A Better Mental Model for Understanding AI

Most of the confusion about large language models comes from a bad metaphor. People reach for “a brain,” or “a search engine,” and both quietly mislead. The brain metaphor makes you trust its judgment and feel betrayed when it fails. The search-engine metaphor makes you expect perfect recall and hard facts, and leaves you baffled when it invents something out of thin air. Both models will disappoint you constantly, because both are wrong.

Here is the model I keep coming back to. Think of it as a very fast, very literal intern who has read everything and only knows what is on the desk in front of it. The intern itself is an amnesiac; the products wrapped around it increasingly are not, because they take notes between your visits and quietly set the desk before the intern sits down. But the desk is still the whole world. Once you hold that picture, the failure modes stop being surprising and start being obvious.

“The context window is its whole world. Everything you want it to know, you have to put on the desk in front of it.”

Three consequences follow, and each one is worth taking seriously:

1. Retrieval beats memory. Do not expect it to “know” your project. Hand it the relevant files, every time.
2. Instructions are load-bearing. A literal intern does exactly what the note on the desk says, including the ambiguous parts.
3. Verification is your job. The intern is confident by design. Confidence is not a signal of correctness. It is a formatting choice.

Let me walk each one, because this is where the metaphor earns its keep.

Retrieval beats memory

The intern has read an astonishing amount, but the moment a task is done, it forgets you were ever there. There is no running memory of your project, your preferences, or the thing you carefully explained an hour ago. Each time you come back, you are talking to someone brilliant and well read who has never met you.

What the intern does have is a desk: a working surface where you lay out everything it needs for the task in front of it. In the tools, that desk is the context window, and it is finite. Whatever you put on the desk, the intern can use. Whatever you leave off, it simply does not know, and it will cheerfully proceed without it rather than stop and ask.

// give the intern the desk it needs
const reply = await model.chat({
  context: loadRelevantDocs(query),
  task: "summarize, then flag risks"
});

The practical upshot is that vague, lazy prompts fail not because the model is dim but because you handed the intern an empty desk and expected magic. If a task needs the brief, the last three emails, and your style guide, then those things have to be on the desk every single time. This is also why the same model can look like a genius one minute and hopeless the next. It is not moody. You changed what was on the desk.

A quick note on the memory features the major products now ship, because they look like a counterexample and are not. When your assistant remembers your dog’s name across sessions, the intern has not grown a memory. The product has hired a second assistant whose entire job is taking notes about you and laying them on the desk before each task. That is genuinely useful, and it changes who assembles the desk, but it does not change the rule. It also introduces a new way to fail: the notes can be stale, wrong, or about the other project, and the intern will trust them exactly as much as anything else on the desk. So the skill has not gone away, it has shifted. Less assembling the desk by hand, more auditing what the note-taker put on it.

Instructions are load-bearing

A green, eager intern does exactly what you say, not what you meant. If your instructions have a gap, the intern fills it with the most statistically ordinary guess, not with your intent, because it cannot read your mind and it is far too eager to pause and check.

This is why “make it better” produces nothing useful and “rewrite this in shorter sentences, cut the jargon, keep it under 200 words” produces something you can actually use. Ambiguity is your problem to remove, not the intern’s to resolve. The most valuable skill in working with these tools turns out to be an old one with new urgency: the ability to specify. Say precisely what you want. Show one example of good and one of bad. Name the constraints out loud. People who are good at delegating to junior humans tend to be good at this immediately, because it is the same muscle. You learned long ago that a fuzzy handoff to an eager junior comes back wrong, so you learned to hand off clearly. Same thing here, minus the coffee run.

Verification is your job

This is the consequence that gets people hurt, so I will say it as plainly as I can. The intern is eager to please and constitutionally unwilling to say “I don’t know” unless you explicitly give it permission. So it produces an answer, always, and it formats that answer in the same confident, well organized, authoritative voice whether it is dead right or entirely fabricated.

Confidence, from these systems, is a style, not a signal. It is not correlated with accuracy the way it roughly is with a knowledgeable human. A person who is unsure usually sounds unsure, and you can read the hedging. The intern’s hedging has to be requested. Left alone it will hand you a fabricated statistic and a real one in identical fonts, and nothing in the delivery will tell you which is which. So you never read fluency as correctness. You build a verification step for anything that matters, exactly as you would double-check the work of a bright intern in their first week. Not because they are stupid, but because they are new, literal, and trying a little too hard to look competent.

Where the metaphor breaks

The intern model is a tool, not the truth, and it is worth knowing where it breaks. A real intern genuinely understands things and grows over time. This one does neither. It pattern-matches without comprehension, and although the products around it now keep notes between visits, the intern itself is not improving from working with you. It is being handed better notes. And no intern in history has read the whole internet or drafted a page in four seconds, so the breadth and speed are genuinely superhuman, and no human metaphor fully captures them. Do not take the intern too literally either. It is a lens for calibrating your expectations, not a claim about what is happening inside.

But as a lens, it is the most useful one I have found. Adopt it and the whole relationship changes. You stop being disappointed that it does not remember you, and you start building the desk. You stop being surprised when it takes you literally, and you start specifying. You stop trusting the confident tone, and you start verifying.

A quick diagnostic

When one of these tools disappoints you, run three questions before you blame the model. Did I give it the desk? Nine times out of ten the answer is no: I assumed it remembered something it never had, or I left the one document that mattered off the table. Was I specific? The next most common failure is a fuzzy instruction that let the literal intern fill the gap with an ordinary guess instead of my intent. And did I verify? If a confident answer turned out wrong, the tool did exactly what it was built to do, and catching the wrong slice was my job, not its.

Almost every bad experience I have with these systems traces back to one of those three, and not one of the three is the model being stupid. They are the three predictable consequences of amnesia, literalism, and confidence-as-style, which is precisely what the intern model told you to expect going in. The diagnostic is useful because it converts a vague “this thing is unreliable” into a specific, fixable mistake on your side of the desk. That reframing is the whole point. With these tools, reliability is mostly something you supply, and the fastest way to get more of it is to stop asking the intern to be different and start running a better desk.

None of this makes the technology less useful. It makes it predictable, which is what you actually need to build on top of it. The people who get enormous value out of these tools and the people who bounce off them in frustration are usually running the same software. They are just running different mental models, and the model you pick is the whole difference between a brilliant, tireless intern and a liar who will not stop talking.

Updated July 2026. The original version of this post said the intern remembers nothing and never will, which was true of the model and already becoming false of the products. I have revised the memory sections accordingly, and yes, I noticed which post the overconfident claims appeared in.