for some extra context, this excellent piece by @andymasley.bsky.social makes a very convincing argument that the environmental impact of LLMs is overstated, all while assuming that LLMs use 3 Wh per query (roughly 10x higher than the actual numbers!)

those numbers are roughly in line with sam altman's claim that the average chatgpt request uses 0.34 Wh.

all of these values are also much, much lower than those assumed in a lot of the discourse around LLMs!

This will be my first time at ACL, and I’m very excited about the opportunity to meet and catch up with other folks doing work at the frontiers of NLP research :) drop me a message if you’re around!

I’ll be talking about the knowledge graph that we’re building at @climatepolicyradar.bsky.social - By weaving together the connections between policy documents from all over the world, we’re hoping to make them easier to explore, understand, and explain

But in this, @aworkinglibrary.com gets straight at the bit which really does worry me.

"AI", not as a technology, but as an ideology, is being used as a prop in the revival of scientific racism, and old, eugenicist hierarchies of intelligence.

It's a very good essay

IMO the more interesting question is whether this sort of tarpit could be used to maliciously steer LLM training, beyond just adding noise. given a large/subtle enough maze, can a more pointed set of disinformation be baked into LLMs' circuits?

not sure one can effectively fight slop by generating even-sloppier-slop