HootersMcBoobies
kiwifarms.net
- Joined
- Jan 7, 2019
Vibe coding is really dumb. Spec code instead. Write your entire requirement doc in a n.md file, tell the prompt to read your AGENT.md or rules.md or whatever and then "implement spec/09-whatever-thing.md"
I wouldn't call this reproducible. But I think it's important to have the prompt checked in. You can also tell it to commit the summary with it, and maybe the token count/cost too if you use something like opencode that measures it.
For my own stuff, I write specs and have the AI demons write stuff all the time. For work ... we inherited a project from a guy who left who LOVED some Claude .. old claude .. shit claude. He checked in all his specs, but there is so much terribly written repeated buggy shit ... I always suspected engineers would have to be rehired en-mass to fix the slop ... but the newer models (opus-5+, gpt astra or hell even GLM-5 or Kimi) are .. pretty good at finding a lot of those slop code issues. I've been relying on them more than I'd like.
The thing is though .. you have to know how to code first. My cognitive coding skills are in decline. I can feel it. I have to force myself to write simple things like Ansible playbooks, just to make sure I'm not losing a lot of critical skills.
I do not know at all how you teach and bring up new developers in this environment. If you don't know how to structure what you want and just give a harness the most basic requirements ... it's just going to build node + react. We're on our way to monoculture. Who will use new frameworks?
The worst are these mid-levels I've had to interview lately who read straight of the fucking phone! I ask them weird esoteric shit like what garbage collectors are in .NET 10 vs .NET 6 and how you would implement cgroups on macos. They start answering! I asked this one pajeet cunt "What LLM are you using?" She said she wasn't using one and looked at her dead in the camera and said "Please do not lie to me. What. Tool. Are. You. Using." She still denied it. After the interview, the other manager said "You almost made her cry!" "What? No I didn't" "Yes you did. Did you not pay attention?"
My manager then said, "Please don't make the candidates cry."
We always do an in-person and we've had several that answered perfectly on the phone screening and knew zero when put in a room where they were forced to use our laptop. I guess it's impressive they can read and regurgitate LLM answers that fast, but they're still fucking lying pajeets.
I've asked around and this is happening everywhere. One of my old managers works for a fully remote shop and they rented out a room and a PO box to fly out candidates and staff to a make shift office for in-person interviews, which also feels dodgy for totally different reasons.
Do any Europeans or Americans do this? I haven't seen it yet. Most of the American candidates are just ... incompetent, but at least they don't try to cheat.
I don't know how we solve the education problem. I don't know how we solve the Pajeet interview problem. I don't know how you grow a new generation of domain experts when the predictive machine spits out code from a massive lossy-compressed store of human creations.
I do know none of this is sustainable, and I'm not sure what we'll see in 10 years.
I don't think DHH was wrong .. but I still didn't like what he was saying.
I wouldn't call this reproducible. But I think it's important to have the prompt checked in. You can also tell it to commit the summary with it, and maybe the token count/cost too if you use something like opencode that measures it.
For my own stuff, I write specs and have the AI demons write stuff all the time. For work ... we inherited a project from a guy who left who LOVED some Claude .. old claude .. shit claude. He checked in all his specs, but there is so much terribly written repeated buggy shit ... I always suspected engineers would have to be rehired en-mass to fix the slop ... but the newer models (opus-5+, gpt astra or hell even GLM-5 or Kimi) are .. pretty good at finding a lot of those slop code issues. I've been relying on them more than I'd like.
The thing is though .. you have to know how to code first. My cognitive coding skills are in decline. I can feel it. I have to force myself to write simple things like Ansible playbooks, just to make sure I'm not losing a lot of critical skills.
I do not know at all how you teach and bring up new developers in this environment. If you don't know how to structure what you want and just give a harness the most basic requirements ... it's just going to build node + react. We're on our way to monoculture. Who will use new frameworks?
The worst are these mid-levels I've had to interview lately who read straight of the fucking phone! I ask them weird esoteric shit like what garbage collectors are in .NET 10 vs .NET 6 and how you would implement cgroups on macos. They start answering! I asked this one pajeet cunt "What LLM are you using?" She said she wasn't using one and looked at her dead in the camera and said "Please do not lie to me. What. Tool. Are. You. Using." She still denied it. After the interview, the other manager said "You almost made her cry!" "What? No I didn't" "Yes you did. Did you not pay attention?"
My manager then said, "Please don't make the candidates cry."
We always do an in-person and we've had several that answered perfectly on the phone screening and knew zero when put in a room where they were forced to use our laptop. I guess it's impressive they can read and regurgitate LLM answers that fast, but they're still fucking lying pajeets.
I've asked around and this is happening everywhere. One of my old managers works for a fully remote shop and they rented out a room and a PO box to fly out candidates and staff to a make shift office for in-person interviews, which also feels dodgy for totally different reasons.
Do any Europeans or Americans do this? I haven't seen it yet. Most of the American candidates are just ... incompetent, but at least they don't try to cheat.
I don't know how we solve the education problem. I don't know how we solve the Pajeet interview problem. I don't know how you grow a new generation of domain experts when the predictive machine spits out code from a massive lossy-compressed store of human creations.
I do know none of this is sustainable, and I'm not sure what we'll see in 10 years.
I don't think DHH was wrong .. but I still didn't like what he was saying.
