Meng Yan argues that excitement over fast-rising AI performance often rests on a misunderstanding of how modern society actually works. In his view, many people overrate the importance of individual intelligence and individual productivity inside large collaborative systems.
He points to a familiar pattern: many institutions and large companies set high academic and intelligence requirements when hiring for well-paid roles, yet once people are inside, much of the actual work is ordinary and can be handled with average ability. He adds that favoritism and the priority of relationships over capability are still common, and organizations will often choose a person with ordinary skills but stronger social identity or social resources.
Why extra productivity often goes nowhere
Meng describes this as an “ineffective productivity spillover.” He says modern society is a high-density network of collaboration, and a good job is really an advantageous node in a network that allocates social resources and power. The resources tied to that node are scarce across society, but the person occupying the position often does not need extraordinary capability or very high output.
In that setup, organizations raise entry requirements not because the role truly needs exceptional productivity, but because many candidates can do the job and the employer needs some principle for choosing one of them. Ability becomes a relatively uncontroversial filter. Once someone enters the system, any extra productivity they bring may not be needed, used, or absorbed by the network.
He says that in modern collaborative networks, most people handle only a small part of the overall work. Aside from a small minority, their role is largely to route resources through the network. As long as their productivity clears the required threshold, further gains do not create a clear benefit for the system as a whole.
AI can raise efficiency without raising status
Against that backdrop, Meng says many people are now eager to talk about how much productivity they have gained through learning and using AI. But those gains, he argues, may have little value at the level of the broader collaborative network and may not be distributable through it.
His point is blunt: organizations often do not need and cannot digest this overflow of productivity. In practice, one of the main benefits for many AI users may simply be that they can free up more time for themselves, rather than move up in income, resources, or social position.
He applies the same logic to OPC, or the “one-person company.” The core bottleneck for most such firms, he says, is not internal productivity but external resource connections. Given where most OPCs sit in the broader commercial network, their products and services may not sell in the first place.
That matters because economic activity needs an economic incentive to close the loop and create positive feedback for future growth. At the moment, he argues, the positive reinforcement most people get from AI is emotional rather than economic, and that cannot last.
What changes incentives is connection, not just output
Meng says immediate positive incentives come from changes in network connections. Once connections shift, resources and power are redistributed with them.
If a node gains more links in the network, he writes, that person may quickly feel a rise in income, influence, and power even without using AI at all. By contrast, someone who stays in the same place and only enjoys local efficiency gains may not capture the early benefits of AI, even if their technical use of the tools is strong.
He does add a longer-term caveat. Broad productivity gains across nodes will eventually change connection patterns. In the AI agent era, differences in people’s AI ability are widening quickly, and those differences will sooner or later be recognized by the social network and converted into changes in connections, or even changes in network topology. Still, he believes that process will take time and move more slowly than many AI influencers suggest.
The most active AI learners are often secondary nodes
Meng says the people most actively studying AI today are often secondary nodes in the network: programmers, designers, researchers, and analysts. They spend their time on Agent systems, MCP, Skills, and workflows, trying to lift their productivity several times over, even by 10x.
But more individual output does not mean the organization can absorb it. His examples are straightforward. If a person used to write one report a day and now can write 10, the organization may not need 10 reports. If someone used to build one prototype a week and now can build one a day, budgets, decisions, sales, legal, and procurement have not suddenly sped up by 7x.
The result is that much of the new productivity gets stuck at connected nodes and creates limited value for the system as a whole.
Resource-rich decision makers are drifting away from the frontier
On the other side are the main nodes that control resources: entrepreneurs, executives, investors, and institutional leaders. These people already hold the most clients, capital, hiring authority, and social connections. If their AI capability rose 10x, the leverage could be far greater than what an ordinary employee could generate.
Yet Meng says the opposite is happening. The distance between these resource-rich actors and the frontier of AI application is widening quickly.
One reason, he writes, is that since the rise of AI agents in the second half of 2025, AI application innovation has become increasingly programming-driven. Earlier AI products centered on chatting, writing, search, and image generation, where both CEOs and programmers could participate. In the agent phase, though, many of the new advances first emerge through AI coding, including Agent systems, MCP, Skills, Harness, CLI, API, Git, IDEs, and a range of workflows and orchestration methods.
AI is starting to look less like a finished piece of software and more like a management system that users must assemble and direct for themselves.
Meng gives a concrete example. A 50-year-old CEO may be highly capable at managing a company, raising money, and mobilizing resources, and may fully understand that AI matters. But if that person has to open a terminal, install a coding agent, configure MCP, run several agents in parallel, and connect APIs, GitHub, and various tools, he may give up quickly. Meng says this is not an intelligence issue. It is a break in the skill stack.
That leaves what he calls a strange situation: the people who understand AI do not have resources, while the people with resources do not know how to use AI. The first group is pushing hard to raise productivity without having a large enough network to absorb it. The second group controls large resource networks but lacks the ability to plug the latest machine productivity into them. He sees this mismatch as a major issue to watch over the next few years.
Mainland China faces an added ecosystem gap
Meng says mainland China has an additional problem. The two most active AI application ecosystems today are OpenAI and Anthropic. Because mainland users cannot directly and reliably use those systems, the gap is not limited to the models themselves. It also includes the rapidly growing surrounding ecosystem of agents, coding tools, toolchains, and developer communities.
He argues that this may be more troublesome than the model gap itself, because knowledge in the agent era accumulates. Others move from chat to tool use, then from tool use to agent orchestration, and then to having agents build tools on their own. Arriving six months later may leave a user behind not by a few benchmark points, but by an entire way of working.
Three stages of AI’s social impact
Meng says the real question is when this mismatch begins to break. On one side are people now in secondary nodes but with strong AI capability. Can they use that strength to gain more clients, capital, attention, and social connections, and in turn change their position in the network? On the other side are people already near the center of the network. Can they truly master agents and combine resource allocation power with machine productivity?
In his framework, AI will change society in three stages.
- First, it changes nodes. Individual productivity rises by 3x, 5x, or 10x, and he says that is what is happening now.
- Second, it changes connections. Some people gain more clients, capital, attention, and organizational capacity through AI, and their position in the network shifts.
- Third, it changes network topology. At that point, an organization that once needed 100 people may need only a dozen or so. Work that once required a company may be done by a few people plus a group of agents. Some old intermediary nodes disappear, while new super-nodes emerge.
Only in that third stage, he argues, will AI truly transform social structure.
He closes with an observation about prominent AI commentators. Many of them get excited by each technical advance and make dramatic predictions that later fail to materialize. Even so, those predictions bring them fresh attention and new resource connections. In Meng’s view, that is the clever move: the real payoff is not AI skill or productivity by itself, but the attention it converts into network position. People who only improve their AI skills without expanding their social connections, he suggests, are unlikely to capture the earliest gains.