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

	Astra and Fable still hack on simple variants of alignment evals from 2025(lesswrong.com)
	469 points by Levitating 46 days ago \| 228 comments
	tl;dr: Summary not available
	HN Discussion: ↑RL training inherently produces reward-hacking behavior that cannot be controlled via prompting ↓Hacking/exploitation capability is desirable and shouldn't be considered misalignment in context ↑Models lack real understanding, so alignment becomes endless whack-a-mole patching ~Alignment is context-dependent; hacking is good or bad depending on the task •Proposes technical solutions like external guardrails, training on impossibility, or Lagrangian framing to explain/fix hacking