENIAC.

Next Action

1. Support for Data and AI Utilization

   "There are various challenges from AI development to utilization. To address these, we plan to conduct representative projects and widely share their results. We will solicit and implement projects in three categories: building datasets, collaborating with data holders, and creating leading AI utilization examples."

2. Support for Providing Computational Resources

   "Since February 15, the development of foundation models by seven companies receiving computational resource procurement support has begun sequentially. In late May, additional development funds were allocated for computational resource procurement, and this development is expected to be completed by mid-August. We plan to provide new computational resource support around October, prioritizing developments focused on social implementation. Detailed requirements will be shared later."

3. Generative AI Hackathon

   "We plan to develop AI applications using unique data and resources like ABCI from the National Institute of Advanced Industrial Science and Technology in a hackathon format. We hope that many talented individuals will gather and create various AI applications, leading to a virtuous cycle where new data generated is utilized for further AI applications."

Introducing GENIAC Developers

The following six developers participated in this event:

• ABEJA Inc.
• Nippon Telegraph and Telephone Corporation (NTT)
• StockMark Inc.
• Preferred Elements Inc.
• Ricoh Company, Ltd.
• Research Organization of Information and Systems (National Institute of Informatics)

ABEJA Inc.

ABEJA Inc. offers the "ABEJA Platform," a suite of software for the continuous and stable operation of generating, collecting, processing, and analyzing data necessary for DX in companies, as well as AI modeling. In addition to the transformation area that creates systems, it also supports the operational area to assist with ongoing management. Emphasizing social implementation, the company aims to enhance RAG and Agent functions to address issues related to information quality and cost performance.

Nippon Telegraph and Telephone Corporation (NTT)

NTT provides "tsuzumi," a lightweight Japanese-language large language model (LLM) with 7 billion parameters developed in-house. It is highly customizable due to its low-cost tuning capabilities and features multimodality, allowing it to handle formats such as chart and table interpretation. It supports three usage environments (on-premise, private cloud, public cloud) and can be combined with three solution menus (CX solutions, industry-specific EX solutions, IT operations support solutions).

StockMark Inc.

StockMark Inc. provides information collection and document preparation support services for companies. The company is working on developing a highly reliable LLM that suppresses hallucinations, which is a challenge for companies using generative AI. Compared to ChatGPT, which has an accuracy rate of 40% on a business domain question set, StockMark's LLM achieves an accuracy rate of 90%. It can clearly output "unable to answer" for questions it cannot answer, providing a highly reliable LLM.

Preferred Elements Inc.

Preferred Elements (PFE) and its parent company Preferred Networks (PFN Group) integrates the AI value chain from chips, computing infrastructure, generative AI/foundation models to solutions and products for industrial applications. In September 2023, they released PLaMo-13B, a large language model with 13 billion parameters for research and commercial use. Currently within GENIAC, they are working on large-scale multimodal foundation models with 100 billion and 1 trillion parameters, with plans to release PLaMo-100B around fall 2024.

Ricoh Company Ltd.

Ricoh has been developing AI since before deep learning and created a 6 billion parameter Japanese LLM in 2022. They successfully developed a 13 billion parameter model strong in Japanese and apply custom LLMs in their maintenance services.

National Institute of Informatics (NII)

NII exhibited at the event, conducting R&D to ensure the transparency and reliability of LLMs. They plan to release a 175 billion parameter LLM with high Japanese language performance, conducting various evaluations and verifications.

Active Opinion Exchange During Matching Time

After the presentations, a one-hour matching time was held where participating companies actively visited each booth and exchanged opinions with developers.

Comments from GENIAC Developers and Participating Companies

Finally, here are the comments from the developers and participating companies who took part in the matching event which marked the first step towards social implementation.

"It is said that 65% of white-collar work is allocated to information gathering and document preparation. The labor cost is 300 billion yen per day and 6 trillion yen per month. Despite such enormous effort, the utilization of generative AI has not yet progressed significantly. I believe the reason lies in the issue of hallucinations. We aim to create high-precision AI that suppresses hallucinations, but for social implementation, it is necessary to create 'killer use cases' that genuinely demonstrate the convenience of use through collaboration between industry, government, and academia. I participated in this event with the expectation that it would be the first step toward that goal."

"At this event, we were able to have in-depth discussions with companies that are actually using AI or developing services using AI. We received positive feedback for creating competitive models from scratch at PFE, with many expressing interest in trying it out, which gave us confidence. It seemed that each company was exploring how to use AI in their specific business. Therefore, at PFN Group, we want to focus not only on providing foundation models but on creating actual use cases."

"We already have examples of using tuned open-source models, and we participated in this event with the hope of utilizing the foundation models developed by domestic companies in the future. While I believe the U.S. is more advanced in some aspects, I have high expectations for the many companies that the Ministry of Economy, Trade, and Industry is supporting to develop LLMs themselves. Although it is still in the early stages, I hope such initiatives will not end after one attempt but will continue repeatedly, thereby boosting the AI field in Japan."

"Given that we are currently promoting the use of generative AI across the company, we participated in the hope of finding services or language models that we could use. My honest impression is that the results exceeded my expectations. It was a great opportunity to understand how cutting-edge companies are moving in the field of foundation models. If there is an opportunity to participate in the GENIAC community as a user company, we would love to be actively involved."

The matching event, the first of its kind by GENIAC, provided valuable time for both developers and application/user companies towards social implementation. The day when we see services and solutions utilizing the foundation models developed by the developers may not be far off. Please continue to pay attention to GENIAC's future activities.

GENIAC Top Page