The search for a better metaphor for LLMs’ inner workings
One user describes large language models as “bags of contextually activated circuits, heuristics, and algorithms”—and wishes for a better metaphor.
TLDR
A user wants a better metaphor for how they think about large language models’ internal workings. In a follow-up reply, they say understanding LLMs through next-token prediction and reinforcement learning is powerful, but leaves out the internal structure learned. They argue that the “bags of heuristics & algorithms” framing explains both LLM capabilities and failures despite massive training.
Combined views
1.5K
3 Sources, first seen 3h ago