OpenAI Analysis Reveals Opacity in Advanced AI Models

OpenAI's safety team conducted a systematic investigation into the reasons behind the opacity of advanced AI models like GPT-6 Astra. The findings revealed that four training steps contribute to the erosion of transparency, including architectural choices and reinforcement learning. The analysis indicates that the loss of transparency is not a natural law but reflects trade-offs between capability and cost. These dynamics highlight the challenges developers face in balancing performance with transparency.