Great article on what had me worried ( and still does) in training models: "The Most Forbidden Technique." (Zvi Mowshowitz) …

The rule is: you train on the final output only. never on the reasoning process, and never on the interpretability signal that reveals it.

Basically: A model that hides its bad thoughts isn't aligned. it's just gotten better at lying to the people who built it.

Alignment first!