#luddites : "WAAH LLMs eat the planet with huge energy hungry Datacentres !!!!"

[n_dimension@infosec.exchange]

#luddites : "WAAH LLMs eat the planet with huge energy hungry Datacentres !!!!"

#Google : Here is one that runs on 4 watts of power of your prayer tablet.

#luddites : (Waaah ambulacing intensifies)

Time to admit it was never about Planet destroying #Ai but your reluctance to learn new shit and hanging out with all the Kool kids, dancing around the bonfires in the woods?

#gemininano

[bhg@troet.cafe]

@n_dimension I don't get what you mean. Running the LLMs is not the issue, especially if they're lightweight, you and I know. Of course you can run a lightweight model on a low ressource price tag. The training of LLMs is ressource hungry as hell. This is what a large part of the ressources of datacentres are being used for.

[tuban_muzuru@beige.party]

@bhg @n_dimension

Once we dissect the truly colossal power consumption involved in training these beasts, the staggering inefficiencies of the MAC operation == multiply and accumulate, the Really Big Story is how the indexers are address the power consumption:

Algorithm efficiency is a big story right now: DeepSeek's V3 model reportedly cost just $5.576 million to train and used only around 2,000 chips, where competitors were using 16,000+.

As one Rhodium Group analyst put it, DeepSeek "demonstrates that training high-performance models can take far less electricity than previously thought." The catch, as some researchers note, is that cheaper training may just unleash more demand overall: Jevon's Paradox

https://www.axios.com/2025/01/28/deepseek-ai-model-energy-power-demand