The existing local models already exist, and new releases will happen because the companies want to sell you inference. They receive no profit or incentive to release the models as open weights but they do it anyway - There is no coherent ethical delimma against using them locally, as it has no effect on whether the companies will or will not be encouraged to do more training.
I’m using it locally, with models my existing computer can handle. I think, but I’m not sure, that it’s better for the environment than buying into some massive model hosted in a data centre. And I don’t want to lose my skills so I still try to do as much as I can without calling on it.
I will say that using a super efficient cloud model (deepseek for me, v4 flash is very cheap) is probably more efficient than using a local model in most cases, especially if you need strong capabilities. The question is whether your own grid or the provider’s grid is more environmentally friendly (hint: china’s energy mix is about 50% fossil fuels while the US’s mix is around 75% fossil fuels overall)
Training is of course expensive still, but Chinese labs have all but caught up to US labs despite much lower compute resources being available to them.
No, and I do understand that training is the most energy-intensive part of the process. I don’t know how exactly it divides up: how much of the work of massive data centres is training and how much is running people’s queries on finished models.
you got me at “responsibly” burning resources.
Not everyone uses big data center powered LLM inference though. Running some models locally uses less power than playing a video game.
now where have these models been trained?
The existing local models already exist, and new releases will happen because the companies want to sell you inference. They receive no profit or incentive to release the models as open weights but they do it anyway - There is no coherent ethical delimma against using them locally, as it has no effect on whether the companies will or will not be encouraged to do more training.
I’m using it locally, with models my existing computer can handle. I think, but I’m not sure, that it’s better for the environment than buying into some massive model hosted in a data centre. And I don’t want to lose my skills so I still try to do as much as I can without calling on it.
I will say that using a super efficient cloud model (deepseek for me, v4 flash is very cheap) is probably more efficient than using a local model in most cases, especially if you need strong capabilities. The question is whether your own grid or the provider’s grid is more environmentally friendly (hint: china’s energy mix is about 50% fossil fuels while the US’s mix is around 75% fossil fuels overall)
Training is of course expensive still, but Chinese labs have all but caught up to US labs despite much lower compute resources being available to them.
That’s a fair point. Where I am is about 28% fossil fuels I believe, so it’s a slightly different calculation.
did you generate your model locally?
No, and I do understand that training is the most energy-intensive part of the process. I don’t know how exactly it divides up: how much of the work of massive data centres is training and how much is running people’s queries on finished models.
maybe ask the vendor to let him pay the actual cost plus a proper profit margin to figure that out?