We're talking about it being invisibly moved to a weaker model if it looks like you're distilling (which is best detected through something that is at least partially a reputational / account metric).

Now, Anthropic stepped away from this, but it highlights one more kind of systemic risk you're exposed to when you're not running the model yourself.

1. Anthropic certainly has the ability.

2. They’re willing to use it silently.

i disagree, but it seems clear, from how you put it, that there's no point explaining the why