The Blank Slate, Part 4

"Find a restaurant near me."

It's the default search - whatever it is, near me. That's how most people are using AI, so it's where this starts.

Zero context, so the machine guesses your location and dumps random spots. Overwhelm and irrelevance. Level up - "American restaurant in Longmont" - still vague, still a mediocre list. Add timing and "top-rated" and you're getting there, but you'll still filter everything manually, because the machine still doesn't know what you're actually picturing.

Now watch the top of the ladder: "Suggest 3-5 top-rated burger spots in downtown Longmont under $35/person with an extensive bourbon selection and outdoor seating, this Friday evening, party of 6 - include ratings and why they stand out. Suggest parking within 500 feet, factor in any events between 6 and 10pm, especially live music. Suggest 3 after-party bars known for an evening-ending vibe. Cross-verify with recent posts. Check your work - and ask me 3 questions to clarify before making a suggestion."

Look at how that ends. The highest rung of asking isn't more adjectives. It's making the machine ask you the questions. You are now the LLM.

Hold that. We'll come back to it, because it's the whole post.

Which rung did you hire?

Part 3 built a ladder out of people - fan, enthusiast, hobbyist, professional. That restaurant search is the same ladder compressed into a single text box. And it exposes the question nobody asks when they open the app:

Not which rung are you. Which rung did you hire it to be?

Because here's the pattern, everywhere I look: people take the most capable machine ever put in their hands and hire it as their enthusiast. Fast, agreeable, confident, free. And we already priced that rung. Part 3 repriced it to zero - which matches the mental tax you paid for it. The free answer is fast because it's worthless - it cost you nothing to ask, it cost the machine nothing to guess, and it's wrong in ways neither of you paid attention to.

Here's what that hire looks like in the wild: search engine 2.0. Type a question, get an answer, close the tab. And honestly - as a faster Google, it's fine. It saves time. Nobody's mad at that. But if you use this thing for work, hear this part: it is not a search engine, and it is not an email polisher. A search engine finds things that already exist. This machine builds things that don't - and it builds them out of whatever you handed it, which, if you asked like a search, was nothing. Treat it like search and you get search-shaped results: confident, generic, vaguely right, specifically useless. And the email it "saves you time on" is the same trap wearing a collared shirt - you hired a suit and asked it to type.

The tool is a suit. It amplifies whoever's inside it. Put an operator in the suit and it lifts what they couldn't lift alone. Put nobody in it and you've got the world's most expensive machine doing donuts in the driveway. You can tell the difference in one sentence: the empty suit cites reality as a constraint. The operator treats reality as a draft with poor version control. [OPTIONAL ADD, your call: "— the scars are all in the history, somebody just has to go read them."]

So this part is the operator's manual. Not for the machine - it doesn't need one. For you.

Get it out of your head

One concept runs under everything else here, and it's a concept, not a prescription:

Everything you don't say, the machine invents.

You know things without speaking them. The budget you'd never exceed. The person you're actually writing for. The way "done" or "good" looks in your head. That knowledge is real, it's load-bearing - and it is completely invisible to a blank slate. Every unstated constraint gets replaced by a guess. The lazy restaurant search doesn't fail because the machine is dumb. It fails because you kept the entire dinner in your head and asked it to read your mind.

If you end up with a useless list of restaurants, it's annoying at most. Let me show you what it costs when the stakes go up.

The Internet and the LLMs Were Wrong

I've been working on an old Jag. Window channels specifically - the felt-lined tracks the glass rides in. The internet is exactly the mess you'd expect for a decades-old car: old-school guys say the replacement is one way, newer guides say something else. Bonus? Neither one is right. Pick either camp and you're doing the door twice - and this isn't a job you get to casually redo.

Here's the thing: the answer was never in the categories. It was in the millimeters. When I finally asked for the actual dimensions - width, thickness, the numbers - the whole argument dissolved. The felt channel is way thicker than what they installed new, on purpose: it's built to absorb decades of body-build variation. The old insert is thin and tight because it was made for a door that matched the drawing. The forums weren't wrong. They were answering for the same doors, at different times.

And the machine? It agreed with everybody. Of course it did. It has never done a window on a Jag - and the people who actually have? They never wrote it down (or posted it on the internet at least). So it trained on the argument, not the answer. Ask it cold and it doesn't amplify wisdom; it amplifies whatever the internet said loudest, with better grammar and more confidence. The internet's wrongness didn't get corrected by the LLM. It got a spokesperson - and a spokesperson's job is to close, not to counsel. A salesman sells you what you walked in asking for. Order two Big Macs, three fries, and a Diet Coke, and nobody behind that counter is going to ask about your heart. You get exactly what you ordered, served fast, with a smile - unless you make the counsel part of the order.

The deciding detail was in my head and my garage the entire time: the dimensions of my door. Nothing on the internet could pick my fork, because the fork wasn't on the internet.

And notice what had to happen: I reached for the dimensions. Nobody asked me for them. Left unprompted, the machine would've picked a camp, insisted it was correct, handed me that confident answer, same as the forums did - and I would've been confused when I did the work, asking more questions with the panel already off. "AI sucks."

The Inversion

So stop relying on yourself to remember what you forgot you know. Delegate the reaching. This is my favorite hack, and it's one sentence:

"Ask me 3 questions. One at a time. Use each answer to inform the next question."

That's it. That's the inversion - the top rung of the restaurant ladder, generalized. You flip the machine from answering to interrogating, and every question it asks marks something it was about to invent. Each answer narrows the next question. Three exchanges later, the thing in your head is on the table where the tool can actually use it.

And the number is a dial. The number of questions is a unit of risk. Dinner gets three. A door panel gets three. An architecture decision gets ten. The number is your bet on how much you haven't said out loud - how expensive it would be to let the machine guess. Low stakes, let it guess. High stakes, make it dig.

The catch

When you get good at this, it takes longer to get an answer.

The tool sold as a time-saver pays out to the people who spend time on it. The free rung is instant - that's search engine 2.0 again, confident vagueness on tap. The operator's rung costs you three questions, ten questions, the dimensions, the intent — it costs you getting it out of your head, and that was never going to be free. Precision, context, and intent win. Ask like you're explaining to someone who does not know what you mean - because you are.

The key

How do you know you've climbed? Same test the whole series has been running, compressed: you've earned the key when you can do three things with the work you're about to ship. Teach it : explain what you're doing to the blank slate, plainly, no jargon armor. Predict it : say where it'll break before anything breaks. Defend it : argue the strongest case against yourself before the mirror does. Explain, predict, defend, every pass. Teaching is the proof you're a mirror and not a tollbooth: if the only thing you can do with your knowledge is charge admission to it, you never had the key - you had the door.

Full circle

Part 1 opened with the sentence this whole series has been walking toward from the other side: the model was never the thing being tested. You are.

Four parts later it should land differently. The suit amplifies whoever's inside - so the only variable worth working on is the operator. Get it out of your head. Make the machine ask. Spend the questions the stakes deserve. And ship your thinking sooner, so the mirror can correct you sooner.

The blank slate was never grading the machine.