SignalTech/AI
Don’t assume composability in AI safety. Deployments need their own testing.
AI safety deployments require independent testing rather than assumed composability.
Context
What happened: MLCommons argues that AI safety scores from a base model do not reliably carry over to deployed systems built on it, and says safety testing must cover the whole deployed system.
Why it matters: The post says many widely used benchmarks are too noisy for reliable comparisons between systems or over time. MLCommons says its Candela product tests the full deployed assembly, and that margins of error would let deployers tell whether a change caused a safety score to shift.
What changed
MLCommons argues that AI safety scores from a base model do not reliably carry over to deployed systems built on it, and says safety testing must cover the whole deployed system.
Why it matters
The post says many widely used benchmarks are too noisy for reliable comparisons between systems or over time. MLCommons says its Candela product tests the full deployed assembly, and that margins of error would let deployers tell whether a change caused a safety score to shift.