Heard It, Seen It, Done It
“ChatGPT writes my legal documents for me,” said my lawyer friend. I pushed back: “But you’re not learning the skill. Take the AI away and you can’t do it yourself.” The friend said “So what? Now I can be 5x more productive”. I nodded, but walked away feeling right.
Last week I was building a new feature in a TypeScript app. Compiler gave me an error. I just told my coding agent “go fix it”. I have no idea what went wrong. I don’t know what was the fix. Yet it worked, it got the job done. I was so much more productive. Faster. But I realized that I became my lawyer friend. Past me was right about the skill, but present me ignored him anyway.
Knowledge comes in 3 tiers:
- “Heard about it”
- “Seen it”
- “Done it”
You heard that Cirque du Soleil exists -> You have seen a performance -> You performed in one. Each level is profoundly different from the previous one. Level 1 is just general broad knowledge. Level 2 is familiarity, you’ve seen it in action and you know what is going on. Level 3 is the expert. You have performed, you have fallen, got scars. Same with engineering. You start with hearing that Kubernetes exists, then you see a senior dev building a cluster, maybe even see it in prod. Then you build one yourself, get paged at 3 am because a dependency failed and your cluster went haywire.
AI is the best compression engine for level 1 -> level 2 transition. It will explain to you what Kubernetes is, how it works, and it will even build a system with 25 microservices and will run it on your laptop. You’ll have real implementations of the patterns you only read about in Martin Fowler’s books. But the same AI power is a heavy anchor for level 2 -> level 3 transition. It creates fake level 3. It may build your microservices app and run it for you. But same as me and TypeScript, you did not struggle while building it, you did not dig deep to figure out how pieces fit together. Struggle is the encoding mechanism for knowledge. Nowadays GitHub is full of new shiny repos. People’s resumes proudly show that they have built this and that. But when I ask during an interview “what happens when a dependency returns 200 and garbage payload”, I only get blank stares. Artifacts showing the experience without the actual experience. Rented knowledge.
The problem is not that devs or lawyers suddenly got lazy. Homo Sapiens just did what they do best - adapted. Adapted to the same old incentive - speed. Hundreds of startups work on the same ideas at the same time. But only 2 or 3 get the market share. Because they were first to claim the market. Speed matters. No manager in a tech company ever got promoted based on the depth of knowledge in his team. But how many features you ship this quarter will decide if you drive a Honda or a Maserati next year. AI is like strapping a big-ass turbo that gives a boost to whatever you are doing. Speed is the name of the game. “Go fix it” is the rational choice. Everyone renting knowledge is exactly what the system rewards.
So what? No one writes assembly anymore. True. Compilers ate assembly for breakfast, cloud made the server room obsolete. And we’re doing fine. We have always gone one abstraction layer up. It has always been the graybeards who have been mourning the lost layer. And every time they were wrong. Maybe Rust is the next assembly. Chat will become the next standard interface into the computer for programmers instead of a programming language. Maybe me feeling uneasy about it is just me becoming that graybeard.
All previous abstractions that vanished had two things in common. First - the knowledge did not vanish - it got absorbed into a deterministic tool. Compiler devs still exist, the compiler does the same thing every single time. The knowledge was concentrated, then sealed into an artifact anyone can trust and inspect. Second - it took time. We stopped reading compiler output after we have seen it many times and started trusting it. AI breaks both - it is a probabilistic tool that generates whatever each time and we stopped reading its output before we started trusting it. And this time the knowledge layer below did not get sealed into a deterministic artifact - now it is encoded into probabilistic models owned by a handful of labs.
The level 3 knowledge did not go away. It just got centralized in a bunch of labs and weirdos who decide to learn things themselves anyway. Everyone else just keeps shipping. Until the tool goes away. Take AI away from programmers now and we cannot do fucking anything. Elon promises post-knowledge abundance. Abundance for whoever holds the scars. Atrophy for the rest. Owned knowledge is becoming scarce, and someone else owns it.
Two years later, I finally have an answer for my friend. Rent whatever gets the job done — speed wins, it always will. But pick a few things to own. The slow way. Scars and all. Because if you own nothing, you’re not using AI. You’re just a meat proxy between the model and the market.