The Scissors Gap, Five Layers of Knowledge, and Why Time Can't Be Compressed
"The limits of my language mean the limits of my world."
— Ludwig Wittgenstein, Tractatus Logico-Philosophicus (5.6)
I've just completed Part I: Epistemology of a book-length project on working with AI. It comprises three chapters that I believe every knowledge worker — especially in tech — needs to grapple with right now.
Here's the essence.
Chapter 1: The Scissors Gap (剪刀差)
AI breaks a fundamental balance in knowledge work: the relationship between production and verification.
Before AI, production and verification were roughly coupled. A human writes code at roughly the same speed they can review it. An editor writes at roughly the same speed they edit.
AI shatters this.
- Production speed → approaches infinity. 24/7 operation, parallel agents, near-zero marginal cost. An AI can generate 2,000 lines of code in 5 minutes.
- Verification speed → barely changes. Verification requires judgment, experience, domain knowledge, and risk assessment. These still need humans. A human reviews that same 2,000 lines of code in about 5 hours.
The gap: ~57x.
This isn't theoretical. METR's 2024 study found that developers with AI felt 20% faster but actually completed 19% fewer correct tasks. Faros AI found that while AI increased commit frequency by 62%, PR review time increased by 91%. Google's DORA report showed AI widened the gap between high and low performers instead of closing it.
AI doesn't buy you a fast ticket to the future. AI gives you a 60x-leverage credit card — spend today, repay tomorrow.
The gap exists everywhere: content creation, data analysis, management decisions. Every domain where AI accelerates production but human judgment remains the bottleneck.
Chapter 2: The Five-Layer Framework (五层框架)
Why does this gap exist? Because AI and human knowledge live on fundamentally different layers.
Layer 1: Application Knowledge — "Knows the Answer"
- Concrete syntax, APIs, standard patterns, known solutions
- AI: Dominates. GPT-4/Claude hit 70%+ on SWE-bench. Any skill that's just "knowing how to use X" is fast-depreciating.
Layer 2: Systems Thinking — "Builds the Right Thing"
- Coupling/cohesion, abstraction boundaries, long-term marginal cost, technical debt
- AI: Can write code that looks correct, but can't understand how code rots over three years. It's seen git commits, but never lived the night after a release.
Layer 3: Meta-Domain Knowledge — "Knows What Good Questions Are"
- Calibrating uncertainty, knowing when to stop searching, designing verification loops. Judgment about judgment.
- AI: Can mimic the form of meta-knowledge. Cannot calibrate uncertainty. It doesn't know what it doesn't know — and it's structurally blind to this.
Layer 4: Meta-Cognitive Creation — "Creates New Frameworks"
- Making frameworks where none existed before. Newton unifying falling and orbiting. Einstein questioning simultaneity itself.
- AI: Can optimize within a given framework (AlphaGo Zero found new Go strategies), but has never created a framework ex nihilo. Four fundamental bottlenecks remain: framework awareness, original creation, credit assignment over long chains, and infinite regress of self-evaluation.
Layer 0: Embodied Grounding — "Lived Experience"
This is the foundation. Split into two sub-layers:
Layer 0b (Instrumental Embodiment): Having a body with sensors and actuators. AI can achieve this — robots learning from physical failure is real. Figure 02, Tesla Optimus, 1X's EVE are all doing it.
Layer 0a (Native Embodiment): Having lived a life. Born, loved, hurt, trusted, betrayed, learned to trust again. This is not a technology problem — it's an existential modality problem. AI doesn't feel code "wrong" in its gut. It doesn't solve problems in the shower. It doesn't know it will die, and therefore doesn't have to make real choices about priority.
The most dangerous thing is not that your work is in Layer 1. It's that you think it's in Layer 3, but six months from now, AI has penetrated it.
Chapter 3: The Limits of Compression (压缩极限)
If AI is climbing these layers, is anything truly incompressible?
Yes — three things, rooted in the fact that time itself carries information.
1. The Sediment of "Wasted" Time
90% of life is "unimportant" time — commuting, staring at ceilings, waiting in line. These moments create the narrative continuity that turns isolated events into a lived life. Creativity emerges from not processing information. The programmer solving the bug while running. The scientist having the insight in the shower. AI has no "offline" mode where it replays fragments without a goal.
Compression can give AI all the events. But it cannot give the blanks between events — because blanks are not events, and they cannot be compressed into events.
2. Long-Tail Failures Across Diverse Contexts
An expert programmer's intuition comes not from a few big failures, but from hundreds of small, never-recorded failures — each in a slightly different context. A null pointer in a single-threaded toy project teaches a different lesson than one in a multi-threaded financial system handling millions of users. The distribution can't be compressed into a single sample.
3. The Time Integral of Trust and Betrayal
Trust can't be accelerated. If someone says "I can make you trust me in 72 hours by simulating three years of collaboration," you don't actually trust them — because knowing the trust is accelerated dissolves the trust itself. Betrayal's weight isn't the information "they left" — it's the path you walked together.
The deepest moat you have: your experience isn't just accumulated in time. It is made of time. And time, so far, has no compression algorithm.
What This Means
These three chapters form the epistemological foundation for working with AI in the age of 60x leverage.
- The Scissors Gap — the structural bottleneck you must design around, not ignore
- The Five-Layer Framework — the map that shows where you stand and where AI is going
- The Limits of Compression — the things that may never be automated because their essence is time
AI is not coming to replace you. AI is coming to tell you where your old moat was — and where your next one needs to be.
This is Part I of a multi-part series. Part II (Strategic Theory) will answer: once you understand the map, where do you go?
The full PDF of Part I is available. If you're a knowledge worker wondering what's truly defensible about your craft in this moment, I'd love to hear your thoughts.
Tags: AI, Philosophy, Career, Software Engineering
Top comments (0)