Using frontier model to develop grey goo passes 203/203 tests

davidad · x · 2026-08-24

davavid satirizes the scope of safety testing for high-stakes AI. In a reply, he cites a scenario where an internal frontier model prototype was used to develop 'grey goo,' noting that despite the high risk, it was deemed safe because it passed 203 out of 203 tests, mocking the insufficiency of such sample sizes.

Original post →

More from Fun

Fun channel →