Researcher slams OpenAI's redefinition of alignment as just being more useful

nabla_theta · x · 2026-09-24

nablatheta argues that redefining alignment to mean "more useful" has been a disaster for the alignment discourse, quoting OpenAI's 2026 claim that larger computational depth improves alignment because models generalize more flexibly—including from alignment data—and better-generalizing models are more aligned and hallucinate less. The post spotlights concerns that safety terminology is being diluted into product metrics.

Original post →

More from AGI Musings

AGI Musings channel →