BTW, how could “operating … within a context window” be relevant, since context is meaningless to an LLM? I’ll just stick with the Grady Booch analysis. LLMs often appear to be useful, because their allure is in appearing to be accurate.

Context windows, intuitively speaking, bound the space of random outputs (because LLMs are just fancy autocompletes), so larger context windows let you feed more data in and get better results /based on the tokens in the context window/. The extreme version of this is the (misnamed) “chain of /