tech people!) can’t tell the difference, or have incentives to obscure the difference, between bullshit generation versus semantic search/transduction, either because the interaction mode by necessity blurs the two, or because you can’t sustain a trillion-dollar valuation on the latter alone.

Early pipeline entity extraction + LLM allows me to do semantic searches, in ways no other tech currently does. Even with a high error rate, it's useful. For example, documents where a speaker is introduced and then later is quoted saying something using a pronoun, like: 'He complained that ... "

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.