dunno. IMHO Commercial LLM as they are now are unethical trained and energy hunger. Maybe if Debian raised it own LLM, trained on its own data and hosted as environmentaly sustainable as possible…
They don’t need their own LLM. They can just use existing open weights models trained far more modestly.
Maybe finetune them for some specific task, but TBH something like Deepseek/GLM flash, Qwen 3.8, or even Nemotron are going to be plenty good for automation and chores.
…But the thing is, I don’t think Debian can dictate an automation setup to contributors, just like they can’t dictate what IDE or other software they use to code. It’s up to contributors to be conscientious.
Playing video games is also energy hungry. If you’re an experienced developer and just use AI a bit to help with some coding tasks (rather than vibe-coding the entire day), your usage would likely use less power than playing a game on a powerful gaming PC for an hour. If you look at Neuralwatt’s pricing for example, smaller models are ~100-300mWh per request on average, while larger ones (GLM, Kimi K3) are ~2-3Wh, at least on Neuralwatt’s infra.
dunno. IMHO Commercial LLM as they are now are unethical trained and energy hunger. Maybe if Debian raised it own LLM, trained on its own data and hosted as environmentaly sustainable as possible…
They don’t need their own LLM. They can just use existing open weights models trained far more modestly.
Maybe finetune them for some specific task, but TBH something like Deepseek/GLM flash, Qwen 3.8, or even Nemotron are going to be plenty good for automation and chores.
…But the thing is, I don’t think Debian can dictate an automation setup to contributors, just like they can’t dictate what IDE or other software they use to code. It’s up to contributors to be conscientious.
Playing video games is also energy hungry. If you’re an experienced developer and just use AI a bit to help with some coding tasks (rather than vibe-coding the entire day), your usage would likely use less power than playing a game on a powerful gaming PC for an hour. If you look at Neuralwatt’s pricing for example, smaller models are ~100-300mWh per request on average, while larger ones (GLM, Kimi K3) are ~2-3Wh, at least on Neuralwatt’s infra.