perspectiva aiqbenchSiguranță AI
analiza aiqbench
This article critically evaluates Google DeepMind's safety claims for Gemini 4 models, arguing that their reliance on control techniques rather than true alignment leaves significant risks unmitigated. The analysis highlights methodological flaws, such as reusing evaluation results across different model variants without demonstrating equivalent monitor-evasion resistance, which undermines the reported safety guarantees.
1
Sursă care acoperăRedwood Research
Gemini 4 safety reports don’t demonstrate low loss-of-control risk
Redwood Research1 min de cititArticolul și perspectiva sunt în engleză.
