[bitwize]

Except it's not really a mechsuit; inasmuch as it is, it's more like an Evangelion: it's got bits of people's souls floating around in there; it only works well with a high synch rate with the user, which is difficult to achieve consistently; and sometimes it pops off and does its own thing anyway.

And the way you achieve and maintain deep knowledge of the shape of the output you want is... well, to write and test it yourself. LLMs are a cognitohazard. If you aren't accustomed to programming without them (or even if you are) it's hard to see the line between using them to remove drudgery and relying on them to think for you.

The Whispering Earring story reminds me of that episode of Rick and Morty in which Morty gets a piece of crystal embedded in his forehead which allows him to perceive the moment of his death; he then begins steering his actions towards the future in which an elderly Jessica (his crush) tells him she loves him as he dies, down to individual articulated phonemes and muscle actions. In the end, it's revealed this future is not what Morty had hoped for.

[rmunn]

That Evangelion analogy is a good one. I don't necessarily agree with all of it (for example, I think the high synch rate is the weakest part of the analogy, and LLMs can be used as tools with some decent success without needing to plunge into them too deeply), but every piece of it can be coherently argued for (for example, perhaps I'm missing an even higher threshold of success that I could achieve with a higher "synch rate" equivalent, and I just don't know about it).

But kudos for coming up with such an evocative analogy.

[bitwize]

"Synch rate" is roughly "you know how to manipulate the model to get it to do precisely what you want". This is nonobvious even for open-weights models and can change as new models are published, often under the same name as the old models.