In July 2026, the AI circle once again entered a cycle of intense, iterative involution. Within just half a month, major vendors released new model versions in concentration: OpenAI launched the full GPT-5.6 family, xAI updated Grok4.5, and domestically Kimi K3, Tongyi Qwen3.8, Tencent Hunyuan Hy3, and DeepSeek V4 successively made official announcements, continuously refreshing parameter leaderboards and performance benchmarks. The geek circle buzzed with excitement; every day brought new rankings, new evaluations, and new "strongest in the whole network" conclusions. The excitement belongs to the big vendors and the tech circle, yet the anxiety is transmitted to large numbers of ordinary users. Watching wave after wave of model updates, many fall into meaningless internal friction: is the model I am using already obsolete? Has someone using the new version already achieved a corner-overtaking? Should I switch tools immediately? I believe the anxiety of the vast majority of people comes from the same question: standing in the user's position, yet insisting on joining the vendors' battlefield excitement. This March, Jensen Huang once used a clear metaphor to break down the underlying logic of the AI business model.
I. Jensen Huang's "Five-Layer Cake": The Model Is Only the Fourth Layer
Jensen Huang wrote in a signed article on NVIDIA's official blog: AI is not some clever app, nor a single model, but infrastructure like electricity and the internet. His original words: "It is not a clever app or a single model; it is essential infrastructure, like electricity and the internet." Jensen Huang breaks the entire AI industry down into a five-layer cake, from bottom to top in order: energy, chips, infrastructure, large models, and applications. He explicitly stated that AI is not a sophisticated application, nor just a single model; like electricity and the internet, it is a new type of infrastructure. The involution of large-model parameters and leaderboard rankings is the fourth layer of the entire industrial structure. This layer is the main battlefield of big vendors, the arena of capital's games, and the coliseum of technology R&D; all arms races essentially revolve around this layer. Yet the vast majority of ordinary people and office workers reside in the fifth layer: the application layer. As users of large models, ordinary people need to care whether it can solve problems, improve efficiency, and create real value. This is like everyone using the internet today — no one compares base-station parameters, computing bandwidth, or server versions every day; people only care whether the network speed is adequate and meets their own needs. Having developed to today, AI has entered the same stage: the basic capabilities of mainstream models on the market long ago exceeded, with margin, 99% of ordinary people's work scenarios. The version iterations and parameter levels that ordinary people obsess over add almost no incremental value to actual work output.
II. The Arms Race Is in the Fourth Layer; Ordinary People Live in the Fifth
Vendors competing on parameters, on benchmark scores, on leaderboards stems from a commercial logic of financing, market share, and industry discourse power — an inevitable stage of industrial development. But ordinary people anxiously following along and frequently switching models is meaningless self-consumption. Jensen Huang once gave a very penetrating example: AI can assist in reading radiological films in the radiology department, but it cannot replace the doctor's core value. Reading films is standardized execution, the kind of thing models excel at; diagnosis, trade-offs, decisions, and risk judgment are where human irreplaceability lies. The core truth of the AI age has never been "the model is not strong enough" — it is that tools are becoming ever more general-purpose while human cognition and judgment are becoming ever scarcer. Many people superstitiously believe that as long as the model is new enough, parameters large enough, and benchmark scores high enough, all problems can be solved; but reality gives precisely the opposite answer. This May, Andon Labs ran a hardcore test: it put the top several large models on the market into real business scenarios, operating businesses independently and achieving profitability — and all ultimately failed. The reasons for failure were not insufficient model computing power or outdated versions, but that all top models could not handle open-ended business judgment: how to control costs, how to acquire users, how to weigh risks, how to achieve long-term closed-loop profitability. Industry data confirms this point as well: Gartner predicts that by the end of 2027, over 40% of AI agent projects will be completely scrapped; 2026 industry statistics show that nearly 90% of AI pilot projects cannot be commercially deployed; attack risks targeting AI agents surged 340% year over year. Large batches of AI projects crash and burn, and not once did they lose on model performance — they all lost on scenario fit, value judgment, and the control of human-machine boundaries.
III. The Application Layer Is the Ordinary Person's Home Turf
I believe today's anxiety about AI essentially all stems from a mismatch of thinking. In the industrial age, the core of personal competitiveness was tool proficiency — the more tools mastered and the more skilled the operation, the stronger the competitiveness; this logic held completely for the past two hundred years. But entering the AI age, this underlying logic has thoroughly failed: the iteration speed of tools far outpaces the individual's learning speed; chasing versions, chasing parameters, chasing the newest models, one can never catch up with the industry's update rhythm. Many people study new models, evaluate new features, and switch tools following the trend every day, yet never calm down to think about what core problem their own business actually needs to solve. Using the newest tools to do the most undifferentiated execution is the most common ineffective effort today. AI can handle 94% of standardized work, including drafting, organizing, statistics, formatting, review, and basic Q&A, but the remaining 6% of non-standard judgment, risk trade-offs, directional choices, and business decisions are the core gap between people — a value that no large model, iterated ten thousand times, can replace.
From my personal experience of using AI, I have three suggestions. First, fix your tools and produce steadily. For routine needs like daily office work, copywriting organization, and data aggregation, one adequate model is enough; frequently switching tools only disrupts your own work rhythm. Second, choose on demand, and do not blindly chase the new. For specific scenarios like complex reasoning, code development, and deep analysis, then specifically match a fitting model — ignore leaderboard hype, look only at whether it fits the business need. Third, humans make judgments and AI executes. AI is an efficiency tool, a digital colleague, a super-strong executor, but never the final accountable party; risk judgment, value trade-offs, external output, and core decisions must be backed by humans. You may use AI to improve efficiency, but must not hand over and depend on AI. I believe the large-model arms race will keep going, parameters will grow larger, versions will update, leaderboards will keep refreshing, and industry hype will keep cycling — but none of this has a direct relation to ordinary users. Vendors compete on the technological ceiling; ordinary people must polish their implementation capability. Tools become obsolete, versions get eliminated, parameters get continuously surpassed — only the understanding of business, the judgment of value, and the control of human-machine boundaries will never be iterated away, never be replaced. AI is increasingly becoming like water and electricity, a social infrastructure. The personal gap of the future has never come from "how new a model you used," but only from whether you can master the tools rather than be led around by their iterations. The ultimate logic of human-machine symbiosis has never changed: humans set the direction and AI executes; humans guard the judgment and AI raises efficiency. Leave the excitement to the industry, keep the capability for yourself; ordinary people need not make up the numbers in the large-model race. Human value has never come from "using the strongest tool," but from "knowing what you want and applying the tool to the blade's edge." The more AI spreads into every layer of life, the more clear-headed humans should be: a model's strength or weakness does not define you; how you collaborate with this thinking partner is what defines you. I believe this is precisely the relationship humans and AI should have — not who replaces whom, but human wisdom plus AI wisdom, mutually inspiring, Harmonized Intelligence, creating new wisdom.
