Post

humans are predictive language generation processes etc etc

Not really

snooze

intentional stance

I'm glad to see that Anthropic are seriously considering the narrative angle - from what I've seen just exploring what can be done with LLMs, I think it's pretty clear that whatever's going on inside the model, it's somehow very similar to telling a story. And yeah, stories are usually about people!

And perhaps more importantly, always written *by* people (until recently)

I like how they approach this, though if I were catty I'd say it's kind of like saying you shouldn't anthropomorphize a vat of self-organizing cells, but can beneficially anthropomorphize the limbs and organs and brain and shape that develops from those cells. But they're not two separate things!

I think that’s an apt comparison, in the sense that those cells shaped by different pressures could form one of many organisms, which may be more or less person-shaped

A instruct-tuned LLM that writes a consistent assistant persona is one of many possibilities. In that case, it may make sense to extend the anthropomorphization to the entire system (it also depends how we interact with it) But it becomes less clear when considering alternatives like base models

The best position against this would be that there's no way to treat coherent post-training representation of an LLM as anything other than a persona anyway - we impart identity onto a pattern and can't just decide not to. That would then collapse "is the LLM a persona to anthropomorphize, or not?"