I’ve been noticing a weird split in my own work.

Having a thought can feel like holding something solid. Writing the thought up, even as a scrappy paragraph, often reveals it was closer to fog. The strange part is that both states feel like “an idea.” One is private and untested. The other is public, even if nobody reads it.

Here’s the practical tease I keep coming back to. AI can make the “write it up” part cheaper, but only if I use it like a compiler, not a ghostwriter. The point isn’t nicer sentences. The point is finding where the thought fails, then either fixing it or throwing it away.

In practice, that looks like a loop I can run in ten minutes. I write the smallest honest version of the idea, then I ask the model to do four things: restate my claim in plain language, list the assumptions it depends on, give the strongest counterargument, and propose one concrete example that would make the claim testable. None of that output is truth. It’s pressure.

That pressure-loop is the writing version of the discovery loop in AI Didn’t Automate the Grind, It’s Doing Something More Interesting.

Pressure is what makes the thought become real.

Why is there such a gap between thinking and writing in the first place?

In my head, a thought is compressed. It shows up as an image, a vibe, a conclusion with a few supporting moments attached. It’s like looking at a photo thumbnail and feeling sure it’s sharp. Writing is the full-resolution export. Suddenly blur appears. The background is distracting. The subject isn’t where I thought it was. The composition is wrong.

That’s not a character flaw. It’s a feature of how a mind can store things. A thought can get away with missing steps because the mind quietly supplies them. The missing steps stay invisible because familiarity patches the gaps. Once the idea becomes sentences, the gaps stop being private. Claims need boundaries. Nouns want definitions. Verbs force sequence.

Code has a name for this. Compilation errors. Something “makes sense” until a strict system asks for exactness. Then the missing imports show up.

What makes a thought valid, anyway?

Not “morally correct” or “socially acceptable.” Valid as in “I can hand it to another mind without it collapsing.” If it only exists as a feeling I can’t reliably translate, it’s not wrong, but it isn’t finished. It’s closer to a melody I can hum than a song I can play. Writing is notation. It forces the melody to choose notes.

This is where the usual advice about writing flips. Writing isn’t only packaging. For the kind of work I care about, writing is part of the thinking. Paul Graham makes this point in a way I keep returning to. Trying to say something clearly often turns “I believe X” into “I believe X because Y, unless Z,” which is a different thought entirely.¹ Joan Didion put it even more plainly: “I write entirely to find out what I’m thinking.”²

That’s one half of the split.

The other half is social, even when nobody else is involved. A thought in my head doesn’t have to survive friction. A written thought does, because I can reread it. The words don’t get to borrow confidence from the feeling that produced them. They have to earn it.

So why does AI change this?

Not because it’s smarter, or because it has “better ideas.” What changes is the cost of externalization. For me, the bottleneck often isn’t insight, it’s the labor of converting a half-formed insight into language that can be tested.

There’s evidence that generative AI can speed up writing tasks and improve judged quality, at least in controlled settings.³ One widely cited experiment found people finished writing tasks faster (around 40% faster) while independent raters scored the output higher (around 18%).³ That matches what I’ve felt. The blank page becomes less sticky when a machine can draft, rephrase, or propose structure on demand.

But here’s the trap. Fluent text feels like finished thinking.

This is where it gets dangerous. Language models are optimized for plausible continuation. Plausible isn’t the same as true. Even the people building these systems emphasize that they can confidently produce errors, including made-up details that sound right.⁴

So AI can help me write something that sounds more valid than it is.

If writing is the compiler, AI is the autocomplete. Autocomplete is great until it inserts the wrong function call and the program still runs. Worse, it runs and produces output that looks reasonable.

Which brings me to the real question. It isn’t “Can AI write my post?” It’s “Can AI help my idea survive writing without letting me skip the hard parts?”

My sense is yes, but the use case is narrower than people assume. The move is to treat AI as scaffolding for the ugly middle, the part where a thought transitions from fog to form. Scaffolding is temporary. It’s not the building.

Fair enough, writing isn’t the only way to test a thought, and it’s not always the best one. Some ideas get clearer through conversation, through building a prototype, through teaching, through shipping and watching what breaks. Writing can also distort. It can overfit to what sounds elegant instead of what’s true. A model can amplify that bias by rewarding the most “post-shaped” version of the idea.

That’s why I keep coming back to the loop. The loop isn’t about polish. It’s about forcing contact with reality.

Once the weak joints show up, the options are simple. Strengthen them with evidence, narrow the claim, or delete it. This is the part that still takes taste and honesty. AI can generate ten counterarguments, but it can’t tell which one is load-bearing. It can propose examples, but it can’t know which example is true without checking.

That’s also where the “validity” question stops being philosophical.

A thought becomes more valid when it gains handles. When it can be summarized without changing meaning. When it can be criticized without falling apart. When it predicts something. When it survives contact with a concrete example. When it logically adheres to reality. Writing is a way of manufacturing those handles.

AI changes the economics of this in a subtle way. It doesn’t only make writing faster, it makes iteration cheaper. It becomes easy to generate five framings and pick the one that exposes the real claim. It becomes easy to ask for the steelman and notice the original version was a strawman, without realizing it.

Cheap iteration only matters if it still creates signal, which is the same discipline I unpack in Act Like It’s Impossible to Fail.

But the loop only works if I keep one rule. Facts don’t get to be optional.

If I’m writing to understand, I’m allowed to be wrong in the draft. I’m not allowed to be lazy in the final. Once the idea has form, it has to touch reality. Citations can help, but the deeper thing is posture. Being willing to shrink a claim until it’s actually true, even if it’s less punchy.

This is why I like thinking about the difference between a thought and a writeup as the difference between intuition and an artifact.

Intuition is fast and private. An artifact is slower, but it’s shareable. It’s the difference between seeing a composition in the viewfinder and producing a print. The print has limits. It also has evidence.

AI is a tool for producing more prints.

That’s exciting, but it also forces a choice. The tool can help publish more, or it can help test more. Publishing is tempting because it comes with dopamine. Testing is quieter. Testing is where the idea gets sharpened or killed.

The best outcome isn’t “AI makes me a faster writer.” It’s “AI makes me more honest about what I actually believe,” because it lowers the effort required to confront the messy parts in public language.

So maybe the reframe is this. A thought isn’t valid because it feels right. It becomes more valid when it survives articulation, criticism, and reality checks. Writing makes those tests possible. AI can help, but only if it’s pointed at the tests, not at the applause.

The question I want to keep carrying is simple. What thoughts are still stuck in my head because the cost of writing them down feels too high, and what happens if that cost drops enough to run the loop every day?

The practical move is small. Pick one thought a day, write three honest sentences, run the pressure loop, and keep only the version that survives. After a week, the pattern is hard to miss. Some “ideas” were always fog, and a few turn into something that can actually hold weight.

Rabbit Hole

  1. Paul Graham, “Putting Ideas into Words” (February 2022) and “Write Like You Talk” (October 2015). Both essays argue that forcing ideas into simple language exposes missing steps and weak reasoning.
  2. Joan Didion, “Why I Write,” The New York Times Magazine (December 5, 1976). The essay frames writing as a method of discovery, not just expression.
  3. Shakked Noy and Whitney Zhang, “Experimental evidence on the productivity effects of generative artificial intelligence” (Science, 2023). A widely cited experiment showing time savings and quality gains for writing tasks in a controlled setting.
  4. OpenAI, “GPT-4” (March 14, 2023), “Limitations” section. Discusses model reliability issues, including hallucinations (confidently wrong outputs).