Which questions can’t we punt?

A new strategic framework published on LessWrong argues that AI strategy researchers must adopt a 'just-in-time' approach, focusing intensely on questions relevant to the early stages of the AI transition while postponing work on many seemingly critical long-term issues. The core thesis, authored by Lizka, posits that humanity's current limited capacity for strategy work will be massively expanded by future 'AI uplift,' making today's marginal research on far-future topics orders of magnitude less valuable unless it informs imminent decisions.

The framework organizes high-priority questions into five clusters: understanding the early trajectory and impacts of AI, preparing for acute near-term risks (like misalignment or bioweapons), identifying choices that set up later periods well, exploring early automation levers, and clarifying foundational concepts. Conversely, it explicitly suggests deprioritizing extensive work on later-stage challenges such as the alignment of superintelligent AI, space governance, making deals with advanced AI agents, and acausal trade, arguing these are more 'puntable' to future, more capable generations aided by AI.

This represents a significant shift in orientation for the AI safety and strategy community, which has historically focused heavily on existential risks from superintelligence. The authors contrast their view with what they call an 'Implementation not strategy' mindset, urging a ruthless focus on what human minds need to understand in the next few years to navigate the initial disruptive phase of AI transformation, rather than attempting to solve all problems in advance with today's limited cognitive resources.