Post by @filiph@mastodon.social

2026-07-21 09:49 UTC

As always, great information from Our World in Data: https://ourworldindata.org/how-much-energy-do-data-centers-and-artificial-intelligence-use My opinion: - using LLMs for simple queries that you could just find yourself in a few seconds is way too inefficient - using LLMs for &quot;fun&quot; (through things like AI dungeon) is hopefully a short-lived gimmick and will die soon - because it gobbles energy like crazy - using LLMs to do actual useful work is not that bad, imho. If you build a useful thing and it costs as much as watching TV for a day, then that&#39;s fine?

Replies (2)

@fuji@ruby.social 2026-07-21 10:06

@filiph@mastodon.social We need to look at the aggregate numbers because that’s the real thing. That’s what’s been added to the world. That’s what causes the additional effects. Once you start to slice and dice it, you can arrive at any numbers that look nice/acceptable but hide the reality.

Open ##5281469
@chebra@mstdn.io 2026-07-21 11:25

@filiph@mastodon.social I don't think people are actually making useful things in one day. Instead I've seen AI hackathons where 200 teams run agentic workflows nonstop for three days, then present, then forget about it. Many new slop projects are abandoned within days. And agentic is definitely maxing out the context windows - even a single auto-complete when you just pause the cursor will send the whole file, so that's thousands of tokens in the context just like that. Single-use plastics all over again.

Open ##5281472