They Skipped The Local Streets

I had a conversation this week with someone early in their career. Still in college, actually.

The rules are about to reverse on them. In school, you get penalized if someone thinks you used AI to write the paper. Then you graduate, and in your first job you're judged on your output. If you don’t use AI, you're going up against someone who does.

I don't know how I'd navigate that if I were starting now. At this point, I'm less concerned with how I think about things and more concerned with how quickly I can get to a good outcome. That's where I see the value of AI. My experience lets me get better results out of these tools. At least, I believe it does.

Three months into my first job, I'm not sure I'd know when it was okay, or what good even looked like.

I wrote in The Career Rubric Disappeared about how the path was laid out through college used to be fairly well defined, and then wasn't. Nobody replaced the scoring system when it disappeared.

That was about not knowing where you stand. This is about not knowing whether the thing you just handed in is any good.

Which got me into a second conversation the same week, about the difference between expertise and experience.

Here's how I've been explaining it: I can put you in an empty parking lot and teach you to drive. Forward, reverse, three-point turns, figure eights. You'd get excellent. You'd know how to get the most out of that vehicle. That's expertise.

The parking lot is only where it starts.

What comes next is backing out of your own driveway. Then quiet local streets, where the worst thing that happens is you stop short and feel stupid. Then a road with a light at every block. Then the first time you merge onto a highway, which is genuinely frightening, and then it isn't. Somewhere in there you learn how far ahead to look, and that the guy drifting in your mirror is going to change lanes before he signals.

Nobody designed that sequence. It's just how it went. Each stage handed you one new variable while you still had room to recover from getting it wrong.

Then someone hands you the keys and says: get to California. From New York.

Now there is weather. Construction. Traffic.  A nearly empty tank and eighty miles until the next gas station. You didn't learn that in the parking lot, because the parking lot can't teach it. It has no traffic. You didn't learn it in one jump either. You got there in increments, at speeds where mistakes were survivable.

You can be excellent at driving the car and still not know how to get there

Some of this I covered in The Second Curve of Expertise, when an AWS engineer spun up in fifteen minutes a database configuration that used to take me weeks of capacity planning. What I'd thought of as expertise turned out to be context. The commands weren't the value. How I thought about systems under pressure was.

That one was about people who'd built something and then watched the ground move. What's happening now is the other end of it.

The people coming up don't get the local streets.

AI collapses the mechanics before they ever practice them. The summary arrives written. So does the analysis, and the deck. First assignment out of school, and what leaves their hands looks like something that used to take years to be able to produce. There's no quiet road in between. No stage where the stakes are low enough that getting it wrong is how you learn something.

So the question isn't whether they'll use AI. That ship has sailed. The question is where the road sense comes from when there are no local streets.

I'm Doing It Too

I skip the parking lot constantly.

I don't build the first version of an analysis anymore. I ask for it, read it, and start reacting. That's a real change in how I work, and it's made me faster in a way I like.

The reason I can get away with it is that I've already driven the route. When something comes back wrong, I notice. Not because I checked every line, but because it doesn't match what I know. That recognition came from roads that don't exist anymore, and there's no way to hand it to anyone.

That's why I didn't have a good answer for them. I'm not worried the tools produce bad output. Mostly they don't. I'm worried about the gap between the person who can tell and the person who can't, and how little the output reveals about which one you're dealing with.

Consulting the Oracle was about how convenience starts to feel like competence. That was about all of us. This is narrower. If you never developed the reflex, there's nothing to lose. There's a thing you were supposed to develop that no job description asks for and no assignment measures.

So Here's What I Actually Do

Thank you to Cheryl Kellond, who asked about my process. I've mentioned in The Inefficiency IS the Creative Process that Claude is a thinking partner for almost every issue, so this isn't a confession. It's the detail.

I start with an idea and write a full draft. Badly, usually. This one I dictated into Evernote using Wispr Flow, which I'm new to and still figuring out — I'm not a paid endorser of either. I just like being able to talk instead of getting stuck typing.

Evernote has been my repository for well over a decade. There's a lot in there.

Then it goes into Claude. I have a project set up with source documents: every issue of this newsletter, split into two files at the one-year mark; every blog post I've written, Medium and my own site; and every LinkedIn post I've ever made, exported. That last one is where Berkson's Bits comes from. Claude pulls five random posts from fifteen years and I pick something short and quippy.

The archive files are there so Claude can check whether I've written this before, and find callbacks to prior issues when I haven't.

I built the project deliberately. It's less a writing tool than a memory I can query

What I Don't Let It Do

Claude has invented callbacks.

Not often, but it happens: a confident reference to an issue where I made an argument I never made. It reads exactly like the real thing, which is the problem. If I didn't know my own archive it would go straight through.

So the rule in my project is simple:  No reference goes in until I've found the actual sentence in the actual file. Not the slug or the topic. The sentence.

That check takes me thirty seconds because I wrote the thing. Someone new to their job, handed a draft full of confident citations to a body of work they've never read, has no version of that thirty seconds. They'd have to know enough to be suspicious first, and being suspicious is downstream of having driven the road.

Then Saniya takes it. She's my editor, and after I finish a draft it goes into a Google Doc where she does a manual pass for clarity and to keep my voice intact — and sets it up in Beehiiv to publish. Charlie Rodriguez does the illustration, working from a style guide we built together, which is why Toni and Gray look like themselves every week.

Berkson's Bits

That feeling when you're on a video call and you can't help wanting to click on a tab in the screen that's being shared that has the red dot that indicates there are new items to view...

What I'm Watching...

In 1985 Sting (of The Police fame) pulled together some of the best jazz musicians of the time, including Branford Marsalis, Kenny Kirkland, and Omar Hakim, to record an album and go on tour. I was fortunate to get to seem them at Radio City Music Hall in NYC. I highly recommend the documentary Bring On The Night, a behind-the-scenes look at them rehearsing for the tour.

The student asked me what they should do. I gave them the honest answer: I don't know. Use the tools. Everyone else will. But find the places where the stakes are low and do it the slow way anyway, not because slow is virtuous, but because that's the only version where you find out you were wrong while it still doesn't cost anything.

That's not much of an answer. It's what I've got.

If you're further along, you have the other half of this problem. Somebody on your team is producing work that looks finished on day one. Whether they built any road sense is going to depend on what you ask them, and how often, and whether there's still somewhere they're allowed to get it wrong.

For founders, CMOs and marketers, that may be the more important question.

How do you build experience in a world where AI keeps removing the experiences?

I'd like to hear how you're handling that. I don't think anyone has this figured out yet.

Looking forward to continuing the conversation...

Alan