Yeah you’d really only say it on the theoretical side of things, I’ve definitely heard it in research and academia but even then people usually point to the particulars of their work first

It depends on the model but I’ve seen image generators range from 8.6 wH per image to over 100 wH per image. Parameter count and quantization make a huge difference there. Regardless, even at 10 wH per image that’s not nothing, especially given that most ML image generation workflows involve batch generation of 9 or 10 images. It’s several orders of magnitude less energy intensive than training and fine tuning, but it is not nothing by any means.

The training is a huge power sink, but so is inference (I.e. generating the images). You are absolutely spinning up a bunch of silicon that’s sucking back hundreds of watts with each image that’s output, on top of the impacts of training the model.

I don’t really have issues with them aside from the youtube mirrors 🤷‍♂️ and even then libredirect lets you set fallbacks so I just stack them and rarely have problems

When I have to use a big tech platform I use a privacy preserving proxy site like invidious for youtube. The LibRedirect extension for Firefox makes it really easy and has a surprising number of sites it can serve mirrors for

That’s not the case for the newer open source drivers from nvidia. They’re only compatible with the last few generations of cards but they’re performant and the only feature they lack is CUDA to my knowledge. Not talking nouveau here

Haskell mentioned λ 💪 λ 💪 λ 💪 λ 💪 λ

I use pfetch-rs in new terminal sessions to add a little bit of decor. It doesn’t do anything but look nice, I just added some custom ascii art and it shows some specs