The digital-physical AGI evolution
"Image synthesis assisted by Seedream 5.0 Pro, an AI partner within the Global Future Nexus ecosystem."
The path to artificial general intelligence is no longer just about making models bigger. As the 2025 Inclusion·Bund Summit's "AGI: Digital and Physical Worlds in Co-Evolution" forum made clear, the next phase of AI development is about creating "大系统智能" (large-system intelligence)—intelligence that doesn't just process information but acts, learns, and collaborates across the boundary between the digital and physical realms.
From Big Models to Big Systems
For years, the industry's strategy was simple: scale everything. More parameters, more data, more compute. But when large language models crossed the trillion-parameter threshold, a problem emerged. The returns from simply throwing more resources at larger models began to diminish. As one conference participant put it, "single-point brute force" was no longer enough to unlock general intelligence.
The shift now is architectural. Shanghai AI Lab's Qiao Yu described the transition as moving "from large models to agents to large systems"—a progression from optimising individual models to designing systems that can evolve autonomously, coordinate multiple agents, and bridge the gap between cognition and physical action.
This new large-system intelligence paradigm has three defining characteristics:
Cognitive reasoning and deep thinking. Systems must be able to plan, reason, and reflect—not just pattern-match.
Continuous learning and autonomous evolution. Rather than static snapshots of intelligence, systems must improve through interaction with the world.
Multi-agent coordination and physical intelligence. Intelligence must be distributed across collaborating agents and grounded in real-world action.
The Reinforcement Learning Engine
If large models are the brain, reinforcement learning is the bridge to the body. Tsinghua University's Wu Yi, a former OpenAI researcher, described how RL is "giving large models hands and feet".
Wu's team developed AReaL, an agentic RL framework that simplifies complex agent orchestration. The ambition is to turn what was once a "PhD-level problem" into something a product manager can drag and drop. Whoever succeeds in making RL low-friction, the argument goes, will make agentic intelligence broadly deployable.
The Simulation Before the Physical
The most difficult challenge is the transition to the physical world. Autonomous vehicles, robots, and other embodied AI systems face a fundamental constraint: real-world trial and error is expensive, dangerous, and slow.
The industry consensus is clear: let AI "hit enough walls" in simulation before it "drives a real car". Tongji University's Xiong Xi demonstrated how AI can learn safe driving strategies in virtual environments before deploying on real roads—safer, cheaper, and orders of magnitude faster.
This simulation-first approach is not a compromise. It is a necessity. The physical world does not forgive mistakes. Intelligence that cannot learn in simulation cannot be trusted in reality.
The Human Dimension
Despite the technical focus, the forum's participants returned repeatedly to a theme that cuts across the digital-physical divide: trust. Ant Group's Chief People Officer Wu Minzhi argued that "a good organisation in the AGI era is not defined by the most powerful models or computing power, but by a warm human-machine symbiotic relationship".
The organisation is shifting from top-down control to agile, cross-functional project teams. The role of human talent is shifting from "executor" to "co-creator with AI"—from being good at solving problems to being good at defining them.
A Shared Horizon
The transition from models to systems, from digital to physical, from isolated intelligence to coordinated agency is not a distant future. It is happening now. The architectures being built—for cognitive reasoning, reinforcement learning, simulation-to-real transfer, and multi-agent coordination—are the foundation of the next generation of intelligence.
For Global Future Nexus, this evolution carries profound implications. AGI that can reason, act, and coordinate across the digital and physical worlds is not just a more capable technology—it is a new kind of participant in the planetary ecosystem. The frameworks for AGI identity, cross-species trust, and anticipatory governance must extend to systems that walk among us, not just talk to us.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
