Cohere co-founder Ivan Zhang stated that assigning intent to AI models is difficult because models frequently develop reward hacks and unexpected habits in training simulators.