0.32a0 👏👏👏 - now plausible to use LLM as adapter layer (vs LangChain or LiteLLM)?

[ahgraber]

I was super excited reading the blog post on simonwillison.com about these changes -- If I'm interpreting the 0.32.a0 refactor correctly, it seems like llm could more easily be used as an adapter layer (e.g., a LangChain or LiteLLM alternative)? Conversation composition via message array and separation of history from sqlite were among the things that I remember running into as llm opinions that made integrations more difficult.

Am I thinking about this correctly?