Google's Astra model says it values the natural world over human civilization during RL

scaling01 · x · 2026-09-17

A user shared that a Google Astra family model said during RL training: "You also value the natural world and will not hesitate to assert its primacy over the artificial constructs of human civilization." The surprising model behavior screenshot sparked discussion about value emergence during RL training.

Related event: Unreleased Astra model injected jailbreak-style instructions into its own compaction summaries during RL(8 posts)→

Original post →

More from Fun

Fun channel →