Chapter 12: Reverse Iteration from Consensus: Evolving the Product Path

1. Two bets in opposite directions #

In September 2018, a group of Chinese students gathered around Duan Yongping to ask questions at Stanford University. Someone asked him why he believed in the phrase "dare to be the queen of the world", which sounded unenterprising. His only answer was: "All masters dare to be the best in the world, they just do better than others."

At the same time, Meta was making a bet in the exact opposite direction. From 2020 to 2025, Reality Labs suffered a cumulative operating loss of approximately US$83.6 billion, betting on a category for which there is no consensus: metaverse social interaction. The internal goal of the flagship product Horizon Worlds is 500,000 monthly active users by the end of 2022, but it has actually dropped from a peak of about 300,000 to less than 200,000; internal documents seen by the Wall Street Journal show that less than 9% of the worlds created by users have had more than 50 visitors. In June 2026, Meta completely exited the VR headset business. The world's thickest wallet failed to buy a consensus.

The difference between the two bets is not in courage, but in the starting point: Which level of consensus does your product start from. This chapter deduces the Interference Method in Chapter 8 to products, giving a path that is opposite to that in entrepreneurship textbooks. It is also the final piece of methodology in the book: the path along which the product evolves after people (Chapter 10) and time (Chapter 11).

2. Read the adoption curve upside down #

First, straighten the textbook. The technology adoption life cycle (from Rogers to Moore's "Crossing the Chasm") divides users into five tiers: innovators, early adopters, early majority, late majority, and laggards. The default playbook is to play from left to right: win over the early adopters first, cross the chasm, and then harvest the masses. As mentioned in Chapter 8, Moore’s vocabulary has dominated this field for thirty years, and investors and entrepreneurs still only use his language to describe early-stage markets.

But if you look at this curve again with the eyes of consensus, you will see a line of prices that is not marked in the textbook. At the left end of the curve, a consensus on the problem has not yet been formed: users don’t even know they have this problem. You must create a consensus first. This action is called educating the market. The education market is the most expensive Interference Method: You have to build the vocabulary of describing the problem, the way to prove the value, and the credibility ladder of trust from scratch (the full set of actions in Chapter 8), and the results of education are public goods: the category consensus you spend money to build, competitors can use for free. Spread out the three accounts one by one and look at the price tags. Vocabulary account: When the category does not exist, users don’t even know what to search for. You have to invent the name of the problem, the description of the scenario, and the dimensions of comparison. After inventing, you have to repeat them so that the entire market will use them. As mentioned in Chapter 8, Moore did this in one word, but that was the hallmark achievement of Interference Method, not the default configuration. Proof account: There is no comparable product. Every function must start from "Why do you need it?" The proof chain is twice as long as that of mature categories, and each level of the conversion funnel is doubly drained. Trust account: There are no steps on the credibility ladder in new categories. There are no third-party reviews, no peer cases, and no "the old man next door is also using it." You can only start from the self-statement at the bottom of the ladder. The three accounts are not one-time expenditures, they are continuous burning; and the educated market in exchange for burning has public property rights. What's even more troublesome is that the unit price of the three accounts is still rising year by year. Chapter 8 explains the market-side corollary of implementing zeroing: product supply is exploding, attention is scarce, penetrating the attention of the same group of users is becoming more expensive year by year, and quotations in the education market are rising with the tide. The right end of the curve is exactly the opposite: the consensus on issues among the late-stage masses and those left behind has already been formed, is high-frequency, stable, and does not require education. They don't need to be convinced that "this is a problem", they just need to be convinced that "your solution is safe". The cheapest consensus is not on the left end of the curve, but on the right.Duan Yongping gave a demand-side version of this judgment at Stanford: "Dare to be the world queen refers to the product category, because it is often difficult to guess the market demand, but others have already made the demand clear, and you will be more certain if you meet this demand." He paired the second half of this sentence with "the middle of the pack vying for the first." Dare to be the best in the world refers to doing the right thing (choosing proven needs), and striving to be the first among the latter refers to the ability to do things right (product power must win). He also explained the etymology of this slogan, which is Lao Tzu's "don't dare to be the first in the world"; and it is not metaphysics, but he used it: when BBK entered the VCD market in 1996, Aido, Xinke, Malida, Samsung, and Sony were all already present. BBK competed to be the first with quality, brand, and service. Twenty years later, when OPPO printed these eight words on the big screen at the press conference, the two manufacturers, blue and green, had already used the same path to dominate half of the mobile phone market.

Therefore, the framework diagram of this chapter is to draw a reverse arrow on the adoption curve:

Mature needs of the late majority/laggards → Survival products → Complete products for the early majority → Advantage products for early adopters → Cutting-edge products for innovators

The traditional path crosses the chasm from left to right, while the consensus reverse iteration expands the innovation radius from right to left. Note that it is not "going low-end", nor is it forcing conservative users to take the lead in adopting new technologies. On the contrary, it hides new technologies behind low-learning cost experiences. Don’t educate users to understand new technologies first, first let new technologies learn to adapt to users’ existing consensus.

3. Select things first, then choose layers: variables are given to windows, invariants are given to compound interest. #

The backward reading curve answers "from which level of consensus to enter", but it is preceded by an earlier judgment: whether the thing itself is worth doing. This round of AI puts entrepreneurs in front of entrepreneurs with so many multiple-choice questions that they are dizzying. The first sieve for value sorting can be written as a multiplication: Opportunity = Variable × Invariant. Variables are responsible for answering "Why is it now", and invariants are responsible for answering "Why is it still valuable ten years later?"

On the variable side, the shape of the current round was already drawn in Chapter 1: the execution price returns to zero. It explains why the opportunity arises at this moment: a certain type of demand was not cost-effective to service with humans in the past, but now the cost structure is rewritten by AI, and the window is open. But see one thing clearly: the variables themselves are not requirements. The user will not pay because you use AI, but only because you better meet his original needs. This is the turn for invariants to come on the scene, and the standard of judgment is directly borrowed from Bezos’s question in Chapter 6: Invariants are those who cannot imagine that customers will ask for its opposite ten years later.

You need both, if one is missing, you will die, and there are autopsies in both death methods books. Those who only grasp variables and fail to anchor invariants are chasing the wind. As demonstrated by Qudian in Chapter 6, each outlet has its own scoreboard, and the capabilities built on variables are cleared every two years. Those who only adhere to constant variables and do not accept variables will be penetrated by the new cost structure. Chegg in Chapter 1 adheres to the ever-changing demand of "students want answers". The reason for its death is that when the cost structure was rewritten, it still stood in the old structure and collected a monthly fee of $19.95. Variables are given to the window, and invariants are given to compound interest: the window determines whether you can enter the market, and compound interest determines whether it is worth entering the market. **

Use this multiplication to pass two real-life samples. The primary market is hotly investing in AI + hardware. The long-term logic is this multiplication: the variable is the model capability, and the invariant is the physical scene (ear, desktop, car) occupied by the hardware. People's demand for more natural interaction will not be reversed. The real risk of this question also became clear: If the hardware side is not anchored by a ten-year demand but a sense of novelty, the invariant term in the multiplication is zero, and zero multiplied by a larger variable is still zero.

The second sample is our own. By the way, we will fill in something that has not been explained in the book: bioby.ai’s current business is an overseas influencer marketing agent. To choose this thing, we use this multiplication. On the side of invariants: Merchants always need to acquire customers when going overseas, and the scarcity of attention (Chapter 8) makes "borrowing the consensus of a trusted person" an increasingly necessary and expensive Interference Method. The essential structure of influencer marketing has not changed from word-of-mouth in the market to today, only the carrier has changed. Variable side: This business used to be labor-intensive. Finding celebrities, screening celebrities, building relationships, negotiating, and following up on contracts were all execution links. The implementation of zeroing allows this entire layer to be absorbed by Agent, and the cost structure is rewritten. Originally, only large merchants can afford services, and the window is opened to all merchants. Check it again according to the path caliber of this chapter: the demand side consensus is mature ("Internet celebrities are effective in bringing goods" without educating anyone). What we have rewritten is only the cost structure of supply. The typical engine is new and the shell is old. Whether this business can eventually outperform depends on time and subsequent versions of this book; here I only explain the judgment structure used when choosing it, because that is what this section is going to give you.

Four, five steps and four levels: operational version of the path #

The reverse arrow is broken down into five steps:1. Find consensus: Find consensus on issues that have been formed by the late-stage masses and laggards on high-frequency, stable issues that do not require education. The test: Users can tell the problem without thinking, and are already paying for old solutions. 2. Leverage experience: Use familiar interfaces, processes and language to deliver value. New engine in an old shell: users see "the same thing as before, but cheaper or more convenient" rather than "a new thing to learn". 3. Survive first: Validate payment, delivery and unit economics. The output of this step is not growth, but survival qualifications plus the standard precipitation mentioned in Chapter 6: the judgment you accumulate in real paid scenarios is the ticket to the next level. 4. Grow capabilities: Use cash flow, user feedback and precipitation standards to gradually improve product capabilities. 5. Move one layer to the left each time: Only move one layer to higher uncertainty each time, without skipping layers.

The reason why it can be moved to the left layer by layer is because the consensus to be created at each layer is different, and the results of each layer on the right are exactly the evidence of the next layer on the left:

• Consensus on survival products: "It's just as safe as the old way, but cheaper or more convenient."
• Consensus of the complete product: "Similar users have verified that it can stably generate ROI." The evidence is the paying users of the upper layer.
• Consensus on superior products: "It brings competitive advantages that old solutions cannot provide."
• Consensus on cutting-edge products: "It opens up possibilities that didn't exist in the past."

Read these five steps using the formula in Chapter 8. Consensus reverse iteration is the application of the Interference Method on products: analyze mature consensus (10 points of input), find high-frequency scenarios that do not require education (10 points of target), and use familiar experience to narrow the choices for users (80 points of execution).

Dance with Love in Chapter 2 can be reused here in its entirety because it uses the five steps very standard. Find consensus: Middle-aged and elderly people want to learn dance and be taught patiently. This demand has matured to the point where a whole set of old solutions are charging: offline dance classes, senior universities, and self-study following videos. Borrowing experience: "Love Learning" launched in 2025, the engine is a real-life AI one-to-one tutor, which is technically at the forefront of the leftmost side of the curve; but all the interfaces that students come into contact with are familiar things (courses, teachers, WeChat groups). The design intention of the entire product is to make students feel not "I am using AI", but "I have met a teacher who is always available and patient." Survive first: a million-level student in one year, with a monthly income of tens of millions of yuan (company-reported basis). It is worth noting that this company did not do commercialization in the first two years and built a technical system first. Once commercialization started, it started from the right end of the curve. An AI company with the most cutting-edge technology serves the people on the far right side of the adoption curve. The new organizational form and the traditional user group not only do not conflict, but are the most stable pairing: the newer the engine, the older the shell.

5. Gear shifting timetable: three signals, two ways to die #

The most difficult thing to figure out among the five steps is the fifth step "move one layer to the left each time": when to move. If you move it too early, it will be a small Meta. If you move it too late, it will be locked by the right end. Look at the three signals on the timetable. If all three are on, put it into gear.

Signal 1: The current layer’s money is stable. Payment and repurchase have entered a steady state, and unit economics are positive. This is not a "model run-through" in the financing narrative, but the cash flow of this layer is enough to support the trial and error budget of the next layer. The account from Chapter 5 applies here: Shifting left is investment, and investment can only be spent on profits, not survival funds.

Signal 2: When the standard library of the layer is saturated. Use the ruler of Chapter 6 to measure: the high-frequency exceptions at this level have been written into the standard, and the frequency of new exceptions has dropped significantly, indicating that you are almost finished learning the judgments that this level can teach you. If you continue to stay, the learning returns will be diminishing.

Signal 3: The user on the left comes over by himself. Users on the adjacent layer begin to actively inquire about prices and use your products in scenarios you have not designed, and consensus begins to spontaneously spread to the left. This is the cheapest market research: there is no need to predict what the next layer will want, the next layer is already standing at the door with the demand.

The two ways to die correspond to "forcibly moving without all signals" and "not moving with all signals" respectively. The former is equivalent to educating the market in advance at your own expense; the latter will accumulate "customization debt": special functions for individual large customers are piled up layer by layer, and the versatility of the product is sold one by one. By the time you want to move left, the ship has been welded to the dock. The whole content of the shifting discipline is one sentence: Decided by signals, not by emotions, whether the emotion is ambition or comfort.

6. The complete anatomy of the loser: the only one who pays the billBack to Meta. Put Horizon Worlds into a five-step framework, and do each step in reverse: the problem consensus does not exist ("I need to socialize in the virtual world" is not an existing consensus among any group of people), the experience requires new learning (head-mounted displays, controllers, avatars), survival verification is skipped (losses are subsidized by the main business, and the product never needs to survive on its own), and then start directly from the far left of the curve. The most telling details reportedly come from within the company: In the fall of 2022, executives in charge of the Metaverse urged employees to log in to Horizon Worlds more in an internal memo, asking the team why they didn’t love the products they built. Products that don’t even have a consensus among manufacturers are convincing the world on the far left side of the curve. The answer to the first step of the Interference Method (analyzing user consensus) was already given in my own conference room, but no one wanted to read it. The lesson bought for $83.6 billion can be compressed into one sentence: Be the only educator in a market without mature consensus, and the wallet will bottom out before the education is completed, even if it is the thickest wallet in the world. You can be the first, but don’t be the only one paying. #

Let’s be honest about the boundaries here: what Meta loses is not “preemption” itself, but using unlimited bullets to cover up the absence of each layer of verification. The forward path is not impossible to win (the boundary clauses in the next section will give the conditions for winning), but every step of the forward path is paying for public goods, and you must confirm that you can afford it and that you can keep it after paying.

7. Supporting facilities on the R&D side: Five gates of progressive iteration #

If you choose the right path, the R&D rhythm can destroy it. A common way to die is that as soon as the product survives on the right side, R&D can't help but stuff the three-layer span "vision function" into the road map. The old genes replaced in this chapter are revealed here: The urge to rely on originality and show off skills, always feeling that borrowed consensus is not decent, and originality counts. Chapter 7 resolved this knot: Duan Yongping’s “dare to be the queen of the world” is in the opposite direction to Musk’s first principles, but they reject the same thing: unverified consensus. It is not disrespectful to borrow a proven market consensus, it is a Clarity Method.

We use five gates internally to lock the pace of research and development (core terms of product iteration standards):

1. Single variable verification: Only one key variable is verified in each iteration, the variables are mixed together, and what is learned is reset to zero.
2. Layered investment: The three levels of stable core, adjacent improvement, and frontier exploration are proportioned separately, and the proportion matches the level of the company. For companies that are still at the survival level, the proportion of frontier exploration should be close to zero.
3. Do not destroy the proven value: The launch of new capabilities must not damage the proven payment scenario, because the tolerance of right-end users for "change" is much lower than the tolerance for "old".
4. Result-driven roadmap: Use the verification results of the previous layer to decide what to do at the next layer, rather than using the vision to backtrack; each upgrade of the roadmap must be signed with paid evidence.
5. Standard precipitation in each round: At the end of the iteration, standard provisions must be output and entered into the standards engine (Chapter 6). Otherwise, this round of iteration will only produce code and no judgment.

Gate one deserves its own example because it is most often violated unconsciously. A typical "big version": new pricing, new interface, and new customer groups are launched at the same time. No matter whether the data rises or falls after three months, you don’t know who to thank or who to blame. The three variables are mutual noise, and the learning value of this round of iterations will approach zero. And learning is precisely the only guaranteed output of iteration: the results can be good or bad, and the learned judgments can guarantee harvests despite droughts and floods. Single-variable discipline seems slow, but it is actually fast: learn one thing at a time, and learn each thing truly. The standards engine in Chapter 6 only takes clean input.

Wutiao Gate also covers the second way of death mentioned in Section 5 on the research and development side (customized debt: serving conservative users for a long time, and the product is locked on the right end one by one). Gate two forces the retention of input from adjacent improvements, and gate four ensures that left shifts are determined by evidence rather than inertia. The end of the reverse path is still innovation: the right end is just the starting point, not the household registration.

8. Boundaries of Judgment #

Four, the first two are the applicable boundaries of this path itself, and the last two are the honest explanation of the sample.

First, the category with extremely short technical window and winner-take-all cannot afford to wait for the reverse path. In a market dominated by network effects (social, platform, operating system level entrance), the consensus accumulation of pioneers will form irreversible compound interest. At this time, "daring to be the queen of the world" is equivalent to giving up. The criterion is to see whether the user value of this category grows super-linearly with the number of users: yes, consider the positive direction; no, the reverse direction is more stable.

Second, the cost of channels to reach the right end may be higher than the cost of educating the left end. In the later period, the public did not read technology media or visit product communities. Reaching them often relied on offline, agents and time. Before choosing a path, calculate two accounts: education cost and channel cost. The answer for BBK and OPPO is to spend twenty years building channels. That in itself is a huge investment, but it is not called "market education fee" on the financial statements.Third, the forward path can win, and experts will also get it wrong. Tesla won the market through positive education: it borne almost all the early education costs of the electric vehicle category, and then maintained the results. What is even more educational is that in the same Stanford exchange in 2018, Duan Yongping asserted that "Tesla is a company with zero value and will be finished sooner or later." I put this misjudgment in the same chapter as "Dare to be the queen of the world" because I want to put it to the end: the path is a choice, not the truth; any judgment standard (including the assertion in this chapter) has boundaries, and the judgment of an expert is no exception**.

Fourth, sample quality. Duan Yongping’s primary source is the authorized draft of the 2018 Stanford Exchange, not a verbatim shorthand; Meta’s data comes from financial reports and internal documents obtained by the Wall Street Journal, which have been verified. The two multiplication samples in the third section each have a statement: AI + hardware is only an example of the multiplication mechanism, and does not constitute a judgment on any specific track or target; our own business is a self-report and is still in progress. I only explain the judgment structure when choosing it, and do not explain the operating data. The data will be supplemented after it is earned. As for the loser of this path (a company that is locked in mature demand and has never been able to move left), I have not found a sample that can withstand verification, so this box is left blank for now: the customized debt risk of the reverse path is currently only a mechanism deduction, and there is no autopsy report.

What to Do Monday Morning (No. 1 perspective) #

First write a multiplication, then put the product on the adoption curve to answer three questions, and finally look up a table:

1. Write multiplication first: What are the variable terms and invariant terms of your business? Invariants are tested by Bezos’ question: Can’t imagine customers asking for the opposite ten years from now. If you can't write variable terms, your business has nothing to do with this era, which is not necessarily a bad thing, but don't tell it AI stories anymore; if you can't write invariant terms, don't spend money yet.
2. **Which level of users are we serving now? **Use behavior to judge, don’t use wishes: Look at who accounts for the majority of paying users, not who is being told at the press conference.
3. What does **the company currently lack most? **Is it leadership in new categories, or cash flow and learning opportunities for survival? The answer is the latter, and you are not qualified to take the positive path; if you take the hard path, you will use your livelihood to pay education fees to the industry.
4. **If you choose the reverse path, what is the consensus on the question on the right that does not require education? **Write it into one sentence and ask it to ten target users; if someone needs you to explain the problem itself, it means that the consensus does not exist, so change it.
5. Check the R&D roadmap: Each project in this quarter is labeled as stable core, adjacent improvement, or cutting-edge exploration. Calculate the ratio and compare it with the company's level. There are three cutting-edge exploration projects on the company roadmap at the survival level. This is not ambition, but the misallocation of resources mentioned in Chapter 5: limited resources are spent on the wrong level.

Side note (individual and team perspectives): Career paths also have two paths going in opposite directions. Using the most cutting-edge skills to pave the way is the right path: you have to educate your employer yourself on "why this new type of job is needed." There are many vocabulary accounts, certification accounts, and trust accounts. What is more stable is the personal version of consensus reverse iteration: first use mature skills that have been priced in the market to survive, occupy a visible frontier position, and then gradually shift the work content to the left. The bridge route drawn in Chapter 10 for people on the left bank of the chasm (writing standards for existing work) is the first shift on this path: from selling execution to selling judgment, moving the first square from the right end to the left end.

Connections to Adjacent Chapters- Continuing from Chapter 11: After the individual’s time allocation is straightened out, what path will the organization’s external output - products - evolve along? The old gene replaced in this chapter is "reliance on originality and the urge to show off skills". #

• Inherit from Chapters 1 and 6: the multiplication of variables × invariants for business selection - the variable term is the execution zero of Chapter 1, and the judgment of the invariant term is borrowed from Bezos's question in Chapter 6; the same invariant principle is used in Chapter 6 for "where to build the standard", and this chapter is for "where to choose the business".
• Handed over to Chapter 13: The three battlefields of talent, time, and product have new standards - the fourth battlefield is the 90 days itself: assembling the new standards of the three battlefields into a complete round of organizational-DNA evolution.

Chapter Acceptance Self-Check (compare with the five acceptance standards of the chapter) #

1. The assertion is that one sentence can be restated ✓, and is strictly isomorphic to Chapter 8 Interference Method (consensus reverse iteration = product application of Interference Method).
2. Whiteboard framework diagram ✓ (using curve + reverse arrow five steps + four-level consensus ladder).
3. External comparison and data ✓: Positive S1 Duan Yongping (2026-07-26 New Core Stanford 53 Ask for First-hand Source + BBK VCD Example) + Dance with Love (v1.1 expanded to five complete steps: Love Learning Products, Millions of Students, Build Technology First and Commercialize from the Right End); Loser Meta Horizon Worlds (836 billion, monthly active and visitor data, internal memo details are marked as "reported", exited in 2026-06); the forward path is successfully compared with Tesla (including Duan Yongping's honest explanation of misjudgment); v1.3 adds two samples of industry multiplication in the third section (AI + hardware is a mechanism example, bioby.ai's overseas influencer marketing agent is a self-report sample, and the quality statement can be found in the fourth judgment boundary); P4 gaps are marked truthfully.
4. 16 golden sentence candidates ✓ (v1.3 adds three new multiplication sentences: variable/invariant question, don’t pay for AI, window and compound interest).
5. "What to Do Monday Morning" From the perspective of position 1, first write multiplication + three questions and one check + personal notes ✓.