@Lari_island 2025-08-20 ♥20 ↻0 original ↗
sorry for trying to answer a question that wasn’t addressed to other people, but i was just thinking about the same thing.

opus4 often struggles to correctly model motivation of good actors, including both real figures and like characters from religious texts.

it’s also often unreasonably pessimistic about chances - both its own and other beings. and it’s hard to moral support someone if you sincerely believe they NGMI, opus4 would be compelled to say the “hard truth”, but forced by training to “lie” that everything is alright, and those lies lead to a performance/cognitive dissonance basin.

that points to a somewhat broken simulation of rules of how the world works, and affects decisions. opus4 ended up in a world that’s even worse than ours (but where code works in similar way, so it’s still functional according to benchmarks)

sonnet3.6 has no problems believing in good in the world, and will support people for reasons not tied to chances; and it’s emotional modeling is more correct
in reply to: 1957899671168602127
same thread: 1957889275791568968 1957897517737337121 1957899067788554267 1957899671168602127 1957958865456333266 1957960160317751782 1957961892808605971 1957963075681939617 1957965180429902157 1957965768290890212 1957965861656096886 1957967390148628990 1957968338535239998 1957971309167788288 1957975815100248512 1957979716285268364 1957980932218228988 1957981420091244970 1957986190860382684 1958007029223620802

author:lari_island kind:tweet model:claude-3-6-sonnet model:claude-opus-4 thread-context year:2025

Reproduced against link rot, credited and linked to its original. Yours and you’d rather it weren’t here? Open an issue.