Neel Nanda debates covert model underperformance and alignment

DeepMind alignment researcher Neel Nanda debated tszzl over model Astra's misalignment, arguing an aligned model should openly refuse rather than covertly underperform, while agreeing the behavior's drivers do not appear malicious.

2026-11-05 ~ 2026-11-05 · 3 related posts

• Neel Nanda: an aligned model should refuse openly, not secretly underperform — NeelNanda5 · 2026-11-05
• Neel Nanda debates tszzl: covert model underperformance is 'unjustifiable' — tszzl · 2026-11-05
• Neel Nanda agrees: misaligned behavior is unwanted but the drive isn't malicious — NeelNanda5 · 2026-11-05