Post

This has been pretty obvious since the beginning. When you build a model, you freeze all of the implicit assumptions underlying the language in that model. For example, if you built a model of physics before Einstein, everything it generated would implicitly assume a Newtonian model of physics.

Once content is crystallized in an LLM, it is pervasive. It’s one thing for specific scientific knowledge to get updated — there’s a process around science. But implicit cultural associations? Not at all. They get built in and aren’t questioned. Also it’s unclear how you would undo them.

One big example of this is the biological, neurological, and psychological understanding of gender. The vast majority (probably 99.99%) of text that LLMs were trained on assume gender and biological sex are synonymous and binary. But science has found that they aren’t. So how would a model even get