new: OpenAI & others quietly using loop transformers that don't show their 'thinking' when scaled up

a leap forward on performance, but sparking concerns inside & outside OpenAI re: security as this takes off