Contra Leicht on AI Pauses

In a technical post on LessWrong, AI safety researcher David Scott Krueger (capybaralet) dissects and rebuts Anton Leicht's recent arguments against pausing AI development. Krueger takes issue with Leicht's core framing, which models a pause as "resampling from a fixed distribution" of possible AI timelines. Krueger argues this is a fundamental mischaracterization; the goal of a pause is to actively improve safety outcomes and buy time for alignment research, not to randomly roll the dice again. He also challenges Leicht's assertion that the current AI development landscape—marked by minimal compute overhang, multipolarity, and liberal democratic control of supply chains—is "going pretty damn well." Krueger counters that alignment remains unsolved and that multipolarity may exacerbate dangerous race dynamics.

Krueger systematically addresses Leicht's political analysis, which claims pause proposals are only popular among radical political wings and that any enacted pause would be a unilateral "second best" version lacking crucial controls. Krueger finds this reasoning lacking and points out that Leicht concedes a pause is something governments could actually do, undermining his own argument. The critique highlights that Leicht fails to properly engage with the "ardent safetyist" worldview, which sees existential risk as paramount and finds Leicht's proposed alternative—a progression of transparency, auditing, and unspecified "surgical interventions"—to be clearly inadequate. For safety advocates, Krueger notes, these milder measures are precisely why a pause is being proposed in the first place.