跪拜 Guibai
← All articles
Frontend · Backend · Programmer

DeepSeek Founder Liang Wenfeng's Path from Math Prodigy to AGI Disruptor

By CodeSheep ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Liang Wenfeng's trajectory from quantitative finance to AGI shows a repeatable pattern: apply mathematical rigor and GPU-scale compute to a domain, strip out unnecessary cost, and open-source the result. For Western developers, DeepSeek's existence means frontier model access is no longer gated by a handful of San Francisco labs, and the cost floor for inference keeps dropping.

Summary

Before DeepSeek shocked Silicon Valley, Liang Wenfeng was a math prodigy who scored 806 points on China's college entrance exam in 2002, topping his region and choosing Zhejiang University's EE program over Tsinghua. He stayed at ZJU for a master's in computer vision, but his real interest had already shifted to financial markets. In 2008, during the global financial crisis, he assembled a team to explore fully automated quantitative trading.

By 2015, at age 30, he co-founded High-Flyer Quant, which became one of China's largest quantitative hedge funds with assets exceeding 100 billion yuan by 2021. The firm's 2016 shift to AI-driven strategies, using deep learning models and GPU compute for live trading, was the direct technical precursor to DeepSeek. In 2023, he founded DeepSeek to pursue AGI, arguing that large language models are the necessary path.

A rare interview with DarkSurge Waves captured 50 of his core positions: DeepSeek exists to explore AGI's essence, not to chase commercial applications; open-source is a strategy to attract top talent and accelerate ecosystem progress; and hiring prioritizes raw curiosity and long-term thinking over conventional credentials. The R1 release broke the monopoly of a few international giants on frontier models, driving down costs through open-source efficiency.

Takeaways
Liang Wenfeng scored 806 on the 2002 Gaokao, becoming the top scorer in his region, and chose Zhejiang University over Tsinghua to match his interests.
He studied electronic information engineering and later completed a master's in computer vision at ZJU under Professor Xiang Zhiyu.
In 2008, during the financial crisis, he formed a team to explore fully automated quantitative trading.
High-Flyer Quant, co-founded in 2015, managed over 100 billion yuan in assets by 2021 and shifted to AI-driven strategies using deep learning and GPUs in 2016.
DeepSeek was founded in 2023 with the explicit goal of exploring the essence of AGI, not just building commercial LLMs.
DeepSeek-R1's release broke the monopoly of a few international tech giants on frontier AI models through open-source, low-cost innovation.
Liang Wenfeng's 50 core thoughts, from a rare interview, cover his views on AGI, open-source as talent strategy, and hiring for curiosity over credentials.
Conclusions

The technical lineage from High-Flyer's 2016 GPU-based deep learning trading systems to DeepSeek's LLMs is direct; the same team spent seven years optimizing large-scale model training on GPU clusters before pivoting to language models.

Liang Wenfeng's decision to skip Tsinghua for ZJU's EE program, and later to skip a conventional career for quantitative finance, suggests a pattern of choosing fit over prestige that carried through to DeepSeek's contrarian open-source strategy.

The claim that DeepSeek's R&D team is 'not large' yet shipped multiple competitive models in a year challenges the assumption that frontier AI requires thousands of researchers and billion-dollar compute budgets.

Concepts & terms
Gaokao
China's national college entrance examination, a high-stakes standardized test where top scorers in each region are widely publicized. An 806 score in 2002 was exceptional and would have qualified for any university in the country.
Quantitative Hedge Fund
An investment fund that uses mathematical models and algorithms to make trading decisions. High-Flyer Quant applied deep learning and GPU computing to automate stock trading, a technical approach that later informed DeepSeek's model training infrastructure.
AGI (Artificial General Intelligence)
AI that matches or exceeds human cognitive abilities across a wide range of tasks, as opposed to narrow AI designed for specific problems. DeepSeek's stated mission is to explore AGI's essence, not merely to productize language models.
From the discussion

The most engaged thread challenges DeepSeek's ethics — training on public web data for free while allegedly charging users — a claim immediately ridiculed by replies pointing out the models are free and open-source. A separate substantive comment lays out DeepSeek's three-part strategy: AGI as the north star, ecosystem before monetization, and distrust of experienced hires as innovation blockers. Other remarks praise the founder's ambition and lament China's stifling tech culture, while a few are spam or low-effort.

DeepSeek's models are free and open-source, making accusations of paid membership baseless.
Training on publicly available web data is standard practice, not an ethical violation.
The company's roadmap prioritizes AGI research, open ecosystem building, and self-driven innovation over immediate commercial applications.
Experienced AI team members can hinder innovation; letting teams self-organize produces better results.
China's domestic tech culture overemphasizes results and suppresses genuine innovation, widening the gap with international competitors.
Liang Wenfeng's combination of capital, technical ability, and moral conviction is rare and necessary for breakthrough progress.
Featured comments
WANGXC 2 likes

Taking web data for training without spending a cent, then charging membership fees — is there no justice!

桜吹雪  · 10 likes

In the year 6202, this is the funniest comment I've ever seen 🤣 Where does DS charge membership? Send me the link 🤣 Is the DS model a paid download? Want me to send you the model download link? 🤣

桜吹雪  · 5 likes

Why don't you buy tens of thousands of H100s yourself? Since you said it costs nothing, go train a model and open-source it for free. Otherwise, you're the one with no justice 🤣

前端养生工程师 4 likes

Actually, DeepSeek's development path has already told everyone: 1. The ultimate goal is to achieve AGI; scientific research and technological innovation come first. 2. Build the ecosystem first, won't go closed-source, won't rush to make applications, and may not even need to make applications. 3. In AI innovation, experienced team members can be a constraint; let the team drive innovation on their own.

超厉害 1 likes

The kind of figure the tech world should have. The domestic tech scene has a very unhealthy culture that stifles innovation — that's where the gap with the outside world lies. When everyone just obsesses over results, people only care about results and lose any sense of innovation.

See top comments, translated →
Source: juejin.cn ↗ Google Translate ↗ Backup ↗