I might start by stating the problem to solve, and a few constraints. Then I’d share how I’ve thought about the problem.

But before we get too far, I’d ask her opinion. She may see something I’ve missed, or have insights or experience I lack.

Together, we’d work out the right approach before anyone gets started in earnest.

But that wasn’t how I used AI at first.

I gave it an assignment — I need a slide deck formatted just so — waited for an answer, and spent time fruitlessly arguing with it when the output inevitably wasn’t quite right.

Then my friend John gave me a better mental model:

Think of AI as a brilliant junior colleague. It’s a fast worker who knows an enormous amount, but it’s new to your field, your organization, and the kind of work you’re trying to do.

That changed how I approach AI.

I stopped treating it as a drone to execute tasks and started treating it as a collaborator.

So I’m less likely to ask for a finished product and more likely to begin with the problem: Here’s what I’m trying to accomplish. Here’s how I see it. What questions do you have? How would you approach this?

The conversation that follows is often where the work gets better. AI can surface ambiguities I hadn’t noticed, suggest approaches I hadn’t considered, or help me make sense of a half-formed idea.

When my AI collaborators make suggestions, I use my judgment about whether and how to apply them. Sometimes they merely show me what not to do — which points towards a better answer.

This is possible because of a continuous process of learning — essentially onboarding a new colleague.

At work, I started by feeding it all our marketing frameworks, legal guidelines, style guides, examples of content that had actually landed well, and details about who else is involved.

On a regular basis I share examples of what’s working and what’s not from what we produce together. We’re constantly refining processes together to make things smoother or address gaps.

I still argue with it plenty. But it’s more constructive now.

Not that I use it regularly. I use macOS rather than Linux. I use ChatGPT and Claude rather than one of the open-weight alternatives. In fact, most of the software I rely on every day is commercial.

What I value most about open source is how it allows thousands of people with different knowledge and interests to build on one another’s work—often enabling better approaches to come to the fore. (The scientific process and Wikipedia work much the same way.)

Consider Firefox. The open-source browser restored competition to a stagnant market dominated by Internet Explorer. It popularized user interfaces like tabbed browsing that are now table stakes for every browser on every platform. And it spurred a revolution in how websites are built by demonstrating what open standards like CSS could actually make possible.

I spent hours exploring the CSS Zen Garden website and digging into its submissions. Dozens of designers took the exact same HTML and, by changing only the CSS, produced wildly different sites. It showed how people could learn from one another by studying, borrowing, and adapting each other’s work.

Nearly every piece of software you and I use was shaped somewhere upstream by code people gave away for others to build on.

It looked great on tours: a flatscreen TV, comfy-looking couches, a little kitchenette, dramatic lighting. And a pool table, obviously. But there was no reason to go there as a resident.

It was jammed into an out-of-the-way space in the basement — an afterthought filled with generic contemporary decor. It felt designed around a checklist, without any thought to why people might want to come.

A lot hinges on that why.

Lively town squares are surrounded by shops and restaurants and cafes. People come for a sandwich and then sit talking on a bench. They come so their kids can play on the swing set and then chat with their neighbors.

In offices, people gather where there’s food or coffee. There’s often more collaboration around the coffee machine than on the weird set of chairs by the elevators that nobody sits in.

Public spaces that work usually aren’t destinations in themselves. They piggyback on things people already want to do.

Anyone can claim a bit of sand by placing a chair or towel. It’s first come, first served, so the sooner you get there the better.

As the beach gets busier, people will fill in around the early arrivals. Everyone has a sense of how close is too close, and there are unofficial norms about how much space you can claim, and how much margin to leave so people can walk past.

If you arrive at midday, you can claim an available space. You know almost instinctively if a space is too small — and if you judge wrong, your new neighbors will let you know.

Nobody designed these rules or posted them on a sign at the beach entrance. They emerge from thousands of individual interactions between people and adjust as conditions change. On a crowded holiday weekend, people squeeze in. On a quiet Tuesday, they spread out.

The rules change without anyone rewriting them.

That made some sense for publishers that ran ads: a click meant an ad impression on the website, and an ad impression meant revenue.

But this became the norm even for senders that didn’t depend on ads, like nonprofits. They built the same teaser-and-click model.

One big reason why? Measurement.

When the content lives in an email, there’s far less to count. No page views, no time on site, no funnel. Even open rates have gotten unreliable as email providers block marketers from seeing who opens. A click on a link is one of the only numbers left that can still be measured with any confidence.

So click-through became the metric — and then, like every proxy that stands in for a goal long enough, the metric became the target. Chasing clicks skewed the actual job of a newsletter, which is to give subscribers something worth reading.

Some writers went a different direction. Ben Thompson was sending full-text emails as early as 2013. Substack made that approach feel normal. A better reader experience, but less data.

That’s a trade worth making. A newsletter’s job is to create value for the people who signed up for it, not to generate a number for the people who sent it.

Until recently, the lot hosted a one-story fencing studio. Teenagers lunging at each other after school, Olympics-themed banners in the windows. I had no idea the neighborhood needed a fencing studio. Turns out it did, at least for a while.

Before that, it was a jewelry shop where I bought our wedding bands. A decade and a half later, it’s still what I think of when I walk down that block.

It’s continuous reinvention. Someone finds a new use for a space, which makes sense until it doesn’t.

I can’t wait to see what’s next.

And at the grocery store, I grabbed a container of watermelon. Also about $10. I didn’t really think about it until I got home.

Why did I agonize over one decision and not the other? It wasn’t price: They cost about the same.

It might be the difference between a purchase and a bet.

The watermelon’s value was obvious before I picked it up: I like watermelon. I’d rather have watermelon than $10. Done.

But the notebook’s value was unknown. It depended on whether a new way of capturing notes would actually change how I think. I wasn’t really buying a notebook but trying to predict whether I’d actually use one.

The agony was about the uncertainty of the payoff, not the price.

I wear it because the sentence is true and worth sharing. It has the secondary effect of advertising FIRE, the free speech organization that sells the shirt.

People regularly stop me to praise the shirt’s message. Sometimes they ask what FIRE is. The shirt never feels like a pitch.

It gives people language for something they already believe, or helps them see a problem more clearly, before it asks for anything.

Jason Fried and David Heinemeier Hansson have followed a similar approach with Basecamp, the collaboration software.

I read them first for their thinking about work: why meetings drag, why busyness isn’t the same as productivity, why calm is better than chaos. I internalized a lot of that before I ever thought about their software.

A shirt slogan, a blog about work. Each appeals to who we want to be. Neither asks for anything. The pitch, if it needs to come at all, comes later.

These stories are often treated as evidence that AI is overhyped and doomed to fail, maybe even that it should be discarded.

They’re better understood as evidence of learning. Of trial and error.

I’ve spent the past year figuring out how AI can help me. At various times, I’ve worked with AI to plan fitness regimens, shape marketing plans, and create elaborate automations on my phone to send morning briefings with the news and my to-do list.

Some of these uses I abandoned within a week. Some were almost right but required revision. And some, like using AI to help explore ideas, have worked from the start.

My misses were the only way to find out what worked for me. I couldn’t predict in advance where a tool would or wouldn’t fit my needs.

Likewise, we can’t derive from first principles how to use AI in customer service, or in management, or in law. The stakes might be larger for a business than for me, but the trial-and-error approach to learning is the same. Christopher Mims’ recent Wall Street Journal piece makes a complementary point: trial and error is how we learn where not to use AI.

We have to try things and watch what breaks.

None of these fireworks are official. Many of them are of dubious legality. All of them are joyfully celebrating the country‘s birthday.

The official fireworks on the Mall a few miles away are visible if you squint. They reflect a patriotism of pomp: Sousa marches, military flyovers, and elaborate displays planned months in advance.

The neighborhood fireworks are America. Informal. Self-organized. Ungoverned. Thumbing their noses at authority.

Here’s to another 250 years.