The Boundaries of Compression

Some knowledge can be serialized, while others cannot. Take a dish of tomato scrambled eggs, for example:

If you want to make a dish of tomato scrambled eggs and need an AI to give you a recipe, you’ll receive an extremely detailed recipe, with time precise to the second and ingredients measured to the gram.

However, from another perspective: if you’re hosting guests at home and have prepared a dish of tomato scrambled eggs, AI cannot tell you if you’ve added too much or too little salt, because it doesn’t know your guests’ tastes or health conditions, nor does it know exactly how much salt is in the dish.

In Nassim Nicholas Taleb’s book “Antifragile,” there are two concepts: Metis and Techne.

• Techne (Technical Knowledge / Serializable): This can be written into manuals, formulas, or algorithms and is highly compressible. As long as you grasp the underlying logic (formulas or scripts), you can reproduce the results infinitely.
• Metis (Practical Wisdom / Non-Serializable): This represents a kind of shrewd, flexible, and hard-to-express practical wisdom that is non-compressible. You cannot convey an old Chinese doctor’s “feel” to others with a string of characters.

AI and Software Development

Recently, Claude Code has been popular, and many classic statements have resurfaced, such as “Programmers will be out of jobs; AI can do a week’s work in half an hour,” and “Programmers will be more in demand because AI-written architectures are unmaintainable.”

Both sides have valid points, and here I want to analyze this classic topic from the perspective of Metis and Techne. In reality, Metis and Techne are not binary but exist on a continuous spectrum. Taking software product development as an example, it can be broken down into these layers:

1. Choice

Which product should be developed? Social software or food delivery software?

This consideration includes user needs, policy trends, platform attitudes, capital cycles, competitive landscape, and other vague and unquantifiable variables. Market data is always lagging, and many key pieces of information may not appear in public data.

The core here is not “analytical ability,” but the ability to make decisions amid incomplete and even contradictory information. The specific environment of a choice has never appeared before and is unlikely to be fully replicated in the future. Many key judgments come from intuition, experience, and understanding of human and organizational behavior, rather than verifiable causal chains. This knowledge is a typical example of Metis.

AI, based on probability and statistics, can only operate within “known distributions,” but the problem at this layer does not conform to a stable distribution and naturally cannot truly bear the responsibility of choice.

2. Architecture

The overall design of the product requires designing the technical route for each module based on constraints such as budget, requirements, and time, while continuously managing risks.

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

• Sensing real but unarticulated user needs
• Anticipating compliance risks, platform boundaries, and policy gray areas
• Gauging investors’ expectations and management’s real concerns

These judgments rely heavily on non-serializable implicit knowledge, falling under the realm of Metis.

On the other hand, at the technical level:

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

These aspects are highly structured, and AI can often provide a “seemingly reasonable” solution, falling under the domain of Techne.

However, real-world architectural decisions are rarely purely technical issues. For example, a misjudgment of QPS (queries per second) might lead to 10 or even 100 times the traffic after going live. At that point, deciding whether to “temporarily limit traffic to maintain stability” or “withstand the traffic and bet on growth” is not a question that technology alone can answer. It involves a comprehensive trade-off of business goals, resource costs, and organizational responsibility distribution.

Therefore, the essence of this layer is:

• Using Techne to generate feasible solutions
• Using Metis to decide trade-offs, boundaries, and risk exposure levels

This also constitutes the current compression boundary line of AI capabilities: AI can assist with architecture but finds it difficult to independently bear the consequences of architectural decisions.

3. Execution

The specific implementation details of each module, such as implementing a specific requirement: a real-time interface for pushing the delivery person’s location, which may require writing a lot of code, including SSE / WebSocket push, location smoothing, time estimation, exception handling, etc.

The common characteristics of this layer are:

• Relatively clear requirement boundaries
• Inputs and outputs can be clearly verified
• A wealth of mature paradigms and historical implementations exist

This knowledge is highly serializable, making it very suitable for learning and reuse by large language models (LLMs). Therefore, this layer is being rapidly encroached upon by AI.

4. Reaction

The specific implementation of each step in the requirements, such as each CRUD interface, form validation, and data handling.

The characteristics of this layer are:

• Highly repetitive
• Almost no need for contextual understanding
• Limited impact of errors

It is closer to “action reflex” rather than “judgment” and has already been largely replaced by scripts and templates. Human advantages in this layer are not significant, and in terms of stability and consistency, they are even at a disadvantage.

Antifragility

Taleb believes that Metis is important because it naturally possesses the characteristic of “antifragility.” Unlike being “fragile” or “merely resilient,” antifragile systems do not attempt to eliminate volatility and uncertainty but instead benefit from them. Mistakes, shocks, and randomness do not simply result in losses but serve as informational inputs that prompt continuous revision and evolution of cognitive models.

The formation path of Metis is clear:

• It is not acquired through one-time learning
• It is gradually accumulated through repeatedly making mistakes, adjusting, and making mistakes again in real environments
• These experiences are often contextual, implicit, and difficult to abstract into general rules

Metis cannot be fully recorded, copied, or transferred; it can only exist in specific people, specific bodies, and specific histories. The key to enhancing Metis lies not in mastering more explicit knowledge but in keeping oneself in environments with high uncertainty and high feedback density over the long term: decisions must have real costs, mistakes cannot be easily covered up, and the environment constantly forces changes in cognitive structures. It is in this continuous exposure that Metis can grow, thus exhibiting antifragile characteristics.

AI’s compression of Techne does not diminish the value of Metis but instead amplifies the difference between the two: the more automated and standardized a field is, the easier it is to be replaced; whereas abilities that rely on contextual judgment, trial-and-error evolution, and experience accumulation will become more scarce in an era of increasing uncertainty.