A recently leaked recording of Liang Wenfeng's investor exchange has fully exposed DeepSeek's AGI technical roadmap. Different from the outside imagination of "video generation + world model + multimodality + embodiment" all cooked in one pot, DeepSeek has drawn a restrained main line: language model → CoT → Agent → continual learning → self-iteration → embodied intelligence. Each step is the next bottleneck that naturally emerges once the previous capability is unlocked — not a product-manager decree. The most noteworthy sentence in this roadmap is: after Agents, the next problem to solve is continual learning. Liang Wenfeng said bluntly: today's Agents look bustling, but they're essentially engineering wrappers around zero-shot inference — parameters don't change at inference time, so no matter how strong an Agent is, it's just "an advanced LLM that calls tools", not a real employee who can "show up for work". Once a model has continual learning, like a human who can take over the job after two months on the desk, the production relations will truly be rewritten. Alongside this, video generation, 3D, and world models are explicitly placed on the "non-main line" — because these directions are important for C-end products and for commercial landing, but they don't have a direct causal relationship to "intelligence upper-bound" improvement. This is the restraint of "not doing technology for the hot money in front of you". On the other hand, their attitude toward Scaling is honest: they believe in it, but acknowledge that "what's stopping us from Scaling is compute". So low-cost training, domestic-compute adaptation, MoE/MLA, in-house compilers — these aren't cost-saving tactics, they're trading engineering efficiency for iteration headroom. The lesson for the industry is clear: when compute is a finite supply, the model ceiling is no longer determined by parameter scale, but by the density of engineering efficiency and training methodology. This is also why, in 2026, "inference-side innovations" like MoE, hybrid attention, and speculative decoding get more attention than simply stacking parameters — intelligence density is replacing parameter scale as the new arms-race dimension.