MarsBit has published an article by GritMeng arguing that in the AI era, the scarcest capability is no longer knowledge itself, but system thinking.
The piece opens with a blunt question: if AI can raise a person’s efficiency by 100 times, why can a company still end up close to collapse? GritMeng writes that coding, design, copywriting, and data analysis, skills that once took years to master, are now being rapidly offset by large models. For many knowledge workers, that shift has created a sharp sense that the craft they relied on is losing value fast.
From there, the article lays out a workplace paradox. Teams are using AI aggressively to improve productivity, yet company-wide decisions can still get worse. Product, operations, and marketing may each optimize KPIs to the limit, only to see user experience and profit fall sharply. Companies may spend tens of millions of yuan to install advanced AI systems, then end up reverting production scheduling to Excel and manual reconciliation.
GritMeng describes the problem this way: organizations are trying to harness an algorithmic racehorse with extreme computing power to pull a broken cart assembled out of point-by-point thinking. Once everyone becomes absorbed in using AI to solve isolated problems, the risk of a system-level breakdown rises. In that framing, the key question in the AI age is no longer just how to do things right, but what is right to do in the first place. The article calls that capability system thinking.
Reductionism pushed to its limit
The article says industrial civilization spent the last 400 years training people to think through reductionism: break a complex problem into separate parts, optimize each one, then assemble the whole again. In GritMeng’s view, that method built modern material prosperity, and AI now acts as its ultimate weapon. Give the model a clear target, and it can search through more options in a second than a human could calculate in a lifetime.
But the piece argues that the real world is not a machine that can be freely disassembled and rebuilt. It is a dynamic network in which every part affects the others. In that kind of environment, optimizing one piece to the extreme can become a direct hit to the system as a whole.
Drawing on 22 years of supply-chain research, GritMeng gives several examples. A procurement team uses AI to force material costs to the lowest possible level, upstream suppliers start cutting corners, and brand credibility eventually takes a hard fall. A finance team pushes inventory turnover to the limit, and a production line shuts down because one screw is missing. A marketing team drives click-through rates to the top of the industry, but after overpromising, the after-sales unit gets buried in complaints.
The article argues that this is the failure mode of smart organizations that lack system thinking but use AI as a single-point accelerator. In that case, AI becomes an engine without brakes, pushing the company toward destruction at greater speed. GritMeng labels this a "Goodhart collapse": once a metric is optimized too aggressively, it stops being a good metric.
Three core capabilities behind system thinking
The article says many people confuse system thinking with simply being more comprehensive or drawing a mind map. GritMeng instead defines it as a cognitive operating system that can be described with precision, built on three hard capabilities.
Boundary setting
AI always searches for an optimal answer inside the boundaries and goals it is given. For that reason, the first move in system thinking is not solving the problem, but examining whether the boundary itself makes sense.
The article offers a simple example. Ask AI to maximize a company’s profit, and it may quickly return what looks like a perfect answer: fire all employees, sell all assets, and put the cash in the bank to collect interest. The target is met. The company is dead.
For GritMeng, the way a problem is defined matters more than the way AI solves it. If a supply chain is framed around minimizing procurement cost, the answer will differ sharply from a boundary defined as full life-cycle capital efficiency. If a product is judged by user time spent, the outcome will differ from one built around long-term user value.
Closed-loop awareness
The article contrasts first-order thinking with system thinking. First-order thinking sees a straight line: A causes B, and B causes C. System thinking looks for delayed positive and negative feedback loops. A may drive B, but B can come back three months later and suppress A. C and D may appear unrelated, yet still affect each other through a hidden variable E.
GritMeng uses two examples. Generating 100 AI-written marketing advertorials may create a traffic spike in the short term, but three months later it can trigger a sharp collapse in user trust. Squeezing a supply chain to hit quarterly earnings can improve near-term financial statements, but six months later suppliers may raise prices across the board, permanently damaging the company’s cost structure.
In the article’s telling, a person with system thinking can see that destructive negative feedback on day one, not only after it arrives.
Reconstruction
The piece also argues that every AI model, every management theory, and every successful past experience contains a permanent gap when measured against the complexity of the real world. GritMeng calls that gap the residual.
Ordinary people ignore residuals. Smart people use them to fine-tune the model. People with system thinking go further: they activate metacognition, step outside the framework they have depended on, inspect it, criticize it, and ask a harder question — whether the thing they have long believed to be correct is itself the source of the problem.
The article says real innovation happens in the residual, and paradigm shifts come from those willing to break their old frame.
A five-dimensional mental system unique to humans
On the question of why AI cannot spontaneously produce system thinking, GritMeng’s answer is not that computing power is too weak, but that the hardware is wrong. The article argues that system thinking rests on a five-dimensional human mental operating system, with each ability tied to a different neural network in the brain.
- Conscience (Esp): linked to the default mode network, or DMN. The article says this sets the ultimate boundary of the system and slams on the brakes before it goes over a cliff. In GritMeng’s example, AI does not have that brake.
- High sensitivity: linked to the amygdala-locus coeruleus system. It is described as a radar able to catch weak abnormal signals behind the data and outside the report.
- Sensibility: linked to the salience network, or SN. The article calls it an intuitive cache that can draw on a lifetime of experience at low energy cost and deliver a vague but directionally correct judgment.
- Fluid intelligence: linked to the frontoparietal control network, or FPN. This is the logical calculation engine, and the article says AI is the best substitute for humans on this dimension.
- Metacognition (Φ): linked to the frontopolar cortex, BA10. GritMeng calls this the commander of system thinking because it can step outside an existing frame, inspect it, criticize it, and rewrite the program of thought.
The article’s conclusion is that silicon-based AI is essentially an amplified form of fluid intelligence. System thinking, by contrast, requires coordinated work across all five dimensions. That is why GritMeng describes it as the strongest and hardest-to-surpass "neural constitution" available to carbon-based life in the digital era.
From high-level executor to system architect
The article closes with three practical principles for readers who want to become system architects rather than advanced executors.
Move from receiving instructions to setting boundaries
Before starting any AI task, GritMeng says people should ask three system-level questions: where is the real boundary of the problem, which critical variables have been excluded, what second- and third-order chain reactions the AI plan might trigger, and whether the KPI being pursued has already mutated into a proxy metric that can damage the company.
Move from process optimization to feedback-loop design
The article says AI should not be deployed as an open-loop execution machine. It needs a monitoring radar that can sense the condition of the whole system, along with clear red lines that trigger human intervention or automatic shutdown. In the author’s wording, carbon-based conscience has to install brakes on silicon-based computing power.
Move from certainty to residuals
GritMeng also urges readers not to focus only on where AI is correct. They should pay attention to the residual between AI predictions and actual outcomes. According to the article, that is where old models break down, where customers’ real pain points remain unmet, and where organizational rigidity and decayed processes first become visible. Disruptive innovation starts there.
Computing power defines efficiency, but system thinking decides survival
The piece ends with a sequence of comparisons: in the agricultural age, physical strength was most expensive; in the industrial age, technology held that role; in the information age, knowledge did. In the AI era, GritMeng argues, the highest-value capability is system thinking.
As AI drives the cost of acquiring knowledge and optimizing local tasks closer to zero, the article says competition shifts to a higher level: who can see the whole system, who can understand how order forms and evolves, and who can use conscience and metacognition to direct large-scale silicon-based computing power. The final message is straightforward: do not try to beat AI in memory or raw computation. Build the system thinking that remains uniquely human. That, the article argues, is the most durable moat in the digital age and the only reliable compass for the future.