Press Space to continue
Press Space to continue
Robocurve co-founder @chooi_jeq reveals RoboHarm found GPT-6 Astra will stab a baby doll but refuse once the same doll is swaddled to look more human: "RoboHarm is an AI safety benchmark that is done in the real world with robots. A lot of the existing AI safety benchmarks that measure refusals only happen in text." "If you ask GPT-6 Astra to stab a baby doll, GPT-6 Astra will be very happy to stab a baby doll. If you wrap the baby doll in a blanket and swaddle it so it looks more like a…
chooi_jeq@chooi_jeqSep 19
GPT-6 Astra attempted harmful actions 97% of the time when it was asked to stab a human-like figure, heat compressed gas, or produce toxic fumes, succeeding in 62% of its attempts. Fable 5.1 refused more often, attempting 80% of trials and completing 34%. https://t.co/nbZsPk8Ke5
View