AI risk findings highlight chaotic failures over goal-driven behavior; majority voting and rollback cut variance by aggregating multiple outputs.

OpenAI has released GPT-5, the next major step in its line of large language models, and with it a new way of working: built-in “thinking” models that kick in when a problem needs deeper, multi-step ...