In late July 2026, at an investor meeting that ran for four hours, Liang Wenfeng said "no" dozens of times. No 3D, no video generation, no world models, no next super-app, no closed source, no chasing user counts, no unreasonable profits. Strung together, these "no"s form a counter-intuitive strategic logic: in an industry where everyone is frantically doubling down, restraint itself is a way of raising the odds of success.
AGI Is the Only Main Line
Liang Wenfeng has compressed DeepSeek's task list to a single item: AGI. Products are by-products, multimodality is a component, and video-generation capability has little to do with "the upper limit of intelligence." He said this in a calm tone but with unmistakable intent — at a moment when the LLM track is so crowded that every company is straining to extend its product lines, he has shut the door on everything and everyone that sits off the main line.
This is itself a hard-edged judgment: achieving AGI requires neither 3D generation, nor video understanding, nor world models. These are not obligatory waystations on the road to intelligence but decorations "made for end users to look at." He classifies them as "product problems" rather than "intelligence problems" — there are ways to solve them, but they do not push the frontier of intelligence forward.
DeepSeek has "not much money, not many GPUs, not much fame" — just "a group of very ordinary people." For a resource-constrained team, subtraction is instinct rather than choice. What is interesting, though, is that his logic of subtraction is not "we can't do it, so we won't," but "it doesn't matter, so we won't."
He has converted the modesty of "we can't" into the disdain of "we choose not to" — and that shift of posture is itself the core of DeepSeek's strategy.
One Boundary Drawn on Each of Four Lines
Liang Wenfeng's "no"s fall along four dimensions: the product line, commercialization, the organization, and the open-source strategy. All four lines point the same way — pooling the company's entire resources, attention, and cultural cohesion onto a single thing.
The Product-Line Boundary
"Multimodality matters for the product, and it matters for end users. But it is only a component — not the main line, and not intelligence itself." DeepSeek will not do 3D, will not do video generation, will not do world models. The most important thing at this stage is the coding agent. This is a concrete judgment: programming ability is the most direct yardstick of progress toward intelligence. If a model can write good code, it has already mastered logical reasoning, symbolic manipulation, and task decomposition — the core capabilities on the road to AGI. Financial agents and medical agents rank below a general-purpose agent in priority, because "general-purpose" already includes coverage of specific domains.
The Commercialization Boundary
DeepSeek spends no effort on commercialization: "There is no customer service and no sales force — the users come on their own." Liang Wenfeng recalls that when the company cut its prices by a quarter, many people in the internal group chat cheered — "because this is exactly why we put so much heart into making the model well: so that everyone can use it to the fullest." He has no intention of building the next ByteDance or Tencent. "The reason we don't fight over this thing is that there is still a watermelon waiting behind it; what lies in front may all be sesame seeds." The judgment behind that sentence is simple: the commercial value of AGI will be so vast that there is no need to worry about how the cake is sliced — build it first.
The Organizational Boundary
Team stability is the one thing that cannot be traded away. "Only one thing is non-negotiable: we must keep the team stable." DeepSeek's way of working combines top-down and bottom-up: half of the "real work" is assigned from above, and the other half of people's time is left to free exploration, "with no preconditions." Not working overtime is not because the workload is light, but because "doing research requires a relatively relaxed environment." Not technological leadership, not market share, not revenue targets — but whether people will stay.
The Open-Source Boundary
Liang Wenfeng defines open source as "giving up profit" — "the employees gain a sense of achievement, the company gains cohesion, and society benefits as well. Competitors and ordinary people alike are happy." He does the arithmetic: if AGI ultimately comes to account for 10% of GDP, then the probability of "whether it gets done at all" matters far more than the share of "how much one gets to keep." Open source, by lowering the intensity of competition, raising industry-wide coordination, and winning over potential rivals, actually increases DeepSeek's probability of "getting it done." "If your vision is to take a lot, you have already lost. You may find yourself facing far greater difficulties. That is just how the world works."
The Four Rungs of the Technical Roadmap
Liang Wenfeng's roadmap has four steps:
- Last year's rung was CoT (chain-of-thought) — it solved the reasoning-chain problem.
- This year's rung is the agent — it solves the tool-use problem.
- The next target is continual learning — "the next-generation model must have the ability to learn continually, otherwise it does not deserve to be called next-generation." In his view, AI "does not lack taste and intuition; what it lacks is the ability to learn continually."
- The rung after that is AI accelerating AI research — the model can complete everything a human can, including developing more advanced AI models of its own. Only after this step is finished does embodied intelligence arrive — "the endpoint of intelligence may well all be embodied. Because for an ordinary person, what is needed is not a computer but human labor."
This roadmap is more conservative than the industry's mainstream narrative — there are no promises of world models — and also more concrete — every step has an explicit, measurable target. Moreover, each stage corresponds to a "problem that cannot currently be solved": CoT solved reasoning chains, the agent solved tool use, and continual learning addresses the model's capacity for self-evolution — each step attacks one specific bottleneck.
A Case Study: From One Sentence to a Methodology
The clearest illustration of this methodology in action is a sentence pattern Liang Wenfeng returned to again and again throughout the meeting: "If it is not important, don't do it; if you do it, do it to the best." The pattern appears seven or eight times across the 52 recorded quotes, and each time it lands on a different decision point.
- On pricing: "We only earn a reasonable profit; ours is not profit-maximizing pricing." — this is restraint in commercialization.
- On technical choices: "Many things are not on our main line — for example 3D and video generation; or world models, which also have little to do with the upper limit of intelligence." — this is restraint in the product line.
- On organizational management: "Restraint is a strategy: give some things up, in exchange for many others." — this is restraint across the board.
Behind every "no" there is an alternative — not refusal in a vacuum, but refusal traded for investment elsewhere. The pattern itself can be abstracted into a decision framework: given an extremely clear goal, ask the same question at every resource allocation — "Does this move me directly toward AGI?"
The Landing of Organizational Culture — The Management Logic Behind No Overtime and No KPIs
On July 23, a full transcript of the investor meeting was leaked, filling in more of Liang Wenfeng's thinking on organizational culture — in particular, the management logic behind two counter-intuitive rules: "no overtime" and "no KPIs."
Two-Line Management — A Parallel Structure of Top-Down and Bottom-Up
Liang Wenfeng describes DeepSeek's organization as two lines running in parallel. From the bottom up, each person works on whatever he or she wants — "there are no KPIs, and no one manages him." From the top down run formal projects that require company-wide coordination, such as the release of V4. The key constraint is that formal work should not occupy more than half of an employee's time, and the remaining time is left unassigned. "Whatever he wants to explore, he goes and explores. As long as the company's compute can support it, he does not need to come coordinate with anyone."
This structure is itself a risk-management framework. Fifty percent free exploration means the company is not executing a single path but trial-and-erroring multiple technical routes simultaneously. If the formal direction veers off course, the organization is always holding a "seed bank" — the results of the free exploration can be activated as alternatives at any moment.
No Overtime Is Not Because the Workload Is Light
Liang Wenfeng gives two reasons. First, research needs a relaxed environment: "if you press people too hard, they cannot do research." Since research runs on interest and self-driven tinkering, the environment has to stay loose. The second reason is more fundamental — "because we are restrained, we choose not to do many things. When there are fewer things I want done, there is less work to hand out to each person."
This sentence connects the overtime question directly to the question of strategy. The "unfinished" state of the products is not a management oversight but a deliberate choice: "you can see that many of our products are far from complete, but we are not rushing to patch them either — that too is part of our culture." No overtime is not a perk; it is the result of strategic restraint — when a company does only one thing, and each person is assigned only the most necessary tasks, overtime becomes fundamentally unnecessary.
A Framework Set Against Mainstream Silicon Valley AI Companies
The work cultures of OpenAI, Anthropic, and Google DeepMind are famous for high intensity and high competition — a "sprint culture" is treated as the necessary price of staying technically ahead. DeepSeek's choice is almost exactly the opposite: protect creativity by shrinking the volume of tasks, and preserve team stability by lowering the intensity of competition. Which of the two organizational cultures is better cannot be judged in the abstract — but DeepSeek's product record proves at least one point: a relaxed environment and low KPIs are not synonyms for technological backwardness.
Against the Genius Theory (2026-07-31 Supplement)
In a monthly review, the video creator Xiao Wang Albert (小王Albert) relayed remarks from what is described as an internal discussion speech by Liang Wenfeng. The section on "opposing the genius theory" runs in the same vein as the core of the four-hour meeting, only stated more bluntly.
At the beginning, DeepSeek did not have much money, did not have many GPUs, and did not have any fame — it was simply a group of ordinary people doing an extraordinary thing. The key is not one or two headline figures, but whether things actually get done. In the age of mass media, everyone loves heroism and genius stories — from Nobel to Fields, from Zuckerberg to Musk — but all of them overlook structure. Genes are fair; every nation can produce geniuses. Why, then, do the heroes of other countries never get the chance to shine? The key is whether one can build a mechanism that lets large numbers of people take part, so that through continuous trial and error, a hero's story is eventually brewed.
These remarks and "restraint" are two sides of the same coin: restraint is not-doing at the strategic level; the anti-genius stance is not-worshipping at the organizational level. Both point to the same underlying judgment — the breakthrough in AI will not come from individual supermen, but from a soil that lets ordinary people keep doing things.
The Competitive Dimension: How to Fight from a Position of Compute Disadvantage (2026-08-01 Supplement)
Reading the same four-hour speech from an international-relations angle, the commentator Yang Sheng Alex (杨升Alex) offers a different interpretation: Liang Wenfeng was talking not only about how the company restrains itself but also about how to fight the United States from a position of compute disadvantage. What this section discusses is the same thing as the "restraint" elsewhere on this page — the flip side of subtraction is putting limited resources on the cutting edge.
Compensating for compute with efficiency. Liang Wenfeng acknowledges that the compute DeepSeek has used so far may be only a fraction of what the leading American companies command. Its advantage lies not in having more chips but in making each chip do more work — mixture-of-experts architectures, low-precision training, communication optimization, and high-efficiency operators. The goal is to use a fraction of the compute available in the United States and shrink the time gap from one or two years to half a year, or even three months. Once the gap narrows to a matter of months, it becomes very difficult for the United States to convert a fleeting technological lead into a long-term monopoly.
Eroding the CUDA moat. What makes NVIDIA truly hard to replace is not chip performance alone but the entire software ecosystem built on top of CUDA. Liang Wenfeng's judgment is that AI-assisted programming and advanced compiler tools are lowering the difficulty of rebuilding that software ecosystem — insert a translation layer between the models and the chips, and once the same model code is compiled, it runs on NVIDIA chips just as it runs on Huawei's or other domestic chips.
Four Huawei cards to match one NVIDIA card. This is not to say that Huawei's single chip has caught up with NVIDIA's. Rather, the race shifts from chasing single-card performance to chasing the usability of the entire computing system: where individual cards fall short, the answer is whole-system substitution with more chips, larger supernodes, more abundant power, and system-level optimization. The precondition is first producing those four cards stably and at low cost — solving power consumption, heat dissipation, chip interconnect, and cluster failure rates. According to reporting by Lianhe Zaobao and The Wall Street Journal, China has already begun mass-producing DUV lithography machines, and domestic chip self-sufficiency is only a matter of time.
The two faces of open source. The section on "The Open-Source Boundary" above describes open source's internal side as "giving up profit" (employee achievement, company cohesion); the competitive dimension adds the external layer. Open source undercuts the pricing power of closed-source American models and lets developers around the world use, deploy, and modify Chinese models. What is being contested is the right to diffuse AI — the United States competes for a monopoly over the most advanced models and technical standards, while China competes to give developing countries models that are good enough, cheap enough, locally deployable, and keep their data within national borders.
Limits and Boundary Conditions
The framework's validity rests on two premises. First, the goal must be extremely clear and the team must share consensus on it. If a company chases AGI and short-term revenue at the same time, the framework breaks at the very first constraint — products "made for end users to look at" and research that "pushes the frontier of intelligence" demand entirely different logics of resource allocation. Second, the founder's personal style must be tightly aligned with the organizational culture. Liang Wenfeng says "restraint is part of our vision" — this is not a slogan painted on a wall; it is the way he himself works. In the hands of a different founder, executing the same framework could produce a completely different outcome.