You Can't Outsource Understanding
Andrej Karpathy's three-layer method, and why your AI keeps handing you shallow answers.
It is late, and the memo is due in the morning. You open a capable AI model, describe what you need, and read back three paragraphs that are fluent, well-organized, and quietly wrong. Not wrong in the facts. Wrong in the way that only you can see, because the model reached a confident conclusion without the one thing the whole memo turns on. You rewrite it yourself. You wonder, again, whether these tools are oversold.
They are not oversold. You are using them at the wrong layer. The gap you keep hitting has a precise shape, and once you see the shape, the fix is clear.
The test that explains everything
Andrej Karpathy, who led AI at Tesla and has spent more time than almost anyone watching these models behave, uses a small question to expose their limit. Ask any current model: I want to go to a car wash 50 meters away. Should I drive or walk? Every model tells you to walk. Fifty meters is close. They all miss that you cannot wash a car you did not bring.
That is the whole problem in one line. These models are brilliant at anything that can be measured and blind to anything that depends on context they were never given. They do not know what they do not know, so when the context is missing they do not stop. They produce a confident answer built on the half of the picture they can see.
Your late-night memo failed for the same reason. The model had the words. It did not have the decision the memo exists to drive, the politics in the room, the number that cannot move. It could not ask for those things because it did not know they were missing.
The line that matters
Here is how Karpathy frames what is left for us to do. You can outsource your thinking. You cannot outsource your understanding.
The distinction is the entire game. Thinking is labor: drafting, summarizing, restructuring, comparing, the work that fills the hours. Hand all of it to the machine. Understanding is different. It is knowing what the work is actually for, what a good outcome looks like, which detail is load-bearing and which is decoration. That cannot be delegated, because it lives in you, and the model has no way to reach it unless you put it there deliberately.
Most people try to delegate the understanding too. They type a one-line request and hope the model will infer the rest. It cannot, so it guesses, and the guess is the shallow answer you keep rewriting. The professionals who get real return from these tools are not better at prompting. They are better at moving their understanding into a form the model can use.
That move is not a trick. It is a small system, and it has three layers.
The three layers, briefly
Karpathy breaks effective AI work into three layers that stack.
The first is what you tell the model before it starts, how you get your understanding into it before it builds anything. Call it the specification. The second is how you check what comes back, against a standard you set rather than trusting it because it sounds sure. Call it the verifier. The third is the workspace you set up once and reuse, so your context, your rules, and your tools are waiting every time instead of being re-explained from scratch. Call it the environment.
Notice that one of these layers will not work for you the way it works for the engineers these tools were built for. When they check their work, they have a compiler: the code runs or it does not. You have none. Whether a memo is right, whether a position holds, is a judgment with no automatic test, and the context that settles it lives nowhere but in you. So for your work the checking cannot be handed off the way the writing can. It folds back into the same act as the thinking. That is not the tool failing you; it is the tool reaching the edge of what it can do, and the edge is your judgment. The guide follows this all the way down.
The full method, and how each layer maps onto real knowledge work rather than software, is the subject of a longer guide. Here I want to show you just the first layer, worked through on an ordinary problem, because it is the one that turns the shallow answer into a useful one, and you can start using it tonight.
Working the first layer: the specification
Go back to the memo. The instinctive prompt is "write me a board memo on the vendor decision." That is a task. It is not your understanding, and it leaves the model to invent the part that matters.
The specification move is to refuse to start there. Instead, you make the model draw the understanding out of you before it writes a word. In practice you tell it: interview me. Ask me what decision this memo has to drive, who reads it, what they already believe, what outcome I am steering toward. Then the model asks, and you answer, and with each answer a piece of what was only in your head becomes something it can build on.
Watch one exchange. Before it drafts a word, the model asks: what does this memo have to make the board do, approve the switch or pause it? And is there a number you are not willing to put in writing? You stop. That number is exactly what your first instinct would have leaked into paragraph two. One question about what you would not write down, and the model has already saved you from a mistake you could not see coming.
Three habits make this work, and none of them are technical:
First, find the goal, not the task. "A board memo" is a task. The goal is the decision the memo is meant to produce, and the model can never decide that for you. Naming it is your job, and the interview is how you get it out.
Second, work in small pieces. The temptation is to hand over the whole job and wait for a finished product. Resist it. Give the model a tight scope, look at what comes back, correct it, and go again. A wrong assumption caught in the first paragraph costs nothing. The same assumption caught in the finished memo costs you the morning.
Third, make it check with you. Tell the model to surface its key choices and confirm them before it commits. Every assumption it makes unprompted is a chance to drift from what you meant. Each one you catch early is a rewrite you never have to do.
This feels like the slow path. It is the opposite. The interview takes three or four minutes. The rewrite you are trying to avoid takes the rest of the night. You pay a small, known cost up front to kill a large, hidden one later.
Notice what has happened. You have not written the memo, and you have not handed off the understanding. You have done the one thing only you can do, decide what the work is for, and you have moved that decision into a form the machine can act on. The thinking is now safe to outsource, because the understanding is on the page.
Why this is a leadership skill, not a tooling skill
It is tempting to read all of this as a productivity tip. It is not. The professionals who pull ahead in the next few years will not be the ones with the best models. Everyone will have the same models. They will be the ones who keep ownership of their own understanding and stay disciplined about transferring it.
The model is the cheapest part of the system. Your judgment about what matters is the part that does not commoditize. That is the strange part: these tools are sold as a way to think less, and for serious work they reward you for understanding more.
The car wash question is funny until you notice you make a version of it every day, every time you accept a confident answer to a question the model did not fully understand. The fix is not a better prompt. It is the decision to stop outsourcing the one thing you cannot.
This is the first of the three layers. The other two, the verifier and the environment, are where the method stops being a technique you have to remember and becomes a workspace that remembers for you, one that stops making you re-explain yourself every time you sit down. The full method, mapped onto knowledge work rather than software, is in Context by Design. If the first layer fixed one memo tonight, the other two are what change how the work feels by next month.
Originally published at omarshraim.com