in this second post i wanted to speak briefly about the arrival of artificial intelligence into a person’s need to produce. i’m talking about myself, but plenty of people reading this will see themselves in it.
by nature, i’m someone with a certain propensity to feel the need to produce. i’m aware it’s a feeling society has, in a sense, imposed on me for being a straight white man who’s supposed to succeed, and that success means producing, contributing to the economy, having a company, launching ventures. knowing that doesn’t make the feeling go away. i have many ideas, many things interest me at once; i try to organise myself to gradually execute the ones that pull at me hardest at any given moment. i drop many of them: it isn’t only productivity, it’s enormous curiosity to explore, and when i get tired i let it go.
where am i going with this? to how the arrival of artificial intelligence has pushed all of it to a limit that borders on unhealthy. i’m a software engineer — i studied physics and chemistry, but i work as a software engineer. i’ve always had several ideas for projects i’d have liked to build, and because of my engineering background they were remote possibilities: i had the aptitude to build an app, a website, a specific piece of software, but often i didn’t have the time or the resources.
the arrival of artificial intelligence is a paradigm shift. all of a sudden it enables me to pursue every avenue i’d kept on the back burner. it lets me make a quantitative leap: all those projects sitting at the back of my mind as a remote possibility start to take shape. i can now work on six or seven projects in parallel that i once thought i’d never find time for. and that delivers a neurological reward: feeling that heightened productivity, doing things i thought impossible, feels great.
but at the same time, that brutal capacity to create enslaves you. it’s great on the one hand, but i feel i have access to tools so powerful that every minute i’m not pouring into them feels like wasted time. what it gives me, it takes from me. that rush of dopamine — the same reward circuit that drugs and social media exploit — is at the same time a comedown, because it generates the obligation to be hooked all day long. and it binds you to the point that right now i’m working on six projects in parallel, trying to dedicate every minute of the day i don’t spend on my job.
and what’s the most serious problem? that it generates a rush, a need to run and run so fast you lose sight of the projects themselves. in my case i use claude code. you start by paying the twenty-euro subscription, you use the sonnet model, you see how it works, it feels great. with those limits you can work on one project at a time: a session is easy to monitor, to follow, to steer, and it speeds things up a lot. you supervise it, you see what it does. but we’re lazy, we tend to procrastinate, to minimise effort. and bit by bit you lose focus, until you’re not even supervising it any more: you accept every change by default and skim what’s happening. it turns into a snowball. you get the high of “look how fast i’m going”, but it’s a false sensation, because there comes a point where you’re doing nothing more than hitting accept.
and this is the real danger. the sensation is so addictive you think: “if i pay the most expensive subscription, i’ll have higher limits, i’ll be able to do more”. you try the two-hundred-euro-a-month one. before paying for it you try opus, the most powerful model, and in just three or four prompts you burn through the five-hour window. but it lasts long enough for you to see a brutal jump in quality between the sonnet you’d been using and opus. and at that point, yes, you go for the two-hundred-euro plan, because it’ll let you use opus more.
it’s the closest thing to a heroin hit i’m ever going to experience.
suddenly you’re using only opus, which programs far better than i do, and i’m only supervising in case it drifts off course. you start to see how much you can get done in a five-hour window, and you try two, three, four, five sessions in parallel. you find yourself with five terminals open, five claude opus sessions working at once for five hours. it’s madness.
working on a single project with five sessions is hard, even with worktrees, because of the merge conflicts. so you end up working on three or four different projects at once, with opus practically on autopilot all day long. you barely review anything. you accept everything, because it’s impossible to review. and the most critical thing: you’re constantly switching focus from one project to another. you fire off a prompt, it runs, you look it over, you steer it, you switch to the other project because it has already finished its task, and on and on for hours. a while back, on an older blog more focused on programming, i talked about how programming with ai had become “tiktokified”: rapid swipes, constant hits of dopamine. and it feels exactly like that. from a productivity standpoint it’s a tiktok of production: i keep jumping between sessions, checking if they’ve finished, pushing them along.
the problem? those projects were mine, my own ideas. before — eight, nine, ten months ago — the workflow let us supervise everything. i had a clear picture in my head: the architecture, the file structure, the use cases, the entities, the database tables. i knew what was right and what was wrong, and if there was a problem i knew exactly which file it was in. i was the owner of it all, i had context on everything. now it’s the opposite: i have five complete projects and i haven’t written a single line of code. it’s like saying i’ve written five books without having read them. i have no idea what they’re about; i know what my idea was, but i haven’t read them. and this is the great problem. i could read them, sure, but how can i, when that slows me down? what i want is to make the most of every second of that five-hour window. there’s no time to stop and read, to test. normally we’d test each feature, each fix; we’d do manual testing, we’d validate. now there’s no time: changes pile on top of changes until you no longer know where you are, what works, what’s been fixed, what’s still pending. you lose the pointer.
that’s why i say it enslaves. artificial intelligence is opening so many doors at once that it has enslaved us to spend all day opening one door after another, without thinking about whether we should. many of these things are within my reach: i could change them, handle them differently. but that feeling forces me not to stop for a second. and what really worries me is what’s going to happen when this ends. today you pay a flat rate of two hundred euros; with those limits, working two or three five-hour windows a day at full throttle, you burn through them. but the real cost to the company is in the thousands. if they charged me what it really costs them, i’d be paying thousands. perhaps it’s marketing — i don’t think it’s only marketing, but there’s something to it. as long as they offer me that flat rate, i feel i have to make the most of it down to the last minute, and i regret every month i went without paying for it. and the trend is not to go slower, but faster: opus 4.6 comes out and i go back over what’s already done to see if it finds any bugs; it ships with a million-token context window; opus 4.7 comes out with different effort levels, and i push it hard. what’ll be next?
this is like the space race, to see who’d reach the moon first, with no idea where it was all going to end. but the ai race is that one cubed: it’s no longer one country against another, it’s companies against companies against countries. and there’s no clear winner: gemini, chatgpt, claude, they’re all going strong. when is this going to stop? at some point it’ll have to. but until it does, it’s madness. and the real victims are us. and the environment.
