What I Thought About When Founding This Company

Photo: A view of the River Cam flowing north-south on the west side of Cambridge (taken in May 2024)

Introduction I'm Kobayashi, CEO of LangCloud Technologies. This is our first tech blog since founding the company. Since this is our first post, I'd like to introduce the direction of the technology development we are pursuing as a company, as well as the background leading up to founding the company and what I was thinking about at the time.

The Shock of Open-Source LLMs Running on a PC As you may know, OpenAI announced ChatGPT on November 30, 2022, and it will soon be two years since then. I had vaguely sensed the progress of generative AI myself, but the emergence of ChatGPT-level AI truly came as a bolt from the blue. I imagine many people felt the same way.

As ChatGPT fever spread around the world, I couldn't just sit still either. A few months after the announcement, I dug out an aging gaming PC that had been gathering dust in the corner of my room (an ASUS laptop with a GTX 1050 Ti) and started running some of the open-source large language models (LLMs) that were beginning to circulate. The GPU was underpowered, so the models responded very slowly, but as I experimented with continued pre-training and fine-tuning on my PC, I was struck by how significant this was becoming. I was especially amazed when LLaVA, a multimodal vision-language model, ran on my own computer.

Amid all this, coinciding with changes in my own life stage, I left my corporate job in November 2023, and on January 22, 2024, I finally founded the company I had long dreamed of. In starting this business, the emergence of ChatGPT was certainly a factor, but there is no doubt that the excellent open-source LLMs that kept popping up like bamboo shoots strongly pushed me forward. Below, I'd like to dig into what I was thinking at the time of founding the company, including these points.

In Search of a New Form of Work Where People and AI Collaborate Looking back, for a little over ten years until last year, I worked as a data scientist in the early days of the field, exploring new ways to leverage data with cutting-edge data science and AI alongside many companies, mainly in manufacturing and finance. My success rate on projects was relatively high, but naturally there were projects that went well and others that didn't, due to conditions not being right. Recently, no-code/low-code data analysis tools have started to spread, greatly advancing the democratization of AI, but even as tools become easier to use, data analysis and AI development can't easily reach their goals unless various conditions are properly in place. For example, various obstacles arise, such as being unable to concretize an approach due to a lack of domain knowledge, trying to solve a problem that isn't actually solvable, or lacking the necessary data.

To address such issues, there have long been calls to train data scientists and data analysts, and I myself have supported the development of numerous data analysis practitioners. Looking back on the state of DX promotion in society over the past decade, no-code/low-code tools have certainly begun to help in the field, but fundamentally, the need for data analysis practitioners on the ground hasn't changed much. People who can properly untangle problems and give the right direction on-site have always been, and remain, invaluable.

With this awareness of the issues, the emergence of ChatGPT struck me as an enormous game changer. A new form of "workplace" where AI with language communication skills works collaboratively with people as an expert or colleague. I realized this could apply not just to data analysis but to any workplace or job site, and that it had the potential to change the very concept of work and labor as we know it. Of course, training human experts is still necessary, but at the same time, I began to envision a future where excellent AI is built and deployed on-site, where people and AI collaborate to carry out work and solve problems.

What seemed like science fiction just a few years ago is now becoming increasingly achievable thanks to the evolution of generative AI technology, exemplified by ChatGPT. Together with a desire to help, even a little, Japan's labor market as it faces an unprecedented labor shortage, my determination to pursue and create new forms of collaboration between people and AI at work became a major impetus and motivation for founding the company.

Open Source, Edge, Multimodal, Agents Today, a variety of generative AI services are beginning to be offered. From OpenAI's ChatGPT, to Microsoft's various Copilot services, to cloud API services like AWS's Amazon Bedrock and Google Gemini, to Meta which continues to develop and provide open-source large language models — the generative AI business among big tech companies is truly a battleground of rival factions. Countless open-source models continue to be uploaded to Hugging Face. In Japan, IT companies are engaged in fierce competition to develop LLMs, many of which are offered to the market as open source available for commercial use. Looking at the most recent trends in the generative AI industry, we can also see the main battlefield beginning to shift from conversational AI to autonomous AI known as AI agents.

Given this landscape, the direction of the technology development we pursue as a company can be summarized in four keywords: "open source," "edge," "multimodal," and "agent."

• "Edge" is nearly synonymous with "local," referring to edge AI that runs on PCs or on-premises servers, as typified by local LLMs. The overall system architecture in which edge AI and various other components operate is what a local network built on a private cloud aims to achieve.
• The reason we're committed to "multimodal" is that in the real world — as typified by workplaces — we need to handle not just verbal communication but also data via vision, hearing, touch, and other IoT- or robot-based interfaces.
• "Agents," currently in vogue, are being turned into services by big tech and major IT companies alike, but we look a bit further ahead and focus on agents operating in the real world. Network-type multi-agent systems, in which multiple edge AIs and cloud AIs communicate with each other over a network to carry out work, are already an established concept, but we want to find real-world solutions built on that idea.
• Perhaps the most essential keyword among these is "open source." Around the world, many talented AI developers form communities, developing excellent open-source software day and night, exchanging it with one another, and pushing each other to improve. There is debate about how important Meta's open-source strategy for AI adoption is compared to the closed strategies of other big tech companies, but for our part, we believe it's important to walk alongside the vibrant, wide-open global open-source AI community and, first and foremost, keep up technologically without falling behind. Furthermore, from the perspective of contributing to society, we believe it's very important to provide low-cost, high-quality AI solutions built on open source to many small and medium-sized enterprises in Japan, to help power their DX efforts.

With that, I'll conclude our first tech blog post. Thank you for reading this far.

• Two other themes I was thinking about when founding the company — predicting future risks to society and the environment, and the growing importance of AI safety — I hope to address in future blog posts as the timing allows.