What is the real AI thinking?

Currently, much of the understanding about AI is in a complete mess. Some people claim that pattern recognition is not a native form of intelligence, while others argue that pattern recognition is a typical AI algorithm and thus should be considered native to intelligence. Similar debates exist regarding unmanned companies, single - person companies, and so on.

A rather unique aspect here is AI thinking, which is very different from Internet thinking. Nowadays, few people talk about it.

So, does mastering AI require a new way of thinking? If so, how should it be defined?

Application Levels of AI

We can always use various large models in the same way we use a more advanced version of Word. In this case, AI serves as a better tool. At this stage, there's really no need for AI thinking; regular use will suffice.

However, AI is obviously more than just a tool. Multi - agent systems can encapsulate entire business processes into their own frameworks. At this point, AI is no longer just a simple tool but becomes the main body of value creation.

Of course, there are different levels between the tool and the main body, roughly as follows:

The further we go, the more a new way of thinking is needed. Otherwise, just as Genghis Khan's tactics couldn't handle light infantry, the more one tries to advance, the more likely they are to harm themselves, and haste makes waste.

Intelligence First

In the book Unmanned Company, the top - ranked new principle to follow is Intelligence First.

Note that it's not the boss, the status quo, etc. that come first, but intelligence.

This is actually the same as AI becoming the main body of value creation and is also a prerequisite for AI to truly be effective.

Some people may ask, what if we can't prioritize intelligence?

Then just use AI as a tool and don't let it be the main body. Otherwise, even if the multi - agent system runs for a short period with some effects, it will gradually decline.

Since the cost of using AI tools is very low (it should be easier than learning Office), the key to making AI truly effective lies not in using the tools but in encapsulating business processes based on the basic characteristics of AI.

The most crucial factor in the combination of the above three aspects and business is AI thinking.

AI Thinking

AI thinking is a brand - new problem - solving methodology that we must adopt when applying the "Intelligence First" (AI First) principle to the organization process of production and services.

It doesn't mean that individuals need to learn to write code or use AI tools. Instead, it refers to thinking and acting in a mode that is inherent in computation and simulation at the strategic and execution levels. Its core essence can be summarized into three points: Virtual - First Simulation, Rapid, Scalable Trial and Error, and Computational Hedging.

1. Virtual - First Simulation: Rehearse Everything Before Action

The traditional business model follows the linear process of "Plan - Do - Check - Act" (PDCA), where each step occurs in the physical world, resulting in extremely high trial - and - error costs. The first principle of AI thinking is "Virtual - First Simulation," which means creating a "World Model" in the digital world that highly corresponds to the real environment before investing real resources and conducting large - scale simulations within it.

This world model, as recently discussed in academia, is an algorithmic proxy for the real - world environment. It can be narrowed down to be relevant only to one's own business. Its core goal is not to generate realistic videos for entertainment but to simulate all possible actions in the real world to support purposeful reasoning and action.

This ability is called "Hypothetical Thinking" in psychology and is what we commonly refer to as "thought experiments" in practice.

AI makes the cost of such vertical - domain thought experiments extremely low.

Whether it's AlphaGo exploring chess moves that humans have never thought of through self - play or an autonomous driving system predicting the future trajectories of all vehicles and pedestrians on the street, the essence is to conduct deductions on countless "possibilities" in a low - cost virtual world to find the optimal solution.

This is the first superpower that AI thinking gives us: seeing the future before taking action.

2. Rapid, Scalable Trial and Error: Explore the Optimal Path with Parallel Computing

Human trial and error is sequential, expensive, and limited by individual energy and experience. AI, on the other hand, can conduct millions, tens of millions, or even billions of parallel trials and errors in the world model with almost zero marginal cost.

A marketing team may take a week to design and evaluate three advertising plans. An AI Agent, however, can generate a thousand combinations of copywriting and images within an hour, test the click - through rate and conversion rate among a virtual user group, and iterate in real - time based on feedback, ultimately selecting the best few plans for the real market.

This large - scale, automated trial - and - error cycle has increased the speed of innovation by several orders of magnitude. It's like changing the time axis.

The basis for this ability is that Virtual - First Simulation can generate countless "hypothetical trajectories," allowing agents to make full use of all "imagined experiences" through methods such as reinforcement learning or imitation learning.

It's worth noting that if the trial - and - error cost is low enough, especially in the digital space, we can skip Virtual - First Simulation.

3. Computational Hedging: Replace Physical Costs with Computational Costs

The economic basis of "Virtual - First Simulation" and "Rapid, Scalable Trial and Error" is "Computational Hedging." This means that we can use relatively inexpensive computational resources (CPU/GPU time, electricity) to hedge and replace extremely expensive physical - world resources (such as time, raw materials, human capital, and market opportunity costs).

In the past, verifying a new drug required several years of clinical trials and billions of dollars in investment.

Today, AI can simulate the interaction between drugs and proteins in a molecular - level world model, pre - screen a large number of ineffective or toxic candidate drugs, and narrow down the scope of physical experiments to a few most likely successful options. Here, millions of dollars in computational costs hedge against hundreds of millions of dollars in R & D failure risks.

Similarly, when a company decides whether to enter a new market, it no longer needs to spend months on expensive market research. Instead, it can operate a "virtual branch" in a world model that simulates the consumer behavior, competitive landscape, and social culture of the market, observe its virtual financial reports for several quarters, and then make a final decision.

If we have to find a unified example for the above points, we can recall the story of developing apps in the mobile Internet era:

You can carefully develop an app in a specific direction;

Or, like a certain company, you can create an array of apps and keep the successful ones.

Obviously, the key to the latter approach lies not in the idea but in the trial - and - error cost and success rate. AI thinking can undoubtedly greatly enhance the universality of the latter approach, not limited to app development.

Unmanned Company: The Ultimate Organizational Carrier of AI Thinking

When the above three AI thinking modes are systematically applied to a business organization, its final form will inevitably evolve into an "Unmanned Company."

An "Unmanned Company" doesn't mean that there are no people in the physical space. It means that the core value - creation chain is dominated by AI agents rather than human employees. The role of humans changes from hands - on executors to goal designers, rule - makers, and value - givers.

In such an organization, AI thinking is no longer just an "icing on the cake" tool but the "operating system" on which it depends for survival.

Its technical core can borrow the blueprint depicted by the latest paper: the general world - model architecture of PAN (Physical, Agentic, and Nested).

Physical: An unmanned company needs to simulate the physical dynamics of the real world. For example, an unmanned e - commerce company's world model needs to understand the complete logistics process of a package from the warehouse to the user.

Agentic: The core of the company is autonomous agents. An unmanned company must support the simulation of multi - agent behavior. For example, how an agent in charge of marketing and an agent in charge of customer service can work together. Its future development direction is to expand from single - agent simulation to the simulation of the collective behavior of an entire business or society.

Nested: The world model of an unmanned company is hierarchical and nested. It can use a structure similar to LLM for strategic planning and conceptual reasoning at a high level and use diffusion models to handle fine physical or sensory details at a low level.

As mentioned before, the physical and agentic aspects may overlap. However, overall, this is a system based on dependency inversion: controlling the real with the virtual.

Take the simplest example. A typical workflow of an unmanned company is as follows:

The human founder sets a business goal (such as "Increase the ROAS of a certain product to 2 this quarter"). This goal is input into the company's "brain." Then, multiple AI agents (market analysis agent, advertising creative agent, budget allocation agent, etc.) conduct a large number of simulated advertising experiments in this model sandbox. They will "pre - calculate and cache various possible world states, feasible actions in these states, and their simulated results." (It's not necessary to simulate; a small amount of real - world operations can also be carried out.)

Finally, the system will select an action plan with the highest expected return and automatically execute it on a real - world advertising platform (such as Google Ads).

From Theory to Reality: The New Business Wave Shaped by AI Thinking

Although the fully mature and widespread "Unmanned Company" is still a future vision, the principles of AI thinking have penetrated into current business hotspots and demonstrated great power.

Industry and Manufacturing: Digital Twin and Virtual Factory

Nvidia's Omniverse platform is a typical example. Before building a new production line, an automobile manufacturer will create a 1:1 digital twin factory in Omniverse. Engineers can simulate every movement of the robotic arm, test the production line rhythm, optimize the logistics path, and even simulate the safety of workers' operations in this virtual factory. This is a perfect embodiment of the ideas of "Virtual - First Simulation" and "Computational Hedging," replacing expensive physical installation and rework with virtual debugging.

Content and Marketing: AIGC and Automated Growth

The traditional marketing model is being disrupted. Nowadays, a "one - person team" can use GPT to generate marketing copy, Midjourney and Sora to generate advertising images and videos, and then use automated tools for full - channel distribution and A/B testing. Behind this is AI thinking driving: generating a large amount of content creatively and testing its effectiveness at a very low cost, which used to require a large team to complete. Although current world models focused on video generation (such as Sora) do not support interactive reasoning, they have shown great potential in content creation.

Science and R & D: AI - Driven Hypothesis and Verification

The essence of scientific research is the cycle of "proposing hypotheses - conducting experiments - verifying conclusions." AI is accelerating this cycle in an unprecedented way. For example, AlphaGeometry2 mentioned in the paper can solve Olympic - level geometry problems, which is essentially an efficient "thought experiment" in a pure mathematical world model. Systems like ReasonerAgent can automatically conduct literature research and information integration on the Internet to assist human researchers in forming and verifying hypotheses more quickly. Chai 2 mentioned above is also a good example.

The Future Belongs to Enterprises That Master the "Simulation - Action" Flywheel

We are moving from a "experience - driven" business world to a "simulation - driven" business world.

The core competitiveness of an enterprise is no longer just about how much capital, how many talents it has, or how much past success experience it has accumulated. The core competitiveness in the future will depend on: how high is the fidelity of your world model's simulation of the real world? How fast can your "simulation - action" flywheel spin?

Mastering AI thinking means mastering the ability to "foresee the future" and "choose the future" at the lowest cost. Based on this, the "Unmanned Company" will have agility, efficiency, and scalability that traditional organizations cannot match. They can more flexibly adapt to the changing market, more accurately capture users' potential needs, and ultimately gain a structural advantage in the competition.

Our ultimate goal is to build an AI system with "the adaptability, resilience, and autonomy unique to human intelligence." This path is long but full of opportunities. Enterprises and individuals who are the first to embrace AI thinking and start building their own "world models" and "Unmanned Company" prototypes will undoubtedly become the pioneers in defining the next business era.

Of course, before taking action, we need to first narrow down the business scope and then determine how many levels of simulation are needed.

We can first observe and learn without rushing. For example, we can read Unmanned Company...

There is some imagination in the above content, but obviously, it will build a real intelligent civilization, which will be a very different world.

This article is from the WeChat official account "Zuo Moshi". Author: Lao Li Hua Yi San. Republished by 36Kr with permission.