Chapter 9: The Consensus Engine: How Shared Context Is Created

1. Work for a slightly different company #

In 2016, Administrative Science Quarterly published a study: two scholars used 76 in-depth interviews to review the entire process of Nokia’s losing smartphone battle from 2005 to 2010. They found neither technical incompetence nor strategic blindness. Nokia engineers have long seen the direction of touch screens and ecology. What they found was a conveyor belt of fear.

The mechanism works like this: the top management is afraid of the outside world (Apple’s offensive, the patience of shareholders), so they put crazy pressure on the middle level to get results, but are unwilling to reveal the full extent of the threat; the middle level is afraid of the inside (the anger of superiors, competition from peers), so they only report good news and filter out the bad news a little bit at each level. The two fears are connected end to end, producing a fatal result: the Nokia that the top executives saw is a company that does not exist: its operating system is in good progress, its technical capabilities are sufficient, it just needs to be a little faster. Each level of management worked for a slightly different company, and the real company was dying in front of everyone.

Two details in the study are worth amplifying. First, it’s not that the middle management doesn’t know the truth. The true state of Symbian, and the true gap between it and the iPhone, is close to an open secret in the middle levels of the organization. Managers in the interviews admitted that they deliberately remained optimistic during presentations because in that organization, people who brought bad news were seen as passive and unprogressive, and bad tempers at the top were intimidating (both points came from interview transcripts in the study). Second, it’s not that the top management didn’t take action. They continue to exert downward pressure and demand faster results, but the pressure just intensifies the filtering: the greater the pressure, the more the middle managers dare to report only the kind of information that can suspend the pressure. People at every level are rationally protecting themselves, and together they kill the company. Here's the study's coldest finding: The process doesn't require a single bad actor, just a structure that makes honesty more expensive than optimism.

What defeated Nokia was not Apple, but the fake context that reached the top after layers of beautification.

This chapter deals with the final assembly of the whole book: the parts (standards engine, Clarity Method, Interference Method) built in the previous three chapters are combined into a complete machine called consensus engine. What it produces is exactly what Nokia lacked most before its death: unified context.

2. The only output of an organization that cannot be outsourced #

Let’s get rid of a deep-rooted belief: the company’s output is products.

Following the derivation of this book, this sentence is no longer tenable. AI is doing the execution of the product (Chapter 1); AI is optimizing the market launch; code, copywriting, reports, and customer service are all on the zero list. Peel off the layers that can be outsourced to AI, and finally there is a core that cannot be peeled off: the consensus of "how do we judge", that is, what is considered good, what not to do, and what standards should be followed when encountering conflicts. There is no model that can produce this thing for you, because it is not information, but a group of people's common belief in the same set of judgments.

Here is a common misunderstanding: Consensus does not mean unanimous agreement. What the consensus engine produces is "a common belief in the same set of judgment standards", not "everyone raises their hand for every decision". The latter is a form of decision-making, and often the slowest. On the contrary, the thicker the consensus, the faster single-point decision-making can be: when the standards and context are unified, the judgment made by one person will probably be the judgment made by others, and there is no need to hold a meeting to align, because it has already been aligned. Consensus is not used to make decisions together, but to allow everyone to still point in the same direction when making decisions alone. This also answers the question of "will unification kill diversity" by the way: unification is a ruler, not an answer. When the rulers are unified, different answers can be compared, and the debates can converge instead of diverge; if the rulers are not unified, the debate will be just two sets of coordinate systems passing each other for a hundred years. So the first half of this chapter’s conclusion: humans are responsible for consensus, and AI is responsible for everything else. If the consensus is good, the product is a by-product of the consensus; if the consensus is bad, AI will only use perfect execution to scale up the bad consensus. The root of the paradox of abundance in Chapter 2 is here.The general assembly diagram of the consensus engine is the second main diagram in the book. The assembly relationship must be explained at once: the standards engine (Chapter 6) is the core component, the Clarity Method and the Interference Method (Chapter 7 and 8) are the two production lines, the consensus engine is the assembled complete machine, and the cultural spillover is the external effect after the continuous operation of the complete machine**. There are no parallel machines, only the layer-by-layer assembly of one machine. There are three layers of progressive functions within the whole machine: the standards engine solidifies judgments, turning the implicit judgments of a few people into testable and iterable provisions; the consensus engine copies judgments, allowing different members to act independently using the same standards in a common context; and cultural spillovers spread judgments. After the judgments continue to produce results, people outside the organization begin to actively use them.

Chapter 4 says that great organizations need and only need two capabilities: unified judgment standards and unified context. This half of the judgment criteria has been covered in Chapters 4 to 8; the protagonist of this chapter is the other half: context. It's easier to overlook than a criterion because it doesn't look like an asset, like a by-product of everyday communication. Nokia has demonstrated ignoring its price.

3. First-hand textbook on context: Context, not Control #

How to create shared context? The most complete first-hand teaching material comes from Zhang Yiming's speech "Being a CEO should avoid rational conceit" at the Source Code Capital Conference in 2017.

Let’s look at his definition first, which is as precise as an engineering document. Context: The collection of information required for decision-making, including what the principles are, what the market environment is, what the industry structure is, what the priorities are, what level of achievement is required, business data and financial data. Control: committee, instructions, decomposition and aggregation, process, approval. Then there is the principle: "We prefer 'Context, not Control' solutions": sufficient Context, a small amount of Control, everyone has complete contextual information and makes business decisions by themselves, and managers only intervene a small amount when necessary.

The most worthy thing to copy into the standard library is the reflex arc: "When encountering a problem, it is often customary to first ask whether the Context is not sufficient, rather than adding Control." When a certain progress has a problem, the first reaction is not to replace it with a higher-level person, but to ask: Have you not shared the industry situation, business data, and past failure cases with him? Put this reflex arc side by side with Nokia, and you will see two machines with opposite directions: Nokia's information is filtered according to fear, and the higher it goes, the more false it becomes, so it can only be compensated with more Control (committees, approvals, pressure), and Control creates new fears; Bytes of information are spread out according to decision-making needs, and the more transparent they are, the less they need to control, and judgment is delegated to the place with the most information. **Control is paying interest on non-uniform contexts, compounded. **

There is another sentence in the speech that directly points out the old genes to be replaced in this chapter. Zhang Yiming observed that some people "use information asymmetry to reflect their own value." This is information privacy: keeping the context in your own hands and making it the only way to go. In the old organization, this is an individual rational survival strategy; in the AI-native organization, it is the sand in the machine, because the output limit of AI is determined by the context fed to it (conversion in Chapter 6). A person who hoards context is equivalent to lowering the output limit of all colleagues and all AIs at the same time.

He also gave the answer at the tool level: the internal tool team of nearly a hundred people, the self-developed OKR system and internal IM are connected, and new employees can see internal information when they join the company. In his words, this is "building the company as a product." Unification of context does not rely on transparent all-employee emails, but on infrastructure that makes it easier to obtain context than hoard it.

4. Three workshops of the machine: meetings, documents, review #

Shared context sounds like culture, but when it was implemented, it was three specific workshops. Every company has these three things. The only difference is whether they operate them as parts of the consensus engine.

Meetings are assembly lines of consensus. Meetings in most companies produce two things: emotions, and to-do lists. Meetings in the consensus engine can only produce one thing, which is an update of judgment: a new standard, a standard revision, or a hypothesis to be tested. The acceptance method can directly convey the logic of the hard rule in the next section: the input must be evidence (data, executable things, users' original words), not opinions; the output must be able to be written into the standard library. If it cannot be written in, it means it has not been opened. Use this ruler to measure your regular meetings. Most companies will find that more than half are idle: neither digesting judgment gaps nor producing judgment updates, but just aligning the time of the same people into a busy ritual.Documents are the storage of consensus. The engineering meaning of "building a company as a product" is: verbal consensus does not count as consensus. It decays every time it is conveyed verbally. By the time it reaches the fifth person, it is already another judgment. Chapter 6 mentioned that standards are how judgments are stored. Here is the other half: the context required for decision-making must also be documented. Otherwise, each newcomer will have to rely on "asking more old colleagues" to piece together the world view, and what they piece together will inevitably be a slightly different company. Nokia's disease started on the first day of employment. MD Anderson in Chapter 2 still owes an echo: The last straw that broke the $60 million project was the medical records system that Watson couldn't read and the organization replaced it. People and machines are completely equal on this point: There is no essential difference in the judgments produced by people who cannot feed shared context and by machines that read the wrong format. **

Review is a correction loop for consensus. The launch of a standard is not the completion of consensus, but the beginning of assumptions. The review is responsible for feeding the execution results back to the standard: if it is verified, it will be solidified, and if it is falsified, it will be revised. Without this loop, the consensus engine will degenerate into a dogmatic printing press: more and more documents, increasingly deviating from reality, and every page is stamped with "Consensus Reached". Nokia lacked neither meetings nor documentation. What it lacked was the most vulnerable component of the circuit: the right of passage for bad news. Of the three workshops, it is always the first to go bad, because it is the only component that requires someone to pay the price first: the person who tells the bad news bets that the organization will not punish honesty. Position 1 has only one way to maintain this component: publicly and observably reward the first person to report bad news. There is also a personal minimal version of this circuit, left to Chapter 11 to unfold: three questions a day, cheaper than any review session.

5. A machine being transformed: Chuanshen #

Byte is a near-native sample, but more readers are in a different situation: How can a company that has been running for twenty years install the consensus engine into the old workshop? The mechanical design of Chuan Shen Yu Lian is worth looking at piece by piece.

The first action is to set up a position for consensus on the organization chart: set up a chief intelligence officer (CAIO) and clearly divide the labor with the CIO: the CIO makes the system run smoothly, and the CAIO makes the business smart. One manages the channel for information flow, and the other manages the density of judgment. This division of labor itself is a vote for the conclusion of this chapter. The second action is to give consensus a decision-making body: the AI ​​Native decision-making committee, so that the judgment of transformation is not scattered among various departments. The third action is a hard rule: If you don’t have a runnable DEMO, you can’t do it. It forcibly upgrades the input of the meeting from opinions to evidence: opinions will quarrel, but running DEMO will not. The fourth action is the most ingenious, called Energy Gold Mechanism: Only when internal AI applications are actually used by colleagues and are satisfied with them, developers can accumulate profits. Judgment power is given to users, decision-making power is given to data, and AI promotion changes from administrative orders to market behavior. The four actions combined are a miniature consensus engine: the committee solidifies judgment, DEMO rules filter the moisture in the context, and energy gold allows the consensus of "what tools are easy to use" to be created by usage data rather than rank. The effect data of this set of mechanisms has not yet been verified by a third party. I only quote the mechanism itself; however, every piece of the design logic of the mechanism can be borrowed directly.

Founder He Enpei has two words that can be used as the nameplate of this transformation machine. One sentence explains why installing machines instead of training individuals: "Instead of waiting for employees to become AI masters, it is better to let the organization develop AI capabilities." Whether the ability grows in the organization or in the individual is the dividing line between organizational judgment and personal judgment in Chapter 4. It is also the only feasible path for the transformation of old companies: You can't wait for every employee, but you can change the machine first. Another sentence adds a "human" version to the execution zeroing in Chapter 1: "There are only two types of people left in the future world - those who produce Tokens, and those who use Tokens to create greater value." Translated into the language of this book: the former stands at the execution level and is listed on the zeroing list; the latter stands at the judgment level and uses consensus to command execution. A company that has been doing translation for 20 years has come to this realization, which is more convincing than the declaration of any native company. It doesn't want to understand it, it understands it by being chased by the zeroing curve.

6. Spillover: When others start to think in your words #

A wonderful thing happens when the machine keeps running: Consensus crosses the boundaries of the organization and begins to multiply outside.

The samples from the three museums all appeared in Chapter 4, and this chapter only looks at their spilled surfaces. Duan Yongping's circle of influence is the cleanest object of observation: the word "duty" has not remained in BBK: it has entered the value ranking and decision-making language of OPPO, vivo, and Pinduoduo, and a generation of entrepreneurs use it to make judgments about their own companies. Musk's business group demonstrates cross-company replication: first principles, rapid iteration, and extreme goals, which are repeated in several companies such as SpaceX and Tesla and have become a common reasoning process for tens of thousands of engineers. The extreme example of scale is the Communist Party of China: judgment standards are compressed into a communicable organizational language, ultra-large-scale decentralized execution is supported by a common context, and practice feedback enters the correction mechanism; my research on this sample is limited to the organizational mechanism itself, and dimensions outside the mechanism are beyond the scope of this book.The three samples jointly define the acceptance line for cultural spillover: It is not "everyone has heard of you", but "everyone starts to think in your language"**. Brand awareness measures attention, and cultural spillover measures the spread of judgment standards: the former can be bought with money, while the latter can only be earned through consistent correct judgment. Moore's word "chasm," which dominated venture capital language for thirty years (just covered in Chapter 8), is a personal version of the same thing. Spillover is therefore the ultimate instrument for testing the consensus engine: if internal judgments do not continue to produce results, no one from the outside will take the initiative to borrow your words. Cultural spillover is the result, not the project. Companies that treat it as a communication project will gain visibility, not spillover.

The spillover is not just honor, it will send back three very real things. Self-screening of talents: When "duty" becomes the language of the industry, candidates who agree with it will go to the BBK system, and those who don't agree will avoid it. The first filter of the recruiting funnel is outsourced to Culture itself at zero cost. Customers with pre-installed consensus: When customers start to use your framework to describe their needs, the first two steps of the Interference Method (analyzing consensus and finding scenarios) have been completed for you; the first of the three valves mentioned in Chapter 8 is always open for you. Pricing power of language: If the industry uses your words to discuss issues, it means that all competitors must first enter the coordinate system you defined before discussing differences. After Moore invented the "chasm," every early-stage company had to answer the question "How do you cross the chasm?" whether they liked it or not. The common form of these three things is compound interest: once the spillover starts, it will reduce the cost of creating consensus every time in the future. **

7. Boundaries of Judgment #

Three.

First, "no outsourcing" is a ranking, not a ban. Humans are responsible for consensus, and AI is responsible for everything else. This refers to the allocation priority of scarce resources: the maximum proportion of human time in the organization should be invested in consensus production (proposing standards, checking context, and testing judgments), rather than investing in the execution that AI can do. It does not mean that products and technologies are not important, it means that when you can only put the best people on one thing, put it on consensus.

Second, shared context does not mean transparent everything. The definition of Context is "the collection of information required for decision-making": spread out according to decision-making needs, not indiscriminately disclosed. Salary privacy, compliance red lines, and unofficial mergers and acquisitions are not included in the "need for decision-making" caliber. The acceptance criterion for shared context is: people making the same decision see the same set of facts, rather than the entire company seeing all the facts.

Third, tell the truth about the quality of the sample. Nokia has 76 interviews with academic research endorsements, which is the load-bearing wall of this chapter; Zhang Yiming’s speech is a full text first-hand, but the effect of Byte’s subsequent practice is the company’s caliber; the expressiveness only quotes the mechanism, and the effect has no third-party verification; Duan Yongping’s circle’s failure and correction list, and Musk’s group’s bilateral data, the accounts owed in Chapter 4 are also owed to this chapter. There is another known gap: I have not yet obtained available evidence for the quantitative contribution of the collaborative tool layer (such as Feishu products) to context unification. The discussion of the tool layer in this chapter is only based on the first-hand expression of Byte's self-developed tools.

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

Three instruments, read in one hour:

1. Read speed: Call up the most important document of the company (strategy, core standards or customer definition). Look at two numbers: When was the last update? Randomly ask people from five different departments, how many of them can recite its key points without looking up the information? If the document has not been updated for half a year, it means that the standards engine has stopped; the repetition rate is less than 30%, which means that the consensus engine has stopped.
2. Check the sand: Find out the three most common information nodes in the company that "you have to ask someone to know". Ask a question: Is this a division of responsibilities, or is the information private? The test method is simple: how long does it take to document that information? The answer is within an hour but has not been written yet. It is that someone is using information asymmetry to reflect value. Document them all this week and see who objects.
3. Look at the spillover: Search whether customers, candidates, and peers actively use the vocabulary and judgment caliber you invented: whether candidates will use your slang in recruitment interviews, and whether customers will use your framework to describe problems when making demands. It’s not a shame if you can’t find any of them, it just tells you honestly: the machine is still building up its first lap speed.

Note (individual and team perspective): Establish a personal version of Zhang Yiming’s reflex arc for yourself: When the subordinates or AI hand over things that are wrong, first ask "Is the context I gave enough?" and then "Is his ability adequate?" If the order is reversed, you will use substitutions to solve all the problems that should be solved with documents.

Quotable Lines1. Humans are responsible for consensus, and AI is responsible for everything else. #

2. Contextually disunified companies, where each employee works for a slightly different company.
3. What defeated Nokia was not Apple, but the fake context that reached the top after layers of beautification.
4. Control is paying interest, compound interest, on inconsistent context.
5. Standards engine solidification judgment, consensus engine copy judgment, and cultural spillover judgment.
6. A person who hoards context is equivalent to lowering the output ceiling of all colleagues and all AI.
7. Contextual unity does not rely on calls for transparency, but on infrastructure that makes acquisition easier than hoarding.
8. Opinions will quarrel, but running DEMO will not.
9. Brand makes others remember you, and cultural spillover makes others start to judge like you.
10. Popularity can be bought with money, but spillover can only be earned through continuous correct judgment.
11. Cultural spillover is a result, not a project.
12. The output of the meeting should not be a to-do list, but a new standard.
13. People at every level are rationally protecting themselves and collectively kill the company.
14. Verbal consensus is not consensus: it decays once it is conveyed, and reaching the fifth person is already another judgment.
15. The right of passage for bad news is the first to break and the most difficult to repair part of the consensus engine.
16. Once spillover starts, it will reduce your cost of creating consensus every time thereafter.
17. Consensus is not used to make decisions together, but to keep everyone pointing in the same direction when making decisions individually.

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

1. The assertion can be restated in one sentence ✓ and is an inference of the core assertion (judgment is scarce → shared belief in judgment is the only output of the organization that cannot be outsourced).
2. Whiteboard frame diagram ✓ (consensus engine assembly diagram - the second main diagram in the book: two production lines + standards engine core parts + dual output + overflow layer; three-layer progressive chain; daily operation diagram of three workshops).
3. External comparison and data ✓: Loser Nokia (first-hand research on ASQ 76 interviews, including details of mid-level fears); positive S5b Zhang Yiming Context, not Control (full text of the first-hand speech in 2017, newly verified this time) + Chuanshen Four Mechanisms (verified, only citing mechanisms) + // Spillover Three Halls (accounts and debts are truthfully explained); MD Anderson contextual details are Chapter 3. Chapter 2 echoes without repeating the load.
4. 17 golden sentence candidates ✓ (v1.1 adds five new sentences: Nokia structural sentence, verbal consensus decay, bad news right of passage, spillover compound interest, and consensus direction sentence).
5. "What to Do Monday Morning" No. 1 perspective, three instruments + personal notes ✓.