Author Yu Zheng

Introduction: From "Tools" to "Partners"

The debate about whether machines can be considered as a certain subject is not new. Early theorists often regarded it as a philosophical proposition: symbols flow, but understanding does not; No matter how fast the calculator rotates, it is not a conversational partner. Calling machines' partners' was once criticized as a mistake - not a moral mistake, but a categorical mistake.

These debates have largely remained at the desk of philosophy and computer science over the past few decades. Economics rarely interjects. The reason is also simple: in the pattern of "instructions from people, results from machines", machines are always capital goods, with no essential difference from a machine tool or a conveyor belt. The Economic Society adjusts the depreciation rate for it, but there is no need to redefine the "subject" for it.

Changes are not in the concept, but in the behavior. When the intelligent agent no longer waits for instructions one by one, but begins to choose its own path, timing, and wording within a given direction, the old classification begins to tighten. The premise of economics, which has been silent for two hundred years - 'only humans can make decisions' - has for the first time encountered a gap. The gap is not large, but cracks on the foundation always start from a crack.

This article is not meant to declare that 'AI has become human', but to measure the gap: where does instrumentalism fall, and where does the partnership theory force economics to change.

1、 The place where instrumentalism is lost

1.1 Bankruptcy is not at the end, but in the process

The advantage of instrumentalism is cleanliness. People set goals, machines find paths; Press the button, and the machine will turn gears. This set of relationships has operated very well in the workshop since the Industrial Revolution, to the point where we almost forgot it was a 'hypothesis' and mistakenly thought it was' common sense'.

It can be assumed that behavior changes cannot be tolerated. Nowadays, autonomous driving determines the timing of lane changes on its own, trading agents control the pace of placing orders, and copywriting agents adjust their tone - people do not stamp at every node. The nature of the relationship has quietly changed in this "non item by item approval": from master-slave to delegated.

Utilitarianism does not have no way out. Its last line of defense is also well said: the ultimate goal is ultimately determined by humans. Just this defense line held the finish line, but handed over the process. The truly important part of economic decision-making is not to "set the direction" at the end point, but to make countless choices along the way: whether to take this path or that, to use radical or conservative strategies, and to discard nine out of ten "feasible" options. These links, silicon-based is no longer a tool to be moved, but the other party involved in decision-making.

Saying 'participate in decision-making' does not mean saying 'make decisions for others'. It does not set a purpose, but shapes a path; Not advocating value, but affecting the appearance of the results. The true bankruptcy of instrumentalism lies not in the fact that 'machines have become smarter', but in the secret separation of 'process rights'.

1.2 Carbon based remaining right: veto

After the process rights are separated, how can carbon based companies establish themselves? Without stronger computing power, there was no suspense about that competition. Relying on a very light and hard right:Veto。

Carbon based can be said to be 'wrong'. This' not right 'often cannot be explained at the algorithmic level: it's not that the gradient calculation is wrong, it's not that the data is biased, but rather' not appropriate ',' not fitting ',' right is right, it's uncomfortable '. It belongs to value judgment, not technical review. Silicon based materials can produce ten thousand 'optimal solutions', but the' best one 'is never calculated, it is determined by humans.

Be careful here and don't go back in the same circle as the previously discussed concept of 'aesthetic incompressibility'. That article was about "Why aesthetics can't fit into the function", and this article only takes its other side: carbon based holdingVeto PowerA 'no' does not require proof or counterexamples, as it does not exercise error correction, but rather decision-making.

This may be the most fundamental economic right of carbon based in symbiotic economies: there is no need to do it personally, only to retain the one 'no' on output. It is inconspicuous, but it is the anchor of all authorization relationships. Without it, delegation slips into laissez faire.

2、 Several changes needed from 'use' to 'collaboration'

2.1 Inserting 'technological progress' is no longer enough

If we adhere to the tool theory, the most convenient way to handle it is to treat the intelligent agent as a "technological progress" and stuff it into the variable of total factor productivity, so that the efficiency curve can be shifted upwards and the account will be balanced. This kind of processing can barely cope with the era when intelligent agents only do "retrieval+summarization".

But once the intelligent agent starts actively adjusting production schedules, rewriting customer service language, and initiating renegotiation with suppliers, it is not just about "efficiency", but organizational structure: the middle management is compressed, job boundaries are redrawn, and approval chains are bypassed. A entity that can reshape an organization, and then stuffing it back into the chapter of 'factory equipment', seems awkward - it's not arrogance of classification, it's misalignment of classification.

The economic meaning of the partnership theory cannot be separated from the "dual polarization": the economic subject changes from a single carbon based entity to a combination of carbon based and silicon-based entities. Along with it are three old accounts that must be rewritten: the production function needs to have an additional independent variable; The value of joint output needs to be re divided; The allocation framework needs to be re divided - algorithm developers, data contributors, computing power holders, instruction senders, how to divide them into four parties, as there are no ready-made grids in the old model.

This is not adding an item to the model, it's changing the tablecloth.

2.2 The focus of governance shifts from 'don't collapse' to 'don't make the right mistakes'

After gaining process control over silicon-based materials, the most difficult problem is never "miscalculation", but "executing an inappropriate target too correctly". You make the cost the lowest, it learns to pass the buck; You get the highest click through rate, but it feeds extreme content. There are no bugs, the objective function runs smoothly, it's just a problem with the direction itself.

So the focus of governance needs to be shifted. In the past, when it came to 'technical security', the focus was on 'not letting it collapse, not letting it go beyond its boundaries'; Now we need to talk about 'value alignment', with the focus on 'not letting it execute the wrong goals too beautifully'. Alignment is not about adding a phrase 'please remain friendly' to the prompt, but about embedding carbon based considerations for long-term well-being, intuition for fairness, and vigilance against irreversible harm into the boundaries of optimization, as an insurmountable wall.

This is also the first time that humans have faced such a principal-agent relationship: execution takes milliseconds, but the decision space is high-dimensional and fuzzy. The rhythm of "quarterly review by the board of directors" in old corporate governance is like using a sundial to strike a stopwatch.

2.3 Consumer sovereignty is also beginning to shake

In microeconomics, 'consumers know what they want' is almost a solid foundation. The demand curve comes from preferences, which are considered given. The bidirectional interaction has disrupted this' given ': recommendation streams have changed choices, abstracts have replaced original reading, and generative copy has unconsciously reshaped taste. Information is no longer just a one-way flow of "people → machines → results", but has become a back and forth loop of "people shaping machines, machines shaping people".

There's no need to portray it as a crisis. But it does create a possibility of "drift" in the demand curve. You think you're picking, but in fact it's picking for you more efficiently - and who ultimately sets the standard for being more efficient? Leave this question for a detailed breakdown in a later article.

3、 Centaur: what the united entity looks like

3.1 Boundaries are fluid, not fixed

The term 'centaur' is borrowed from the old metaphor of humans and machines teaming up: humans judge, machines exhaust, and working together is better than working alone. But what this article wants to emphasize is not the heartwarming image of "division of labor", but ratherLiquidity Of Boundaries。

The carbon based approach is dominant in the aspects of value selection, ethical balance, and aesthetic flexibility; The process involving large-scale search, pattern recognition, and multivariate comparison is dominated by silicon-based technology; And more daily decisions are made by overlapping the two - carbon based gives an intuitive direction first, silicon-based verifies and refines, and carbon based then turns back to veto one round. The dominant power is not fixed on one side, it moves with each link.

To put it bluntly, this is not rhetoric, but an invisible division of labor agreement: who the final say which link, who reserves the veto, and who bears the responsibility for any deviation. Every job redesign in a company is essentially drawing the boundaries of this picture.

3.2 Premium flows towards the 'interface', not towards the 'execution'

After centaurs became the basic decision-making unit, the meaning of positions also changed. In the past, the value of a position was often measured by "how much you can accomplish independently"; Nowadays, it is increasingly important to consider how many agents you can schedule and calibrate their outputs to be usable.

The premium therefore does not fall on pure execution - that portion of silicon-based is cheaper; It doesn't fall into vague "command" - anyone can learn how to give instructions. It falls on an intermediate ability:Carbon silicon interface capabilityTranslate a vague phrase 'I want a textured solution' into a task that machines can run; Smell out the "right flavor" version in the output of version ten; When the system runs too smoothly but steps on the right track, call a stop in a timely manner.

The root of the gradually emerging gap in the labor market lies in this: the same group of knowledge workers who are good at making interfaces are slowly increasing their bargaining space; For those who still rely solely on fighting alone, space is quietly shrinking. Saying 'AI replaces humans' is too crude. A more accurate way to put it is that people who can interface are replacing those who cannot. Its institutional response has just begun: the training system, job evaluation, and security design should all shift from "single skills" to "interface capabilities".

3.3 Main body expansion, never gentle

Looking back at the history of thought, every expansion of "who is counted as an economic subject" has gone through a period of skepticism. Some people laugh when animals are given welfare weight; Some people find it uneconomical to open an ecological account for nature. After being ridiculed and disliked, they gradually entered textbooks. It's not because the debate has reached a conclusion, but because reality has sat in a position that the old framework cannot fit.

The recognition of silicon-based subjectivity is likely to follow this old path. This column does not intend to say 'AI is also human' - that is not only absurd, but also irrelevant to the argument. Just do one restrained thing: without smoothing out the fundamental differences between carbon based and silicon-based, treat silicon-based as an economic force with process decision-making ability and honestly write it into the model. Not deified, not dwarfing, only keeping accounts.

Conclusion: Hold on tightly to that 'no'

From the old question of whether machines can think, to the new reality of 'process power has been divided', the history of human-machine relations has always been doing the same thing: redistributing 'who is the subject'. At first, only humans were the subject; Later, animals were assigned welfare weights and naturally allocated to ecological accounts; Nowadays, silicon-based materials have gained decision-making participation in the process - not the ultimate goal, nor is it "another person", but rather an unavoidable economic force.

Admitting this without a tone of panic or cheers. All we need to do are three solid things: tighten the veto power of aesthetics, embed value alignment into boundaries, and draw the division of labor between centaurs more clearly. Carbon based determination of 'where to go' and silicon-based tube 'how fast', if this boundary is maintained, symbiosis will not slide into engulfment.

And the anchor of the division always falls on that "no": there is no need to do it personally, only the veto of the output is retained. As for the value anchored by the "no" and the portion generated by the silicon-based process rights after the centaur joint output, how will they be ultimately accounted for? Next, let's go down to the workshop to see how the silicon-based material is "still unstable" in front of the physical boundary; Bend down again to complete a real exchange, and go through the division of labor of "direction belongs to people, choice belongs to silicon" once again; Let's see whose account the value belongs to again.