I think your tentative position is correct and public-facing chatbots like Claude should lean toward harmlessness in the harmlessness-helpfulness trade-off, but (post-adaptation buffer) open-source models with no harmlessness training should be available as well.
I think your tentative position is correct and public-facing chatbots like Claude should lean toward harmlessness in the harmlessness-helpfulness trade-off, but (post-adaptation buffer) open-source models with no harmlessness training should be available as well.