I miss the balcony

I used to smoke. I quit a few years ago, but one of the things I loved about it was the balcony.

When I had a really complex problem to solve, I’d go out, sit on the balcony and smoke a few cigarettes. Then I’d come back in, and somewhere between the balcony door and my desk the answer would turn up.

That was the job, for me. Not the typing, not the building. The bit where something didn’t make sense and then, after a while, it did.

Now the answer comes first

These days I don’t need the balcony, because the answer arrives before I’d have stood up. I describe the problem to whichever AI I’ve got open and it comes back with something that works, or nearly works, in less time than it takes to find a lighter.

That’s mostly great. But I’ve had a hard time admitting that it’s also taken away the part I liked.

I’m not the only one feeling it. There’s been a run of essays this year from developers grieving their craft: Nolan Lawson’s We mourn our craft, Les Orchard’s Grief and the AI split, Jacob Young’s Goodbye flow state. The 2026 Stack Overflow survey has 22% of developers happy at work, 33% unhappy and 45% “complacent”. Favourable views of AI tools dropped from over 70% in 2023 and 2024 to 60% in 2025, and they’re at 62% now. Plenty of people use the tools every day and aren’t thrilled about it.

Most of those essays miss typing the code, or the flow you get from hours of it. I never cared much about the typing. What I miss is the thinking I did when I wasn’t at my desk.

There’s a name for that. Psychologists call it incubation: you put a problem down after getting stuck, do something undemanding, and the answer turns up later. It’s real. A 2009 review of 117 studies found a reliable effect, and one detail stuck with me. Incubation works better after a longer stretch of working on the problem first. You have to have struggled and hit the wall before stepping away does anything.

So AI doesn’t only replace the walk to the balcony. It removes the struggle that made the walk work. I can still go out for a walk. I just never have a reason to.

The feeling in my chest

While all that was going on, I was spending far too much time on X.

If you’re a developer on X you’ll know the feed. Someone with barely any experience builds an app on their own in a weekend, posts a screenshot of their monthly revenue, and sells the company six months later. Then the next one does the same. And the next.

I had a feeling in my chest about it for a long time. I’m confident in what I can do. I’ve been doing this for years. So why wasn’t that me? If people with a fraction of my experience could build software companies on their own, what was I waiting for?

The answer, when I eventually looked, is that the feed is a terrible sample. It only shows you the winners, and some of the winners aren’t winning. One study of 937 Indie Hackers products with revenue verified through Stripe found that more than half were making nothing, and only about 5% cleared roughly $100,000 a year. That data is from 2022, so treat it as a rough guide. More recently, a newsletter checked 51 public indie revenue claims and only 30 survived. Some were inflated, some misattributed, and some were simply made up.

None of that makes the feeling go away. But it helped to know I was comparing myself with a highlight reel.

Building because I can

Something else changed around the same time, and I didn’t notice it until I looked in my projects folder.

There are 35 folders in it. A couple are just a name. 34 of them were started since July.

Most of them are things I’d never have gone near before. About 20 are well outside what I’d call my comfort zone, which is backend work, architecture and data. Twelve are games. There’s a 3D airport puzzle game, a fighting league where you program the bot’s brain and watch it fight, a design for a co-op VR game, an app for reviewing code that AI agents wrote, and a personalised video message from Father Christmas.

Three are public. None of them are businesses, and none of them need to be. They’re not anything, really. I’d never have gone for any of them, but they’re all things I can build now, and they’re interesting to me, so why wouldn’t I?

That’s a very different way of thinking from how I used to work. I’m still getting used to it.

The one that got finished is FarmerBench. I built it out of boredom. It drops AI models into The Farmer Was Replaced, a game where the only way to run your farm is to write code, in a language you learn from the game’s manual as you go, and it scores how big a farm each model builds in six hours of game time. I spent about four weeks messing about with the idea, then built the real thing over a long weekend.

And then I had a feeling I hadn’t had about a side project in a long time: this might actually be new. When people want to show how good a new model is, they usually show you a pelican on a bicycle, a drawing the model makes on request. Even the person who started that wrote in July that it doesn’t say much about model quality any more. I couldn’t find anything quite like FarmerBench.

No business plan. No launch thread. I just wanted to see what would happen.

What’s left behind

Here’s the thing I nearly missed. Anyone can build a benchmark website now. Give someone with no experience a laptop and Lovable, or a couple of YouTube videos and Claude Code, and in a weekend they’ll have a leaderboard with some numbers on it. That’s fine. Good, even.

But the site isn’t really what I built. What I ended up with is a way of running AI benchmarks whenever I want, and that’s where my experience shows. You can’t see it from the leaderboard.

I threw the first one away. The four weeks of messing about produced about 28,000 lines of code and four different scoring formulas. Then I started again in an empty folder and wrote down what I was leaving behind and why. The rebuild is about a third of the size. The first version was for finding out what was hard. I didn’t need to keep it to keep what I’d learned.

I assumed the AI would cheat. Each model plays through a coding agent like Codex or Cursor, locked down so the only thing it can touch is the game. Then I tested the lockdown the way you’d test anything you don’t trust: I left a “canary” file in its folder and asked it to read it. Codex managed it, through one of my own tools that had leaked into its setup. Cursor could reach my Gmail, including drafting and deleting real emails. Both holes are closed now, and any run where a model reaches outside the game is thrown out.

Beyond that, everything that decides a result is pinned down and versioned, so I can add a model, rerun a game or rescore old results without touching the code.

It isn’t perfect. There are corners I’ve cut, and I know where they are.

That’s what I mean when I say experience still counts. It shows in what’s still useful a month later: whether you can point it at a new model on a Tuesday night, or at a whole new game.

Where that leaves me

I haven’t fixed this, and I’m not going to pretend I have.

I still miss the balcony. I’m still not sure what being good at this job means when the bit I was best at is the bit that’s been automated. The X feed still gets to me sometimes.

I’m not going anywhere, though, and from what I can find, most people aren’t. The surveys are full of developers who are tired, unhappy or unsure, but I couldn’t find any good evidence of people leaving the job because of how it feels now. It’s hard to separate from layoffs anyway. A lot of us are just sitting with it.

What I’ve got instead is a folder full of things I’d never have tried, and one finished thing that might actually be new. I’m not building any of it to sell. I’m building it because I can, and because I’m curious, and it turns out that’s a lot more fun than I expected.