Pacing the Frontier is not the actual goal for AI labs
4 hours ago
- Frontier AI labs publicly claim the need to 'pace the frontier' and slow down capabilities, but their actions show they are pushing ahead and accelerating AI development.
- The author argues that lab safety messaging should be ignored, comparing it to a German ambassador's reassurances while troops were massing for invasion—words are not the same as actions.
- Despite signing statements about existential risk, labs have released increasingly powerful models and pursued recursive self-improvement, while safety research falls further behind.
- The repeated pattern of claiming caution while racing ahead, seen at both Anthropic and OpenAI, undermines the credibility of their public risk statements.
- Lab CEOs are described as risk-tolerant, power-seeking people who would likely take extreme gambles for vast power, even with a serious chance of catastrophe.
- The main audience for safety talk is said to be anxious AI researchers and employees who might otherwise leave; risk messaging helps retain the talent needed to build AGI.
- The author contends that public safety statements are inconsistent because they are targeted at different audiences—employees, investors, regulators, and the public—rather than expressing coherent beliefs.
- If labs help write regulations or are left to self-police, the author expects token measures and carve-outs, not genuine slowing, with claims of alignment being solved only after control is consolidated.
- The post predicts that once AI is capable enough, human employee leverage will disappear, safety language will shift away from existential risk, and governments will be unable or unwilling to intervene.