Summary of Key Points
Liang Wenfeng, the founder of Deepseek, has invested 20 billion yuan in AGI (Artificial General Intelligence) and shared 20 key insights at a closed-door investor meeting, covering four main areas: AGI technology pathways, computing power and domestic chips, open-source commercialization strategies, and organizational management. He emphasized the importance of "ordinary people working together to achieve great things," highlighting both the shortcomings of China's AI industry (lack of computing power) and its strengths (low-cost reasoning and local expertise). Deepseek also clarified its long-term strategy of avoiding short-term trends, focusing on open-source development, and taking a steady approach.
I. AGI Technology: Focusing on Core Issues
Liang Wenfeng believes that AGI research should focus on the fundamentals and avoid being distracted by secondary technologies:
- The larger the model, the better the results—although China lacks computing power: The industry's belief that "larger models lead to greater intelligence" has not yet reached its peak; the reason China cannot create larger models is a lack of sufficient computing resources (specifically, GPUs). While Silicon Valley claims to have already achieved this level of scaling, China still has a way to go.
- The biggest challenge is continuous learning: AI currently does not have the ability to accumulate experience like humans (for example, employees can learn new tasks quickly after two months on the job, but AI cannot). This is a critical hurdle that requires global efforts and cannot be solved by a single technology.
- Multimodal capabilities and video generation are secondary: Technologies like 3D video and world models are not directly related to the core goal of making AI think and learn like humans; Deepseek will not pursue these areas for now.
- Product-related issues, such as hallucinations (random speech): Although AI's occasional nonsense can be improved through training, it is not the most urgent area of focus.
- AI evolution is a gradual process: It will first develop the ability to think logically (step-by-step reasoning), then become an intelligent entity capable of completing tasks independently, followed by solving the problem of continuous learning, and finally achieve "embodied intelligence" with a physical body.
II. Computing Power and Chips: Domestic Alternatives Have Potential, but Time Is Needed
Computing power is essential for AI development. Liang Wenfeng's view on domestic chips is practical:
- NVIDIA's monopoly is weakening: NVIDIA dominates the market with its CUDA system, which makes developers accustomed to using its chips. However, since AI can now generate code automatically, Deepseek's self-developed TileLang language allows them to quickly replicate NVIDIA's ecosystem and break this monopoly.
- Huawei chips can compete, but there is a gap: Huawei's 950 super nodes can replace NVIDIA's GB200/300 chips, but four Huawei cards still lag behind one NVIDIA card in performance (the latest Huawei chips are equivalent to those from two years ago).
- The gap between China and the US lies in computing power, not talent: China lacks GPUs, which limits the number of experiments that can be conducted and hinders talent development. This gap can be narrowed through algorithm optimization (making AI more efficient with less computing power) and the advancement of domestic chips.
- The market for large models will be reshaped: Many companies in China are developing foundational large models, but only a few will retain complete R&D capabilities, leading to more competitive prices.
III. Open Source and Commercialization: Sharing Honestly, Not Seeking Quick Profits
Liang Wenfeng's approach to commercialization is cautious and focused on long-term benefits:
- Open source is a voluntary choice: While some companies were forced to open source, Deepseek does so out of a belief that AGI is a global endeavor that requires collaboration.
- No hidden versions: The open-source models provided to the public are identical to those used internally; no better versions are kept back.
- Avoiding short-term gains: They willingly give up some immediate profits to increase the chances of successfully developing AGI.
- Not trying to monopolize the market: The AI market is too large; attempting to control it alone will lead to exclusion by the industry. Instead, they aim for a win-win situation with customers and society.
- Fair pricing: Their API fees are designed to recoup the cost of servers within ten months, ensuring a balanced approach that benefits both parties.
IV. Team and Purpose: Working Together with Ordinary People
Liang Wenfeng values his team and their shared vision:
- Team stability is crucial: The company's core goal is to maintain a stable team; as long as the team remains intact, AGI will eventually be achieved.
- Decisions are made through consensus: Major decisions require everyone's agreement, not just the boss's decision.
- No need for a written vision: While there is no official mission statement, the company's approach and attitude towards the world reflect a common goal.
- Starting a company was not about going public or making money: The motivation was to do something beneficial for humanity.
- Belief in ordinary people: He believes that a group of ordinary individuals can achieve extraordinary things together, rather than relying on a few geniuses—this aligns with their commitment to moderation and kindness.
Liang Wenfeng's investment of 20 billion yuan represents more than just financial support; it also reflects a long-term commitment to AGI, focusing on practical solutions, avoiding short-term gains, and believing in the power of ordinary people. This approach may represent another potential path for China's AI industry to make breakthroughs.