I wouldn't be surprised if more resources were put into abliteration resistance the more capable open weight models become. It's something you don't need at all to start hosting the model on your own, but something you need to take care of before you release the weights (if you do care about it at all).

I'm curious what they mean by that too... They might be trying to weaken the cyber capabilities... Or I guess they might mean safety evaluation and hardening of the open source (and perhaps closed source Chinese) software ecosystem...

They need to make money. Let them do it. They deserve it. Also, this is what inference engines like vLLM want to have "zero day" supporr