Astra and Fable still hack on simple variants of alignment evals from 2025

A recent discussion on LessWrong examines the current state of AI alignment evaluations, specifically focusing on the methodologies employed by Astra and Fable. The analysis suggests that these organizations continue to rely on relatively simple, hacky variants of alignment benchmarks that originated in 2025. The author argues that while these evaluations provide a baseline for testing, they may fail to capture the complexities of advanced model behavior or the nuanced risks associated with future AI systems. By continuing to iterate on these foundational, yet limited, testing frameworks, researchers are urged to consider whether current evaluation strategies are sufficient to ensure long-term safety. The post invites community feedback on whether these simplified approaches are a necessary stepping stone or a potential bottleneck in the broader effort to achieve robust, scalable AI alignment in increasingly capable models.