All AI does for me is to give me more work to do. Its code is terrible. It has horrible taste. Its work is not modelled after the task, but after “must look complicated”
Every now and then, when zeal comes upon them, my bosses email me a bunch of risk assessments and policy documents that have obviously been written by AI, probably just from one or two prompts. I have actual work to do, I don’t have time to be reading through that shit - so I get AI to summarise it for me.
For this we are boiling the oceans. Sometimes I think our species is just one big Darwin award waiting to happen.
Thx, will take a look! GLM 5.3 (flash and no flash), mistral large 4 were the best i tried. The worst were chatgpt 5.6 sol, claude sonnet (a new one, no idea which one). Deepseek v4 flash was ok-ish, qwen coder was nice if i was too lazy to paste stuff.
I used them via mistral’s CLI, served via Mistral’s API, and an academic cloud vendor. I have the cli run in an extremely strict jail (podman-based), so the stuff they can use is super limited (e.g. no bash).
You got access to much bigger models than I have! Nice. I use the gondolin sandbox for Pi. It traps it in a minimal Alpine Linux VM with only packages you define and the only directory it can see being the dir you cd’d into.
How can someone have a skill issue with AI? Isn’t the whole point that it does everything for me? Never understood this.
People are like “you’re gonna get left behind!” and I’m like. How could I be left behind by the magic machine that does everything for me? I’ll just start using it one day and be just as good as everyone else because that’s explicitly how they are designing the software to be used, right?
I think you do have a point. I think the overarching problem with discussions about “AI” is that people rarely refer to the same things when they talk about it and they end up talking past each other. There is a difference between using a custom harness with (a) carefully optimized local LLM(s) and other models you glue on top yourself for other-than-text tasks and using server-side models with a ready-made CLI and UI served to you on a golden platter with all processing being done elsewhere. “You’re gonna get left behind” is not the argument here. If the problem is “the code is too complicated”, that is a style issue that can be mitigated. It’s not set in stone. Sure, there are many problems with LLMs and ML models in terms of where they source their training data from and the environmental impact of data centers. I’ve also had my fair share of literal slop pop out. Some problems are just there and won’t go away (yet), but some problems are a choice, whether you know it or not.
I also think that some people like the idea of being on the designing end rather than just being on the using end. Why stop at skill.md? Why not post-train your model on your own code, style and preferences? I’ve trained a tiny model on a small subset of data I may legally use because it’s… fun. It’s my hobby. I just tinker around with it. The model itself is absolute poop, because it’s super small. But it works. I also locally rebuilt llama.cpp with a TheRock ROCm build for better support for my graphics card and got a very significant performance improvement. It doesn’t even need to be useful in any meaningful way (it often is anyway), as long as I got fun understanding how it works.
“I’ll just start using it one day and be just as good as everyone else because that’s explicitly how they are designing the software to be used, right?”. Yes. You’re still going to need to be capable enough to express what you want with precision, but if you can do that, you’re absolutely right, especially considering future models. But maybe you don’t want to be absolutely dependent on big companies serving you those models.
It’s with anything. It’s a tool. Like if you buy a synthesizer but you don’t press any buttons and complain you don’t get the sound you want.
Like go on a image generation site and say “dog” sure you’ll probably get a picture of a dog but it’s probably not the breed, color or setting or pose you want. If you are more spesific in your prompt you’ll get a more spesific thing.
Skills etc are just prompts. And with better prompts you’ll get better results. It’s very simple
All AI does for me is to give me more work to do. Its code is terrible. It has horrible taste. Its work is not modelled after the task, but after “must look complicated”
Every now and then, when zeal comes upon them, my bosses email me a bunch of risk assessments and policy documents that have obviously been written by AI, probably just from one or two prompts. I have actual work to do, I don’t have time to be reading through that shit - so I get AI to summarise it for me.
For this we are boiling the oceans. Sometimes I think our species is just one big Darwin award waiting to happen.
Get Ponytail https://github.com/DietrichGebert/ponytail and an auto-linter such as pi-lens if you’re using Pi harness. And a goal/to-do (pi-goal-x) extension.
What model are you using?
All of this crap for a single markdown file 🙈 or am I overlookong something
Markdown is the code of the LLM world. It works, if you are precise.
For absolute obedience, throw in a §.
§ 1 Speak like a pirate.
Thx, will take a look! GLM 5.3 (flash and no flash), mistral large 4 were the best i tried. The worst were chatgpt 5.6 sol, claude sonnet (a new one, no idea which one). Deepseek v4 flash was ok-ish, qwen coder was nice if i was too lazy to paste stuff.
I used them via mistral’s CLI, served via Mistral’s API, and an academic cloud vendor. I have the cli run in an extremely strict jail (podman-based), so the stuff they can use is super limited (e.g. no bash).
Still learning. Could really be a skill.md issue?
You got access to much bigger models than I have! Nice. I use the gondolin sandbox for Pi. It traps it in a minimal Alpine Linux VM with only packages you define and the only directory it can see being the dir you cd’d into.
Gondolin looks good! Would love to have a secure environment on my Working Laptop though. There is agentic sandbox, but it does not support Vibe CLI.
I think I’ll try to improvenmy podman solution…
Skill.md issue
How can someone have a skill issue with AI? Isn’t the whole point that it does everything for me? Never understood this.
People are like “you’re gonna get left behind!” and I’m like. How could I be left behind by the magic machine that does everything for me? I’ll just start using it one day and be just as good as everyone else because that’s explicitly how they are designing the software to be used, right?
I think you do have a point. I think the overarching problem with discussions about “AI” is that people rarely refer to the same things when they talk about it and they end up talking past each other. There is a difference between using a custom harness with (a) carefully optimized local LLM(s) and other models you glue on top yourself for other-than-text tasks and using server-side models with a ready-made CLI and UI served to you on a golden platter with all processing being done elsewhere. “You’re gonna get left behind” is not the argument here. If the problem is “the code is too complicated”, that is a style issue that can be mitigated. It’s not set in stone. Sure, there are many problems with LLMs and ML models in terms of where they source their training data from and the environmental impact of data centers. I’ve also had my fair share of literal slop pop out. Some problems are just there and won’t go away (yet), but some problems are a choice, whether you know it or not.
I also think that some people like the idea of being on the designing end rather than just being on the using end. Why stop at skill.md? Why not post-train your model on your own code, style and preferences? I’ve trained a tiny model on a small subset of data I may legally use because it’s… fun. It’s my hobby. I just tinker around with it. The model itself is absolute poop, because it’s super small. But it works. I also locally rebuilt llama.cpp with a TheRock ROCm build for better support for my graphics card and got a very significant performance improvement. It doesn’t even need to be useful in any meaningful way (it often is anyway), as long as I got fun understanding how it works.
“I’ll just start using it one day and be just as good as everyone else because that’s explicitly how they are designing the software to be used, right?”. Yes. You’re still going to need to be capable enough to express what you want with precision, but if you can do that, you’re absolutely right, especially considering future models. But maybe you don’t want to be absolutely dependent on big companies serving you those models.
It’s with anything. It’s a tool. Like if you buy a synthesizer but you don’t press any buttons and complain you don’t get the sound you want.
Like go on a image generation site and say “dog” sure you’ll probably get a picture of a dog but it’s probably not the breed, color or setting or pose you want. If you are more spesific in your prompt you’ll get a more spesific thing.
Skills etc are just prompts. And with better prompts you’ll get better results. It’s very simple
Honestly yeah likely, I have extreme lack of knowledge re: vibe coding.
I am much faster per pedes right now