To all the pathfinders and pragmatists who persist in their quest amidst the AI entrepreneurial wave:
If you have ever been drawn to the concise power of "first principles" and attempted to use it to reshape an industry, then over the past two years, you must have witnessed or personally experienced that sense of setback when reaching from the high ideals down to the ground of reality.
The problem may not lie in the spirit of pursuing essence, but in whether we have misread the essence of "complex systems."
When the irreducible interactive network of the business world meets a thinking model that attempts to reduce everything to basic units; when the dependence and lock-in of technological paths meet an ideal blueprint of "calculating from scratch"—the gap between them is precisely the abyss where many idealists have fallen.
What is "First Principles"?
When Musk's Mars thinking meets China's complex business soil, why does the same methodology roar in the hardware field but fall silent in the software ecosystem? Have we misread the true spirit of "first principles"?
The so-called "first principles" was first elevated to a pedestal by Musk and became a catchphrase for countless entrepreneurs and innovators. Its core appeal lies in a purity akin to physics: setting aside all existing and conventional practices, returning to the most basic conditions of a thing, breaking it down into its most fundamental elements, and then reconstructing and innovating from the source.
In the fields of rockets and electric vehicles, the victory of this engineering thinking is stunning—it directly confronts physical laws and manufacturing costs and has indeed achieved breakthroughs. However, when this thinking is enshrined unchanged as the golden rule of the business world, especially when applied to entrepreneurship in complex systems like AI, a dangerous simplification begins.
The problem first arises from the fundamental difference in dimensions.
The physical world is to a great extent reducible, its variables are relatively limited, and causal relationships are more direct.
You can calculate a rocket's thrust and material strength, and you can decompose a battery's chemical composition and cost.
But the business ecosystem, especially the modern business ecosystem driven by digital technology, is a typical complex system.
It is full of irreducible "emergent" properties: the interactions of countless individuals (users, developers, competitors, regulators) produce macro patterns that cannot be predicted from the behavior of any single individual.
There is no single "fundamental particle" here; there is a dynamic, entangled network. Applying the thinking used for reducible physical systems directly to an irreducible business-social system is tantamount to using Newtonian mechanics to explain a panic in the financial market—its one-sidedness and risk are inherent.
Thus we see a widespread misreading: many entrepreneurs turn the "first principles" methodology into a slogan, regarding it as a "get-out-of-jail-free card" that can circumvent all complexity. They believe they can ignore the existing industrial landscape, user habits, and collaborative networks, and with a superior technical solution or a more "essential" concept, overturn everything from scratch.
Behind this is an overconfidence in "rational design," ignoring the system's own powerful self-organization and path dependence.
This thinking is particularly prominent in the AI entrepreneurship boom, and particularly prone to failure.
AI does not operate in a laboratory vacuum; it must be embedded in the existing socio-economic-technological system to create value. A typical dilemma is "technological path dependence and ecosystem lock-in."
For example, even if your algorithm is more elegant and efficient from a "first principles" perspective, if it is incompatible with current mainstream development frameworks, cannot leverage the existing massive computing infrastructure, or is difficult to connect to the data ecosystem monopolized by giants, then its commercial prospects will immediately dim.
OpenAI's development path itself is an excellent counterexample: it did not overturn Google from the start with a perfect "first principles" blueprint, but went through a complex process from non-profit to capped-profit, from open to partially closed, and continuous trial and error and evolution in technical routes (such as from reinforcement learning to brute-force large models). This is a story of continuous interaction, adaptation, and gaming with a complex ecosystem, not a blueprint designed purely from scratch.
In the Chinese market, this complexity is compounded by a unique layer of "implicit constraints."
A strong regulatory environment means that policy is a core variable that must be internalized in the business model, not an external disturbance. Unique user habits (such as reliance on super apps and differences in trade-offs between privacy and convenience) shape completely different forms of product demand.
Deeply ingrained supply chain networks and industrial clusters mean that hardware innovation must consider coupling with the existing manufacturing system, rather than simple replacement. These are not backdrops that can be easily "disassembled" and "reconstructed"; they are rigid structures that shape the space of entrepreneurial possibilities. Here, "from scratch" thinking that ignores the ecosystem is more likely to hit invisible walls.
Therefore, the fallacy of worshipping "first principles" lies not in the principle's spirit of pursuing essence, but in its simplistic application without reverence for "complex systems."
It attempts to use thinking tools designed for limited variables and linear causality to deal with a real world of infinite variables and nonlinear interactions. When AI entrepreneurship falls into this trap of physics worship, the outcome is often not a disruptive victory, but a tragic failure of being out of place in the real business ecosystem.
The real challenge lies in how to combine the thinking that pursues essence with the wisdom that embraces complexity.
The Essential Divide Between Hardware and Software
Musk's success has been oversimplified.
People see him breaking down battery costs and restructuring rocket manufacturing, and assume that "first principles" is a master key that can unlock all business problems. But this thinking has a severely overlooked boundary of applicability: hardware and software follow two fundamentally different underlying logics.
We can understand hardware-dominated fields as a "convergent system."
Whether it is electric vehicles or reusable rockets, the core challenges ultimately converge on a series of clearly definable physical laws and engineering problems: energy density, material strength, aerodynamics, manufacturing cost. Although these variables are enormous challenges, their nature and interrelationships are to a considerable extent known and measurable.
Tesla's success is essentially the ultimate optimization and system integration of known variables under a set of stringent physical and cost constraints. Here, physical laws are indisputable, delineating a relatively clear and calculable track for innovation. You are gaming with steel, chemistry, and Newtonian mechanics; the opponent's rules are constant.
However, once you enter fields centered on software and ecosystems, especially the commercial application of artificial intelligence, the rules of the game undergo a fundamental shift. This is no longer a "convergent system" but a "divergent system."
Your opponents are no longer silent physical laws, but "people" full of uncertainty—users' fickle preferences, the inclinations of developer communities, competitors' dynamic reactions, regulators' changing perceptions.
The core feature of this system is no longer optimizing the known, but responding to "emergence."
Network effects create monopoly barriers out of thin air, user habits form stickiness that is hard to migrate, and the victory of a technological path often stems from a small accidental advantage in the early stage, which is then sharply amplified by the system.
Here, the optimal technical solution often loses to a product that happens to occupy a better ecological niche but is technically suboptimal. You cannot precisely calculate the final form of a user interaction pattern or a data flywheel, just as you cannot calculate battery energy density.
In China's unique AI market, the complexity of this "divergent system" also manifests as a "dual-track" split.
One track is the laboratory race pursuing cutting-edge technological breakthroughs, benchmarking against the latest global papers; the other track is the industrial arduous journey seeking commercial implementation, searching for application scenarios in factory workshops, government halls, and the cracks of super apps.
There is often a huge gap between the two.
A technology may attract much attention at global academic conferences, yet be completely unable to integrate into China's existing industrial chains, data governance rules, or consumers' willingness to pay. This gap is precisely what the "first principles" of pure hardware thinking finds most difficult to understand and overcome. It cannot incorporate "non-technical constraints" such as policy trends, local protection, and data silos as basic variables into its initial calculation model.
This brings enormous risks in cross-context transplantation.
The narrative of "disruptive innovation" originating from Silicon Valley often undergoes profound mutations in the Chinese market.
Pure technological idealism here often needs to be combined with a profound "ecological realism." Observing the growth of ByteDance or Tencent, you will find a wisdom very different from Musk-style declarations—an evolutionary wisdom based on trial and error, feedback, and rapid iteration.
They rarely claim to reconstruct an industry from physical first principles; instead, they keenly perceive the gaps and traffic depressions in the existing ecosystem, quickly adapt and embed like organisms, and then in turn transform the environment. Douyin did not invent short video from "first principles"; rather, it found an explosion point at the intersection of China's unique mobile internet infrastructure, content creator ecosystem, and user entertainment needs, and nurtured it through precise algorithmic operations.
Tencent's WeChat is even more a super organism that naturally grew and continuously "emerged" new functions amid complex social relationships and communication needs.
The core of this wisdom lies in acknowledging that in the "divergent system" of software and ecosystems, you cannot design the ultimate form on day one. Success comes from building an adaptive organization capable of continuously sensing the environment, absorbing feedback, and adjusting its own structure.
Then, it is no longer the conquest of physics, but the symbiosis of biology.
When technological idealism encounters ecological realism, the victor is often not the engineer with the most perfect blueprint, but the species with the greatest environmental adaptability.
In China's vast and unique digital jungle, understanding this is more important than worshipping any single mental model.
An AI Startup Framework Under Complex Systems Thinking
So, should we completely abandon the pursuit of essence? Not so.
The real problem lies in the underlying mental models we rely on.
What we need is no longer a "physics first principle" suited to a stable, reducible world, but a "biology first principle" capable of coping with dynamic complexity. This is not a negation of first principles, but a profound paradigm upgrade: from reductionism to complex adaptive systems thinking.
Its core lies in three key cognitive leaps.
First, it is a shift from pursuing the creation of perfect solutions in an ideal vacuum to seeking a "niche" in reality where one can survive and develop. You must first answer: in the existing technological landscape, industry alliances, and regulatory frameworks, where is there a gap that allows you to sprout, acquire initial resources, and build defenses? Entrepreneurship is not about building a utopia from scratch, but about finding your combination of soil, sunlight, and water in an existing rainforest.
Second, you must understand the power of "co-evolution" and abandon the romantic fantasy of "one-way disruption." Your product and technology will co-evolve with user habits, competitors, complementary suppliers, and even regulatory rules. Your success does not depend on your unilateral "disruptive power," but on whether you can initiate and guide a multi-party, mutually beneficial evolutionary process, like the symbiotic relationship between coral and algae.
Third, you must accept the logic of "emergent adaptation" rather than clinging to the sense of control of "total design." You cannot preset all details and the final form on day one. The right approach is to build an initial system with core adaptive capabilities, set clear evolutionary rules and feedback mechanisms, and then, through continuous interaction with the real environment, let better strategies, forms, and models "emerge" on their own. What you design is the rules of evolution, not the results of evolution.
Based on this, a new decision architecture emerges, containing at least four layers that must be continuously monitored and interacted with. The
outermost layer is "environmental layer sensing." In the AI field, this especially means you must model China's unique AI policy cycle as a core variable.
Policy fluctuations are not noise, but climate change that defines the game's landscape.
At the same time, the rhythm of capital flows and fluctuations in talent structure together constitute the macro hydrological conditions that entrepreneurship must comply with. Moving inward, the next layer is "ecosystem layer analysis."
You need to scan, like an ecologist, the existing open-source frameworks, cloud service platforms, hardware supply chains, and major players' API interfaces—what kind of "technology rainforest" do they constitute? Is your innovation forcing an invasion as a disruptive species, or can you find a point of entry for "parasitism" or "symbiosis," first gaining the right to survive and then seeking development? This determines whether you make enemies on all sides early on or can leverage momentum to move forward.
Only based on the above understanding can you carry out "path layer design."
The core here is to build a "minimum viable ecosystem unit" (MVE). It is not just an MVP for testing product features, but a test of whether your solution can complete a minimum closed loop in the real ecosystem—whether it can attract initial users, whether it can establish data or value exchange with at least one key partner, and whether it can form a sustainable value stream in a tiny scenario. It is the reliable starting point of your evolutionary process.
Finally, it is for the entire organization to establish an "adaptation layer mechanism."
This means the decision architecture must be agile and feedback-driven, able to quickly translate massive signals from the environment, ecosystem, and path practice into strategic and tactical adjustments. Your plan is not a blueprint to be strictly executed, but a set of hypotheses that need constant revision.
To practice this framework in China, one must also skillfully use a set of unique "variable toolkits."
The first is to understand and harness the "government-market" dual-driven logic. Opportunities often appear at the intersection of industrial tracks guided by national strategy and spontaneously formed market demand, such as specific scenarios under "smart manufacturing" or "smart cities."
Second, be good at leveraging the powerful "industrial chain cluster" effect in China.
Your AI algorithm advantage must be deeply integrated with the consumer electronics supply chain in the Yangtze River Delta, the rapid hardware prototyping capabilities in the Pearl River Delta, or the data service industry clusters in the Beijing-Tianjin-Hebei region, in order to be dramatically amplified and transformed into hard-to-replicate delivery cost and speed advantages.
Of course, you must also deeply understand the "semi-open nature of China's data ecosystem."
Within the boundaries of security and privacy protection, how to find compliant channels for value flow among enterprises, industry platforms, and public data is itself a core entrepreneurial capability.
Practical Paths for AI Entrepreneurship in China
Based on evolutionary thinking, we can roughly identify some practical paths for AI entrepreneurship in China. The first task is to clearly recognize and avoid a localized variant of the "innovator's dilemma."
The core of this dilemma lies in blindly pursuing head-on technological confrontation at the general level or on the consumer side when facing internet giants with massive data, capital, and traffic gateways.
A more pragmatic path is to execute a thorough differentiated niche strategy.
This means entrepreneurs need to abandon the grand narrative of "replacing whom" and instead think deeply about "which unmet gap can I fill that giants cannot quickly respond to."
Looking back at the journeys of early AI companies such as SenseTime and Megvii, their path choices deserve re-examination. They did not try to build a brand-new digital world from the start, but in vertical fields such as security and finance, where there is rigid demand for technology and traditional internet giants have not penetrated deeply, they first became indispensable "components" of the existing industrial ecosystem by providing key visual recognition capabilities. The essence of this strategy is to find a micro-environment with relatively independent sunlight, water, and nutrients in the shadow of giants, and take root and grow there.
To realize this path, one must firmly grasp three pragmatic strategic priorities for evolution.
First, "depth of scenario embedding" must be placed above "purity of technological theory." No matter how many model parameters you have or how impressive your papers are, it is not as good as improving detection accuracy by 0.5% on a specific production line and enabling both workers and factory managers to use it seamlessly. Deep embedding means understanding all non-technical constraints of the scenario: workflow, procurement systems, and personnel capabilities.
Second, pursuing "speed of ecosystem compatibility" should take priority over pursuing "degree of functional disruption."
Can your solution quickly connect to mainstream cloud platforms? Can it be compatible with common industrial protocols? Can existing developers get started with as little training as possible? Achieving "plug-and-play" in the ecosystem often has far greater commercial value than a "disruptive feature" that requires overturning everything and educating the market.
Third, "density of iterative feedback" must replace the obsession with "completeness of top-level design."
Do not spend months designing a flawless ultimate product architecture. Instead, put even a rough minimum viable product (MVP) into real scenarios as soon as possible and establish a high-frequency, short-cycle feedback loop. The real complaints and needs obtained from frontline users and partners can guide the correct direction of evolution far better than a perfect blueprint in the conference room.
Looking ahead to the next two to three years, the landscape of AI entrepreneurship in China will show several clear evolutionary trends.
In terms of commercial returns, "small vertical-domain models" will form closed loops significantly faster than "general large models." In professional fields such as healthcare, law, and industrial quality inspection, the requirements for data privacy, domain knowledge, and reliability are extremely high, and general models are often inadequate. Small models optimized for specific scenarios, with controllable costs and the ability to be directly embedded in workflows, will be the first to achieve large-scale revenue.
The competitive focus of AI-native applications will shift from comparing whether features are cool to measuring "ecosystem affinity." This means whether an application can keenly integrate into super gateways such as WeChat and DingTalk, and whether it can efficiently collaborate with local chip and framework ecosystems such as Huawei and Baidu, will be more critical than breakthroughs in a single feature.
On the other hand, we speculate that regional differentiation will intensify. The Yangtze River Delta and the Greater Bay Area, relying on their deep foundations in manufacturing and commercial services, will respectively evolve two distinctive paths: "AI + manufacturing" and "AI + services." The former goes deep into factories, focusing on process optimization and supply chain intelligence; the latter connects merchants, reconstructing marketing, customer service, and cross-border trade.
So, has a potentially huge "second growth curve" emerged?
For example, the comprehensive digital and intelligent upgrade of state-owned enterprises.
This is not IT procurement in the conventional sense, but deep demand involving management process reengineering, data governance, and industrial collaboration.
AI solutions that can understand the operating logic of state-owned enterprises, meet security and trustworthiness requirements, and transform technological value into improvements in management efficiency will usher in a sustained and massive market.
These trends together point to one conclusion—in China, the success of AI will increasingly depend on the ability to precisely embed into and co-evolve with complex business ecosystems, rather than on isolated technological peaks.
The golden age of AI in China belongs to doers, and even more to thinkers who make good use of tools.
Now, start by clarifying the ecosystem in which you are located.
This article comes from the WeChat public account "Dongzhen Shanglue", author: Dongzhen Shanglue
