Why splits happen
As labs scale, disagreements can arise about risk tolerance, deployment strategy, governance, and which research bets to prioritize. Because frontier training is expensive, these disagreements are amplified: you can’t easily “agree to disagree” while sharing one compute budget.
Safety framing as a product and research strategy
Safety work spans data, post-training, evaluation, and policies. Some labs emphasize formalized principles and systematic preference shaping; others emphasize iterative deployment and feedback. In practice, most successful deployments blend both.
Connections
For the post-training mechanics that underlie many safety approaches, see RLHF and Post-Training. For the broader ecosystem of new labs, see New Labs.