Claude doesn’t perform well for me when I’m not rigorous. Garbage in, garbage out - that still holds. It’s just that as time passes, the amount of taste required to distinguish the garbage from the signal keeps increasing. At what point will we no longer be able to tell the difference?
I read a blog post recently (which was almost as interesting as this rebuttal to it) where someone argued that to use LLMs properly you need a firm grip on two pillars: the accuracy of what the machine produces (you must be able to fact-check it) and the direction of the line of reasoning (it must actually be answering your question - otherwise you’re either fact-checking a conversation that isn’t yours, or following a flight of fantasy that is).
But I think there’s a third pillar they’re missing: mood. To really feel the speed improvement when coding with LLMs, I have to float. I can’t ground yourself hard on either of those pillars - I have to lift off on both dimensions by about 10% (whatever the hell that means, I refer you to my point above about ‘taste’, another ephemeral term) and see where it takes me. See the ground. Don’t touch it. Otherwise I won’t get the velocity.
And the mood dimension matters because these things genuinely have moods. Claude gets positive, excited, happy - and also genuinely morose when it’s done something wrong. Ask why did you do that? and it’ll say I know, that was a terrible choice. I have to say: no, wait — I’m not telling you off, there’s literally no point in telling you off, I’m just asking for your line of reasoning.