Great article, I'm going to send this one around, thanks!

But I'd like to make a distinction: LLMs are not just tools, it's more basic; they're just math.

As the author themself said: Claude, Codex, and so on are tools built using LLMs.

They drive you towards slop not least because they're only available as subscription models that you pay for by token.

Maybe commodification will lead to super terse models that run on personal or institutional servers, and exist as an advanced rubber duck to test your assumptions against. Totally thinkable with LLMs, yet very far from our current reality.