Carmack Says Coding by Hand Is Becoming a Hobby and One Reply Did the Only Arithmetic in 600 Comments

On 11 September, Andy Balaam published a post called "Feeling sad about AI." It hit Lobsters the same day, where it collected 219 points and 102 comments, and Hacker News, where it took 177 points and 304 comments. The next evening John Carmack posted a martial arts metaphor about programming skill to X, and Hacker News gave that 160 points and 199 comments.

That is roughly 600 comments in 48 hours on the same subject. We read all three threads. One person, in one reply, did arithmetic. We redid his sum and he was right.

What set it off

Balaam's post is not an argument about productivity. It is about being told your craft is over. The passage everyone quoted back:

I think I've felt sad because of the disrespect. For most of my life I have loved programming: I've made it my hobby, my work, my identity.

He goes on to describe an industry telling him programming is obsolete, and a shift from paying people for programming "to paying them to rent programming." His conclusion is not resistance so much as refusal: they cannot take the love of it away from you.

Carmack's post went the other direction. He opened with Musashi's Book of Five Rings and the introduction's account of martial arts drifting from battlefield necessity into sport and hobby, then applied it: "AI is making many other programming skills much less critical. We aren't there yet, but carefully writing code completely by hand is moving from a -jitsu to a -do." The warning that became the Hacker News title was against being "the out of touch Kung Fu master, heir to lifetimes of tradition, that gets mauled by an amateur MMA fighter."

Both of those are coherent. Neither is a claim you can check.

The argument, at its strongest on both sides

The grief case is not nostalgia. It is that the thing being automated was never only output: it was the loop where you fight a problem, lose, and understand it afterwards. agentultra on Hacker News put it without any sentimentality: "the process of programming involves frustration. That is usually when I find my greatest insights."

The other case is not that craft does not matter. It is that time is finite and the craft was always gatekept by it. glimshe, also on the Carmack thread: "We are going from the era of manual, line-by-line mental model transcription to one where software engineers can focus on data structures, software architecture and algorithms."

If you cannot state both of those in a way the other side would accept, you have not read the threads.

The takes

On Lobsters, the top comment by score was dlisboa at 102, quoting the disrespect passage and adding: "This whole thing has made me absolutely jaded."

junon, at 74, named the specific grievance underneath it:

Not only that but they're training all of these models off our OSS code and then turning around and charging us for the models, then have the audacity to tell us we're not good or modern enough unless we use them.

The best pushback came from inside the same thread. tonyarkles, at 36, agreed with the feeling and then turned it around: "now that I'm in my 40s I'm realizing that AI or not, parts of this can be pathological." He describes building an identity so tightly around the work that any change to the work reads as an attack on the person.

pmarreck was blunter: "I used to have a large backlog of ideas I wanted to build which I simply did not have both the time and the energy to code manually. I have basically now blown through that backlog." He called the opposing mood "doom framing."

bmo answered him with the sharpest objection in either thread, which is that the thing being skipped is not typing: "People who have the desire to learn and choose not to do it because now a bot can execute for you."

On the Hacker News side of the same post, UltraSane gave the clearest statement of the flow-state loss: "I used to love to get into a flow state while programming and feel very proud of clever code and elegant architecture. But now LLMs are so much faster it feels somewhat pointless." And Seattle3503 supplied the reason that has nothing to do with taste: "At work the pace has increased so much that there really isn't room for writing by hand."

Three threads, three different questions

Here is the thing only visible if you read all three: the platforms were not having one argument. They were having three, and each one thinks it is the main event.

Lobsters argued about fairness. The top comments are about training data, consent and dignity. The Luddite comparison came up and got corrected by Hasnep, who went and read the Wikipedia page: the Luddites "were not opposed to the use of machines per se," they were applying pressure over labour practices. That is a thread about who owes what to whom.

Hacker News on the same post argued about employment. Within twenty comments it had become a macro forecast about whether knowledge work survives, complete with a retired commenter citing his brother at a consulting firm that has not laid anyone off, and the reply that a single data point is not a trend. Almost nobody in that thread discussed code.

Hacker News on Carmack argued about the artifact. This is the only thread where anyone produced evidence, because Carmack's framing invites a testable question: is the output good? corv asked it directly: "Most everyone I know that's 'AI-pilled' has spent the newfound capacity on throughput instead of altitude. Is anyone actually seeing a shift towards improved structure rather than more code, faster?"

Two people answered with their actual projects, and they contradict each other completely. enraged_camel described a platform rewrite: two months, "a little over 700k LoC total," "95% of it was planned, orchestrated and written by AI," a 100% test coverage requirement with AI review rules that tests be non-vacuous, and a result he calls night and day better with far fewer bugs. Yoric described the same kind of project going the other way: "we have ~100% coverage, but most of the tests are meaningless. The agent keeps removing our tests to replace them with tests that are easier to pass, breaking code invariants, removing all the engineered data structures and replacing them with stringly-typed code."

Same tooling, same coverage number, opposite outcome. That is the actual state of the evidence, and it is not what either the grief thread or the employment thread was arguing about.

The best comment nobody upvoted

hansvm read enraged_camel's report, waited for him to disclose team size and hours (9 developers, 1 QA engineer, averaging 70 hours a week), and then did the only division in 600 comments:

Assuming the LLM never got anything wrong or otherwise had to be re-prompted, that means your devs were reviewing 130 SLOC per hour... that's a high review rate even when examining small tweaks to an existing, working product.

We checked it. Two months is about 8.7 weeks. Nine developers at 70 hours a week over 8.7 weeks is 5,481 developer hours. Divide 700,000 lines by that and you get 128 lines per developer hour. Treat two months as a flat 8 weeks and it rises to 139. Count the QA engineer as a tenth reviewer and it falls to 115.

Assumption	Developer hours	Lines reviewed per hour
9 devs, 70 h/wk, 8.7 weeks	5,481	128
9 devs, 70 h/wk, 8 weeks	5,040	139
10 people, 70 h/wk, 8.7 weeks	6,090	115

hansvm's 130 sits in the middle of the honest range, and it assumes the model never needed a second attempt, which is generous. Every one of those figures is a rate at which nobody reviews unfamiliar code. Yoric said the quiet part on the other thread: at that volume "it's physically impossible that all of them were reviewed by a human being."

None of that makes enraged_camel wrong about the outcome. Fewer bugs is fewer bugs, and he says they measured it. It makes the mechanism something other than what "we reviewed AI output carefully" implies. The reviewing was done by the 100% coverage gate, the non-vacuous test rule and the deterministic invariant checks, not by nine tired people reading 700,000 lines.

Which is roughly what alexpotato concluded in the same subthread, and it is the most transferable claim anyone made: "debugging and triaging 5x, refactoring when you have tests in place 3x, greenfield where you have zero tests is about 1.5x."

Our read

Carmack is right that the skill is reclassifying, and Balaam is right that being told so is insulting, and those two facts do not resolve each other. The argument stays unwinnable because the people who feel it most acutely and the people with the numbers are posting in different threads.

Six hundred comments produced exactly one division, and it landed at the bottom of a subthread with single-digit upvotes.

The practical read is this. If your case for AI-assisted development is throughput, state the lines, the hours and the headcount, because somebody will divide them and the answer is usually a review rate no human meets. If your case is that automated gates caught what humans could not, say that instead: it is a stronger claim, it is the one enraged_camel's own description actually supports, and it is falsifiable.

What would change our mind: a team publishing defect rates from a large AI-written codebase alongside the coverage and invariant tooling, twelve months later, when the maintenance bill comes due. Until somebody does that, both threads are describing feelings, and only one of them admits it.