AI Coding Tools Made Us Faster. They Also Made Programming Lonely.

#mentalhealth #career #ai #discuss

When did a human coworker last lean over your desk and say "huh, that's weird, let me look"?

If you had to think about it, two essays published on 9 October 2026 were written about you. One is by Glyph, the other by Fernando Borretti, and together they pulled well over 200 comments on Lobsters and Hacker News. Borretti grants the tools might mean "higher productivity, though at the cost of a messier codebase." Glyph doesn't grant even that. Both argue we're losing what made us good at this job, and I'd add that much of it came from the people around us. We're dropping them because the model replies sooner.

Two essays, one bad feeling

Glyph Lefkowitz, who started Twisted (the Python framework whose Deferred fed into the async/await style you probably write every day), wrote Programming Isn't Special. His question is blunt:

"Artists understand that AI is bad for art. Why don't programmers understand that we are artists?"

He doesn't pretend every function is a masterpiece. "Most code could not make a lady weep," he writes; it's "just copy-writing, if you will." But the mundane work is where skill grows. He puts the chance that any given boring project turns into something great at about 0.1%. If you hand off the boring projects, you also lose the rare one that would have mattered.

Fernando Borretti's No Man Is an Island comes at it from the social side, and I think it's the sharper of the two:

"individual intellectual activity can only be sustained in an intellectual community of other humans"

He planned to let agents do the job work and keep his intellectual hobbies for fun. In his words, that second half "has not quite worked out." Software discourse got "narrower, shallower, and more repetitive." Write a blog post and its most reliable reader is the next training run, which will ingest it, "marginally improving AI capabilities." His phrase for this is "your own private garden of code", where you never have to go to the bazaar to trade with people.

What asking a person used to get you

Asking a person for help used to be inefficient, and it gave you a lot more than the answer:

You learned who on the team knew the payment code.
They learned that you were stuck, which is how mentoring starts.
The awkward "can I grab you for five minutes" built the trust you needed later, when production was on fire.
Someone reviewed your PR and left one comment about naming that you still remember three jobs later.

A model gives you the answer and none of the rest. An HN commenter, xg15, put it in the thread (316 points and 206 comments when I read it): "Ask the AI. Instant answer, question solved, but also, nothing else."

Another HN commenter, arctic-true, called the API "a digital cloister". The early web was shared by default: you posted a question and strangers saw it, answered it and argued about it. An LLM session is private by default. Nobody else learns anything from the question you asked it.

The comments pushed back, and they have a point

Neither comment section simply agreed, and you shouldn't either. michaelhoney on HN said what plenty of people think: "Stack Overflow died in large part because of the gatekeeping assholes." I can't argue with that. "Closed as duplicate" taught a whole generation to never ask anything in public. If your choice was a smug senior developer or a patient model, the model wins, and I won't pretend otherwise.

On Lobsters, nemin argued that Borretti "picked a conclusion that hasn't come to pass yet." Engineers still own and maintain the code. singpolyma's view was that programming is "just a job after all."

On the Glyph thread, kornel pushed back on the art framing: "You're conflating skill with artistry, and quality with art." In that view, most code exists to solve a problem, the same way most paint exists to keep the rain off a house.

All fair, but none of it touches what actually worries me, and whether code counts as art barely matters to it. The scaffolding we used to turn juniors into seniors was made of other people, and we're taking it down without much discussion because the replacement replies faster. The best reply in the Borretti thread came from ocharles, objecting to the "just a job" line: "Why care about nice type systems when the all-seeing LLM just gets everything right?" If that sounds absurd to you, good. Somebody still has to care, and caring is something people learn from other people.

What you get faster, and what goes missing

I'm not going to tell you to uninstall your agent. The speed is real, and pretending it isn't is how you lose the argument with your manager. Here's how I'd split it:

What you get faster	What goes missing
Boilerplate, migrations, test scaffolding	The junior who used to write it and learn from it
Answers to "why does this throw"	The colleague who'd have said "oh, we hit that in 2023"
PRs that pass CI on the first push	Review threads where someone teaches you something
Shipping alone	Anyone else knowing how the thing works

The right-hand column never appears in a velocity metric, so no one will defend it unless you do.

Check how alone you've been shipping

Here's a one-liner that counts human co-authors on your recent commits. It skips AI co-author trailers, since agent tools add those too:

git log --since="90 days ago" \
  --format='%(trailers:key=Co-authored-by,valueonly)' \
  | grep . \
  | grep -viE 'claude|copilot|cursor|\[bot\]' \
  | sort | uniq -c | sort -rn

I tested it on a scratch repo with git 2.47.3. Trailer key matching is case-insensitive, so Co-Authored-By and Co-authored-by both count. Compare that with git log --since="90 days ago" --oneline | wc -l. If the first number is close to zero and the second is large, you've been shipping alone, and only you can say whether that's a problem.

If it is, these are the cheapest fixes I know:

Ask a person first, at least once a day. Before you open the chat panel, post the question in the team channel. Then answer it yourself if nobody replies in 15 minutes.
Pair with a human for one hour a week. Pick a messy bug for it, since that's where people end up talking.
Leave a real review comment on every PR you approve. One sentence about why. A bare "LGTM" doesn't count.
When the agent solves something non-obvious, write it up where people can read it: a TIL in the repo, a team wiki page or a blog post. That's the public part that private chats took away.
Add Co-authored-by: when a human actually helped. It's a small act of credit. Borretti's complaint is that the models take your public work and solve someone's problem with it "without crediting you"; at least inside your own repo you can do better.

So why do we still do it?

Glyph's answer is that "Programming isn't special": it's "just Art", like everything else people make. I'm not sure I'd go that far. Borretti is closer to my view: people mostly kept programming because other people were around. You read their code, they read yours, someone said "nice" in a review. Take that away and, in his words, "you don't get isolated individuals each working on their own things: you get nothing."

I don't think we're there yet. Still, it's worth noticing how comfortable it has become to never bother anyone.

So tell me: when did you last ask a human instead of an AI for help with code, and do you miss it? And if you think this is nostalgia for a gatekept past that deserved to die, say so in the comments.