A scientist tries to become a better CEO | WAVES

Text | Shi Jiaxiang

Editor | Liu Jing

Yuan Jinhui and OneFlow were once the luckiest ones after the explosion of ChatGPT. After Wang Huiwen posted his "call for heroes," OneFlow was acquired by Light Year Beyond. This six-year-old company suddenly became the team with the best resources at the table. But just two months later, everything disappeared dramatically.

After four months of ups and downs in the spotlight, Yuan Jinhui and the OneFlow team decided to start a new business. From having the most abundant resources in the industry to suddenly disappearing, while other players have reached high positions, the psychological impact is undoubtedly huge. But Yuan Jinhui said that if they still want to stay at the table, starting a new business is the only choice.

The reason is that after the news of Wang Huiwen's illness spread, some companies rushed to the Light Year Beyond office to poach people. The core members of OneFlow also received annual salaries of over ten million, but no company could keep everyone at the same time. Starting a business is the only way to avoid the team being dismembered.

Before Light Year Beyond, Yuan Jinhui's entrepreneurship also had its ups and downs. Different from most companies, OneFlow was like a laboratory at the beginning. It had no income in the first four years and relied on external investment. What they were doing sounded incredible at that time: to subvert the deep learning system framework heavily invested in by giants (such as Google's TensorFlow).

"Non-mainstream Entrepreneurship Notes," Yuan Jinhui named his WeChat official account when he started recording his entrepreneurial journey. This was a self-aware "atypical entrepreneurship."

Confronting giants and not thinking about short-term commercialization, this entrepreneurship, which seems almost impossible to succeed from a common-sense perspective, Yuan Jinhui persisted in for six years, relying on the thought of success to hold on. "I've experienced all the pain that one can experience." He also got a lot more white hair. "It's the price for this 'performance art,'" he joked.

Yuan Jinhui admitted that all the previous failures and attempts were self-inflicted. Before being acquired by Light Year Beyond, OneFlow was always struggling on the verge of survival. In the most difficult times, they chose to cut salaries and lay off employees. He later added that OneFlow had been in such a state for six or seven years.

But this time, Yuan Jinhui hopes to give an account to those who continue to start a business with him. They believe that although AI infrastructure is not as competitive as models or chips, there is still a chance to reach for the stars. Slightly different from before, the focus of the new company has shifted from training to inference deployment.

In just a few months, he changed four companies, from OneFlow to Light Year Beyond to Meituan, and finally left Meituan to found a new company. Yuan Jinhui said that he didn't really recover until this year.

A scientist, after experiencing the baptism of technological changes and the business world, begins to try to be a better CEO.

The Calculated Plan

2012 was an important year for Deep Learning. That year, the Imagenet dataset exploded, and deep learning proved for the first time that in object recognition, machines could surpass humans.

Two years later, deep learning continued to develop in fields other than images. At that time, the mainstream models only had tens of millions of parameters. Yuan Jinhui, who was still working at Microsoft Research Asia, judged that future models would become very large, and the original architecture would be reshaped.

This forward-looking judgment was based on a solid foundation. In the second year after Yuan Jinhui joined Microsoft Research Asia, he developed LightLDA. He cleaned 15 billion high-quality web pages from the full web library and used 200 billion tokens from them to train an LDA model with 1 trillion parameters. This was one of the largest models at that time. Looking back further, when he was studying computational neuroscience at Tsinghua University, he deeply studied brain science. The number of synaptic connections in the human brain far exceeded the deep learning models at that time. Perhaps, scale was the key to human intelligence.

At that time, the existing deep learning systems were not designed for large models. Yuan Jinhui thought, "When large models become a reality and we have the best 'hammer,' not to mention making money, from a technical perspective, it's enough to make a name for ourselves." One morning, he suddenly woke up and felt that he had to start a business. "Once this thought occurred, I couldn't do anything else."

OneFlow was founded in such a context. At first, it was a "friends and family" startup. The investors included Yuan Jinhui's relatives, friends (including his former colleague Su Hua and Su Hua's businessman brother). Later, it successively received support from Jihe, THORS, Kuaishou, the alumni fund, Hillhouse, and the Haidian District.

Before the birth of large models, the most important technological opportunity in Yuan Jinhui's eyes was the framework. "The models above change every day, but the framework below is stable." OneFlow was one of the earliest teams in the world to believe that models would become large. All the system architectures they studied were based on the assumption that models would grow in size.

He was solely focused on making the best "hammer" without thinking about where the "nails" would be. In the first four years, OneFlow had no commercial income. Its technical goals surpassed its commercial pursuits, and its operation style was like a laboratory. Most of the work was writing code. The characteristics of the employees can be summarized as: loving only to write code and discuss technical issues. They were lively when discussing technical problems, and nothing else could excite them. Except for money, of course.

Yuan Jinhui compared his then commercial vision to a "calculated plan" - to create the fastest and most user-friendly framework, capture the developer ecosystem, and then pursue commercialization.

Put simply, OneFlow tried to find a more "elegant" architecture to solve all problems once and for all: any deep learning model (even if it's not a Transformer), no matter how large it becomes, can run on the developed architecture. The name OneFlow also reflects their pursuit.

In 2020, Zhongguancun Guide interviewed Yuan Jinhui and evaluated that OneFlow had lofty goals but was not a Quixotic delusion. But Yuan Jinhui described his state at that time as "fighting against windmills."

OneFlow kept its work secret for six years. Few people believed in it, and no one came to compete. They also thought that if they built a good tool but no one made the models large, they would do it themselves.

But this remained just an idea. By the end of 2022, OneFlow was on the verge of survival, with access to at most a few hundred GPUs. The difficult situation was not caused by an accidental black swan event but was the result of accumulated problems. At that time, the domestic investment environment was in a downturn, and the next round of financing was scarce. In the last two years, the company had to obtain income by doing projects, but these projects also focused on framework research and development. "It couldn't really be called commercialization." Around the time of ChatGPT's release, some core employees received offers from large companies and chose to leave.

This was almost the darkest moment for Yuan Jinhui. That year, he hardly had any weekends. In the most difficult times, he could only choose to cut salaries and lay off employees.

From the perspective of technological progress, OneFlow's exploration was undoubtedly a success. They were the first to make distributed multi-GPU programming as easy as programming on a single GPU. The distributed concepts and designs they developed were also followed by today's mainstream deep learning frameworks.

But the real revolutionary was OpenAI later. Even though the "hammer" they used was just patching the existing architecture and could only support the Transformer model, the actual result was that the Transformer was so effective that the revolution happened faster.

"Whether it's the caprice of fate or the law of society, making more people directly see the value will make things happen faster, rather than our pursuit of elegance or even some personal quirks." Yuan Jinhui mentioned this statement to Waves several times: "OneFlow's experience is a bit like a performance art."

Four Months of Peaks and Valleys

At the end of 2022, the unexpected explosion of ChatGPT pushed OneFlow off its originally envisioned track. Yuan Jinhui was already one of the Chinese scientists closest to the progress of large models. He knew that Professor Tang Jie (Yang Zhilin's teacher), Google, and OpenAI were all working towards large models, but he still underestimated the speed of change. "I believed it would happen, but maybe it still needed some time, and we still had time," Yuan Jinhui said.

Joy and hesitation welled up in his heart at the same time. On the one hand, the judgment he had believed in for years finally came true. Large models had changed from being against the consensus to becoming a super-consensus. On the other hand, the market consensus on AI formed so quickly that it would soon enter an arms race of competing for resources, financing, and brand recognition - competing to be the first to have the most money, computing power, and GPUs - and the opportunity might not belong to OneFlow, which had almost given everything (and after the model structure converged to the Transformer, a general solution was no longer a necessity).

Almost at the same time, Yuan Jinhui saw Wang Huiwen's "call for heroes" with $50 million in capital. He thought that OneFlow might catch his eye. Yuan Jinhui met with Wang Huiwen. At the dinner, Wang Huiwen told him that he wanted to form a new team. At first, the cooperation didn't succeed.

Soon, an intermediary facilitated the two parties to return to the negotiation table. Within a month, the acquisition was completed, and OneFlow moved from Tsinghua Tongfang Science Plaza to Sohu Network Building.

Meanwhile, the valuation of Light Year Beyond exceeded $1 billion, and it suddenly became the team with the best resources at the table.

But just two months later, Wang Huiwen took sick leave and left his post, and Meituan took over Light Year Beyond. There was so much unwillingness to give up entrepreneurship when the wave was coming. After only a month of hesitation, Yuan Jinhui decided to start a new business.

This was also the only way to keep the original team together. After the news of the turmoil at Light Year Beyond spread, some companies rushed to the office to poach people, and some OneFlow members received job offers with annual salaries of over ten million. In this situation, no company could offer an incentive to keep everyone. But if they continued to start a business, most people would stay together.

On the first anniversary of the founding of SiliconFlow, they described their state at that time as: like a helmsman struggling to control a ship in the surging waves. After the course was set, there was no time to catch a breath, and they even held their breath and plunged into a new battlefield again.

Yuan Jinhui told the team that if they could survive such a major crisis without falling apart, there would be no difficulty they couldn't overcome in the future.

Consensus and Open Cards

Yuan Jinhui missed two opportunities to join giants in the early stage. The first was in early 2013 when Zhang Yiming invited him to join Toutiao, but he said he wanted to go to Microsoft Research Asia more. The second was when Su Hua invited him to join Kuaishou, and he also chose to stay at Microsoft because "he still had some ideas in technology."

A few years ago, he was more focused on the success of OneFlow rather than the success of the company. And in such a way, "there was only one way out, which was acquisition; otherwise, it would fail."

But this time is different. "My technological curiosity was satisfied in the previous stage, so the unmet and unfinished pursuits are mainly at the commercial level," Yuan Jinhui said in an interview.

During the Q&A session of a sharing meeting, a viewer asked Yuan Jinhui if he could propose another technological idea like OneFlow that would happen in the next few years but that most people would disagree with. Yuan Jinhui said that he was ashamed because he now spent most of his time on non-technical matters and couldn't come up with such an idea anymore.

The name of the new company was inspired by ChatGPT: SiliconFlow. Different from OneFlow, which blindly tried to subvert the existing framework, SiliconFlow works backward from the business perspective and more pragmatically chooses a large model inference framework that is not as technically exciting as OneFlow but has greater market potential: if the data volume scanned by the training framework is like a swimming pool, then the data volume scanned by the inference framework is like an endless river. Different from OneFlow, this is a path with a high level of consensus.

The large models have given rise to the product form of Model as a service (MaaS). As long as services can be provided according to standard APIs, the originally underlying and difficult-to-commercialize technologies can be quickly monetized. The progress of OpenAI in inference has verified Yuan Jinhui's judgment on the inference framework. "The computing power demand for inference has increased by an order of magnitude."

During the four months in the eye of the storm at Light Year Beyond, Yuan Jinhui quickly made up for the lessons OneFlow had missed - this is an experience that can only be gained through actual battles. The company's scale expanded from 35 people to over 60 people, with the addition of product managers and colleagues in charge of commercialization. SiliconFlow now has three co-founders with commercial backgrounds.

In the first few years of starting a business, Yuan Jinhui still wrote code himself. Later, he stopped because he found that "it's irresponsible for a CEO to write code." He has become more realistic, practical, and more like a CEO. He doesn't seek for a stroke of genius but only hopes for a successful outcome through a well - thought - out plan.

Yuan Jinhui later thought that OneFlow was a gamble and an experiment from the start. Fortunately, he gave an account to the investors, and the employees who left after the acquisition cashed in about 20 million in stocks. But the colleagues who never left and started a new business with him still haven't received any economic returns. There are things to account for in the past and lessons from being too idealistic and impractical.

When he started his first business, Yuan Jinhui named his WeChat official account "Non - mainstream Entrepreneurship Notes." Different from most entrepreneurs who talk about application scenarios and implementation, he told Waves that he was not interested in anything other than researching new frameworks at that time - he only loved solving difficult problems and didn't want to do ordinary things. "I just thought this was the most exciting thing to do, and the more unexpected and surprising, the better."

At that sharing meeting, Yuan Jinhui reviewed his seven - year "atypical entrepreneurship." In the end, he said, "Now, there is an urgent need for a typical entrepreneurial success."