Recently, I attended a small closed-door AI meeting at WeChat headquarters. Everyone there shared hands-on experience and industry judgments. Let me talk about my takeaways.

1. AI Is Creating Deflation: The Cruel Other Side of a Technology Revolution

The most counterintuitive and weighty judgment of the whole meeting: AI is creating deflation.

Most people's default understanding of a tech revolution is: AI raises production efficiency, goods and services costs fall, things become cheaper, everyone lives better. That only looks at the supply side, not the demand side.

The real transmission chain is far more brutal than imagined:

1. Enterprise side: AI sharply lowers marginal production costs. To maximize profits, companies replace labor with compute and start layoffs and cost reduction.
2. Resident side: Employment shrinks, income expectations weaken; those still employed dare not ask for raises, the unemployed lose income, overall purchasing power drops.
3. Market side: Supply becomes increasingly surplus, demand grows increasingly weak. Even with lower prices, nobody buys, not because things got cheaper, but because people can no longer afford them.

This is typical 'technology-driven deflation': productivity advances, but production relations do not improve in sync. The dividends concentrate into capital and compute, the distribution share of ordinary workers is squeezed, and in the end total demand is dragged down.

This judgment matters because it helps us jump out of the 'AI can do everything' frenzy and see the current position of the industry cycle.

It looks a lot like the 2000 internet bubble: everyone knew the internet was the future, capital poured in, valuations went through the roof, but sustainable paying demand lagged far behind. When the bubble burst, the NASDAQ fell 80%, countless companies died, and the survivors built the later internet age.

AI now stands at almost the exact same node: everyone is all in, compute investment grows exponentially, but the players who truly run a viable business loop are few. A bubble will come, and it will burst.

The implication for us: Don't blindly quit your job to chase AI opportunities, and don't all-in on high-risk startups. In the early phase of a technology revolution, protecting cash flow, keeping stable income, and controlling debt matters more than chasing uncertain dividends. When the tide rises, everyone flies; when it recedes, only those with rations survive to the second half.

2. Entry Points and Commercialization: The Real Deciding Factors for AI Products

The second core recognition is about entry-point logic and commercialization truth in the AI age.

Every era has its own traffic entry point: the browser in the PC era, super apps in the mobile era. In the AI era, entry points are becoming 'efficiency workstations': Codex and Claude Code for coding, Doubao and WorkBuddy for general scenarios, all first stops where users access AI capability.

But most people step into the same trap: they always want to build their own entry point.

The truth is, general entry points are always the battlefield of giants, competing on compute, capital, and ecosystem; small teams have no chance. For the vast majority of practitioners, the right posture is never to build an entry point, but to make your product grow on top of an entry point.

Two real data points from the meeting happened to confirm this logic:

1. The essence of Hongguo's success: AI is a tool, demand is the business

Many people look down on Hongguo, thinking it is just a short-drama app, not a 'real AI product'. But it is precisely the one that became the first truly scaled AI-related super product, with monthly active users and time spent exceeding generic conversational products, and more importantly, it already achieves large-scale profitability.

Its core logic is not 'doing AI', but 'using AI to amplify content supply and ride the strong demand of mass emotional consumption'. AI lowers production cost and raises capacity for short dramas, while users are willing to pay for emotional value, so the business loop naturally closes.

2. WorkBuddy's user truth: nobody uses AI to chat; everyone comes for results

The internal data shown at the meeting: WorkBuddy's top three usage scenarios are stock trading, writing papers, and doing self-media.

This data is very representative: ordinary users do not care how high your model parameters are or how fast your inference is. They use AI for one purpose only, getting concrete results.

• Stock trading: helping themselves make money.
• Writing papers: helping themselves save time and get a degree.
• Doing self-media: helping themselves produce content and monetize.

All are strongly goal-oriented and benefit-oriented; none is 'chatting for fun'.

The implication for us: When building an AI product, or using AI to do things, don't obsess over technical parameters. First answer one soul question: What perceptible concrete result can a user get from you? If you cannot find that answer, no matter how fancy the feature is, it is self-indulgence. For small teams, rather than fighting over general large models, it is better to root in one vertical scenario, graft AI capability onto an existing entry point, and solve one concrete problem. That is the path with the highest certainty.

3. Private Domain and IP: The Most Unkillable Personal Assets in the AI Age

The third recognition, which I think is most worth learning for ordinary people, is about the underlying logic of private domain and IP.

The most direct consequence of the AI age: content production cost approaches zero. Batch writing articles, scripts, and posters, anyone can do it with AI. This means content itself is no longer scarce; what is scarce is 'the person', your personality, your label, your judgment, and other people's trust in you.

This is why the meeting repeatedly stressed: in the AI age, 'human flavor' is the core competitiveness.

Many people fail at private domain because they think their operations skills are insufficient. Actually it is not that. The meeting hit the essence with one sentence: most private-domain problems are not in operations, but in product design.

Here I distill the private-domain methodology from the meeting into three actionable underlying recognitions:

1. A community is a product, not a group

Don't call a WeChat group private domain. A real private-domain community is itself a product: you need to design its entry threshold, value delivery system, user tiering rules, and even lifecycle. How rules are set, what is output each day, what users get, how violations are handled, all are part of product design. If design is bad, even great operations skills cannot save a dead group.

2. Threshold is the lifeline of a community; filter people with 'outputs'

A free group without a threshold will eventually become an ad group and a chat-only group. Without filtering, there is no same-frequency, and therefore no value. The right way for a high-quality community: entering must cost something, but not necessarily money. It can be an assignment, a product you made, a GitHub link, or an original article. Filtering with 'action outputs' is more precise than filtering with money. You do not want people who pay; you want people who act. Refuse free-loaders, keep only same-frequency people, and the community will generate value on its own.

3. The three tiers of private domain are essentially a progression of trust

Selling information gaps and atmosphere at low price, selling accompaniment and results at mid price, and selling circles and capital at high price. This is not a simple pricing strategy; it is a deepening of trust relationships layer by layer. From information syncing to accompaniment and implementation, to resource matching, every layer of trust a user has in you allows you to provide higher value, and naturally higher pricing. Do not start with high prices, and do not stay stuck in low-price competition forever.

And all of this rests on your personal IP. Interestingly, the meeting mentioned a counterintuitive pricing logic: in the IP economy, raising prices is actually a form of promotion. Ordinary goods attract users by lowering prices; IP filters users by raising prices. The higher the price, the higher the recognition and circle attributes of the users who stay, forming a positive trust loop.

Finally, two practical reminders, both lessons learned from stepping on pitfalls:

• Do not put community QR codes in public, crawlable places like GitHub or WeChat official account articles. Black/grey industry scrapes and bomb the groups at very low cost; one hit and the community atmosphere is ruined.
• The latest quoting logic for official accounts is 'active user triggering is risk-free, passive induction is fully restricted'. Attracting users to follow actively through content is the only long-term viable path.

In Conclusion: In Uncertain Times, Protect the Basics

Besides the three points above, the meeting also had two very practical lines, for all practitioners: First, individuals and small teams should not touch To C AI businesses. Making money from C-end is a privilege of large companies; our opportunity is in To B. Second, before starting a business, do three things: buy insurance, get a health check, and keep regular sleep and diet. If your body cannot hold up, do not force entrepreneurship; having a regular job is not embarrassing at all.

The biggest feeling after the whole meeting: the industry has finally cooled down from the AI frenzy and started to return to the essence of business, surviving and making money.

AI is a profound production revolution, but it will not make everyone a winner. It raises efficiency, but also creates deflation; it creates new jobs, but also eliminates old ones; it breeds new giants, but also pricks countless bubbles.

For ordinary people, rather than chasing one fleeting hotspot after another, it is better to calm down, understand the underlying logic, polish your own strengths, and manage your own trust assets.

The tide always goes out. Before that, first make yourself the one who survives to the second half.