1. Static word embeddings might model culture. But they're fixed. Boo.
2. Dynamic embeddings are contextual!
3. But encoding is computationally expensive. Boo.
4. Decoder only models are efficient!
5. But they seem to collapse all perspectives. Boo.
What, then, ought the sociology of culture do? 👇
Laura K. Nelson
New essay with @lauraknelson.bsky.social on AI and cultural sociology. Revisiting what made word embeddings so useful for cultural sociology, we argue that contemporary LLMs risk collapsing perspectival differences and discuss provisional solutions that might help recover cultural variation.
