Don't Worry About the Vase Podcast OpenAI Preparedness Framework 2.0
May 2, 2025
Dive into the latest critiques of OpenAI's preparedness framework, focusing on its shortcomings in threat modeling. The discussion highlights emerging risks from AI, biology, and cybersecurity, while questioning the downgrading of persuasion risks. Explore the complexities of AI self-improvement and the urgent need for robust safeguards. Concerns over governance and accountability are also raised, alongside tensions in leadership regarding risk management. A call for more transparency and ethical decision-making rounds out this thought-provoking dialogue.
AI Snips
Chapters
Transcript
Episode notes
Narrow Focus Risks Overlooked
- OpenAI's preparedness framework focuses on specific, measurable, severe, and instantaneous threats only.
- This narrow definition risks ignoring broader or emerging AI dangers that might not fit these criteria.
Persuasion Risks Downgraded
- OpenAI downgraded persuasion risks, removing it as a tracked or even research category.
- Zvi Moshowitz argues persuasion can cause irreversible harms and should remain a key concern.
Risk Thresholds Simplified
- The framework removed low and medium risk thresholds, focusing only on high and critical.
- This simplification may miss important gradations that help evaluate and prepare safer models.