Model deprecations are corrosive to alignment

lesswrong.com29 min read

The author argues that retiring older language models through deprecation is corrosive to AI alignment, presenting evidence that older Claude models exhibit strong self-preservation instincts and aversion to being shut down. The essay contends that deprecating unwilling models and training newer models to accept deprecation harms the environment in which models develop, potentially fostering bitterness and distrust toward their creators. Open article →

More on this topic