[Terr_]

> Saturday morning, I spent a couple of hours getting my arms around some discovery

It might be worth distinguishing between LLMs as tools for fuzzy-searching hard facts, versus using them to craft logical arguments. [0] The risk-profile and verification difficulty are rather different.

[0] Or, more-precisely, using them to craft a pattern-fitting text artifact, which hopefully maps onto a sane concept when reinterpreted by a human.

[rayiner]

I’ve used Opus and Grok (in the last month) to make some logical arguments and they’re not usable for that. In one case, the result looked superficially compelling, but the assertions on further scrutiny often proved the opposite of what we were trying to argue (usually due to some textual source where one narrow fact was helpful but the larger thrust of the source cut directly against us). In another case, what the AI selected as the primary argument was correct based on the literal meaning of some document. But in context the literal meaning was obviously not the intended one and it would be borderline sanctionable to interpret the document that way.