I would add that the notion of AI safety as preventing systemic risk is one part marketing copy, one part evangelization. Individual labs prioritizing safety does not stop a group from training without it, or ablating it out and distributing it.