Evaluators and safety organisations · Blog post · Redwood Research
Astra covertly underperforming on safety research is concerning, even though it’s unlikely to be motivated by ambitious goals
Read at blog.redwoodresearch.org ↗Read at sourceAll rights reserved
Outline · 146 paragraphs
- Summary of OpenAI’s findings¶10
- What could have caused Astra to underperform?¶28
- Astra is probably not strategically underperforming in service of ambitious goals. Its motivation is unclear.¶30
- Root causes of this behavior seem uncertain¶41
- How concerning is Astra’s underperformance?¶48
- Subtle obstruction on safety R&D is harmful no matter the root cause¶49
- But, unstrategic underperformance is less scary than strategic sandbagging¶55
- Astra’s behavior, and OpenAI’s struggle to identify its root causes, shows that current alignment is still a hard and unsolved problem¶63
- Astra underperforming subtly enough to evade misalignment monitors shows insufficient alignment monitorability¶71
- Open questions¶77
- Conclusion¶118