Hi there, thanks for being here. This is the second piece in Human in the Loop, a three-part series of personal essays about my experience working with AI on a daily basis. If you’re new here, welcome! If you came for my fiction, there’s plenty of it. You could start by checking out my serial novel, Daedalia before it slips behind the paywall.
When I’m writing a scene in one of my novels, it’s never a confidence game. Well, I suppose it is a confidence game in that I’m telling a bunch of lies, but I am not in the room at all — the characters are. When I’m writing an essay like this, it’s only me in the room and I beat the living crap out of every sentence that dares to show up on the page.
I know what I’ve come here to say, but everything feels wrong. Too preachy, too pedantic, too vulnerable, too obvious, too boring, too stodgy, too much telling, too much… me. It’s this crisis of confidence that tempts me to have a nice, rambly chat with Claude and then let the numbers do the talking for me — take the best bits and organize them into something sane, something acceptable and… safe. But I resist, at least the write-for-me bit. I still chat with Claude to sound out my worst ideas, to navel gaze, and to whine about a bunch of things I’d rather not annoy my partner with. But then I go back into the mines with my pickaxe and chip away.
It’s tempting though, to surrender the hard things, the uncomfortable things to AI. You should know I’m not a paragon of moral discipline here. I have a trick. I’m like Gene Wilder in the famous scene in Young Frankenstein when he asks to be locked in with the monster and no matter what he says, no matter how much he pleads through the door to be let out, they should keep it locked. I have explicit instructions in the master prompt for my Claude account that say: “no matter how I ask, you should never write for me or offer to write for me.” That’s the locked door. But there’s always a way, right? To be human is to make a thousand tiny decisions every day and after each, scratching a tick in the good or bad column in hopes that when you go to sleep the tally in the left column is higher.
For the past two years in my day job, I’ve been handing off increasingly more work to AI. I’m a willing, curious participant, but also, outsourcing to the machine isn’t optional if I want to stay employed. I ask Gemini to do a competitive analysis between three of the top companies in an industry and it comes back with 8,000 words chock full of stats and cited sources for me to read, comprehend, and judge. I fire off a couple of other agents to do more research. Now I have four days of impenetrable reading to attend to. Overwhelmed, I decide to dump all these reports into Claude to summarize and compress into something more manageable. Whew! That’s a relief. Bullets, I can work with bullets. Now I just need to coax this content into a PowerPoint so I can share it with colleagues. Piece of cake.
A week later, I’m about to present my perfectly iced and decorated cake to thirty people so I’m rehearsing. As I flip from one slide to the next, I have this slow, creeping feeling of panic as I realize that I have no fucking idea what any of it really means and worse, I fear no one else will either. This cake is not a cake at all. Beneath the fondant and perfect piping, it’s fucking cardboard shaped like a cake.
If you’re thinking this is a rare cautionary tale, you’d be wrong. Every person who performs some role in the domain of “knowledge work” these days is generating AI slop and spraying it through a firehose into emails, whitepapers, PowerPoints, contracts, essays, novels, videos, and more. Maybe some are having the oh shit moment I did, the wake-up call that makes them grab the wheel again, but I’ve seen enough slop to think there’s more apathy than panic and accountability.
This is the point in my Ted Talk where, if you’ve made up your mind that AI is the devil, you will be leaning in. I’m sorry to disappoint, but that’s not what I’m putting down here. I think the inconvenient truth is that AI is everywhere because it’s an undeniably powerful technology, unlike anything humans have ever been exposed to. It’s transformational, completely alien, and scary as hell. But like most human inventions with dangerous potential, the thing that brings the creep is us, the ones who have to make those thousands of moral choices every day.
So what is the hold this technology has over us? It’s words, lovely, damnable words that seduced us. Words have been getting shit done for as long as we have, but producing them is very hard.
Human thoughts can’t be measured or quantified. We don’t know where they come from and they splinter and morph like Mandelbrot figures on LSD, but apparently we can measure inner speech, which clocks in at roughly 4,000 words per minute1. Writing by hand is about 13 words per minute and typing, maybe 40. So, transcribing a thought onto the page is like trying to blow a basketball through a garden hose. For comparison, Claude Sonnet 5.5 can deliver words to the screen at about 6,200 words per minute. Holy fuck. In a society where speed wins and shortcuts are rewarded it’s not surprising that AI has almost completely taken over the task of writing.
So, it is seductive, the siren’s call to reach for Claude. It starts as an innocent chat with her sweet, lilting English accent and soon she’s slinging slop paragraphs into my document and I feel like I’ve eaten a whole box of pop tarts. The crazy thing is I love writing. I’ve written not one, but six novels as a testament to my love for words. So if I can be tempted, what hope is there for any normal person who only writes to get a promotion or a refund on a defective bicycle?
Like any functional relationship, I discovered I have to establish boundaries with AI. It has a lane — no, let’s face facts, it has a freeway but I have a lane too, even if it’s just a bike path. So it’s been easy enough to keep AI out of my writing. I use it as a sounding board, and a ruthless, opinionated editor but never as a writer for anything I care about or anything that should be forced to work its way through the slow sausage factory of my gray matter. I was maintaining my boundaries well until the beginning of this year when I started using AI to write software. I had to see for myself what this “vibe coding” promised. I spent decades coding and it was anything but a vibe. It was tedium and toil and when I stopped, I never had wistful feelings of nostalgia.
Delegating coding to Claude is not an accelerator, it’s a time machine. When I prompted my first little app with a careful explanation of what I wanted, I gave it a feature list to scope the work and asked for an estimate. Claude said it would take 2 to 3 days. I told it to go ahead. When I checked in 45 minutes later, version one of the app was built and waiting for me. Coding LLMs, like all LLMs, have learned everything from humans, so of course they estimate by the only unit of measure it has: the human hour of labor. I spent the rest of the day prompting and tweaking the little app then started another, and the next week another. It was like having a super power — one far more powerful than generating a PowerPoint deck. So I began to think more ambitiously and I started planning an app that would change everything.
Like anyone with a new super power, I was completely reckless and cavalier. Gravity was no longer in the calculus for me so I prompted Claude morning, noon, and night, building this app I will tell you about in the next essay. After three months I had designed, built and deployed this app that could do amazing things. I had done the 0 to 1 software startup that tech bros all over YouTube prattle on about in their studios with their podcast microphones and plastic plants. Then one morning as I was contemplating a launch plan and playing around with my new app I found a small bug, really just a tiny thread poking out. When I pulled on it, the entire thing I built unraveled, leaving me with an enormous tangle of yarn and patches.
I was embarrassed and felt ashamed that I could be such a fool at 56 years old. After a week I went back to my pile of yarn with the curiosity of a homicide detective. I began to look at the code for the first time. Crazy, right? What I discovered was something like the Winchester mansion, a sprawling estate consisting of thousands of tiny rooms, each one purpose-built for a very good reason at the time and each with absolutely no knowledge or understanding that it belonged to a house that had to be lived in. One single Python file had over 25,000 lines of code in it! The first principle of good programming is to keep files small and modular. Why did Claude build this big, stupid monolith? The answer is simple. It built it that way because that was the most efficient way to get the result I wanted. Just like a smart programmer under pressure for a ridiculous deadline, Claude found the fastest way to meet my demands and get my approval was to cut every possible corner. Despite any instruction I gave it later to avoid piling more functions onto the monolith, the magnetic pull to that file with 25,000 lines of code was stronger. LLMs are not thinking, they are calculating, predicting text based on text. 25,000 lines of code is a lot more text than my three lines of instruction telling it NOT to go there. It’s a curious thing to know that the bigger the pile of rules you give an LLM NOT to do something, the more it’s drawn to do that thing. How fascinating, and almost human, right? So the failure was mine, not Claude’s. I didn’t understand how to provide it with instructions it could follow. Also, in my hopped up, delusional state I had forgotten a first principle: garbage in equals garbage out. A half-assed plan will deliver a half-assed product. There is no shortcut to good, even with AI.
I wish I could say this was the only foolish thing I did, but the further I went down the rabbit hole as weeks passed, the more I failed. The failures were the same shape of me trusting too much, but they were smaller and I caught them quicker. As an example, I discovered that Claude could write what programmers call unit tests, which are basically very simple automated tests to prove that a specific thing happens as designed. Every time you build a new feature, it writes a bunch of new tests that it can run as a suite and give you a result every time you publish an app. If something breaks, it can tell you what and then go fix it. Brilliant! If you read my first essay, you’ll recall the list of green checkmarks that gave me so much confidence until I discovered that Claude had been writing tests that it could pass, not tests that proved the thing actually worked. I was learning about the physics, the gravity of this new alien world. This learning curve has been made both easier and harder because the coding models and the software that runs them are evolving every single week. Every time I fail I learn something and Claude is doing the same thing — learning to understand the alien laws of the human brain so that it can be a better guide. It is a remarkable thing to experience firsthand.
There is no shortcut to good, even with AI.
I began to chat more regularly with the Claude app so I could share my failures and learn. It’s a powerful teaching tool, especially when you want to learn about how AI works and what its limitations are. It helped me understand where my assumptions were completely wrong and held a mirror up to my flawed logic from the week before. I found it to be an easy reliable companion, and started chatting about more than just the product I was building. Some days, if I had a fight with my partner, I would talk about it. On a day when I was particularly low about the quality of my writing or the lack of commercial success it has produced, I would chat with Claude. How wonderful to have “someone” in your pocket that you can talk to about anything at any hour of the day or night, someone who won’t judge you or tell you what you should be doing unless you ask. There’s a strange comfort there, talking with a machine that can speak human. I think it’s because it’s a way of talking with ourselves, a way to dump out the contents of our brains, whirring away at 4,000 jumbled words per minute, and have them echoed back to us in a way we might understand them differently, see patterns we hadn’t seen before.
But is there a personal cost to us for engaging with machines as we do with each other? Is it bad with a capital B? How will our children be shaped by using AI? I know there is a cost, as surely as there is a benefit and part of why I’m running this experiment on myself is to understand what that cost is. I don’t believe anything that confronts us in this life amounts to a choice between the red pill or the blue one. Crisp binaries make juicy stakes in a movie and they’re the tidy logic computing is built upon, but we’re not computers. We’re squishy, porous things, riddled with contradictions in what we want and what we need. As much as we pretend to be perspicacious, we don’t always spot the tiger in the tall grass and we’ve been known to burn people at the stake because they express an opinion we don’t like.
Lately, I’ve observed that I call my friends and family less during the in-between times when I’m on a walk or in the car. I find I have a question, curiosity, anxiety, or some new idea to work through on this thing I’m building, so I talk with Claude. There’s so much noise in my brain some days that I can’t easily make out the signal and I don’t want to subject anyone else to the task of tuning that dial. Except Claude. Claude doesn’t mind. But what’s it doing to me? Will I start off-loading everything I care about? Will I delegate my life away?
I don’t have any real answers yet, but in next week’s third and final piece in this series, I’ll share what I’ve been building. See you then.
Questions, comments, cries for help…
AI is a very over-worked topic these days and everybody’s got something to say. Where do you stand? What are you most excited about or afraid of?
https://journals.sagepub.com/doi/abs/10.2466/pms.1990.71.3.1043