Post

datacenter inference is and will probably always be less environmentally destructive than local inference btw. the environmentalist option is not 'local AI'.

aly@aly.codes

"soon most AI will be local" yeah sure, the way most gaming is candycrush and subway surfers. what is the point here

5 reposts98 likes

this is why actually regulating datacenters is a lot more important than banning them. banning them is stupid. please just set reasonable floors for % renewables!

fwiw, I am not saying no to local AI there are a million reasons why local AI is desirable in some capacity. I am an ops freak and love making my computer fan go spinny running a model it has no business running. I want my phone's features to work offline. I don't like daddy jensen etc.

but you guys are actually deluding yourselves. like you are just not correct on the face of it.

i think the biggest delusion is expecting the capabilities of the average consumer's device to leap forward by such a massive amount within 5 years, even considering any potential optimizations in locally-hosted models

It's actually both. Local AI is better for several reasons over data centers for more than just environmental ones but also because if we use compute many users already have available and were already basically going to use anyway then it might actually be better to do so locally rather than not.

For example, we have local AI on the phone on your keyboard 🎹. This is much better there than in datacenter. Or voice dictation. Or small always on agents. There's a point when that calculus changes but it's usually with larger non routine tasks that are difficult like solving the Reimann Hypothesis

If people were already going to buy a laptop than slight marginal increase in power they would get from using the model isn't going to be larger than the compute they would use remotely.

Local AI is inferior in a lot of ways if you're just a person, it's true. It's definitely not cheaper. Hyperscale is done for a reason, right? For me really, it's the third party doctrine in American law, corporate greed, and personal control - so I'll always want an option, but there are tradeoffs

not just ewaste but energy sourcing + distribution is worse to residential lines, consumer batteries are certainly worse, and work cannot be pooled for efficiency gains (e.g. your framework desktop has a floor in wattage even when you're not using it that is basically wasted)

tbh i can't wait until gpus have more specialized hardware for inference (we have compute cores yes but i'm thinking about something like a tpu in your gpu) so that you can run small models for translation, accessibility features (better ocr screen readers), transcription, etc

yeah this is pretty much efficiency / energy / sustainability 101 and true of ...many things. like the issue with all the big conglomerates is more down to policy and biz and them choosing to not do better, not that they're necessarily worse environmentally. that said, how we get to a better place

it puts me in this awkward place where i want demand for inference to go down overall, which is probably the least elastic/influencable part, or, and this is probably the winner, a novel ownership model for a datacenter they seem extremely amenable to coops actually

Idk. Our grid in Aotearoa is mostly renewable and when I run local models that's what I pull from. As well no industrial cooling n I'm trying to just get it on mini solar. But it's more about data sovereignty than diseconomies of scale imo.

I mean sure but you're also not getting nearly the efficiency of just throwing 0.001% of a hyperscalar at a problem for 30 seconds with a much bigger model. yeah it's great it's all renewable but that doesn't change the efficiency equation

Efficiency is an economy of scale question which I think is pretty settled in yr favour but just on the net environmentally destructive question I feel like there's some nuance to be had. Esp in the mesh pooling and co core era post u biquitous home solar.

this is true 100% and doesn't really negate what was said though, the second post already notes the push should be requiring more sustainability efforts on corpos. like generally corpo/scale sustainability efforts are also FAR more meaningful than individual ones