A Scientist's Attempt to Become a Better CEO | WAVES

By Jiaxiang Shi

Edited by Jing Liu

Jinhui Yuan and OneFlow were, for a moment, the luckiest players in the post-ChatGPT gold rush. After Wang Huiwen posted his "hero's call," OneFlow was acquired by Lightyear Beyond — catapulting the six-year-old company to the best-resourced team at the table. But just two months later, it all dramatically evaporated.

After four months of white-knuckle highs and lows under the spotlight, Yuan and the OneFlow team decided to start over. Going from having the most abundant resources in the industry to suddenly having nothing, while other players had already secured their positions — the psychological blow was undeniable. But Yuan said that if they still wanted a seat at the table, re-starting was the only option.

The reason: after news broke of Wang Huiwen's illness, companies rushed to Lightyear Beyond's offices to poach talent. OneFlow's core members received offers exceeding 10 million yuan in annual compensation, but no single company could retain them all. Starting a company was the only way to prevent the team from being dismembered.

Before Lightyear Beyond, Yuan's entrepreneurial journey had already been turbulent. Unlike most startups, OneFlow began as something resembling a lab — no revenue for the first four years, sustained entirely by outside investment. What they were working on sounded absurd at the time: disrupting the deep learning frameworks that tech giants had poured massive resources into building (such as Google's TensorFlow).

"Notes from a Non-Mainstream Entrepreneur" — that's what Yuan named his public account when he started documenting the journey. It was a self-aware "atypical startup."

Battling giants while ignoring short-term commercialization — a venture that by any conventional wisdom seemed destined to fail. Yuan persisted for six years, sustained by sheer conviction that it could work, "experiencing every possible form of suffering." His gray hair multiplied. "The price of performance art," he joked.

Yuan candidly admits that his past failures were self-inflicted. Before the Lightyear Beyond acquisition, OneFlow perpetually teetered on the edge of survival, resorting to pay cuts and layoffs during its darkest moments. He later added that this had been OneFlow's state for nearly six or seven years.

But this time, Yuan hopes to deliver for those who continue the journey with him. They believe that while AI infrastructure may not attract the same frenzy as models or chips, it too offers a path to something vast. The difference: the new company's focus has shifted from training to inference deployment.

In just a few months, he changed companies four times — from OneFlow to Lightyear Beyond to Meituan, then finally leaving Meituan to start anew. Yuan says he only began to recover this year.

A scientist, after weathering technological upheaval and the trials of the business world, is now trying to become a better CEO.

Wishful Thinking

2012 was a pivotal year for deep learning. That year, ImageNet exploded onto the scene — for the first time, machines proved they could surpass humans at object recognition.

Two years later, deep learning continued advancing beyond images. At the time, mainstream models had only tens of millions of parameters. Working at Microsoft Research Asia, Yuan made a prescient call: models would grow enormous, and existing architectures would be reshaped.

This forward-looking judgment rested on solid ground. In his second year at Microsoft Research Asia, Yuan built LightLDA — cleaning 15 billion high-quality web pages from the full web corpus, using 200 billion tokens to train a one-trillion-parameter LDA model. It was among the largest models of its era. Looking further back, his doctoral work at Tsinghua in computational neuroscience had immersed him in brain science: the human brain's synaptic connections vastly outnumbered then-current deep learning models. Perhaps scale was the key to unlocking human intelligence.

Existing deep learning systems weren't designed for large models. Yuan thought, "When large models become reality, having the best hammer — forget about making money — from a technical standpoint, that's enough to make your name." He woke up one morning with the sudden conviction that he had to start a company. "Once that idea took hold, I couldn't do anything else."

OneFlow was born in this context. It began as a "friends-and-family" venture — investors included Yuan's personal network (including former colleague Kuaishou co-founder Su Hua and his business-oriented younger brother), later expanding to include Joy Capital, TRS, Kuaishou, an alumni fund, Hillhouse Capital, and Haidian District government support.

Before large language models emerged, the framework was the most important technical opportunity in Yuan's eyes. "The models above change daily, but the framework below is stable." OneFlow was among the earliest teams globally to believe models would scale, and their system architecture was built entirely on that assumption.

He was fixated on forging the best hammer, never considering where the nails might be.

For the first four years, OneFlow generated zero commercial revenue. Technical goals trumped business pursuits; the operating style resembled a lab, with most work being code. The team's defining trait: they loved writing code and discussing technical problems. When discussing technical problems, they came alive; nothing else excited them. Except money, of course.

Yuan likened his commercialization thinking at the time to "wishful thinking" — build the fastest, most usable framework, capture the developer ecosystem, then monetize. In plain terms, OneFlow sought a more "elegant" architecture that would solve everything once and for all: any deep learning model (even non-Transformer), no matter how large, would run on their architecture. The name OneFlow (Yiliu Technology) reflected this ambition.

Zhongguancun Guide interviewed Yuan in 2020, judging OneFlow's goals as ambitious but not Quixotic delusion. But in our conversation, Yuan described that period as "battling windmills."

OneFlow worked in relative secrecy for six years. Few believed; no one competed. They even considered building large models themselves if no one else did, once their tools were ready.

But this remained mere speculation. By late 2022, OneFlow was nearing survival mode, with access to at most a few hundred GPUs. The predicament wasn't a black swan event — it was cumulative. The domestic investment environment had sunk into a trough; the next funding round was uncertain. The company spent its final two years taking on projects for revenue, but these remained framework-development-focused, "hardly commercialization." Around ChatGPT's release, core employees received offers from major tech companies and left.

This was among Yuan's darkest periods. That year, he barely had weekends. At the most difficult moments, he could only resort to pay cuts and layoffs.

From a technical progress standpoint, OneFlow's exploration was undeniably successful. They were the first to make distributed multi-GPU programming as easy as single-GPU programming, and their distributed design concepts were later adopted by mainstream deep learning frameworks.

But the true revolutionary was OpenAI. Even though their "hammer" was merely patched onto existing architectures, supporting only the specific Transformer model, the result was that Transformer worked so well that revolution arrived faster.

"Call it fate, or call it social dynamics — letting more people directly see value accelerates things, rather than pursuing elegance or personal preferences." Yuan repeated a phrase to An Yong Waves several times: "The OneFlow experience was a bit like performance art."

Four Months of Peaks and Valleys

In late 2022, ChatGPT's unexpected eruption yanked OneFlow off its intended trajectory. Yuan was already among the Chinese scientists closest to large model developments — he knew Professor Tang Jie (teacher of Yang Zhilin), Google, and OpenAI were all pushing toward large models — but still underestimated the speed of change. "I believed it would happen, but thought we had more time," Yuan said.

Joy and apprehension arrived simultaneously. On one hand, a long-held conviction had finally materialized: large models had shifted from contrarian to consensus. But on the other hand, market consensus formed so rapidly that it would quickly devolve into an arms race of resources, funding, and name recognition — a competition for who could amass the most money, compute, and GPUs fastest — an opportunity that might not belong to OneFlow, which had sacrificed nearly everything for this moment (and once model architectures converged on Transformer, a universal solution was no longer essential).

Almost simultaneously, Yuan saw Wang Huiwen's "hero's call" — $50 million, capital included. He thought OneFlow might catch Wang's interest. Yuan met with Wang, who told him over dinner that he wanted to rebuild the team from scratch. Initially, the collaboration didn't materialize.

Soon, intermediaries facilitated, and both sides returned to negotiations. Within a month, the acquisition closed. OneFlow moved from Tsinghua Tongfang Technology Plaza to Sohu Internet Plaza.

Meanwhile, Lightyear Beyond's valuation broke $1 billion, instantly becoming the best-resourced team at the table.

But just two months later, Wang Huiwen stepped away due to illness, and Meituan took over Lightyear Beyond. Abandoning entrepreneurship as the wave crested — there was too much reluctance. Hesitation lasted only a month before Yuan decided to start again.

This was also the only way to keep the original team together. After news of turmoil at Lightyear Beyond broke, companies rushed to the offices to recruit; OneFlow members received offers exceeding 10 million yuan annually. No single company could make offers that would retain everyone simultaneously. But if they continued building a company, most would stay.

On SiliconFlow's first anniversary, they described their state thus: beneath surging, turbulent waves, like a helmsman struggling to control the ship's course — once the heading was set, no time to catch breath, even silently holding their breath, plunging once more into new battle.

Yuan told the team that if they could survive such a monumental crisis without falling apart, no future difficulty would be insurmountable.

Consensus and Open Cards

Yuan twice missed early opportunities to join tech giants. First in early 2013, when Yiming Zhang invited him to join ByteDance; Yuan preferred Microsoft Research Asia. Second when Su Hua invited him to join Kuaishou — he declined, citing "still having some technical ideas," and remained at Microsoft.

Years ago, he pursued OneFlow's success more than the company's success. With that approach, "only acquisition or death" awaited.

But this time is different. "Technical curiosity was satisfied in the previous phase, so what's unfulfilled, what's incomplete, is mainly on the business side," Yuan said in an interview.

During a sharing session Q&A, an audience member asked Yuan if he could propose another technical idea like OneFlow's — something that would happen in coming years but most people disagreed with. Yuan replied, with some embarrassment, that he now spent almost all his time on non-technical matters, so he couldn't.

The new company's name came from ChatGPT: SiliconFlow. Unlike OneFlow's singular focus on disrupting existing frameworks, SiliconFlow works backward from business needs, more pragmatically choosing a large model inference framework — technically less thrilling than OneFlow's pursuit, but with greater market potential. If training frameworks process data equivalent to a swimming pool, inference frameworks process a ceaselessly flowing river. Unlike OneFlow, this is a path of high consensus.

Large models have spawned Model-as-a-Service (MaaS). As long as services can be delivered via standard APIs, previously obscure, hard-to-monetize infrastructure can rapidly generate revenue. OpenAI's o1 advances in inference validated Yuan's bet on inference frameworks: "Inference compute demand has jumped another order of magnitude."

During those four months in Lightyear Beyond's eye of the storm, Yuan rapidly completed the coursework OneFlow never offered — insights only battle could provide. Company size expanded from 35 to over 60, adding product managers and commercialization leads. SiliconFlow now has three co-founders from business backgrounds.

In SiliconFlow's early years, Yuan still wrote code himself. Later he stopped, realizing that "a CEO writing code is irresponsible to the company." He's become more pragmatic, more CEO-like — not seeking divine inspiration, just a success built on solid, unspectacular moves.

Yuan later felt OneFlow was a gamble from the start, an experiment. Fortunately, it gave investors an exit; employees who left after the acquisition received roughly 20 million yuan in stock.

But colleagues who never left and started again with him have yet to see financial returns. There are things to account for, and lessons from excessive idealism.

In his first startup, Yuan named his public account "Notes from a Non-Mainstream Entrepreneur." Unlike most entrepreneurs discussing application scenarios and commercialization, he told An Yong Waves that he was interested in nothing beyond researching new frameworks — only loved solving hard problems, unwilling to do ordinary things. "Just felt this was the most exciting thing, and the more unexpected, the better."

At that sharing session, Yuan reviewed his seven-year "atypical entrepreneurial journey." At the end, he said, "What's urgently needed now is a typical entrepreneurial success."

Image source | IC Photo