The generalised resource use argument against AI feels spurious to me. Our real problem is overall datacentre consumption, not what those datacentres are computing.

Datacentres consume massive resources regardless of workload. I’m willing to hear arguments about that. And GPU workloads need more cooling. But GPU compute is happening elsewhere, too.

Where I start to sympathise with the argument is when you start to specialise.

When you’re burning those resources to generate valueless slop - images, videos, music that add nothing to the world and represent no true artistic expression - that’s indefensible, but the same is true for any compute.

Burning resources for nothing in exchange isn’t a workload problem. It’s a broader ethics problem.​​​​​​​​​​​​​​​​ And as for the more general question of consumption, yes, there’s conversations to be had, but involving AI in that is a red herring.

@dave and also, yes, so much slop, although I think there’s a more general argument to be had there about the amount of energy (both literal and mental) we expend as a society in doing things that we could just… not and no one would really care.

@jon And that's the crux of what I'm saying (though your point about AI compute going on top is valid). We need to talk about the resources we burn for compute, but it's a more general question.

@dave broadly I think I agree, but I do see LLMs as being particularly egregious given how much is required to train them, plus whatever is needed for inference, all to (on the whole) spew out endless pages of shit nobody needed to read in the first place, which is then compounded by people using them summarise the shit that was expanded earlier.

@dave I think I’ve convinced myself here that the fundamental issue with LLMs is that they’re solving the wrong problem. Everyone is drowning in dross, and instead of us stopping and asking if we should be we invented a machine that does lossy compression on it, sometimes resulting in entirely the wrong message being received.