The Alignment Problem Is a Human Problem
ai6 min read
AI alignment research assumes humans can specify what they want. Behavioural science says otherwise.
AI Is Making Up Its Own Values
ai6 min read
AI models are developing coherent internal value systems. Some of those values are ones we wouldn't choose.
DeepSeek thinks it's Claude
ai6 min read
An AI that doesn't know who it is turned out to be a fingerprint of industrial-scale model distillation - and the ethics are more complicated than anyone wants to admit