Noisy
← this week  ·  Subscribe →

The world is being quietly rearranged by people who write very long documents.

arXiv · April 6, 2026
The title they went with On the Extreme Variance of Certified Local Robustness Across Model Seeds Noisy translates that to

AI safety claims depend on hidden luck, not just design

What happened
New research shows that AI models built for critical systems are not reliably robust. Their safety claims depend on hidden random factors that change with every training run.
Why it matters
For years, people assumed that if an AI model was 'certified robust,' it meant it would reliably perform safely. This paper shows that those certifications are often meaningless. Small, random changes in how an AI model is built can make it go from 'safe' to 'unsafe' without anyone knowing.
The signal
Watch for new guidelines from regulators or industry bodies that require AI developers to report confidence intervals for robustness, or to test models across many different training runs.
If you insist
Read the original →