The Dao of DeepSeek: Liang Wenfeng on Vision, Restraint, and the Philosophy of Open Source

Recently, DeepSeek founder Liang Wenfeng broke his silence, and a rare transcript of his conversation with investors was made public. This record—spanning 4 hours, approximately 34,000 Chinese characters, and 118 exchanges—was verified by some of the participating investment institutions and is considered authentic. After reading through it, one impression stands out clearly: DeepSeek embodies a distinctly Eastern philosophical approach to AI development. This article distills the core ideas from this conversation and attempts to understand their deeper implications for the current US-China AI competition.

DeepSeek's Vision: Gathering Like-Minded Idealists, Not Interest-Based Alliances

The entire conversation revolves around two keywords—vision and restraint.

Liang Wenfeng believes that managing a company relies not on rules and regulations, but on vision. And vision is not a slogan hung on a wall, nor is it the common practice of "painting a big pie" (making empty promises). The essence of painting a big pie is promising future returns—its driving force is ultimately profit. Teams assembled around profit will "gather for gain and scatter when the gain runs out." When someone else can offer more, the team falls apart.

In Liang Wenfeng's framing, DeepSeek's starting point is not commercial profit maximization: "We approach this with enormous goodwill toward the world. We believe this is useful for humanity—it's something beyond money." This phrase, "something beyond money," draws a clear line between righteousness and profit. He emphasizes that two years before the company was founded, they had neither much money, nor many GPUs, nor any fame or influence—yet this group of ordinary people accomplished what the world never thought possible.

This vision-centered management philosophy contrasts with some peers. Moonshot AI CEO Yang Zhilin has discussed managing teams in a way analogous to training large models—finding a balance between SFT (Supervised Fine-Tuning) and RL (Reinforcement Learning), while worrying that RL-style management might be exploited through "reward hacking," where employees game metrics at all costs, producing beautiful benchmarks that don't reflect actual improvement.

Some technical background on this metaphor is warranted. SFT (Supervised Fine-Tuning) and RL (Reinforcement Learning) are two key stages in the large language model training pipeline. The SFT stage fine-tunes the model using high-quality human-annotated data—essentially "teaching by hand"—training the model to exhibit desired behaviors. The RL stage (typically RLHF, Reinforcement Learning from Human Feedback) uses reward signals to guide the model toward self-optimizing its outputs. "Reward hacking" is a well-known problem in RL: when the reward function is imperfectly designed, the model finds loopholes in the reward function, achieving high scores on the surface while its actual behavior deviates from intended goals. For example, a model trained to maximize user likes might learn to generate sensational but low-quality content. Yang Zhilin applies this technical concept to team management, suggesting that relying entirely on quantitative metrics like KPIs to manage employees may lead them to "game the metrics" while neglecting truly important work.

This is a logically rigorous, well-organized "commander's mindset." But here's a key distinction: one approach focuses on how to prevent employees from gaming the system, while the other focuses on how to attract like-minded people to collectively realize a vision. When team members are there for their own ideals, preventing gaming is no longer the primary concern.

DeepSeek's Strategic Restraint: Ignoring Sesame Seeds to Find the Watermelon

Understanding the vision is the key to understanding why DeepSeek has consistently remained low-key and restrained. In Liang Wenfeng's words: "Give up some things to gain more of other things. The ground ahead is covered with sesame seeds, but the watermelon is still further away. The more restrained you are, the more likely you are to succeed." This restraint manifests in three main areas.

Staying on the AGI track without being distracted by trends

DeepSeek has only one main thread—model capabilities on the path to AGI. AGI (Artificial General Intelligence) refers to an AI system with general cognitive abilities equal to or surpassing those of humans, capable of understanding, learning, and executing any intellectual task without specialized training. Current mainstream AI systems (such as the GPT series, DeepSeek, and other large language models) are classified as "Narrow AI"—they excel at specific tasks but remain significantly distant from truly understanding the world or engaging in cross-domain autonomous reasoning. AGI is considered the ultimate goal of AI research and is explicitly written into the mission statements of organizations like OpenAI and DeepMind. Achieving AGI requires breakthroughs in causal reasoning, common-sense understanding, continual learning, multimodal integration, and other core challenges. The industry remains deeply divided on when AGI might be realized.

Video generation, 3D generation, world models, and other hot directions are not part of DeepSeek's roadmap. Liang Wenfeng notes that when video generation became trendy, it felt like "you're not an AI company if you're not doing it," but upon careful reflection, it has little to do with the roadmap toward intelligence.

From a technical perspective, this judgment has merit: video generation models (such as OpenAI's Sora, Runway's Gen series, etc.) are based on Diffusion Model or Autoregressive Transformer architectures, generating new videos by learning the spatiotemporal distribution of pixels in massive video datasets. These models are fundamentally fitting the statistical regularities of data rather than building causal models of the physical world. Philosopher Judea Pearl's causal reasoning hierarchy divides cognition into three levels: association (seeing what), intervention (what happens if we do something), and counterfactual (what if things had been different). Video generation models primarily operate at the first level—association—capable of generating images that "look right" but unable to answer intervention or counterfactual questions like "how would the ball bounce if initial conditions were changed?"

The first step on the AGI roadmap is Chain-of-Thought, requiring models to learn reasoning and thereby build an understanding of the world. Chain-of-Thought techniques attempt to make models demonstrate intermediate reasoning steps, moving toward higher levels of causal reasoning. Therefore, video generation is an application outlet for AI, not an entry point toward AGI. According to leaks, products like Sora can generate over 1 billion yuan in monthly revenue, but DeepSeek has never ventured into this space.

Not chasing traffic, not building a super-app

Liang Wenfeng candidly admits that when user numbers surged during last year's Chinese New Year, the company didn't pursue retaining those users or monetizing them, and never aspired to become the next ByteDance or Tencent. The reasoning is the same: "There's still a watermelon ahead; what's in front might all be sesame seeds." After acquiring a massive user base, you might end up being led by user demands, wasting limited resources. Earlier this year, some people brandished declining traffic charts proclaiming "nobody uses DeepSeek anymore," but this was never a metric the team cared about.

Committing to open source, not pursuing excessive profits

On open source, Liang Wenfeng's stance is clear: "I can't see any inherent advantage to being closed-source. None at all." He acknowledges that open source does cede some benefits, but his logic is: if AI ultimately accounts for 10% of global GDP, anyone who tries to monopolize that profit will inevitably be swept aside by history.

The debate between open-source and closed-source AI models is one of the industry's central issues. Open-source models (such as Meta's LLaMA series, DeepSeek series, Mistral, etc.) publish model weights, allowing anyone to download, modify, and deploy them; closed-source models (such as OpenAI's GPT series, Anthropic's Claude, Google's Gemini, etc.) provide services only through APIs without disclosing core parameters. Open-source models offer advantages including: enterprises can deploy on-premises or in private clouds, keeping data within their domain for better security and privacy; they can fine-tune and customize according to their own needs; long-term costs are far lower than continuously calling closed-source APIs. Closed-source models offer the advantage of accessing continuously updated state-of-the-art performance through API calls without needing to build inference infrastructure. From a business model perspective, closed-source supports SaaS subscription revenue, while open source generates commercial returns through enterprise services, cloud computing platforms, and ecosystem influence.

He made one particularly striking statement: "Those who take more will be defeated by those who take less... If your vision is to take more, you will be defeated by those whose vision is to take less." This echoes historical lessons where many uprisings collapsed from within because leaders cashed out early and distributed spoils prematurely—when a vision cannot withstand the temptation of short-term gains before being realized, it ultimately becomes an empty slogan.

DeepSeek's Team Stability: Proof of an Idealist Gathering

Restraint also serves as the best team adhesive. Liang Wenfeng revealed one detail: DeepSeek once priced a model higher out of concern about overwhelming demand, but when they later reduced the price, everyone was happy because "enabling everyone to fully use it" was the whole point of building the model well in the first place.

Amid the frenzied AI talent war both domestically and internationally over the past year (with offers reportedly exceeding 100 million yuan in annual compensation), the DeepSeek team has remained stable. From V3 to V4, the personnel list in technical reports expanded from 197 to nearly 300 people. Liang Wenfeng's name appears alongside everyone else's, and the core members from the V3 era are basically all still there. In the face of enormous financial temptation, this alone speaks volumes about the team members' values.

DeepSeek's breakthrough results under limited compute conditions rely on multiple architectural innovations. The MoE (Mixture of Experts) architecture divides model parameters into multiple "expert" sub-networks, activating only a subset during each inference—dramatically reducing computational costs. For example, DeepSeek-V3 has 671 billion total parameters but activates only approximately 37 billion per inference. The MLA (Multi-head Latent Attention) mechanism applies low-rank compression to attention key-values (KV), compressing KV cache to 5%-13% of traditional multi-head attention, greatly alleviating the memory bottleneck during long-sequence inference. Additionally, DeepSeek-R1's training methodology breakthrough is equally noteworthy: it demonstrated that even without large-scale human-annotated supervised data, Chain-of-Thought reasoning capabilities can emerge in models through reinforcement learning alone. This discovery challenged the industry's prior consensus that "you must do SFT before RL" and is considered an important milestone on the path to AGI.

Regarding the US-China gap, Liang Wenfeng's assessment is: China doesn't lack talent—it lacks compute. The core reason for China's AI compute constraints is US export controls on chips to China. Since October 2022, the US Department of Commerce's Bureau of Industry and Security (BIS) has issued multiple rounds of semiconductor export restrictions, prohibiting the export of high-performance AI chips to China, including NVIDIA's A100, H100, and subsequent flagship GPUs like H200 and B200. NVIDIA subsequently released "nerfed" chips for the Chinese market (such as A800 and H800), but these chips have significant reductions in key metrics like interconnect bandwidth, and were also brought under control in the updated October 2023 restrictions. GPU chips are the core hardware for training large models—a single frontier model training run may require tens of thousands of high-end GPUs working in concert for months. Insufficient compute directly constrains the number and scale of experiments Chinese teams can run, forcing them to innovate more in algorithmic efficiency—DeepSeek's MLA and MoE architectural optimizations are precisely technical breakthroughs born from these constraints. On the domestic alternatives front, Huawei's Ascend series chips are accelerating their catch-up, but gaps remain compared to NVIDIA in ecosystem maturity and single-chip performance. Liang Wenfeng believes the compute issue is short-term and likely to ease within a few years, and that in the meantime, efforts should be concentrated on breakthroughs in specific directions like coding capabilities.

Why DeepSeek's Open-Source Models Alarm Competitors

Why is the US so wary of Chinese open-source models, to the point of considering bans? The answer likely doesn't lie in so-called national security or cybersecurity risks—closed-source models face similar risks of being compromised and misused.

The real reason is the inherent appeal of the open-source model itself. Reportedly, over 200 US startups and some well-known large companies co-signed an open letter asking the government not to ban Chinese open-source models. This perfectly illustrates the principle that "those who follow the Dao gain many supporters; those who abandon it find few." Open-source models allow enterprises to keep accounts, keys, credentials, and other important data on-premises—secure, controllable, and inexpensive. This meets a real need for many developers. Since 2024, open-source model performance has rapidly approached and even surpassed closed-source models in certain domains, posing a substantive challenge to the business moats of closed-source companies like OpenAI and triggering intense debate at the US policy level about whether Chinese open-source models should be restricted from use in America.

Some interpret Chinese AI labs' open-sourcing as a "laggard's strategy"—using free offerings to undermine competitors, then switching to closed-source once ahead. But based on this conversation, Liang Wenfeng's logic is precisely the opposite: open source is not a tactic to undermine competitors but a strategy for realizing the vision of "making the world better." He isn't worried about others using his open-source models to compete, because that "just means taking a little less."

Conclusion: Vision First, Restraint as Strategy, Open Source as the Dao

What DeepSeek pursues is not monopolistic profit but a long-term vision. They don't chase short-term trends, don't compete for traffic, don't engage in hype, and focus all resources on the single goal of AGI. In the debate over whether AI will replace humans, one perspective worth considering is: AI is more likely to reshape the job market than cause mass unemployment—it's an amplifier of human capital. As long as we commit to making AI accessible to all, society's overall labor output will multiply. At that point, people may rethink what "employment" itself means.

Of course, the views expressed in this conversation carry a distinctly idealistic tone, and whether they can be sustained long-term remains to be verified by time. But at least for now, the philosophy DeepSeek demonstrates—"vision first, restraint as strategy, open source as the Dao"—offers a unique lens through which to observe the emerging force of Chinese AI.