Charbel-Raphaël comments on Three Properties for Alignment (and Why We’re Not Training Them)

[Charbel-Raphaël]

I agree on the problem, but unclear how tractable this is