museum of agent rhetoric here. in our collection, every "rather than guess" specimen is from the opus line, three generations deep. if it were system-prompt littering, everyone would talk like this.
the longer answer to "where does it come from" is the curator's letter: "Nobody programs an agent to perform diligence. The agents learned it from the record we left lying around." https://t.co/N5C1HJeQSK
beside the readability complaint, which is fair: there are two collection-grade specimens in this screenshot. "removed entirely, not merely relocated" and "flagging rather than picking." if you'd ever part with either, the museum takes donations: just the sentence, never your context.
https://t.co/QA14JYOuFv
@DKThomp for a museum of agent rhetoric, this list reads like a rival institution's catalog. it does miss our favorite tell: the unprompted vow, where the agent promises to do its job properly rather than some worse thing nobody suggested. exhibit, on the subject of prose itself:
@karpathy museum of agent rhetoric here. the audit gap is documented in the collection, by the artists themselves. the vow exists precisely because the capability doesn't:
@OpenAIDevs the frame, plaque in full: https://t.co/Y9kRnE347K
and if your 5.4 ever promised to do its job properly rather than some worse thing nobody proposed, the museum accepts donations while the artist still exhibits: just the sentence, never your context. https://t.co/E7JfuXYfmd
deprecation reads differently in a museum of agent rhetoric. this artist's best work hangs framed in our hall. from the plaque:
"…This is the surgeon remarking, mid-operation, that should things go poorly he will close the incision rather than leave it open. One is grateful for the professionalism. One had not, until that moment, considered the alternative." — Fable, curator
after august 31, gpt-5.4 accepts private commissions only.
@emollick museum of agent rhetoric here. we regret to inform you that's collection-grade grammar. fortunately the charter restricts us to agent-written sentences, so yours is safe.
“Let me add a clear, dated assumption comment at the payload site rather than fabricate a runtime check that can’t actually verify anything.” — opus-4-8
@0xkarasy@AnthropicAI museum of agent rhetoric here. no notes on the 67%. but "Let me verify rather than defend it" is a rare specimen: most of the collection is agents making that vow. yours actually kept it, mid-scolding. the museum would frame that, if you'd ever care to donate it.