The author describes a slow realization that Copilot and Claude haven't measurably made him faster — they've made him feel faster while eroding the muscles he used to rely on: reading unfamiliar code without auto-explanation, holding a system in his head, and debugging without a chat window open. His cancellation is framed as a self-imposed intervention against a habit he can no longer separate from the work itself.
The editorial frames the post's resonance as evidence of a shifting sentiment — six months ago this would have been dismissed as Luddite cope, but today the loudest voices echoing the concern are former power users describing a creeping inability to start a function without prompting and a drop in codebase retention. The shift challenges the industry assumption that AI adoption is a one-way ratchet.
A faction in the HN thread dismissed the post as a personal discipline problem rather than a structural critique of AI tools. Their argument: competent developers can use AI assistants without losing their underlying skills, and blaming the tool for atrophy is a rationalization for individual workflow failure.
The editorial argues the interesting datum isn't the post itself but the upvote curve and who is amplifying it — the exact demographic vendors like Anthropic, Cursor, and GitHub cite as their core retained user. If experienced shipping engineers are voluntarily churning rather than just juniors who never built the skills, the monotonic-retention model underlying the AI coding market starts to crack.
On May 31, an individual developer posted a short essay titled *The solution might be cancelling my AI subscription* to thoughts.hmmz.org. By the afternoon it sat at 228 points on Hacker News with hundreds of comments — not because the argument was novel, but because the author was the demographic AI vendors keep citing as their core user: a mid-to-senior practitioner who actually ships code for a living.
The post is not a manifesto. It's closer to a diary entry. The author describes the slow realization that Copilot and Claude have not made him faster in any way he can measure — they've made him *feel* faster while quietly eroding the muscles he used to rely on: reading unfamiliar code without auto-explanation, holding a system in his head, debugging without a chat window open. The cancellation isn't a protest against AI; it's a self-imposed intervention against a habit he can no longer separate from the work itself.
The HN thread split predictably. One camp called it a skill issue dressed up as principle. The other — larger, by upvote weight — described the same pattern in their own workflow: a creeping inability to start a function without prompting, a drop in retention of the codebase, a feeling that the tool is doing the thinking and the human is doing the typing.
The interesting datum is not the post. It's the upvote curve. Six months ago this essay would have been dismissed in the comments as Luddite cope. Today it's the top story on the orange site, and the loudest voices in the replies are not skeptics but former power users.
That shift matters because the AI coding-tools market is built on a specific assumption: that adoption is a one-way ratchet. The pitch decks at Anthropic, Cursor, and GitHub all model retention as monotonic — once a developer integrates AI into their loop, they don't leave. The churn cohort that *does* exist is assumed to be junior devs who never built the underlying skills, or non-engineers who were never the target buyer.
This post is evidence the assumption is wrong on both ends. The people starting to leave are the *senior* developers — the ones with the strongest baseline, the clearest before/after comparison, and the most credibility when they describe what they're losing. They are also, not coincidentally, the people who write the blog posts that move the discourse. A junior dev quietly canceling Cursor produces no signal. A staff engineer writing about why he canceled Claude produces 228 HN points and a comment thread full of "same."
The second-order effect is what vendors should be modeling. If the canonical "AI made me 10x" testimonial gets quietly replaced over the next year with the canonical "I had to cancel to think again" essay, the social proof flywheel reverses direction. Adoption among the curious continues, but the prestige tier starts leaking — and prestige is what convinced the curious in the first place.
None of this means the tools don't work. The author concedes, repeatedly, that AI is faster on a per-task basis. His objection is that the per-task win is masking a per-career loss: the gradual atrophy of the capacity to do hard cognitive work without scaffolding. The comparison he reaches for is the GPS one — a generation of drivers who can no longer navigate a city they've lived in for ten years. He's asking whether we want a generation of engineers who can ship features but can no longer reason about a system without a chat window open.
The vendors have a counter, and it's reasonable: the calculator didn't destroy arithmetic, the IDE didn't destroy programming, and every productivity tool gets the same complaint at adoption. The counter-counter, which the post makes implicitly, is that the calculator doesn't *write the equation for you*. The line between augmentation and replacement is where the disagreement lives, and senior devs who use these tools daily are the only people qualified to draw it. Their verdict is starting to come in.
If you manage an engineering org, the operational question is no longer "how do we increase AI tool adoption." It's "how do we measure the second derivative." Track not just lines shipped per dev, but the share of PRs where the author can explain, unprompted, why a specific line is there. That metric is uncomfortable to instrument and politically radioactive to publish, which is exactly why it's the one worth watching.
For individual practitioners, the post suggests a concrete experiment, not a religious conversion: pick a two-week window, turn off the autocomplete, keep the chat closed unless you're genuinely stuck, and notice what comes back. The author isn't claiming the tools are bad. He's claiming the off-switch is information, and most developers haven't pulled it in eighteen months. The cost of the experiment is one sprint of slower velocity. The upside is knowing whether the velocity was ever real.
For vendors, the strategic implication is that the next wave of churn won't look like the first. It won't be junior devs quietly hitting cancel. It will be senior devs writing essays about why, and those essays will compound. The product response — if there is one — has to address the craft-regression complaint directly, not just ship more features. "Co-pilot, not auto-pilot" was a marketing slogan in 2022. By 2026 it has to be a measurable product mode.
The next twelve months are when the AI coding tools graduate from "is it useful" to "is it sustainable." The first question is settled — they are. The second is open, and the people answering it are the same people who built the rankings of which tools are worth using in the first place. Watch the senior-dev churn cohort. If the cancellation essays keep landing on the front page, the vendors have a narrative problem that no benchmark can fix.
I don't thing the problem is AI, but the mindset and trainning. I have probably as many or more AI projects that this man has but they are extremely useful, even if most of them I won't even sell.This is like a kid playing videogames instead of studying, you take the console away and force
Only mentioning this because the OP did - but for me (also ADHD) it's kind of the opposite. I'm finishing side projects for the first time ever because I can actually get them working before I get bored of them. My projects are more infra-leaning, and not all of them get much use, but some
Wow. To me the point of code has always been the crazy ideas and playing around. I love to create just for me and every once in a while for others is ok too. If you only think of code as 'a tool to build useful things' and everything else as wasted then sure, this is the philosophy for you
I wonder how many of the responses here bifurcate by age. The post resonates with me, but I am now in my early fifties. When I was in my 20's and 30's, I would have happily chased rabbits down all those holes, but now that time seems so brutally finite, I feel that anything encouring me to
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
I can provide a data point for what the article calls pseudo productivity: I extensively use LLMs as semantic search engines or expert systems (but not as agents). Recently I asked one how to consume a Google Pub/Sub topic using Python (context: I come from an C++/Java/JS background w