If a CEO "distills" themselves into an iterable agent, making their decision-making logic, cognitive framework, and value judgments fully explicit and handing them over to an agent to execute, is the organization still in need of them? To go a step further: if this agent can calculate faster and see more comprehensively than they can, working 24/7 without fatigue, who should ultimately hold the decision-making power?
On the evening of September 16, over a dozen top executives gathered for a three-hour closed-door session themed "The Decision-Making Model for Top Executives in the AI Era."
The speaker, Tan Liang, former President of GE China and former Global Executive President of Alibaba, now teaches at a university in Hong Kong. The course framework has been fully published on his WeChat official account, so I will not repeat it here. Instead, I will only write about the three things this class forced me to think through clearly.
The Prerequisite for Distillation is Seeing
The essence of distillation is mechanization: transforming the judgments a CEO previously made based on intuition, experience, and "I feel" into a set of recordable, iterable, and verifiable disciplines. It is not about eliminating the CEO as a person, but requiring this person to code their cognitive framework. And the prerequisite for writing code is that you must first see it clearly yourself.
Looking back at the blackboard notes from that night, Tan Liang noted that the two most frequently appearing words were "seeing." Seeing the filters in the inputs: first-hand information and second-hand conclusions are two different things. Seeing the truth in the results: making a decision is not the end; the decision must be verified with real outcomes, and the subject of the review is the boss's own cognition and assumptions.
Seeing oneself in the assumptions: for every major decision, one must be able to write down a clear sentence stating what it is based on, what judgment was made, and what results are expected. What can be articulated is called a high-quality assertion.
Pruning is Judgment
"Distilling the CEO" is often understood as copying oneself into an agent and then retiring.
That is not the case.
Tan Liang mentioned Recursive Self-Improvement (RSI) in class: the distilled CEO agent faces questions from everyone inside and outside the enterprise every day, holds the most first-hand information, and in turn helps you correct judgments. I raised a question on the spot: Will recursive self-improvement lead to involution due to a lack of externalities?
To answer this, let us first return to a more fundamental framework. Zhang Lei from Innoangel Fund shared a judgment: the capabilities of humans and AI are orthogonal. Delegating judgment to AI is inefficient: the model is too large, and when making decisions, it considers things too comprehensively, wanting to expand every branch and calculate every step. Therefore, it needs pruning.
Pruning is judgment.
This judgment is evidenced by a real conversation. Previously, while communicating with Zhao Zhongxia, a world model entrepreneur, we discussed a question: if all human knowledge (including tacit knowledge) were turned into a cloud brain, it would be somewhat like an omniscient and omnipotent God. So, would this God have a personality? A style?
I said that the God of Go has actually already been born, in the years when AlphaGo emerged. The so-called playing style is the pruning made by human players based on their limited computing power, an aesthetic expression of trade-offs; whereas AlphaGo's style is unrecognizable to humans—it is too powerful for human comprehension. God has no style because God does not need to make choices.
After AlphaGo, the way professional players study game records changed: memorizing and deciphering AI's moves became the most important task, much like the Wuliang Sword Sect in "Demi-Gods and Semi-Devils": on a bright moonlit night, the shadow of an immortal wielding a sword appears on the Wuliang Jade Wall, the swordsmanship exquisite and unpredictable, with generations of masters gathering outside the wall to study it for decades.
Go has a 19x19 boundary and an endgame, yet humans still cannot fully understand it; enterprises face infinite possibilities. The stronger the AI computing power and the more permutations and combinations, the more its trade-offs become a black box that humans cannot read.
Xie Saining previously said in an interview on Zhang Xiaojun's podcast: the representations learned by language models from internet texts are projections and reflections of human experience, not an understanding of the world itself. Distilling a CEO learns this top executive's representation of the world; and such a representation does not exist on the internet, it can only grow from the real business interactions within one's own enterprise. If the representation is wrong, the stronger the computing power, the more mistakes it will make.
Therefore, the CEO's judgment is to help AI prune: telling it which branches do not need to be expanded, what kind of enterprise we do not want to be; and telling it which direction to take at critical nodes.
Returning to the issue of externalities. The representation of a general model is a projection of others' experiences, while the evolution of the agent is based on internal enterprise data and feedback; its cognitive boundary is the enterprise's cognitive boundary.
Tan Liang said that AI cannot solve the problems of human nature; it only amplifies it. In the AI era, higher demands are placed on people. After a CEO distills their social role, their job as a human CEO is to act as the reformer of their own agent, using critical judgments to make corrections and provide externalities. Voices that do not belong to this class, this discourse system, or this interest structure are not in the projection of any text; the agent will not proactively introduce them, they can only be brought into the loop by humans.
Taking Responsibility and Signing for AI
In the latter half of the discussion that night, an entrepreneur who had been in enterprise-level software for over twenty years threw out a radical viewpoint: "The person under you, if you give him 50 million units of computing power and 20 million tokens a year, and he does it better than you, then why not let him do it?" Another immediately countered: "Then why is he working under you?" Someone in the audience chimed in: "Because they can use the company's tokens to do private work."
The so-called "doing private work" is essentially the boss taking on responsibilities that do not belong to them. Tan Liang reminded that night: in the AI era, taking responsibility is precious. Those willing to take it will be rewarded by the times; but those who take on responsibilities that do not belong to them due to unclear boundaries will also be punished. The proliferation of "doing private work" is precisely a footnote to the latter half: it is not that employees suddenly became bad, but that the organization did not iterate in time, and the boundaries of responsibility were not rewritten.
This actually answers the tension at the beginning. The value of humans in the workplace and in economic activities comes down to just two things: making judgments to help AI prune, and taking responsibility to sign for AI.
On September 7, the Supreme People's Court issued the "Opinions on Hearing Dispute Cases Involving Artificial Intelligence in Accordance with the Law." From that day on, whether you are a supplier or a user in the AI industry chain, you must answer for taking responsibility in the governance system. Technology is technology, and responsibility is human. Algorithms cannot take the blame; the person who ultimately signs must still be a human.
Liberated from "The Person is the Position, and the Position is the Person"
I recall Teacher Tan Liang saying in a previous interview that "the person is the position, and the position is the person" is a kind of alienation in modern society, a byproduct of our wonderful modern life.
From the standpoint of human subjectivity, the very existence of a person is value in itself.
The "distilling oneself" discussed in this class is about distilling the most structured layer of social identity: work identity. Liberating people from this alienation, handing the position over to the agent, and returning the person to being a person. This is the ideal state.
But those who are too deeply alienated will instead become anxious because of this. The question at the beginning, "Is the organization still in need of them?" is precisely this anxiety.
On a deeper level, work identity cannot exist independently of human sociality. As mentioned earlier, the agent needs the CEO to provide externalities; here it can be reversed: the more developed a person's multiple identities are, the more they can become the externalities of the work identity, reducing its entropy. The information, perspectives, and judgments that do not belong to the position are all grown from life outside the position.
Externalities are not just the CEO's obligation to the agent, but also the nourishment of human sociality to the work identity.
In the end of distillation, what is distilled away is the position, and what is liberated is the person.
(The author is a representative of Guancha.cn and attended the "Liangjian" closed-door salon for entrepreneurs on September 16. "Liangjian" is an entrepreneurial learning and co-creation platform created by TTC, serving top executives and focusing on the three most hardcore propositions in the AI era: corporate strategy, organizational transformation, and leadership.)