The Boundaries of Compression

Some knowledge can be serialized and some cannot—take a plate of tomato and scrambled eggs:

If you want to make a plate of tomato and scrambled eggs and ask an AI for a recipe, you will get an extremely detailed one, with timing precise to the second and ingredients precise to the gram.

But look at it from another angle: you are hosting guests at home, and after you have made a plate of tomato and scrambled eggs, the AI cannot tell you whether there is too much salt or too little, because it knows neither your guests’ tastes and physical states nor how much salt is actually in the dish.

In Seeing Like a State, James C. Scott borrows two concepts from ancient Greek thought: Metis and Techne.

  • Techne (technical knowledge / serializable): can be written into manuals, formulas, or algorithms; it is highly compressible. As long as you grasp the underlying logic (the formula or the script), you can reproduce the result an infinite number of times.
  • Metis (practical wisdom / non-serializable): represents a quick, flexible, hard-to-articulate practical wisdom; it is incompressible. You cannot transmit an old Chinese-medicine doctor’s “feel” to someone else with a string of characters.

AI and Software Development

Claude Code has been hot lately, and many classic claims have resurfaced—for example, some say, “Programmers are about to lose their jobs; AI does a week’s work in half an hour,” while others say, “Programmers will be more valued, because the architecture AI writes is unmaintainable.”

Both the pro and con sides have a point. Here I want to analyze this classic topic from the angle of Metis and Techne. In practice, Metis and Techne are not a dichotomy but a continuous spectrum. Taking software product development as an example, it can be broken down into these layers:

1. Choice

Which product should we build? A social app, or a food-delivery app?

The considerations here include user needs, policy direction, platform attitudes, capital cycles, the competitive landscape, and other fuzzy, unquantifiable variables. Moreover, market data is always lagging, and many key pieces of information never even appear in public data.

The core of this layer is not “analytical ability” but the ability to make decisions amid incomplete and even contradictory information. The specific environment in which a given choice is situated has never appeared before, and it is almost impossible to ever fully reproduce in the future. Many of the key judgments come from intuition, experience, and an understanding of how people and organizations behave—not from a verifiable causal chain. This kind of knowledge is a textbook case of Metis.

AI, being grounded in probability and statistics, can only work within “distributions it has seen,” yet the problems at this layer do not obey a stable distribution in the first place—so naturally it cannot truly bear the responsibility of choice.

2. Architecture

The product’s overall design requires designing the technical approach for each module under constraints like budget, requirements, and time, while continuously managing risk along the way.

At this layer, Metis and Techne begin to be deeply intertwined. On one hand, at the product and business level:

  • Sensing real user needs that have not been clearly articulated
  • Anticipating compliance risks, platform boundaries, and policy gray zones
  • Inferring investors’ expectations and management’s real concerns

These judgments rely heavily on tacit knowledge that cannot be serialized—they fall within the domain of Metis.

On the other hand, at the technical level:

  • Technology selection
  • Architectural layering
  • Scalability and performance design
  • Common bottlenecks and best practices

This material is already highly structured, and AI can often produce a “seemingly reasonable” solution—it falls within the domain of Techne.

But in reality, architectural decisions are rarely purely technical problems. For example, a single misjudgment of QPS could lead to a surge of 10x or even 100x the traffic right after launch. At that moment, whether to “temporarily throttle for stability” or to “tough out the traffic and bet on growth” is not a question technology can answer on its own—it involves a comprehensive trade-off among business goals, resource costs, and the allocation of organizational responsibility.

Therefore, the essence of this layer is:

  • Using Techne to generate viable solutions
  • Using Metis to decide trade-offs, boundaries, and the degree of risk exposure

This also constitutes AI’s current compression boundary line: AI can assist with architecture, but it can hardly independently bear the consequences of architectural decisions.

3. Execution

The concrete implementation details of each module—for example, implementing a specific requirement: a screen that pushes a delivery rider’s location in real time, which may require writing a large amount of code, including SSE / WebSocket push, location smoothing, time estimation, exception handling, and so on.

The common features of this layer are:

  • The requirement boundaries are relatively clear
  • Inputs and outputs can be explicitly verified
  • There are plenty of mature paradigms and historical implementations

This knowledge is highly serializable and very well suited for an LLM to learn and reuse. As a result, this layer is being rapidly eroded by AI.

4. Reflex

The concrete implementation of every step within a requirement—for example, each interface of CRUD, form validation, and shuttling data around.

The characteristics of this layer are:

  • Highly repetitive
  • Requiring almost no contextual understanding
  • Limited blast radius for errors

It is closer to a “motor reflex” than to “judgment,” and has already been replaced en masse by scripts and templates. Humans hold no obvious advantage at this layer—and are at a disadvantage in stability and consistency.

Antifragility

Borrowing the concept Taleb introduces in Antifragile, Metis is inherently “antifragile.” Unlike “fragile” or merely “robust,” an antifragile system does not try to eliminate volatility and uncertainty—it benefits from volatility. Errors, shocks, and chance do not simply cause loss; they serve as informational inputs that drive the cognitive model to continually correct and evolve.

The path by which Metis forms is clear:

  • It is not acquired through one-time learning
  • Rather, it accumulates gradually through a process of repeatedly making mistakes, adjusting, and making mistakes again in real environments
  • These experiences are often situational and tacit, hard to abstract into general rules

Metis cannot be fully recorded, copied, or transferred; it can only exist in a specific person, a specific body, and a specific history. The key to improving Metis lies not in mastering more explicit knowledge, but in keeping oneself, over the long term, in an environment of high uncertainty and high feedback density: where decisions carry real costs, where errors cannot be easily covered up, and where the environment relentlessly forces the cognitive structure to change. It is precisely under this continuous exposure that Metis can grow—and that is why it exhibits antifragile properties.

AI’s compression of Techne will not diminish the value of Metis; on the contrary, it will amplify the difference between the two: the more automated and the more standardized a domain is, the more easily it gets replaced; whereas the abilities that depend on situational judgment, trial-and-error evolution, and the accumulation of experience will instead become even scarcer in an era of intensifying uncertainty.

  • Essays