Sure, there are times in a familiar codebase that it would take Claude or any model longer to get its bearings and chew on all the random nonsense it wants to think about even for small changes than it takes me to open vim and make a one line change.
And, sometimes, I know models will write it in a way I don't like, over-engineering is common.
But, I mostly use models for more than few lines of code or code that isn't in a familiar project.
I am building Aphorio. I am wrapping up work on a new universal proxy rule convention designed for simplicity, speed, and maximum flexibility. It already has the fastest WebSocket library that can push messages at a speed of 6,000,000 per sound on a single socket in a memory unsafe way or 800,000 messages per sound in a more conventional way.
The app is entirely written by a human without AI in TypeScript. I am writing this because data integration is a ridiculously expensive commercial market ripe for disruption. There are so many companies that get hooked into automation solutions like MuleSoft only to bleed cash and beg to get off of it but don’t know how.
I've edited Bash scripts almost every day (including weekends) for decades and I'll continue doing so. Bash is the language I use to communicate with computers. I don't differentiate between writing Bash scripts and using Bash interactively. It's all the same. For me, using AI for Bash scripting would be like using AI to write emails.
In my experience, very few people who don't write shell scripts (Bash, Zsh, POSIX, etc.) know how to use a Unix shell effectively. For example, if a person never writes loops when using a Unix shell interactively, they have not yet achieved personhood.
Yes. I work on legacy systems where the training data was minimal. The code that I get back from the LLMs is confused - it pulls objects and methods from other languages, makes up APIs that don't exist, etc. It also doesn't understand the bigger picture of what all is possible in the system, so it gives me over-engineered code that could be done really simply using other tools.
Yes, mostly at work. But that's mainly because we're an MSP and our clients have a no-AI policy, but even otherwise, I don't imagine our clients will be very happy with us running agents against their systems/data or introduce unvetted slop into the environment.
At home, I manually create utility apps for personal and family use using niche languages like GAMBAS (a VB6 clone), and I've found that AI isn't really good at it. Besides, with the kind of simple utilities I make, it's far more easier to just do it myself than to fight with the AI (and then there's also the issue of long-term maintenance - if it's something I coded, it's far more easier for me to fix it, update it etc than to rely on AI for eternity).
I've designed a few languages of my own, it doesn’t get more niche than that. And models handle them surprisingly well, even ones from a year ago. The trick is to give the model a short cheatsheet of the syntax and some link to docs. Today I would create a skill for that. Might be worth trying with GAMBAS too.
It is sad that GAMBAS is niche these days, it was very popular back in 2000s.
I build erp addons and module rewrites for clients and I barely type code anymore. Your clients' worry is right one, but in my experience it's also solvable usually. I've built tools to allow only read-only access to client ERPs and other enterprise systems and also giving agent's their own credentials and identity so work done by them across the application landscape is auditable.
Because they're quick/easy/fun. My main requirement was a RAD environment to quickly spin up GUI apps for my family - apps which are reasonably small and memory efficient (so no Java / Electron etc).
In the case of GAMBAS, I originally picked it up because I missed the days I used to code in VB6. I recall chruning out small utilities like it was no big deal (and I was still in school at the time)! So GAMBAS, as a spiritual successor to VB6, was a natural choice for me - and it's also cross-platform, which is great because some of my family members use Linux, whereas others use Windows.
Before GAMBAS I used to use Delphi, which had a similar RAD environment with Lazarus. But Lazarus is a bit clunky, compared to the polished experience of Embarcadero Delphi (which is Windows-only, unless you pay for cross-platform). So when GAMBAS came along, I first played around with it just for fun, but then over the course of time it became pretty solid and capable, so I switched to it full-time for my GUI needs.
I don't code at all now, but I did before, in trading and algo development. I was never a very good coder, but I put in the effort to get the job done. Later I hired and worked with developers on algos and software, until AI coding got good enough that my own output with AI overtook the devs I worked with. It didn't happen right away. It took me a while to get up to speed, alongside the models getting better, but that's been the real transition for me.
Could you please clarify your premise and thought process? This is not making sense to me. I wanted to answer "yes", but I don't understand the question on a second look.
I appreciate that the OP is asking hn, but I'm curious sbout selection bias here.
I'm writing slightly less code thsn I used to - but only slightly. The difference is mostly tests and refactors and semi-boilerplate stuff. I still live in an IDE.
So am I just a dinosaur at this point, and the entire sw industry has just relocated to agent harnesses and software factories? Because I see comments on hn quite regularly where people claim not to have written code for six months or whatever, and give the impression that traditional coding is anachronistic.
Or are there, in fact, plenty of (very many) people outside the hn bubble who are "still writing code" for various reasons - and who probably will be for quite some time?
Sure, there are times in a familiar codebase that it would take Claude or any model longer to get its bearings and chew on all the random nonsense it wants to think about even for small changes than it takes me to open vim and make a one line change.
And, sometimes, I know models will write it in a way I don't like, over-engineering is common.
But, I mostly use models for more than few lines of code or code that isn't in a familiar project.
Yeah that’s why I put in “meaningful”, I’ve done one liners but these days I barely do that, it’s almost easier to ask the model to find the files
I am building Aphorio. I am wrapping up work on a new universal proxy rule convention designed for simplicity, speed, and maximum flexibility. It already has the fastest WebSocket library that can push messages at a speed of 6,000,000 per sound on a single socket in a memory unsafe way or 800,000 messages per sound in a more conventional way.
The app is entirely written by a human without AI in TypeScript. I am writing this because data integration is a ridiculously expensive commercial market ripe for disruption. There are so many companies that get hooked into automation solutions like MuleSoft only to bleed cash and beg to get off of it but don’t know how.
I've edited Bash scripts almost every day (including weekends) for decades and I'll continue doing so. Bash is the language I use to communicate with computers. I don't differentiate between writing Bash scripts and using Bash interactively. It's all the same. For me, using AI for Bash scripting would be like using AI to write emails.
In my experience, very few people who don't write shell scripts (Bash, Zsh, POSIX, etc.) know how to use a Unix shell effectively. For example, if a person never writes loops when using a Unix shell interactively, they have not yet achieved personhood.
Yes. I work on legacy systems where the training data was minimal. The code that I get back from the LLMs is confused - it pulls objects and methods from other languages, makes up APIs that don't exist, etc. It also doesn't understand the bigger picture of what all is possible in the system, so it gives me over-engineered code that could be done really simply using other tools.
Yes. I was also considering Ask HN-ing that we should add a new flag in Who Is Hiring threads, like NOAGENTIC or similar. @dang
> I can’t find a reason to anymore that’s actually necessary
The huge environmental and societal cost of LLMs, maybe?
Why do these questions implying writing code is a thing of the past keep coming?
Anthropic pretend in their S1 filling that they spend $3B in marketing, there's no need to help them really.
All my work and personal coding is still by hand. The software we build just isn't suited to LLM coding.
Sure. There's some problems I'm using it to solve - aggregating RSS feeds from places that don't offer them, mostly, and a series of backup scripts.
The next project is going to be a browser-based character sheet for an RPG game a friend of mine is trying to resurrect.
Yes, mostly at work. But that's mainly because we're an MSP and our clients have a no-AI policy, but even otherwise, I don't imagine our clients will be very happy with us running agents against their systems/data or introduce unvetted slop into the environment.
At home, I manually create utility apps for personal and family use using niche languages like GAMBAS (a VB6 clone), and I've found that AI isn't really good at it. Besides, with the kind of simple utilities I make, it's far more easier to just do it myself than to fight with the AI (and then there's also the issue of long-term maintenance - if it's something I coded, it's far more easier for me to fix it, update it etc than to rely on AI for eternity).
I've designed a few languages of my own, it doesn’t get more niche than that. And models handle them surprisingly well, even ones from a year ago. The trick is to give the model a short cheatsheet of the syntax and some link to docs. Today I would create a skill for that. Might be worth trying with GAMBAS too. It is sad that GAMBAS is niche these days, it was very popular back in 2000s.
I build erp addons and module rewrites for clients and I barely type code anymore. Your clients' worry is right one, but in my experience it's also solvable usually. I've built tools to allow only read-only access to client ERPs and other enterprise systems and also giving agent's their own credentials and identity so work done by them across the application landscape is auditable.
Why use niche languages if you don’t mind me asking?
Because they're quick/easy/fun. My main requirement was a RAD environment to quickly spin up GUI apps for my family - apps which are reasonably small and memory efficient (so no Java / Electron etc).
In the case of GAMBAS, I originally picked it up because I missed the days I used to code in VB6. I recall chruning out small utilities like it was no big deal (and I was still in school at the time)! So GAMBAS, as a spiritual successor to VB6, was a natural choice for me - and it's also cross-platform, which is great because some of my family members use Linux, whereas others use Windows.
Before GAMBAS I used to use Delphi, which had a similar RAD environment with Lazarus. But Lazarus is a bit clunky, compared to the polished experience of Embarcadero Delphi (which is Windows-only, unless you pay for cross-platform). So when GAMBAS came along, I first played around with it just for fun, but then over the course of time it became pretty solid and capable, so I switched to it full-time for my GUI needs.
I don't code at all now, but I did before, in trading and algo development. I was never a very good coder, but I put in the effort to get the job done. Later I hired and worked with developers on algos and software, until AI coding got good enough that my own output with AI overtook the devs I worked with. It didn't happen right away. It took me a while to get up to speed, alongside the models getting better, but that's been the real transition for me.
Could you please clarify your premise and thought process? This is not making sense to me. I wanted to answer "yes", but I don't understand the question on a second look.
I appreciate that the OP is asking hn, but I'm curious sbout selection bias here.
I'm writing slightly less code thsn I used to - but only slightly. The difference is mostly tests and refactors and semi-boilerplate stuff. I still live in an IDE.
So am I just a dinosaur at this point, and the entire sw industry has just relocated to agent harnesses and software factories? Because I see comments on hn quite regularly where people claim not to have written code for six months or whatever, and give the impression that traditional coding is anachronistic.
Or are there, in fact, plenty of (very many) people outside the hn bubble who are "still writing code" for various reasons - and who probably will be for quite some time?
Jobless and interviewing, hence I am writing some code.