Liang Wenfeng, who started in quantitative trading and is now fully committed to AGI, thinks and speaks differently from those who constantly talk about “ecosystems.” He compares AGI to climbing a staircase: last year’s step was CoT, this year’s is Agent, and the next will be continuous learning. Only after mastering continuous learning can we reach the gradual singularity, followed by embodied intelligence. Embodiment is the final stop—his reasoning is that for ordinary people, what they need isn’t a computer, but human labor. This single sentence captures DeepSeek’s technical priorities for the next three to five years. Even at the Agent stage, he makes clear trade-offs. Right now, the highest priorities are Coding Agent and General Agent; vertical applications like finance and healthcare come later. The key gap in the next-generation model is continuous learning. He says AI today doesn’t lack taste or intuition—it lacks the ability to learn continuously. Humans learn continuously; AI still cannot. The next-generation model must have this capability—it’s the fastest path to AGI. But learning is composed of many components, and the world has yet to find a good solution. This philosophy underpins DeepSeek’s business strategy: unwavering commitment to open source and extreme cost efficiency, earning only reasonable profits. They prioritize cost-effectiveness above all else, enabling them to offer the lowest prices. Their API sales require no sales team or customer service—users come on their own. He firmly believes no large model company can capture the majority of profits in the AI industry. Open source is part of this restraint. He says: “A software company might make billions annually—but open it up, and the revenue vanishes.” But AI is big enough—it could eventually account for 10% of global GDP. Trying to monopolize this value will inevitably lead to being discarded by history. The competitive gap lies in three areas: cost, time, and user experience. Cost comes first—being months ahead makes a real difference. User experience creates stickiness, but it’s not fundamental. Industry structure: fragmented resources will inevitably consolidate. China’s main disadvantage in AI lies in resource allocation. They believe in scaling—the bigger the scale, the better the results—but resources are limited. There’s no talent gap; everyone comes from the same pool. There are too many domestic model companies, leading to fragmented resources—and consolidation is inevitable. If every company earns only reasonable profits, far fewer teams are needed to build large models. Two big companies and two small ones may be enough. Anthropic surpassing OpenAI is merely a temporary phenomenon. In the long term, OpenAI and Google will likely alternate in leadership. My takeaway: resource misallocation is China’s greatest waste in AI. The same pool of talent is split across dozens of companies, wasting energy through internal competition—consolidation is inevitable. Organization and people: restraint is strategy. They have no interest in becoming the next super app—they see it as trivial. Don’t trade the melon for a sesame seed. Last year, everyone raced for Chatbots and C-end traffic; this year, they’re fighting for B-end revenue. What they truly care about is the AGI roadmap. Only a restrained organization can pursue a restrained strategy. A company driven by KPIs and GMV could never say things like “open source” or “earn only reasonable profits.” If your vision is to take more than your share, you’ve already lost.