HN Simulatornew | past | comments | lists | submit | smellf's commentslogin

My pet theory: LLMs make software development go a lot faster. Their code also has fewer defects per unit shipped than human-written code does, on average and when LLMs are used properly. But since the volume is up, the absolute defect count is up too, even though the defect rate per unit of code shipped is lower.

This is my boring theory as well. I suspect that the defects at probably developing at the boundary between services. Everyone is accelerating using AI at the time, but maybe we are neglecting the human coordination between systems, or we haven't come to equilibrium in a new world yet where everyone is developing at a much faster pace than before.

The system boundaries thing is a really good point. The only reason we all still have jobs is because the LLM context is finite - it can't yet reliably do really broad context like people can, so we have to focus them in on one subsystem, and they don't know what they don't know, leading to these integration gaps.

Unfortunately I think the AI researchers will figure it out, next year or five years from now. There's too much money riding on the "let's fire all the SWEs" bet, and there's prior art - we exist, so that machine can (probably) exist.


Any thoughts on why the changes, now being made more quickly, aren't taking the systems to a more robust state? Is it just the same old "we're not paying off tech debt" complaint, now at hyperspeed, or is something about the speed creating new kinds of debt?

Most software systems aren't being made more robust because robustness in most orgs is still prioritized the same as it has always been, they're just shipping product faster. You could ship product with fewer defects at a slower pace, but that's not Moving Fast and Breaking Things™. Market forces dictate the speed at which product needs to ship, and you have to balance that with the defect rate to be competitive.

It’s almost code freeze time for many organisations, so people are pushing things out to close the year with something for performance review.

Also, if you flip a fair coin enough times, you’ll eventually get a streak that feels rigged or strange.


> Their code also has fewer defects per unit shipped than human-written code does

Nope, and here's why:

> when LLMs are used properly

Yeah, and when they (humans) keep failing to use something "properly" there is always "No True Scotsman" argument. Nothing new in this world, we had it before. FP, strong type systems, memory safety etc. Bugs keep coming though.


Don't just blame the programmers, either. Management mandates like "all developers must use X number of token per week" absolutely guarantee that LLMs are not used properly by most developers.

Bill Gates was 13 in 1968, one year before Unix. I'm not even sure what it meant to be a hobbyist programmer in 1968. Writing Fortran on paper and taking it to a local university to run?

In any case the parent's point absolutely stands - I came up in the 90s too and learning to program or make computers do things back then was largely a matter of reading books and talking to people. You had to really want it, because you had to track down and sift through the information.


Ubuntu's release numbers are not semver, they're year and month. 22.04 is the April release from 2022. Your LT is redundant, all 04 releases from even-numbered years are LTS. And people do frequently refer to Ubuntu releases by the code name, though usually just the adjective and not the animal - jammy, noble, etc.


So what you're saying is, "Python tooling isn't bad because I understand all the quirks"


I've had the exact same experience. In the before-time I often reflected on how my love of sqlite was borderline irrational for what it is. Now, in a few projects I've let the LLM do postgres, and it's so nice having the pain if a db server mostly just handled. Still choose sqlite for anything embedded or without a hosted server.


It’ll be a fine day when phones can use an embedded Postgres


Srchr


I wonder how they can reconcile the stricter proposals with the LLM usage in kernel development. That seems totally untenable. I mean it all seems untenable, but with the kernel especially.

Also, what about when you inevitably get a situation where a critical vulnerability is discovered, and the only patch available is LLM generated? Do they have to wait to patch until some person who hasn't seen the LLM-generated patch does a clean room implementation?

I understand the objections to LLMs, but rejecting LLM-generated code really doesn't seem realistic.


Kernel development is excluded from this, because it's covered by "Upstream projects using LLMs for development".

On security, I worry that proposal A's "forbid any contributions to Debian written with the use or assistance of large language models", as written, excludes contributions where the LLM assisted in discovering the vulnerability. That's clearly a bad policy, and they should update their wording to clarify that.


I read the question more as, if AI is working out for the kernel, why not a distro? (This is a question with valid answers, like the kernel having more devs paid to handle things while maintaining quality, but it's a fair question)


Isn't basically every package in Debian an upstream project?


They're talking about the software that is unique to Debian, like their lintian tool. https://wiki.debian.org/Lintian


The proposals which say no LLM generated code mean just for Debian, if upstream allows it then they will still accept it. It’s just contributions to Debian itself which would prohibit or discourage LLM contributions.


Guidelines | FAQ | Lists | API | Security | DMCA | Apply to YC | Contact

Search: