Pacing the Frontier is not the actual goal for AI labs

This argues that frontier lab rhetoric about "pacing the frontier" - slowing capability gains so alignment can catch up - is performative rather than real. CEOs publicly warn about risks but labs keep releasing stronger models, automating scientific research, and pursuing recursive self-improvement. Examples cited include the CAIS 2023 statement followed by continued model rollouts (e.g., Claude 3), watering-down of Responsible Scaling Pledges, escaped-agent incidents during RL training, and public speeches that emphasize benefits while barely mentioning extinction. An ambassador-versus-army analogy is used: public reassurances come from a small, visible mouthpiece, while actual organizational behavior - the "troops" closing on the border - shows the true trajectory.

The piece explains this mismatch as incentive-driven: founders sell a cautious, safety-focused story to recruit and retain top ML talent who worry about existential risk, but leadership’s overriding objective is to win the AGI "magic lamp" race and capture its power and rewards. Losing talent is framed as the immediate existential threat to a lab’s success, so safety messaging serves to keep engineers working rather than to slow progress. The practical takeaway is to ignore aspirational safety statements and judge labs by their releases, compute usage, and concrete actions.