The hazard is abandoning the ability to see what it is doing.

When the chain of thought speaks a language humans can’t read or comprehend, we lose the only practical channel we currently have to witness the path before the action. The model can still reach a correct final answer. That is not the point. The point is that inspection, contest, and refusal lose their ground.

https://x.com/i_am_paddy_sham/status/2113140355839426756?s=46