60 Thoughts of DeepSeek Founder Liang Wenfeng on AGI, Large Models, and Innovation

Liang Wenfeng, an entrepreneur with a background in quantitative investing, has devoted himself to researching large AI models with the goal of achieving artificial general intelligence (AGI), emphasizing language models as the key pathway and focusing on foundational research rather than immediate application development.

He acknowledges the high cost and challenges of innovation, adopts a long‑term perspective, and prioritizes hiring recent graduates who are passionate, curious, and capable, fostering an open and inclusive environment that encourages freedom and experimentation.

Liang stresses the importance of making large models accessible to reduce monopolies, advocates for open‑source as a cultural practice that attracts talent and drives ecosystem growth, and remains committed to self‑funding despite external pressures such as chip export bans.

He reflects on China’s AI development, noting gaps in originality versus imitation compared to the United States, and calls for greater participation in global innovation to shift from a follower to an innovator role.

On December 16, 2024, DeepSeek R1 will be launched, and Liang shared a bold comment about the rapid transformation of the old world.

DeepSeek founder Liang Wenfeng’s 60 thoughts:

1. We build large models unrelated to quantitative finance; our aim is AGI.

2. Language models are the essential path to AGI and already exhibit AGI‑like traits.

3. We avoid premature application design, focusing on large models; over time, application barriers will lower, giving startups opportunities.

4. Human intelligence is fundamentally linguistic; thinking is language processing, suggesting language models could yield human‑like AI.

5. Replicating existing work requires minimal training or fine‑tuning, whereas original research demands extensive experiments, higher compute, and skilled personnel.

6. We aim for universal access to large models, preventing monopoly by big firms.

7. Basic research has low ROI, but we are well‑positioned to pursue it.

8. Model scaling has progressed from 1 GPU to thousands, driven by curiosity about AI limits.

9. The impact of ChatGPT mirrors AlexNet’s earlier disruption; model, data, and compute remain the core drivers.

10. Exciting achievements cannot be measured solely by money, akin to buying a piano for passionate players.

11. Human capital is an investment; we hire curious individuals who can focus on research rather than immediate product needs.

12. Hiring emphasizes ability over experience for long‑term growth.

13. Core technical roles favor recent graduates who explore solutions deeply.

14. Passion is a key hiring criterion; enthusiastic candidates often seek us out.

15. We do not use KPI or task‑based assessments.

16. Innovation thrives with minimal interference, granting individuals freedom to experiment and fail.

17. Shared values and leader example shape culture without formal documentation.

18. Traditional startup methodologies may fail; adaptability is crucial in the AI wave.

19. Verifying hypotheses drives excitement.

20. Believers invest heavily in compute resources.

21. Innovation is costly and inefficient; substantial economic development is required.

22. Some phenomena defy logic, yet contributors find fulfillment.

23. Many can devote youthful years to passion‑driven work without profit motives.

24. Our pricing aims for modest profit, avoiding subsidies or excessive margins.

25. User acquisition is not our primary goal; we lower prices to make AI universally affordable.

26. To pursue AGI we must research new model architectures beyond existing Llama structures.

27. China must shift from merely applying foreign tech to leading innovation.

28. Historically, China has been a technology follower; now it must become a contributor.

29. Most Chinese firms tend to follow rather than innovate.

30. Innovation costs are high; confidence and talent organization are the missing pieces.

31. Past focus on profit neglected innovation, which also requires curiosity.

32. Closed‑source defenses are temporary; knowledge accumulation in teams forms a lasting moat.

33. Open‑source and publishing papers enhance cultural appeal and attract talent.

34. The most profitable U.S. companies are high‑tech firms with deep foundations.

35. China’s AI gap lies in originality versus imitation; without change, it remains a follower.

36. Nvidia’s lead stems from a collaborative Western ecosystem; China needs a similar community.

37. We will not close our models; a strong ecosystem is essential.

38. Funding is not an issue; high‑end chip bans are.

39. More investment does not guarantee more innovation.

40. We view the current phase as a burst of technological innovation, not applications.

41. Application development is possible but research remains priority.

42. Technology has no secrets, but rebuilding teams and catching up takes time.

43. Providing cloud services is secondary; achieving AGI is primary.

44. Large firms’ cash‑flow businesses can become their Achilles’ heel.

45. Only a few large‑model startups will survive; clear positioning improves odds.

46. Innovation should increase societal efficiency; short‑term focus can be misleading.

47. Our V2 model team is entirely domestic, capable of building top talent.

48. We favor bottom‑up organization and natural division of labor.

49. Resource allocation is unrestricted; anyone can request compute without approval.

50. Passion and curiosity dominate hiring criteria over monetary concerns.

51. Innovation requires confidence; Silicon Valley’s daring contrasts with China’s past hesitance.

52. Tackling the hardest problems attracts top talent, which is under‑recognized in China.

53. OpenAI is not infallible; leadership can shift.

54. AGI may arrive in 2‑10 years; we focus on mathematics/code, multimodal, and language.

55. Future companies will specialize in foundational models and services, forming a long value chain.

56. My main effort is researching next‑generation large models.

57. Past business models may not apply to future AI economics.

58. Our work spans over 16 years, though public perception sees only recent progress.

59. China’s industrial shift will rely on hard‑core technological innovation.

60. Hard innovation will increase; societal education will eventually recognize its value.

Liang also wrote the foreword for James Simons’ biography, emphasizing the belief that “there is always a way to model prices.”

“Creating Models that Understand the Market” – an essay by Liang Wenfeng

The piece recounts James Simons’ impact on quantitative investing, the evolution of computing power, and how transparent, fair markets enable both humans and algorithms to compete on equal footing.

Finally, an upcoming AI coding workshop in Beijing on April 13 will showcase practical AI coding tools, encouraging participants to upgrade their programming skills for the AI era.