Pacing the Frontier is not the actual goal for AI labs

In a recent post, AI researcher Dario argues that the goal for AI labs should not be to "pace the frontier," but rather to slow the pace of AI model advancements. He believes that the rapid pace of AI development, driven by recursive self-improvement, could surpass our ability to understand and control these systems. Dario asserts that AI labs, including Anthropic and OpenAI, have not actually implemented the strategy of pacing the frontier, despite claiming to be concerned about the risks.

He points out that these labs have continued to release more advanced models and even ventured into automated biologic research. Dario's impression is that these AI companies aim to remain slightly behind the latest advancements, focusing on safety research using reasonably close models. He also notes that public messaging from AI labs often emphasizes the importance of pacing the frontier, while their actions indicate otherwise.