Liang Wenfeng founded DeepSeek as a research venture backed by High-Flyer. These lessons reflect his interviews and coauthored model research, with historical positions dated rather than treated as permanent commitments. — 36Kr Interview (2023).

Infographic for "Lessons from Liang Wenfeng of Deep Seek".

Vision for AI and China's Role

  1. The Ultimate Goal is AGI: He chose language models as a starting point toward AGI, not merely financial applications. — 36Kr Interview (2023).
  2. China Must Innovate, Not Imitate: China should contribute original technology rather than only commercialize others’ inventions. — 36Kr Interview (2024).
  3. Innovation is Driven by Curiosity, Not Just Profit: Curiosity can justify fundamental research even when immediate commercial returns look unattractive. — 36Kr Interview (2023).
  4. AI is in a Phase of Technological Explosion: In 2024, he prioritized technological breakthroughs over an application boom. — 36Kr Interview (2024).
  5. Solving the Hardest Problems is the Core Driver: Difficult research problems can attract exceptional people. — 36Kr Interview (2024).

On Open Source and Competition

  1. Closed-Source is Not a Sustainable Moat: He locates durable advantage in team learning and innovative culture. — 36Kr Interview (2024).
  2. Open Source Builds Ecosystems: Share foundational results so others can build affordable applications. — 36Kr Interview (2023).
  3. Openness Fosters Achievement: Sharing research earns recognition and cultural appeal as others build on the team’s work. — 36Kr Interview (2024).
  4. Efficiency Can Outmaneuver Brute Force: The V2 report records stronger performance with lower training costs than DeepSeek 67B—not a comparison of total company budgets. — DeepSeek-V2 Technical Report.
  5. Pricing Should Be Based on Cost, Not Hype: He said the 2024 price-war reaction was unexpected: prices reflected costs plus a modest margin, not subsidies or profiteering. — 36Kr Interview (2024).

On Company Culture and Talent

  1. Maintain a "Bottom-Up" Structure: Let researchers bring their own ideas and form roles naturally; allocate resources from above once an idea shows promise. — 36Kr Interview (2024).
  2. Provide Unlimited Access to Resources: He described researchers accessing training clusters without prior approval. — 36Kr Interview (2024).
  3. Trust in Young, Local Talent: For long-term research, prioritize ability, creativity and passion over experience; capable local graduates can grow into the work. — 36Kr Interview (2023).
  4. Innovation Requires Confidence: Confidence matters: young researchers may be readier to attempt frontier work. — 36Kr Interview (2024).

From Quantitative Finance to AI

  1. Technology-Driven Innovation Builds Confidence: High-Flyer strengthened his confidence in technology-led innovation—not a universal guarantee. — 36Kr Interview (2024).
  2. The Transition Was an Unconventional Move: He established DeepSeek separately from the fund, accepting basic research’s uncertain commercial payoff rather than treating it as financial modeling. — 36Kr Interview (2023).
  3. Early GPU Investment Was Strategic: Compute investment accumulated gradually; curiosity about AI and testing new hardware preceded the rush to build chatbots. — 36Kr Interview (2023).
  4. Finance Was a Stepping Stone to a Harder Problem: After exploring AI in finance, his team sought another demanding challenge: AGI. — 36Kr Interview (2023).
  1. Invest in Tools People Want to Use: He compared buying GPUs to buying a piano: affordability matters, but so do people eager to use the instrument. — 36Kr Interview (2023).
  2. Persist in Building Models: When research became difficult, Liang drew encouragement from Simons’s conviction that prices could be modeled; this is inspiration, not a guarantee of profitable prediction. — Liang’s Simons Foreword.

On Business and Strategy

  1. Money is Not the Main Bottleneck: In 2024, he reported no near-term fundraising plans: access to advanced chips, not money, was the constraint. — 36Kr Interview (2024).
  2. More Investment Doesn't Always Equal More Innovation: Larger budgets alone do not guarantee more innovation. — 36Kr Interview (2024).
  3. Focus is Key: His stated early priority was model research, not prematurely designing applications around it. — 36Kr Interview (2023).
  4. Design for Resource Constraints: Reducing compute and memory demands requires architectural work: V2 combines compressed attention with sparse expert computation. — DeepSeek-V2 Technical Report.
  5. Avoid Overthinking the Business Model: Judge direction by social efficiency, not transient business-model fashions. — 36Kr Interview (2024).
  6. When Enabling Technology Catches Up: In his foreword about quantitative investing, Liang connects the field’s early setbacks with the later maturation of computing hardware and software. He treats progress as depending on the available tools as well as researchers’ persistence—not proof that every unsuccessful approach will eventually work. — Liang Wenfeng — Creating Models to Understand the Market
  1. Adapt Rather Than Follow Fixed Rules: Changing markets reward adaptation; inherited organizational rules and experience can become constraints on a new company. — 36Kr Interview (2023).

Sources