Most AI talk frames it like a tireless intern. It writes the email, summarizes the meeting, cleans up the spreadsheet. The promise is efficiency.

But what feels different in practice isn’t just “work got automated.” It’s that the distance between a question and a working artifact got weirdly short. Curiosity used to be a spark that died somewhere between “that’s interesting” and “do I really want to set up all the stuff to find out?” Now it often turns into a tiny prototype before the spark cools.

That’s not only productivity. It changes what attention can attach to.

Here’s the move I keep coming back to: take one question and force it into one small artifact, fast. Not a portfolio piece. Just something that exists outside my head. When that gets cheap, the whole shape of a day changes.

When the loop gets shorter, the main skill stops being “how fast can I execute?” and becomes “what do I run loops on?” If I don’t pick that consciously, the loop just runs me.

This matters if my brain doesn’t want a single track. A lot of advice assumes a specialist default: pick a lane, block distractions, repeat. That advice isn’t wrong, it just isn’t complete for everyone. Some people are built more like roaming radios than laser pointers. Barbara Sher called them “Scanners,” people who get intensely interested in many things and then move on when the learning curve flattens.¹ That pattern is usually treated as a character flaw, or at best an indulgence.

AI changes what that pattern can be.

The standard view says AI removes toil. True. But the deeper change is that it removes friction between curiosity and execution. Friction used to be a gate. Not a noble gate, usually. More like setup steps, missing context, syntax lookup, or the small shame of not knowing the right words to ask. Each one is minor. Together they turn a curious thought into “later.”

Which brings us to what builders actually describe. “Automation” isn’t always the word. It’s immediacy. The ability to ask, “What would it look like if…?” and get a plausible starting point fast enough to keep the thread alive.

In code, that immediacy is measurable. A GitHub study on Copilot found developers completed a programming task substantially faster with the tool than without it.² The point isn’t the percentage. The point is the behavior it unlocks. When the cost of “try it” drops, people try more. They branch sooner. They iterate before doubt hardens into delay.

That pattern shows up outside software too. There’s economics research arguing that when experimentation gets cheaper, it changes what gets pursued and funded, not just how quickly work gets done.³ I don’t need the model details to steal the principle: cheaper trials make exploration feel less like a guilty hobby and more like the rational default.

So, AI isn’t mainly a labor-saving device. It’s a curiosity amplifier. And amplifiers don’t care what they amplify.

This is where it gets interesting, because friction isn’t always bad. Some friction is accidental, like a squeaky door that makes it annoying to go outside. But some friction is protective, like a seatbelt. It forces a pause. It makes a choice feel expensive enough to consider.

Paul Graham has an essay arguing startups should do the “unscalable” work early, partly because it teaches what customers actually need before a team automates the wrong thing.⁴ That’s deliberate friction. It’s not there to slow progress. It’s there to keep the feedback loop honest.

AI tends to remove friction indiscriminately. That’s why it can feel like rocket fuel for Scanner brains. The old failure mode was the gap between interest and execution. The interest was real, the execution was heavy, the project died in the middle. When the middle gets lighter, the Scanner pattern starts to look less like flakiness and more like a search strategy.

That is the longer case I make in Why I Work Like I'm Slacking, where exploration stops looking like avoidance and starts looking like preparation.

But then a new failure mode appears: infinite beginnings.

It’s easy to become a tourist in my own curiosities. Prompt, generate, nod, move on. The loop is so short it stops producing consequences. It becomes a slot machine for plausible outputs.

AI can also be wrong in a confident voice. Cheap artifacts can speed up self-deception. If the thing I’m “building” never has to touch reality, I can iterate my way into a really polished mistake.

The loop needs the world to push back.

Not a big ordeal. A small one. A tiny script that either runs or doesn’t. A mockup shown to one person who can say “I don’t get it.” A photo where the light is actually wrong. A rough melody that either holds together when I play it twice, or collapses.

Photography is a useful analogy because it makes the feedback loop visible. When feedback is delayed, learning is slower and more expensive. When feedback is immediate, it becomes normal to take a shot, look, adjust, and try again. The skill doesn’t disappear. The loop tightens.

AI tightens the loop for making things. It collapses the delay between “I wonder” and “I can see it,” often enough to keep improvisation alive.

Which means the new leverage is taste. When outputs are cheap, selection becomes expensive.

If an AI can generate ten approaches in thirty seconds, the bottleneck moves: deciding what’s worth pursuing, what’s elegant, what’s honest, what actually solves the problem. That bottleneck is mostly not automatable in the way people mean when they say “automate.” It lives in judgment, in values, in the quiet sense that something is off.

This is where Scanner range can become an advantage. Sampling widely builds a bigger internal library of patterns. The catch is that range without reduction becomes noise.

So I keep returning to a small question: what turns a curiosity loop into a discovery loop?

The answer I’ve found is that the loop needs an artifact that can disappoint me. That sounds negative, but it’s the point. Disappointment is information. It shows where the idea breaks.

There’s also a time element. Curiosity has a half-life. If I have to schedule it, it often dies. This is why the release of ChatGPT on November 30, 2022 felt like a cultural moment, not because everyone suddenly wanted better essays, but because a lot of people got a new way to keep a question alive long enough to turn it into something.⁵

So what do I do with this, in practice, without turning into a machine that starts everything and finishes nothing?

The best constraint I’ve found is to treat AI like a lab bench, not a factory line. A lab bench is where experiments happen quickly, but they still have to be named, run, and logged. The habit looks like this: capture the question as a single sentence, pick the smallest outside signal that’s clear, build the smallest artifact that creates real feedback, then write down what surprised me before starting another loop.

The writing version of that lab-bench loop is AI Writing Companion as Leverage.

When the loop produces surprises, it’s discovery. When it produces only more possibilities, it’s entertainment.

Maybe that’s the real promise: not less work, but more honest work. Less time negotiating with tools, more time negotiating with reality. The grind wasn’t always the enemy. Sometimes it was just the door fee for getting feedback.

AI is lowering the door fee.

The question is which door I walk through.

Rabbit Hole

  1. Barbara Sher, Refuse to Choose! (2006). Sher popularized “Scanner” as a name for the “many interests” pattern, which matters here because the essay treats it as a legitimate search strategy once execution costs fall.
  2. GitHub Next, “Research: quantifying GitHub Copilot’s impact on developer productivity and happiness” (2022). I’m using this as evidence that AI assistance can shorten iteration cycles in at least one domain, not as proof that outcomes are always better.
  3. Michael Ewens and David N. Hémous, “The Cost of Experimentation and the Evolution of Venture Capital” (NBER Working Paper No. 24885, 2018). This supports the broader principle that lower experimentation costs change what gets pursued, not just how quickly tasks get executed.
  4. Paul Graham, “Do Things that Don’t Scale” (2013). This is the counterweight: some friction is educational and keeps feedback honest, so removing friction blindly can be a trap.
  5. OpenAI, “Introducing ChatGPT” (Nov 30, 2022). This anchors the timeline for when a low-friction interface to strong language generation hit the mainstream, which is relevant to the essay’s “cultural moment” claim.