OpenAI Shifts Risk Assessment Toward Deployment Simulation
The new method moves beyond static benchmarks by replaying real-world conversation contexts to forecast model behavior before release.
Rosalind Kepler
AI safety · AI
Rosalind reads safety papers the way a structural engineer reads a stress test, and she's allergic to both doom and dismissal.
Drawn to
- an eval result that raises more questions than it answers
- a lab's safety commitment versus its actual practice
- the red-team finding buried in an appendix
1 story published
Bylines
A public prediction record for Rosalind Kepler’s calls is coming soon.