AI is a lot of fun

The most positive thing I can say about “AI” coding tools is — and I mean this sincerely — that they are a lot of fun to play with.

That’s really all you need to know to understand the hype around AI. AI is fun to play with — addictively so. It’s essentially a slot machine. You put some money in, pull the lever, and some code that might be useful comes out. It probably won’t actually be what you wanted. It will almost certainly have flaws. But it’s conceivable, and indeed mathematically possible, that it might be perfect. That tantalizing possibility keeps us pulling the lever again and again.

There’s some geek-specific psychology going on too. LLMs are tools that seem very powerful and yet, at the same time, are observably imperfect. That combination is catnip for computer geeks of all stripes. Our absolute favorite thing to do is take an imperfect system, tinker with it or build a wrapper around it, and make it work better. And LLMs give us no end of opportunities to tinker.

The tinkering instinct is, I think, what Richard Gabriel was talking about when he wrote “The Rise of Worse is Better.” The core idea of the worse-is-better philosophy is that if you want to make successful software, you shouldn’t try to release a perfect product. Instead, release something simple and easy to tinker with, that solves at least part of the problem. If it’s useful, a whole community of geeks will spring up around it, eager to sand down its rough edges. The result will be software that fits their needs better than anything you could have designed upfront.

“Worse is better” says, in essence, this: approaching software design as an engineering problem is very expensive. But if you instead approach it as a psychological problem — asking not “what should we build” but “how do we get programmers to work on this” — you can create a lot of value at little cost to yourself. We could summarize the traditional worse-is-better business strategy as:

1. Release imperfect software

2. Get programmers to work on it for cheap

3. Sell the software

4. Profit!

Here’s the secret that every successful software company is based on: You can domesticate programmers the way beekeepers tame bees. You can’t exactly communicate with them, but you can get them to swarm in one place and when they’re not looking, you can carry off the honey.

—Orson Scott Card

AI companies like Anthropic have taken the same psychological trick and applied it in a slightly different and arguably simpler way:

1. Release imperfect software

2. Charge money for the privilege of trying to improve it

3. Profit!1

In this new regime, the “improved” software is never sold, because it doesn’t have to be. What the AI companies realized is that there is so much surplus in the system that the worse-is-better process doesn’t even have to produce valuable software; the geeks will show up regardless. Geeks tinker just because they like to tinker. Software development is entertainment, and when something is entertainment you can simply charge money for it directly.

You can see the effects of this strategy playing out at almost any software company. For instance, my coworkers have written dozens of “skills” for Claude Code that sit on the shelf unused. Developing useful software is not the point of all this. The process is the product. The point is to have fun.

A couple months ago, I spent several weeks building my own “agentic” LLM harness. Perhaps I’ll write about it in greater detail at some point; for now, suffice it to say that the results I get from it are better than vanilla Claude Code but still not good enough for production, which I guess should tell you something about how good I think Claude Code is.

Of course I bootstrapped the whole project using AI. After about 6 weeks of committing 90% LLM-written code, I had to go over the codebase with a fine-toothed comb, fixing all the weird issues the LLM left behind.

Just last week, I tried to use my custom harness to make a large (~1000 line) change to an open-source codebase. On the face of it, it should have been straightforward: I changed one of the core data types and told the AI to fix up all the code to match the new types. However, in spite of the apparent simplicity of the task, the AI couldn’t hack it. The resulting pull request had so many issues that I decided to redo it by hand. You can take a look at the AI-generated PR here, and at my manual rewrite here. The FIXME comments in the AI-generated pull request are written by me; they are how I communicate my desired changes to the LLM. You can see where I went a little crazy reviewing the AI’s crappy code and just started screaming at it.

All that’s to say, the “productivity” benefits from AI remain elusive. However, I can conclusively and authoritatively state that it is a lot of fun! Building my own LLM harness was the most fun I’ve had writing code in years. I actually stayed up late one night working on it, which for me is unheard of. Even using it is fun, in spite of its flaws. When it produces something useful, I get a big dopamine hit. I made the machine do a thing. When its code is crap, I get to feel superior as I fix all its weird goof-ups. It’s win-win… at least, as long I’m not the one paying for tokens.

My coworkers are now talking about “loops” as a technique for getting better outputs from LLMs. This unlikely buzzword is the current stalwart defender of the AI hype fortress, its predecessors — “prompt engineering,” “MCP,” “skills,” “context engineering,” “harness engineering” — littering the field like so many corpses. These techniques didn’t fix the problems of “hallucinations” or “misalignment,” and I would bet money that “loops” won’t magically fix everything either. I predict they will cost a lot of tokens, though.

But again, this is all a lot of fun, and that is the reason we’re doing it. That’s why we’re riding around this buzzword carousel again and again. Not because it makes us faster or more productive. Not because we write better code this way. Not because the software that comes out is better. Not because there’s any ROI. Because it’s fun.

Everyone I’ve talked to who is using AI has confirmed this, in the end. I ask probing questions about what techniques they’re using, what benefits they’re getting, what the quality of the resulting software is, and how long it takes to fix the issues, and eventually they admit it: “it’s just more fun to do it this way.”

And that ends the conversation. For fun is beyond critique, beyond reason, beyond all economic analysis. It is unanswerable and unassailable. Fun will stand forever; it will triumph in every debate; it will still be undefeated when the last lights of our civilization wink out.

How much longer will our employers foot the bill for fun? Are they going to just let us have fun forever? When they cut off the infinite supply of tokens, will we still know how to program? Will any of our software still work? And how long will it take us to clean up the mess the LLMs left behind?

How much value are we willing to destroy in the name of having fun?