I joined Preferred Networks in December 2016 and leaving after eight years of work. I’d like to take a look back at what I was involved in.
For me, PFN can be described in one word as “a place that built up my confidence as an engineer”.
I majored in physics in university and had no experience in software development, such as competitive programming, and had no achievements in software development. However, looking back on the eight years at PFN, I feel that I had the opportunity to experience various things and grow a lot through working with talented people.
Worked on various collaborative research projects
I was first assigned to the new team as 2nd member (meaning there was only one “team member” before I joined), and we collaborated with various partner companies, mainly in the manufacturing industry, on cutting-edge issues.
It was a time when Deep Learning was emerging rapidly and gaining attention, and it was fun and challenging to work on the forefront issues from the top companies in all industries in Japan. It was also interesting for me to develop software that interacts with the real world rather than the closed in the web.
The projects were carried out with short periods of milestone setting, and as a result, the PDCA cycle turns very fast. Continuation of the project due to success or discontinuation of the project due to verification of technical uncertainty frequently occurs, and I feel that this also contributed to my growth.
Above all, my colleagues were extremely talented, and it was always stimulating to work with them, discussing and working together as a team to solve difficult problems.
OSS release and team launch
I was interested in applying Deep Learning to the field of science rather than popular field such as images or language. At that time, I participated in the IPAB Drug Discovery Contest, where we applied our domain knowledge and the then-popular Graph Neural Network (GNN) technology to achieve the Grand Prix, and we released the GNN implementation used for the contest as an OSS called Chainer Chemistry.
It was a great confidence booster for me to release an OSS, and I remember how much fun it was to implement various Neural Networks and verify their accuracy at that time. These activities led to the launch of several projects in the field of chemistry, and I was also involved in the launch of a new team.
Commercialization and JV launch 〜 Global expansion 〜
Around 2017, a joint research project to develop PFP, a foundational model for general-purpose atomistic simulation, was launched. Achieving commercialization from a seed research with cutting-edge technology to a product was something I had always wanted to achieve, and although there were various difficulties, the project finally resulted in the product release of Matlantis after two years of joint research.
Currently, it has been introduced to nearly 100 organizations, and global expansion has begun. As for the current generative AI, many technologies are at the forefront in the US, but I had always wanted to challenge the world with a SaaS product from Japan, especially in the field of science. Matlantis is a product that fulfills this desire, and I was fortunate enough to be involved in this project, which has a high degree of social contribution and I sympathize with the mission of PFCC, “Achieve a sustainable world by enabling the creation of innovative materials and materials.”
Development of a domestic LLM
Recently, I was involved in the post-training of the PLaMo-100B model.
It was a valuable experience to train a model of this size, which was almost unprecedented in Japan at the time. Even though I felt the pressure for this high target goal to make it usable, but it was also a unique experience to develop a model with the cooperation of many people.
(->A service based on this development, PLaMo Prime, has been released. Please check!)
Conclusion
I feel that I have been a generalist who connects various industries with the field of Deep Learning, rather than a specialist with specific industry expertise.
I was able to move to various fields, depending on my motivation, and I am grateful that I was able to experience various things.
The company’s phase has changed now, such as the commercialization of Plant, the release of Preferred AI, the release of Misemise, and the sales of PFCP and MN-Core, and I feel that the gears have completely shifted to the commercialization phase.
I will support the company from the outside as a fan, and I am looking forward to seeing its further progress in the daily news.
I am really grateful that I was able to work at this great workplace and appreciate those who worked with me, thank you very much!
* The document is a translation of original Japanese blog, translated by PLaMo (and modified a bit manually).