Who Is Defining the “Last Mile” of AGI? A Shenzhen-Based AlphaSpaceX Gives Its Answer: Space Itself

In July of this year, three events within the AGI landscape converged within the same week, sketching a multifaceted picture of the path toward ultimate intelligence. DeepSeek founder Liang Wenfeng made it clear in an investor meeting that his company would not pursue 3D generation, video generation, or world models, channeling all resources into the cognitive intelligence core. Zhipu AI founder Tang Jie issued an internal memo announcing a "Touch High" initiative, with strategic investments aimed directly at the next frontier of AGI. Meanwhile, "AI godmother" FeiFei Li's World Labs secured $1 billion in funding at a valuation of approximately $5 billion, doubling down on spatial intelligence. Three distinct paths, all converging on the same destination—AGI. Behind these divergences lie different dimensions of general intelligence, each being unlocked in its own way.

Liang Wenfeng's strategy is built on a philosophy of "restraint." He explicitly rules out major investment in 3D generation, video generation, and world models, and has no intention of building verticalspecific Agents or pursuing userscale and profit maximization. In his view, directions like video generation and 3D generation "have little to do with the ceiling of intelligence," while multimodality "is merely a component, not the mainline or intelligence itself." His roadmap is a staircase: last year's step was ChainofThought; this year's is Agents; and beyond that lies the challenge of continuous learning. "Everything before is sesame seeds; the watermelon lies ahead." DeepSeek's opensource models have consistently ranked at the top of global AI model invocation volumes, proving that focusing on the cognitive intelligence core can generate enormous industry impact.

In contrast to Liang's "subtraction," Tang Jie has chosen to "break upward," asserting that "the wave is here; not reaching the summit is failure." Zhipu's "Touch High" plan centers on longhorizon task capability, fully autonomous agent systems, complete selftraining, and extreme safety governance. The company plans to invest billions of yuan in mechanistic interpretability research, seeking to understand the internal mechanisms behind model judgments. Tang Jie believes that AGI is not the genius of a single individual but the sum of all human intelligence, and that it must possess the ability to generate original knowledge at the level of the theory of relativity. The two paths are not opposites—the divergence lies in the route. Zhipu expands the boundaries of intelligence upward, while DeepSeek solidifies its foundations from within.

FeiFei Li offers a fundamentally different diagnosis: large language models cannot lead to AGI; spatial intelligence is the optimal path. Her logic rests on two progressive judgments: language has only about 500,000 years of history in human evolution, while spatial intelligence—vision, touch, and physical interaction—has been evolving for over 500 million years. If AI understands only linguistic symbols without grasping the physical world, it will remain trapped in digital pixels. AGI requires "understanding," not mere "description"—AI should possess the innate abilities of an infant: to comprehend depth, occlusion, and the physical laws of threedimensional space. World Labs secured $1 billion in funding in February 2026, with participation from NVIDIA, AMD, and a16z. In July, it announced the acquisition of robotics simulation company SceniX, signaling its next bet on embodied intelligence.

From this panorama, a clear conclusion emerges: AGI is not any single company's exclusive answer—it is the combination of answers across all dimensions. A 2026 paper from Google DeepMind notes that the evolution from AGI to ASI does not necessarily depend on a single model or a single technical path—it is more likely the result of scale expansion, paradigm evolution, selfimprovement, and multiagent collaboration working together. Peking University professor Zhu Songchun further points out that "large models are neither the end point of AGI nor the only path"—the true path to AGI lies in social intelligence. Pure cloudbased text large models have inherent limitations. A truly mature AGI system will inevitably be the product of multiple technical routes converging: longterm memory, strong reasoning, longhorizon autonomy, physical world cognition, and realworld spatial deployment—all are indispensable.

Beyond the discourse on cognitive intelligence, spatial intelligence, and social intelligence, a longoverlooked dimension is emerging: space itself. AlphaSpaceX's concept of "DigitalIntelligent Space" points to the "field ceiling" of AGI: general intelligence requires physical space as its carrier. Space itself—not a single device—is the fundamental unit through which intelligent agents interact with the real world. Behind this is a trillionyuan realworld scenario: China has 7,312 existing shopping centers, 7,188 museums, over 16,500 Agrade scenic spots, and more than 35,000 corporate showrooms. The digitalintelligent transformation of these physical spaces does not need "smarter AI"—it needs an operating system that lets spaces get things done on their own. AlphaSpaceX's proprietary "SpatioCultural Token" sixdimensional data framework—spatial coordinates, virtual imagery, temporal sequences, semantic labels, physical constraints, and action instructions—is designed to structure the complexity of physical space into computable data. Its "Trinity Space Operating System" enables unified orchestration of the three types of agents within a space: humans, digital avatars, and robots.

The end of cognition is action; the end of action is space. All thinking, decisionmaking, and execution ultimately must close the loop within physical space. When AGI discussions center on "model capability" and "perception capability," a more fundamental question has been overlooked: where will intelligence ultimately run? The answer is not in the cloud, not on a chip—but in every inch of physical space itself. From Liang Wenfeng's strong reasoning to Tang Jie's longhorizon autonomy, from FeiFei Li's physical world understanding to AlphaSpaceX's digitalintelligent space—these seemingly opposing paths are actually projections of general intelligence across different dimensions. They will converge at the end state of AGI, collectively forming a complete intelligence that can think, decide, act, perceive, and create. The end of intelligence is space. And "making spaces speak, making cities living organisms" may be the final piece of the AGI narrative.（Author: Hui Siqi）

标签：