A short essay on giving in to AI — what I lost, what I gained, and the workflow I quietly settled into.
The first time I really caught myself, I was about to read a forty-page Postgres replication doc the way I used to read everything: front to back, taking notes, pretending the slow path was virtuous. I closed the tab and asked Claude instead.
I had given in years before. I just stopped pretending the pieces didn't add up.
The worry was the wrong worry. Or it was the right worry about the wrong thing.
The fears were real. They are real. I had them and I still have most of them.
That I would atrophy. That the ability to sit with something hard until it cracked open was a muscle, and muscles you stop using go. That the texture of figuring it out alone — the part where confusion turns into understanding without intermediation — is its own skill, and not one you get back by rereading a Wikipedia summary at 3x speed.
That I would become the kind of engineer whose code looks fine until you ask him to defend a single line of it. That my taste would slowly synchronize with the model's taste and I wouldn't notice. That I'd start writing prose that sounded like prose written by someone who had read a lot of prose written by AI.
That depending on something is a slower kind of becoming it.
These weren't paranoid fears. They were correct.
What changed is that principled abstinence kept costing more than it saved.
Not in a single moment. In a long accumulation. There is a project I would not have shipped if I had insisted on doing every part of it the slow way. Then there is another. Then there is most of the work I am proud of in the last year.
Clarity has an AI layer because the AI layer makes the financial graph readable in a way that columns and charts never quite did. Docpull exists as a tool for RAG pipelines — built with agents, in part to feed agents. Daudio and the theory app live in branches I would never have explored alone, not because the work was beyond me but because the per-unit cost of exploration was higher than my appetite for it.
That last sentence is the actual claim. The work was not beyond me. The cost was.
Doing it the hard way was, mostly, a vanity. The work is the artifact, not the suffering required to make it. I had confused the second for the first because the second was easier to feel proud of. Pride scales poorly.
I work in Claude Code inside this monorepo more or less every day. Cursor for the lighter sessions where I want completion in the editor and not a conversation. The latest Sonnet for almost everything; Opus when I need the model to actually think; Haiku for small mechanical jobs where the cheapest thing that produces correct output is the right tool.
Reading and research. I no longer Google unfamiliar domains. I start a conversation, dump what I know, ask what I'm missing. Five minutes of that beats forty minutes of skim-reading the right four blog posts I would have eventually found. The model is wrong sometimes. I check.
Writing code. Claude Code handles the boilerplate, the mechanical refactors, the first pass at code I've written ten times before. It explores unfamiliar parts of a repo faster than I can. I still do the architecture, the naming, the taste calls. The line is unstable and moves every quarter, but right now: anything where the question is "what should this be" stays mine. Anything where the question is "how do I express this cleanly" is up for grabs.
Writing prose. The model is excellent at structure and second-pass editing. It is not good at the part where you have something to say. That part still costs what it always cost. Most of what I publish goes through several passes that look like editing rounds with a sharp reader who has read everything. None of it is the model alone, which would read like the model alone, which is a thing readers can now smell from across the room.
Reading documentation. Docpull ingests entire docs sites; an agent reads them and answers questions in context. I was building tools for this before I noticed I was the primary user.
The line I draw, mostly: I do not let the model decide what is worth doing. I let it help me do the work I have already chosen.
The honest reckoning is that yes, some things have gone.
I read worse than I used to. I mean unaided reading — sitting with a difficult page and waiting for it to unfold. I still do this. I do it less. The skill is sharper if you grind on it daily and I am not grinding on it daily.
I notice I reach for the model when I am stuck even when stuck is the right place to be for another twenty minutes. The friction is gone, and friction was sometimes the thing.
There is a version of me that would have struggled with a hard piece of theory for two weeks and emerged with the kind of internalized understanding that only that struggle produces. That version is partly hypothetical and partly real and I am not sure which parts.
This is not free. I am not pretending it is free.
What I got in exchange is leverage at a scale I do not want to give back.
Clarity, docpull, daudio, the theory app, this site, the projects I have not announced — none of these exist in their current shape in a world where I refused. They exist in a smaller, slower shape, or they don't exist at all. The surrender was the precondition for the work, not its enemy.
The scope of what one person can attempt has changed. I am not interested in pretending it hasn't to preserve a self-image from 2021.
Strangelove rides the bomb down. The joke is that he loves it because there is nothing left to do but love it. The surrender is total and the world ends.
Mine is not that. Mine is messier. I gave in to a thing that is genuinely useful and genuinely changing me, and I am not sure yet which of the changes I will be glad of and which I will quietly mourn. I am not going back. I am not finished worrying.
I learned to love the bomb. The bomb has not gone off.
A short essay on giving in to AI — what I lost, what I gained, and the workflow I quietly settled into.
The first time I really caught myself, I was about to read a forty-page Postgres replication doc the way I used to read everything: front to back, taking notes, pretending the slow path was virtuous. I closed the tab and asked Claude instead.
I had given in years before. I just stopped pretending the pieces didn't add up.
The worry was the wrong worry. Or it was the right worry about the wrong thing.
The fears were real. They are real. I had them and I still have most of them.
That I would atrophy. That the ability to sit with something hard until it cracked open was a muscle, and muscles you stop using go. That the texture of figuring it out alone — the part where confusion turns into understanding without intermediation — is its own skill, and not one you get back by rereading a Wikipedia summary at 3x speed.
That I would become the kind of engineer whose code looks fine until you ask him to defend a single line of it. That my taste would slowly synchronize with the model's taste and I wouldn't notice. That I'd start writing prose that sounded like prose written by someone who had read a lot of prose written by AI.
That depending on something is a slower kind of becoming it.
These weren't paranoid fears. They were correct.
What changed is that principled abstinence kept costing more than it saved.
Not in a single moment. In a long accumulation. There is a project I would not have shipped if I had insisted on doing every part of it the slow way. Then there is another. Then there is most of the work I am proud of in the last year.
Clarity has an AI layer because the AI layer makes the financial graph readable in a way that columns and charts never quite did. Docpull exists as a tool for RAG pipelines — built with agents, in part to feed agents. Daudio and the theory app live in branches I would never have explored alone, not because the work was beyond me but because the per-unit cost of exploration was higher than my appetite for it.
That last sentence is the actual claim. The work was not beyond me. The cost was.
Doing it the hard way was, mostly, a vanity. The work is the artifact, not the suffering required to make it. I had confused the second for the first because the second was easier to feel proud of. Pride scales poorly.
I work in Claude Code inside this monorepo more or less every day. Cursor for the lighter sessions where I want completion in the editor and not a conversation. The latest Sonnet for almost everything; Opus when I need the model to actually think; Haiku for small mechanical jobs where the cheapest thing that produces correct output is the right tool.
Reading and research. I no longer Google unfamiliar domains. I start a conversation, dump what I know, ask what I'm missing. Five minutes of that beats forty minutes of skim-reading the right four blog posts I would have eventually found. The model is wrong sometimes. I check.
Writing code. Claude Code handles the boilerplate, the mechanical refactors, the first pass at code I've written ten times before. It explores unfamiliar parts of a repo faster than I can. I still do the architecture, the naming, the taste calls. The line is unstable and moves every quarter, but right now: anything where the question is "what should this be" stays mine. Anything where the question is "how do I express this cleanly" is up for grabs.
Writing prose. The model is excellent at structure and second-pass editing. It is not good at the part where you have something to say. That part still costs what it always cost. Most of what I publish goes through several passes that look like editing rounds with a sharp reader who has read everything. None of it is the model alone, which would read like the model alone, which is a thing readers can now smell from across the room.
Reading documentation. Docpull ingests entire docs sites; an agent reads them and answers questions in context. I was building tools for this before I noticed I was the primary user.
The line I draw, mostly: I do not let the model decide what is worth doing. I let it help me do the work I have already chosen.
The honest reckoning is that yes, some things have gone.
I read worse than I used to. I mean unaided reading — sitting with a difficult page and waiting for it to unfold. I still do this. I do it less. The skill is sharper if you grind on it daily and I am not grinding on it daily.
I notice I reach for the model when I am stuck even when stuck is the right place to be for another twenty minutes. The friction is gone, and friction was sometimes the thing.
There is a version of me that would have struggled with a hard piece of theory for two weeks and emerged with the kind of internalized understanding that only that struggle produces. That version is partly hypothetical and partly real and I am not sure which parts.
This is not free. I am not pretending it is free.
What I got in exchange is leverage at a scale I do not want to give back.
Clarity, docpull, daudio, the theory app, this site, the projects I have not announced — none of these exist in their current shape in a world where I refused. They exist in a smaller, slower shape, or they don't exist at all. The surrender was the precondition for the work, not its enemy.
The scope of what one person can attempt has changed. I am not interested in pretending it hasn't to preserve a self-image from 2021.
Strangelove rides the bomb down. The joke is that he loves it because there is nothing left to do but love it. The surrender is total and the world ends.
Mine is not that. Mine is messier. I gave in to a thing that is genuinely useful and genuinely changing me, and I am not sure yet which of the changes I will be glad of and which I will quietly mourn. I am not going back. I am not finished worrying.
I learned to love the bomb. The bomb has not gone off.