I Found Something More Fun Than Gaming
He turned down a top-school postdoc and offers from leading AI labs to start his own company. From a gaming addict who nearly flunked out in his sophomore year to the founder behind one of the earliest multi-agent open source frameworks, he says agent research was the first time he found something more fun than gaming.
One of the earlier open source multi-agent frameworks built around large language models. The core code was written in one night.
L is the founder of an AI startup overseas. In 2023, when GPT-3.5-Turbo had just come out, he judged that large models were the future and sat down to write this framework. Later he brought in two junior schoolmates to round out the research, and they published a paper.
The concept of agents has actually been around for a long time. As early as Minsky’s The Society of Mind, the mind was already understood as a system made up of many simple agents.
We talked about why he chose agents as his direction in the first place.
“I was a gaming addict growing up,” he said. “I chose to do a PhD because agents were incredibly good at games back then. Using reinforcement learning to train an agent that played better than I did, that fascinated me.”
In those years, DeepMind was using agents to play StarCraft, and OpenAI was using them to play Dota. L himself was deep into Dota. He played more than 5,000 matches in college, but only at an average level, barely reaching 1900 on the ladder. At this point he laughed at himself: he didn’t have much talent for gaming back then, while quite a few friends around him played at near-professional level, 2100.
But the addiction was real. In his sophomore year, he failed several courses and nearly got expelled.
In his junior year, after a brief and bruising attempt at entrepreneurship, he decided to go to graduate school.
When he applied to college, he picked telecommunications, not computer science, because telecom’s admission score was higher than CS’s. For graduate school he applied within the same major and got into a research institute, where telecom and computer science sat in the same school. As an undergraduate he skipped classes, sank into games, and barely studied; it was only in graduate school that he wrote his first line of serious code.
His introduction to AI came from a very famous computer vision scholar. One course this professor taught got L hooked on deep learning: the slides came from Stanford’s Fei-Fei Li, the teaching assistant who made them was Andrej Karpathy, and the course was CS231n, Convolutional Neural Networks for Visual Recognition. It was the most cutting-edge course of its day, and the assignments required writing every model layer by hand. L put his final project answers on GitHub, and people still reference them today.
That was ten years ago.
During graduate school, L sent an email to the founder of an AI unicorn.
A cold email from a graduate student earned him an interview. The interview went well, and he interned there for several months doing reinforcement learning. The atmosphere was good, and the project shipped successfully.
Then he left.
The main reason was that the project’s results could not be published as papers. His undergraduate GPA was low, and if he wanted to apply for a PhD, papers were his only way in.
L applied to a university overseas, first interning in the professor’s lab for half a year, then applying to stay on as a PhD student. The school’s committee initially rejected him because of his GPA. It was his advisor who wrote two crucial recommendation letters and pulled him back in.
All of these opportunities, the email, the interview, the internship, the two recommendation letters, he went out and got himself.
L says that before he started doing research, his life didn’t really have a goal. Agent research was the first thing he ever actively wanted to do, and the first time he found something more fun than gaming.
Then came an internship during his PhD. He discovered that pretrained language models could generalize to robotics tasks without any additional training.
“That was the first time I saw the generalization power of language models,” L said.
Later, ChatGPT burst onto the scene. It was also during this period that he wrote the framework from the beginning of this story.
After finishing his PhD, L received a postdoc offer from a top university in the United States. But because of the research direction and some personal reasons, he ultimately didn’t go.
Several leading large-model companies reached out to him around that time, and investors were pushing him too. In the end, he started his own company.
I asked him whether he regretted it. Joining at that moment could have made him one of those companies’ earliest employees.
“Of course not,” he said. “Our work already has real influence in the open source community, and our research has earned commercial recognition. That was my intention from the start.”
As a founder, L is free to choose his research direction and his company’s focus. And choosing research directions is exactly what he was good at during his PhD. In his own words: “I’m fairly good at defining problems.”
I asked him what he had learned from running an AI startup.
He said: you need conviction, you need a vision, and you have to deliver results.
At the end of our conversation, I asked him: what do you think AI agents are still missing today?
Autonomy, he said. In his view, the agents of the future will be a vast, complex system woven into human work and life. And autonomy means an agent can take on longer-horizon work: days, weeks, even months.
I pressed further: if models can handle work over such long horizons, without any human involvement in between, what will be left for people to do?
“Harder work. Or more interesting work.”