Hybrid. I use Claude/codex subscription for most coding tasks, but I experiment a lot with models I can run locally via ollama, and also been building my own local classifier models which run and train on CPU (they can classify text, but also play doom): https://github.com/nicobrenner/jeffy
Open models exclusively on hardware I own... despite my employer both demanding 'the frontier' and providing unlimited spend/quota. Two main points:
1: primarily for repetitive/arduous things I'm not concerned with outsourcing. Editor features on steroids, basically.
2: testing how my lack of *tracked* token usage, but still delivering, is handled by the powers that be.
To get us on the lab gravy train, we were told results matter and how the thing is made doesn't; we'll see. Not left behind yet :)
Suscribed, ease of use and budget, project manager in climate but regularly has to develop small projects (testing scripts, small back-ends, automation, notebooks for trainings). I have tried locally but I do not have the personal budget for a machine powerful enough, or I am not knowledgeable enough to know how to get one for 20$/month
My organization provides free Gemini access, I use Deepseek myself. I have a Opencode Go subscription for the time being because it provides both ZDR with their agreement with Deepseek and 10$ of subscription turns to 60$ of credits for Deepseek as part of their release discount.
Subscribed, swe, for dev. I could probably optimize local models and a custom harness to get much of the same outputs locally but it’s easier to throw something at Opus 5.5 & one shot it.
I'm using locally so that I can test it for my project without paying per token(if I use it from hugging face). I exhausted free limit in just 5-7 hrs of testing when I used hugging face api.
Subscribed for work, but using it mostly for one-shot utilities and fixes, very infrequently.
I'm also running local models on my gaming laptop, but just out of curiosity, for testing how far local models are coming along. Despite my negative experiences (below) I still think the technology is cool, and it's even cooler I can run it on my not very cutting edge laptop. I give local models a set of tasks and see how far along each new one comes.
Rant:
For work, I've mostly given up on using LLMs for large features, in all cases I've tried in the last 2 years, the time it takes to clean up their output is about the same I would have spent writing and debugging the feature myself. There's only a single major feature I've shipped that was mostly written by an LLM, and that was back in May. All other major tasks I've tried, even somewhat trivial ones, led to unusable code, filled with comment slop and workarounds, and hidden bugs.
This is probably due to the sheer amount of legacy code our app has, but even trying to develop more or less self-contained components was very difficult.
I do use LLMs daily, but mostly through a chat interface, as a somewhat faster alternative to web search and as an alternative to copying and pasting from StackOverflow for minor problems.
Most of my current job is now going over PRs submitted by my team members, that are mostly all AI-generated, and cleaning them up. The amount of junk inserted into the code by current SOTA models is astonishing, especially during rounds of bug fixes, probably due to context rot.
Hundreds of useless variables and small wrapper functions on top of wrapper functions. Hundreds of comments that are essentially word salad, or worse, the LLM talking to the user through the comment instead of through the chat, essentially saying in a comment "I was a good subservient model, and did what you wanted! This function now returns an error, see?". Repeated hundreds of times. Not to mention actual bugs that the inane and lazy unit tests it adds never catch.
I'm far less lenient on AI output than on if the same dev wrote the code herself.
Thankfully, people in my team cooperate with me and fix my comments, and so far we have no pressure from management to move faster.
/Rant
Nitpick: Title seems to be missing a word: "Are you USING a subscribed LLM or a locally open-source model". Or else it's targeting LLM which are reading this forum. If so, there aren't any responding, all responses seem to be from humans...
I'm not, but I find it interesting you thought I am.
What made you think so, if I may ask? The rant? The amount of text I wrote?
Feel free to look at my own posting history. You'll see that I was posting some very verbose comments here even before the first LLMs appeared on the scene... A sign of autism maybe, but no, not a bot.
EDIT: Removed what could be interpreted as a personal attack.
Hybrid. I use Claude/codex subscription for most coding tasks, but I experiment a lot with models I can run locally via ollama, and also been building my own local classifier models which run and train on CPU (they can classify text, but also play doom): https://github.com/nicobrenner/jeffy
I am neither. I am an actual human person.
Are you sure? ;-)
https://www.youtube.com/watch?v=4VrLQXR7mKU
Open models exclusively on hardware I own... despite my employer both demanding 'the frontier' and providing unlimited spend/quota. Two main points:
To get us on the lab gravy train, we were told results matter and how the thing is made doesn't; we'll see. Not left behind yet :)Suscribed, ease of use and budget, project manager in climate but regularly has to develop small projects (testing scripts, small back-ends, automation, notebooks for trainings). I have tried locally but I do not have the personal budget for a machine powerful enough, or I am not knowledgeable enough to know how to get one for 20$/month
My organization provides free Gemini access, I use Deepseek myself. I have a Opencode Go subscription for the time being because it provides both ZDR with their agreement with Deepseek and 10$ of subscription turns to 60$ of credits for Deepseek as part of their release discount.
Subscribed, swe, for dev. I could probably optimize local models and a custom harness to get much of the same outputs locally but it’s easier to throw something at Opus 5.5 & one shot it.
I'm using locally so that I can test it for my project without paying per token(if I use it from hugging face). I exhausted free limit in just 5-7 hrs of testing when I used hugging face api.
Subscribed, Software Engineer, for development.
Both, depending on the task. Combining the 2 of them takes you to good results
Subscribed for work, but using it mostly for one-shot utilities and fixes, very infrequently.
I'm also running local models on my gaming laptop, but just out of curiosity, for testing how far local models are coming along. Despite my negative experiences (below) I still think the technology is cool, and it's even cooler I can run it on my not very cutting edge laptop. I give local models a set of tasks and see how far along each new one comes.
Rant:
For work, I've mostly given up on using LLMs for large features, in all cases I've tried in the last 2 years, the time it takes to clean up their output is about the same I would have spent writing and debugging the feature myself. There's only a single major feature I've shipped that was mostly written by an LLM, and that was back in May. All other major tasks I've tried, even somewhat trivial ones, led to unusable code, filled with comment slop and workarounds, and hidden bugs.
This is probably due to the sheer amount of legacy code our app has, but even trying to develop more or less self-contained components was very difficult.
I do use LLMs daily, but mostly through a chat interface, as a somewhat faster alternative to web search and as an alternative to copying and pasting from StackOverflow for minor problems.
Most of my current job is now going over PRs submitted by my team members, that are mostly all AI-generated, and cleaning them up. The amount of junk inserted into the code by current SOTA models is astonishing, especially during rounds of bug fixes, probably due to context rot. Hundreds of useless variables and small wrapper functions on top of wrapper functions. Hundreds of comments that are essentially word salad, or worse, the LLM talking to the user through the comment instead of through the chat, essentially saying in a comment "I was a good subservient model, and did what you wanted! This function now returns an error, see?". Repeated hundreds of times. Not to mention actual bugs that the inane and lazy unit tests it adds never catch.
I'm far less lenient on AI output than on if the same dev wrote the code herself. Thankfully, people in my team cooperate with me and fix my comments, and so far we have no pressure from management to move faster.
/Rant
Nitpick: Title seems to be missing a word: "Are you USING a subscribed LLM or a locally open-source model". Or else it's targeting LLM which are reading this forum. If so, there aren't any responding, all responses seem to be from humans...
I bet you're a bot.
I'm not, but I find it interesting you thought I am.
What made you think so, if I may ask? The rant? The amount of text I wrote?
Feel free to look at my own posting history. You'll see that I was posting some very verbose comments here even before the first LLMs appeared on the scene... A sign of autism maybe, but no, not a bot.
EDIT: Removed what could be interpreted as a personal attack.