Hacker News Reader

Comment by toponijo | Hacker News Reader

Given this, you can conclude that writers should be putting in at least at much effort as readers, whether or not they use an LLM. What really seems to be the problem is writers that don't at least check their own work, and pass that burden onto the readers. This is easier than ever with LLMs.

This is toxic behavior that unfortunately rewards a selfish writer. I'm worried the AI push incentivizes this too much, to where in corporate situations a reader can't say no to doing work for a selfish writer.

View on HN

Its exactly this. I have had a few LLM coding sessions where I reviewed the resulting work and thought "I don't think my team can safely PR this." I then went back and broke it down into smaller PRs, still using LLMs but at a size that is easy to review. And I reviewed the output myself before I asked a reviewer to commit their time.

The problem is that this is increasingly seen as a non-productive workflow slowing everyone else down, so the pressure is growing for writers to just shove massive PRs out the door and reviewers to use LLMs to make that tractable. I suppose those advocates have more faith in LLM output compared to humans than I do.

> I suppose those advocates have more faith in LLM output compared to humans than I do.

Some of this is the funny situation where the faithful will state: "This writes better code than I do!" and miss the irony of: "yes, yes it does"

Some of this is the funny situation where the faithful will state: "This writes better code than I do!" and miss the irony of: "yes, yes it does"

"Blessed are the humble ..."

> "This writes better code than I do!" and miss the irony of: "yes, yes it does"

I guess it depends on what you consider "better". I've tried using LLMs to write code over the past couple of weeks with extremely mixed results.

The LLM certainly writes more interesting code! They like their cute ASCII/unicode animations, don't they?

It definitely writes a lot more code, none of it actually correct but some of it functionally similar to correct code.

If you like lots of code then I guess that's better. I like less code.

I find it can often write correct code but not maintainable, performant, or reviewable code without additional human guidance. The "solution" frequently given is that humans don't need to maintain it anymore so its not actually a problem. But the agent can't be accountable for mistakes, so unless that changes or the risk of a defect is close to zero, one still has to put forth effort to keep the code maintainable.

To be fair, there are plenty of situations where throwaway code is perfectly fine and/or defect risks are low enough to make the trade-off worth it. I don't think a lot of developers are thinking about it in that context, though.

(No unit tests aren't enough)

Yes I basically meant those folks weren't very good developers to begin with and now extrapolating to: "wow this is better than all devs!", when it's more like "it's you, dude"

This sounds awfully like the people who think that self-driving cars and even auto-braking systems will eliminate all accidents, because everyone else is as bad a driver as they are.

Someone pointed out[1] a while ago that LLMs look good at things you are bad at. Which is I think one of the best explanations of why so many people disagree about how good they are at programming. There are a lot of people really bad at programming, and they will look at the output if an LLM and say “Wow, it’s so much better than my code!”

1: https://news.ycombinator.com/item?id=48315309

> They like their cute ASCII/unicode animations, don't they?

One of the few global Cluade directives I have setup is to never use emojis - and it never has, either in chat output or in code. Don't blame the tool when you don't spend 30 seconds configuring it. It's even easier with AI since you don't have to go digging for some obscure .vimrc snippet - it's literally just plain English.

It can write correct code, but it's still really not good at that, at least without lots of prompting. In my experience, that prompting is usually just doing the design/planning work I'd do with or without AI.

I've said this before on here, to much derision as it turns out, but I write my best code in the car. Nowhere near a computer, no distractions, just point the car along the grey stuff and away from the green stuff and think about what I want to write.

Then when I get home, it's just a case of typing it in, which is the bit I'd love to automate away.

My experience with LLMs has been a bit like rubberducking code with someone who's *really* fast at looking stuff up on StackOverflow.

> the pressure is growing for writers to just shove massive PRs out the door and reviewers to use LLMs to make that tractable

Even in these move-fast envs, it should be reasonably apparent for people to realize that the author should be using the LLM to make the PR tractable, not solely using the LLM to shovel out a giant PR + slop PR description.

And the LLMs can often do this - if you ask to restructure or break up a big change differently, they can often make quite reasonable suggestions and help with it. That's just not what you're gonna get if you're lazy. If you want a small LLM-generated change, often you have to start with a big one then ask it to figure out what it can get rid of, since many times it doesn't have perfect model of all the code in it's "head" before it starts spitting stuff out. The big companies have been doing their best to automate this for the last couple of years vs the even-more-blind attempts you used to get, but there's still the issue of the models+tools following generic advice aimed at median codebases vs being intimately familiar with this codebase.

You can go fast without being lazy. And when going fast, in some ways, it's more important than ever to put in that effort to not blowing things up.

It should be but often isn't. There's been a lot of threads on HN where the response to huge PRs wasn't "Don't do that, use AI when authoring better" but "The reviewers are actually the problem, they're missing the AI train". And I see this in industry too.

Thats the thing with giant PRs. They never really needed to be reviewed anyway. In cultures with strong review culture I have worked at, if you send me a thousand line PR and ask me to review it, I will look at the giant blob of text, and immediately fire off a "it's too long, can you cut it into smaller PRs?".

Because I don't trust myself to review a giant PR. It takes too much cognition to properly review it.

And now that people are making PRs with AI, this is even more important. If the AI was good enough to have coded it, please instruct it to make the changes in reviewable chunks.

Isn't this just a continuation of the performance art of the modern corporate environment though? There's an entire industry producing pages of documents which aren't read, aren't responded to, but need to be at least X lines long for anyone to take them "seriously".

Then suddenly LLMs happened and it's like the mask is off: no one's reading them still, but also no one is writing them either.

Which is perhaps a drop in the ocean of the insanity which is "we need you to work on the Jira tasks" as basically a job title.

You're absolutely right! The modern commercial sector has been writing bullshit, bullshit, and more bullshit, and become completely disconnected from the actual outputs of its work. And it has to be, because if only useful work was done, two thirds of the population would be unemployed without benefits and would revolt so they didn't starve.

I've had a guy once reply in an email with a bug report generated by ChatGPT telling me that some piece of software I wrote wasn't working. He just plopped right there the discussion he had with the confirmation bias machine confirming 100% that what he had in front was sending spurious messages. With all the information in the world at its disposal, the LLM did not consider informing the reporter that maybe his code should flush the serial device pipe before starting his processing. I stopped short from facetiously replying to him that maybe he should use another model, since his seems to be broken.

I think that this is the essence of the argument. I spent two weeks of long nights across a few different sessions with millions of tokens generated, to produce a 5 page proof. I think we have come to the age of the aesthetics of curation. At least I don’t feel like I broke this silent contract. The effort you put in driving the torrent of words before distilling it is a new art. First time I tried it and it felt more creative than slop. The judgement nevertheless lies on the eye of the beholder.

Reply on news.ycombinator.com