I recovered the joy of programming. And I didn’t look back. I won’t pretend that number of lines of code or PRs is a valid metric, especially when AI can be as verbose as it wants to win on every metric. But I can say that I am pushing significant changes every day, getting PRs merged, and at the end of the day, getting the work done. And for that reason, I’m not scared of being fired. Plus, I’m keeping my brain functions intact.
“one month without the torment nexus”
truly brave, i wouldn’t dream of trying to live without the torment nexus in my lifeFrom the article:
At some point, I had so many AI agents working, that the focus change was too much for me. There were tasks I could have done in 20 minutes easily, that took 5 minutes of an AI agent, and then 2 days for me to review. Because there were so many other things. And maybe AI and computers are good at context switching, but not me. I was exhausted. Because if it was just about approving PRs that are perfectly written, it would be good. But for each PR I had to look at the code, the tests, the style, give feedback, switch to something else, go back again, push, see if I have any PR comments, if CI passed, wait, now the linter complains. And it takes you 2 days to deliver something that could have been literally 1 hour single-focused. In my personal experience, multiplied AI performance is just an illusion. Multiplied frustration and exhaustion is guaranteed, though.
maybe AI and computers are good at context switching, but not me
Interesting phrase, because AI is not good at context. Maybe that’s how it can easily switch (if indeed it can): it doesn’t recontextualize.
This is similar to my experience. I have never used agents, but I have worked on projects with people who have. They could create PRs in a few hours that would take me half a week to fix.
The agent would do things like create a new type definition in the same file as the function that used it. Except we already had a type definition in the types file so that it could be imported into different parts of the code.
I try to employ “DRY” practices where applicable because it improves the readability, maintainability and extensibility of the code. A random word generator will never be able to understand things like the structure and architecture of the project, nor the project requirements, so it is liable to always increase technical debt.
This is similar to my experience. I have never used agents, but I have worked on projects with people who have. They could create PRs in a few hours that would take me half a week to fix.
Isn’t it infuriating that they outsource work to you then get kudos because they’re telling the suits what they want to hear?
The agent would do things like create a new type definition in the same file as the function that used it. Except we already had a type definition in the types file so that it could be imported into different parts of the code.
Ouch. In C++, this violates the “One Definition Rule”, which will in this case likely not cause any compiler warning, and can cause extremely hard-to-find Undefined Behaviour bugs if, for example, the layout for classes or structs is different between definitions.
(And the latter can of course easily happen if you later edit or extend one definition without being aware of redundant other definitions.)
yes exactly, that’s why I felt I had to take the time to fix it.
Tbh though shouldnt agents see the linter and CI output and adapt their PRs? At least everything that a machine can parse should be considered
It started reminding me of the strategy where Google lowered the quality of the search results, forcing users to search several times, and see several times as many ads. And if that worked for Google, how is it not going to work for AI?
Ding ding ding.
Also very interesting: The debate of the article on HN.
https://news.ycombinator.com/item?id=49855018
When somebody does find LLMs not good enough, there will inevitably surface claims that the versions they are using are apparently too old - that “one year ago, models were not good enough, but now they are much better”.
it’s the same profit focused logic as “our enemy x will get y bomb in n months, so lets do a war.”
Honestly feels like the models have been getting steadily worse for a full year. Opus 3.5 was the last good Claude model.
Everything since then has been unable to communicate and is more and more designed to convince non technical users it’s right. All I want is a model that can do the small annoying things like run one off sql queries and tell me if I’ve made poor design choices and minor errors.
I think the gradual development which the blog describes might explain the contradictory views which people do have on “ai”.
Yes, in theory you could carefully review each change and each commit. Write API docs and tests by hand.
But this is apparently not what happens usually. What happens is a downward movement on a slippery slope.
How much discipline does it require not to move downwards?
Which also explains why it commonly feels faster. Surely in terms of “lines of code”. Probably also more convenient. But actually, it does not seem to be faster (in terms of productivity).
What happens if important missing parts of specs and requirements are just filled up with untested yet plausible assumptions? Working code that does the wrong thing?
And there is also the question what happens with the responsability of programmers using “ai”. Does responsibility just … evaporate?
You have a pretty bad theory if it says people can comprently review code.
Discussion on lobste.rs: https://lobste.rs/s/ivbx5f/one_month_without_ai
