Liang Wenfeng on DeepSeek: 52 Views on AGI, Open Source, and Restraint

“There may be a watermelon farther ahead. What lies in front of us may only be sesame seeds.”

It is one of the most memorable metaphors attributed to Liang Wenfeng after a four-hour investor meeting. Chatbot traffic, enterprise contracts, and the race to build a super app are the sesame seeds. The watermelon is continual learning—and, eventually, artificial general intelligence (AGI).

That metaphor turns 52 seemingly scattered views on products, pricing, open source, talent, and management into one coherent question: does this choice increase DeepSeek’s probability of reaching AGI?

The account below reorganizes a compilation published by elsewhere on July 22, 2026. Elsewhere says the remarks were collected from multiple accounts and that some wording may differ from what was said in the room. This is a thematic, faithful paraphrase—not a verbatim transcript or an official DeepSeek statement.

When you could build anything, choosing what not to build matters more

1. Product work should not maximize revenue at this stage. Products are a step toward AGI and a by-product of that journey, not a reason to spend most of the company’s energy on conventional consumer or enterprise competition.
2. Areas such as 3D generation, video generation, and world models are outside the present main line because they may not determine the upper bound of intelligence.
3. Multimodality matters greatly for consumer products, but it is a component rather than intelligence itself.
4. Hallucinations need solutions, yet DeepSeek treats them largely as a product problem rather than its most important research priority today.
5. Coding agents are the immediate priority. In China, building a strong general agent makes more sense than first optimizing vertical agents for finance or medicine.
6. If the AI era creates several trillion-dollar companies, it would be enough for DeepSeek to become one of them.

Agents are the next step; continual learning is the harder one

7. Today’s AI does not primarily lack taste or intuition; it lacks the ability to keep learning.
8. People accumulate experience while working, whereas AI often needs the full context supplied again. Until that changes, it cannot truly replace an employee.
9. DeepSeek wants its next model to help its own researchers first, rather than merely becoming easier for everyone else to use. AI accelerating AI research may be the fastest route to AGI.
10. No one has yet found a satisfactory approach to continual learning, partly because “learning” consists of many interacting mechanisms.
11. DeepSeek’s long-term destination is AGI. If the route is a staircase, chain-of-thought was last year’s step, agents are this year’s, and continual learning comes next.
12. Once continual learning works, models may approach a gradual singularity: performing most human tasks and helping develop more advanced AI systems.
13. Intelligence may ultimately become embodied, because people usually need capable labor in the physical world, not merely another computer.

Low prices are an architectural consequence, not charity

14. DeepSeek aims to earn a reasonable profit, not to price for maximum profit.
15. When one model’s price was cut to a quarter of its initial level, many employees celebrated: broad use was closer to the reason they had built it carefully.
16. Low cost is an architectural result and a way to train larger models. When compute is scarce, efficiency can matter more than simply adding resources.
17. An API can be maintained by a small team without a large sales or support operation, while users arrive on their own. That makes API sales less compelling as a corporate goal.
18. DeepSeek already commercializes some work, but commercialization is not the objective; a full commercial turn remains far away.
19. If the eventual opportunity is large enough, participation will be possible. DeepSeek is better understood as a product of its era and circumstances than as an imitation of another company.

Why open source has not destroyed DeepSeek’s business

20. Restraint means yielding some immediate advantage in exchange for employee pride, organizational cohesion, public goodwill, and a higher probability of reaching AGI.
21. AI could become far larger economically than traditional software. Trying to monopolize value on the scale of a meaningful share of global GDP may invite historical rejection.
22. DeepSeek says the model it releases openly is the same model it deploys, rather than a deliberately weakened edition.
23. Liang does not fear others deploying DeepSeek models to compete. Small startups often lack the resources, while large companies struggle with organization; DeepSeek’s current scale sits between them.
24. Open source does not damage a model built around reasonable profits. It threatens models that depend on extraordinary profit multiples.
25. DeepSeek does not want to become the enemy of either incumbents or startups and is willing to help peers including Alibaba, Zhipu AI, and Moonshot AI improve.

China may close the gap through compute efficiency first

26. China’s AI narrative should shift toward closing the frontier gap to six or even three months while using a fraction of the compute.
27. Scaling still works: larger scale generally brings better results. Current model sizes reflect resource limits, not a belief that they are already sufficient.
28. The talent gap is small and the underlying pool is often similar. Shortages of a particular kind of specialist are usually temporary.

In the model race, cost comes before experience

29. Anthropic’s current lead over OpenAI may be temporary; OpenAI and Google are more likely to alternate at the frontier over time.
30. China has too many model companies repeating similar work and dispersing resources. The market should consolidate; with reasonable profits, perhaps two large and two smaller model companies could be enough.
31. Model developers are unlikely to capture most of the AI industry’s profits.
32. Competition will ultimately turn on cost, time, and user experience, in that order. A few months of lead time matters, while user experience creates some stickiness but is not the fundamental barrier.

No super-app race while the “watermelon” is still ahead

33. DeepSeek does not aim to become the next super app, ByteDance, or Tencent.
34. It avoids fighting over today’s prizes because larger opportunities may lie ahead—the “watermelon” beyond the “sesame seeds.”
35. The industry chased chatbots and consumer traffic last year and enterprise revenue this year. DeepSeek’s internal attention remains on the AGI roadmap and the next technical breakthrough.
36. Its sudden popularity around last year’s Lunar New Year was not part of the plan.

The asset DeepSeek cannot afford to lose is its team

37. The one issue on which DeepSeek cannot compromise is team stability. It was a major risk, substantially eased by the financing round.
38. Avoiding enemies and enabling peers also helps create a friendlier, more stable environment for DeepSeek’s own team.
39. The organization runs both top-down and bottom-up. Assigned “proper work” should ideally leave substantial time for researchers to explore what they independently consider important.
40. The team generally avoids excessive overtime because research needs a relaxed environment and because focus lets the company tolerate products that remain imperfect.
41. The organization will change as it grows and may add necessary structure, but it should not become a conventional hierarchy or abandon mission-driven work.

A vision matters only when it changes daily decisions

42. The earliest employees did not join primarily to become rich or take a company public. They believed the work could help humanity and approached it with goodwill.
43. DeepSeek is driven by vision rather than KPI targets. That brings disadvantages as well as advantages, but it remains a defining feature.
44. The vision is not necessarily written down. It appears in methods and attitudes toward the world; individual interpretations can differ while the broad direction remains shared.
45. Liang once admired former GE chief executive Jack Welch and now considers many of Welch’s ideas outdated. One principle still stands: vision is a company’s most important asset, expressed through conduct rather than slogans on a wall.

Restraint is ultimately a probability calculation

46. AGI offers the largest potential return. Other work gets done only when capacity permits; restraint is part of the mission itself.
47. AI’s opportunity is so large that even a small share could be enormous. The more restrained the company remains, the likelier it may be to finish the journey.
48. Beyond its vision, DeepSeek does not claim many inherent advantages.
49. At its founding, the company lacked abundant money, compute, fame, and recruiting power. Liang prefers the story of ordinary people doing something extraordinary to a myth about geniuses.
50. Open source and low prices are both forms of restraint. Higher prices might lift short-term revenue, but the long-term result is uncertain, making restraint a strategic choice.
51. Open source and low prices give employees pride, strengthen cohesion, and benefit peers and the public. Together, those effects may raise the probability of achieving AGI.
52. A mission centered on taking the largest possible share may place a company at a disadvantage from the beginning.

Taken together, these 52 views do not describe a company that has no interest in making money. They describe a company that wants profit, traffic, and products to remain means rather than ends.

Lower prices encourage use. Open source earns ecosystem support and goodwill. A looser research environment helps retain the team. Building fewer products leaves scarce compute and attention focused on models. Each act of “taking less” is supposed to purchase the same thing: a better chance of reaching AGI.

Whether the wager works is unknown. Continual learning remains unsolved, vision does not remove compute constraints, and open models with low prices can create long-term commercial pressure. But the meeting sketches a remarkably legible DeepSeek strategy: do not rush to occupy every market already in sight when the larger prize may still be farther down the road.

Primary source: elsewhere, “Liang Wenfeng’s Four-Hour Investor Meeting,” July 22, 2026. This article is a thematic bilingual paraphrase, not a verbatim quotation. Claims about Liang Wenfeng’s remarks or DeepSeek’s strategy should ultimately be checked against official releases and verifiable primary material.