AI Across Industries: The Current Status and Case Studies from Japan 岡野原大輔 (PFN)

Key Concepts

• Preferred Networks (PFN): A Japanese AI company founded in 2014, now a unicorn, focusing on advancing AI for practical applications across industries.
• Matlantis: PFN’s subsidiary specializing in an AI-driven universal atomistic simulator for materials discovery and development.
• PLaMo: PFN’s large language model (LLM).
• AI Chip Development: PFN’s investment in developing its own AI-specific chips, focusing on inference capabilities.
• Vertical Integration: PFN’s strategy of controlling the entire AI stack, from chip development to application solutions.
• Domain-Specific AI: Focusing AI development on niche industrial applications where Japan has a competitive advantage.
• AI Inference vs. Training: Understanding the differing computational requirements and priorities for AI inference (application) versus AI training (model development).
• Code Sovereignty: The desire for Japan to develop its own AI models to maintain control over data and intellectual property.
• Japanese Semiconductor Landscape: The current state of Japan’s semiconductor industry and PFN’s position within it.

The Current Status and Case Studies of AI from Japan: A Summary of Dr. Daisuke Okanohara’s Presentation

Introduction & CJEB Overview

Dr. Daisuke Okanohara, Co-Founder and CEO of Preferred Networks (PFN), was introduced by David Weinstein, Director of the Center on Japanese Economy and Business (CJEB) at Columbia Business School. CJEB, established in 1986, is the only North American research center dedicated to the Japanese economy and operates solely on donations. The presentation focused on the current status of AI across industries, with specific case studies from PFN and Matlantis, and a perspective on the future of AI in Japan and globally.

Dr. Okanohara & Preferred Networks (PFN)

Dr. Okanohara co-founded PFN in 2014 with a focus on translating AI research into practical applications. He recently transitioned to CEO from CTO on November 26th. PFN distinguishes itself as one of Japan’s 11 unicorns. Dr. Okanohara is also a member of the study group on guidelines for AI business in Japan, contributing to the development of safe and appropriate AI usage guidelines. He is a prolific author in Japanese on AI topics, including diffusion models and symmetry in AI.

PFN’s Business Model & Strategy

PFN’s mission is to make the real world “computable” through computing technologies, sensing, analysis, and automation. Unlike typical startups, PFN was initially funded entirely by business partners (Toyota, FANUC, Kodansha, Toei, and others) rather than venture capitalists. This has shaped its diversified business structure, encompassing four layers: AI chips & computing infrastructure, foundation models, AI solutions, and AI products. This structure evolved organically over ten years, stemming from initial collaborations in autonomous driving, drug discovery, and robotics. PFN’s strategy is not to focus on a single business but to integrate these four layers for competitive advantage.

AI Semiconductor Development

Recognizing the increasing importance of computational power for AI, PFN began developing its own AI chips in 2016. The company is now focused on AI inference chips, as demand for inference is growing at 80 times the annual rate compared to 4.6 times for training. Dr. Okanohara highlighted the key differences between training and inference: training prioritizes computing power and network density, while inference prioritizes memory speed, cost-effectiveness, and simpler software requirements. PFN is pursuing a novel 3D stacked DRAM on logic chip architecture, aiming to provide fast inference on standard devices rather than requiring large data centers. PFN aims to be a fabless chip designer and AI cloud data center provider within the Japanese semiconductor landscape, which currently lacks a startup presence in these areas.

Large Language Models (LLMs) & PLaMo

Japan is actively developing LLMs domestically, with PFN being a key player. The motivation is to achieve “code sovereignty” – the ability to control AI models and data, particularly for sensitive applications like defense and government operations. PFN is collaborating with the Japanese government to develop LLMs using government data. PFN’s LLM is called PLaMo.

Matlantis: AI for Materials Discovery

Matlantis, a PFN subsidiary, is a rapidly growing business focused on using AI to accelerate materials discovery. Traditional materials research relies heavily on experimentation, but Matlantis’s AI-powered simulator can reduce simulation times from weeks to milliseconds. This allows for rapid testing of new materials. Matlantis is analogous to DeepMind’s AlphaFold (for protein folding) but focuses on materials science. Current applications include battery development, semiconductor process optimization, and the discovery of new Metal-Organic Frameworks (MOFs). Matlantis’s success is attributed to Japan’s concentrated industrial ecosystem, allowing for close collaboration with key customers like Toyota and Tokyo Electron. Customers are currently simulating several trillion atoms per month using Matlantis.

Plant Automation & Robotics

PFN also develops AI solutions for plant automation, exemplified by its work with Eneos Kawasaki refinery. The AI system learns to optimize plant operations, even adapting to changes in crude oil composition. PFN’s robotics division produces the “Kachaka” robot, initially targeted for personal use but now finding success in industrial, clinical, and retail settings due to labor shortages, particularly in rural Japan.

Japan’s AI Strategy & Future Outlook

Dr. Okanohara outlined Japan’s AI strategy, focusing on addressing labor shortages, supporting existing industries, and establishing a neutral position in the global AI landscape. He emphasized the need to find niche areas where Japan can compete with the US and China. He also highlighted the importance of power efficiency in AI development, given Japan’s limited energy resources. He noted the challenges of raising venture capital in Japan compared to Silicon Valley.

Concluding Remarks

The presentation concluded with a Q&A session. Dr. Okanohara emphasized PFN’s unique position, resembling a combination of Palantir and a semiconductor company. He also discussed the importance of translation AI as a potential area for Japanese innovation. The event was closed with thanks to Dr. Okanohara and CJEB’s sponsors.