GENIAC Community Kickoff Event!

Strengthening Development Capabilities of Foundation Models is the Key to Japan’s Innovation

The event started with a video message from Ken Saito, Minister of Economy, Trade and Industry. In the video, Minister Saito said that generative AI is a key technology for the next technical evolution, like cars, computers, and smartphones, and that it is important for the public and private sectors to strengthen their development capabilities and create sustainable technical evolution. He continued by saying “To promote new industries, the key is to foster collaboration among the parties involved and actively incorporate new ideas and different perspectives from outside, transcending the boundaries of organizations and national borders, rather than looking inward. I want you to take on challenges without fear of failure. I would like to extend my greetings to all of you at this event, hoping that you will go out into the world and become new leaders of society."

Next, Takuya Watanabe, Director of the Software and Information Services Strategy Office, Ministry of Economy, Trade and Industry, explained the GENIAC program. Japan is currently facing many challenges, including natural disasters, traffic accidents, food loss, economic disparity, long working hours, labor shortages, educational disparities, medical issues, and a growing digital deficit. AI is expected to help solve these issues. According to Watanabe, "Japan will safely use AI in a wide range of fields, continue to create innovation, and become a prosperous and proud country. To that end, the development capability of generative AI is crucial, and highly skilled software development personnel such as those gathered here are indispensable. The purpose of this program is to raise the level of Japan's development capability in generative AI and to create an environment in which each business entity can be creative and take on challenges in order to ensure the sustainable development of generative AI in Japan."

After that, Watanabe explained the importance of computational resources, data, and knowledge as elements for strengthening the development capabilities of the generated AIs, and that support will be provided for these elements.

Main support

1. Computing resources

   Supporting the procurement of GPUs and so forth that must be secured to develop a foundation model, which is the core technology of generative AI.

2. Data

   Promoting collaboration with users and other data holders through support programs for data utilization.

3. Knowledge

   Through events and communication tools, we promote exchanges between domestic and foreign developers and with various government system officials. An event featuring Google's foundation model developers is scheduled for late March to early April, with the aim of communicating with developers in other countries.

Gathering Domestic and International Knowledge in the Community to Improve Japan's Development Capabilities

The key to generative AI is the development of a foundation model. A foundation model is the core technological foundation that supports various services that utilize generative AI, and its development capability will determine the availability of generative AI in Japan and the scope of innovation to be created.

To drive the development of the foundation model throughout Japan, it is necessary to share knowledge beyond the boundaries between businesses that are essentially rivals. To this end, within the community, GENIAC will (1) provide computational resources, (2) support matching with companies and data holders, and (3) support collaboration with global tech companies and share knowledge among developers.

The community consists of two main groups.
In (A) A Community of Selected Companies, participants will discuss how to improve the development environment and how to efficiently conduct performance evaluations, with a focus on sharing development know-how within the provision of computational resources at GENIAC.
In (B) A Development Operator Community Including Non-Selected Companies, GENIAC aims to raise the level of foundation model development capabilities in Japan by widely sharing development know-how for foundation model development.

The three concepts of the GENIAC community are to create ways to share knowledge, create ways to form partnerships, and create ways to make rules. Each of these is an essential element in bringing about innovation from Japan.

Three concepts in the community

1. Creating ways to share knowledge

   Sharing information and knowledge among developers and actively disseminating information to the community.

2. Creating ways to form partnerships

   Holding matching events with companies that use AI and companies that develop applications based on the foundation model.

3. Creating ways to make rules

   Considering regulations and rules that will be hurdles in promoting social implementation and making recommendations.

Based on these three concepts, the GENIAC community will take the following initiatives.

Community Initiatives

1. Holding seminars inviting global tech companies and academia

   Inviting companies that are developing foundation models to share and discuss the outlook for the kind of foundation models they are developing.

2. Mutual support through sharing and discussion of know-how among developers

   Holding individual discussion forums and debriefing sessions to present the results including not only the successes obtained in the series of developments, but also the challenges and failures encountered.

3. Collecting information on best practices

   Using Slack to share and discuss knowledge and know-how. Accumulating information through those interactions as best practices and in Notion.

4. Matching developers with users, etc.

   Holding matching events with end users and developers in preparation for social implementation.

A Wide Variety of AI Development Efforts by Selected Businesses

After an explanation of the role of the community, representatives from the seven businesses selected for this year's program talked about their current initiatives and prospects.

ABEJA, Inc.

ABEJA provides comprehensive support for the promotion of DX by companies, from introduction to operation. It provides processes from data generation, collection, processing, analysis, and AI modeling necessary for DX. With the ABEJA PLATFORM at its core, it helps companies to transform and operate more efficiently.

According to Masafumi Kinoshita, General Manager, Corporate Strategy Management Department, "We hope to contribute to society by working together with all our employees to realize our company's management philosophy of creating a prosperous world. Our company is aiming to improve the accuracy of Japanese LLM as well as cost performance. By doing so, we hope to accelerate the implementation of LLM in society."

Sakana AI K.K.

Sakana AI was founded by top-level researchers who are well known in the field of deep learning. By linking many small-scale models, the company aims to develop adaptive AI that flexibly responds to the environment.

According to Takuya Akiba of Sakana AI, "In this project, we’d like to work on achieving high performance with a small model. Another thing we’d like to do is go one step further than something based on conversational text, so that the language model can actually act and solve problems on its own."

National Institute of Informatics /ROIS

According to Sadao Kurohashi, Professor at Kyoto University and Director of The National Institute of Informatics (NII), "We’re working on the construction of an open and Japanese-proficient GPT-3 large language model (LLM)." In May 2023, NII launched the LLM-jp (LLM Study Group), which is open to anyone who agrees with the purpose of the group, and it has been working to make all models, data, tools, technical materials, and so on available to the public, including the process of discussion and failures.

"This time, we’d like to take on the challenge of creating an LLM with 175 billion parameters, and we’ll disclose all the models we have built, source code, development know-how, and failures. We would also like to share our knowledge with companies and universities at study groups and contribute to raising the level of Japan's development capabilities," said Kurohashi.

Stockmark Inc.

Founded in 2016, Stockmark provides business decision-making services using natural language processing technology. The company released a 13 billion parameter Japanese LLM in October 2023.

According to Kosuke Arima, CTO of Stockmark, "Eliminating last mile hallucinations (in generative AI) is a very important issue. By thoroughly inculcating concepts that need to be differentiated as a human society, we will deter hallucinations and create a foundation model that can actually be used in work. We’d also like to create something that’s not only accurate, but also allows you to come up with your own ideas."

Turing Inc.

Turing is a start-up company that is developing EVs that are fully autonomous and EVs without steering wheels. The company is focusing on the development of multimodal AI that combines LLM and visual information.

"We want to create an AI that drives in the same way that humans drive. Like a brain and eyes that understand conditions as a human does. There are many rare situations (long tails) in the driving environment. Since there are limits to what can be handled on an algorithmic basis alone, this project will create a multi-modal AI that is particularly suited to the Japanese traffic environment," said Yu Yamaguchi, AI development officer at Turing.

The University of Tokyo

According to cofounder Yusuke Iwasawa, the Matsuo-Iwasawa Lab U Tokyo set a goal to create an ecosystem of technological powerhouses that can compete globally by cycling through basic research, lectures, social implementation, and incubation.

He went on to say that "In this project, we’ll work on developing an open foundation model that aims to improve a variety of Japanese language skills. There are two main points. One is to develop a variety of smaller foundation models ranging from 14 to 50 million parameters. The other is to select members from those who have completed the course and train those with experience in LLM development."

Preferred Elements Co., Ltd.

Preferred Elements, Inc. develops LLM and multimodal foundation models. The Preferred Networks Group, to which the company belongs, owns supercomputers and develops chips that serve as the basis for computing, aiming for a vertically integrated business structure that can provide everything from infrastructure to applications.

According to Daisuke Okanohara, Chief Executive Officer of Preferred Elements, "In this project, we will develop a multimodal foundation model with 100 billion parameters and aim to achieve the world's highest level of performance by supporting images and audio in addition to language. Another goal is to advance the development and validation of a 1 trillion parameter language model. We hope to contribute to raising the level of development capabilities by publishing pre-trained models and development know-how."

Deepening Friendships and Sharing Knowledge Through Theme Talks and Open Discussions

In the second half of the event, participants had the opportunity to interact. In the World Cafe Talk part, each table decided on a theme, and participants talked freely about topics of interest. A lively discussion took place at each table.

At the party that followed, the business representatives had more frank conversations with each other and deepened their relationships.

The GENIAC community will continue to carry out various activities to raise the level of AI technology in Japan. The day when world-class innovation will be born from Japan may be near.

GENIAC Top Page