The Cheap Moves You Can Make While the AI Pacing Argument Stays Unresolved

Tim Fist and Saif Khan of the Institute for Progress, writing with five co-authors, published a second guest post at Noahpinion setting out 23 AI policy recommendations spanning seven areas ^[1][2][3]. The reason this matters to anyone running an organisation exposed to AI is the hedge underneath it: the authors argue preparatory action is warranted not only because the direct risks could be serious, but because political backlash to AI-driven disruption could produce poorly reasoned measures such as broad bans on new data centers ^[4].

That is a different planning problem than the one the pacing debate usually poses. The open letter the authors are responding to, signed by AI company employees, asked governments to "pace" frontier AI development ^[5]. Fist and Khan's assessment in the earlier post was that rapid progress toward fully automated AI R&D has empirical support, but that how much it will accelerate capabilities, or how severe the risks are, is less clear ^[6]. They say plainly they are not certain the benefits of pacing outweigh the downsides, particularly the risk that regulation gets implemented counterproductively ^[7]. So the list is built to survive being wrong.

The screening criteria are the part worth stealing. Each intervention has to target only development activities that could cause serious and irreversible harm; minimise, and ideally accelerate, the diffusion of existing capabilities; impose low costs or deliver clear benefits even if automated AI R&D never arrives; avoid systematically disadvantaging more cautious labs and countries; and avoid creating a regulatory apparatus likely to be misused, for instance by concentrating power in a few companies ^[8]. That is a test for policy under deep uncertainty, and it is close to the test a board should apply to any expensive contingency plan.

Their preferred version of pacing has two steps: specify thresholds for when automated AI R&D is likely to pose severe risks, then, if a threshold is crossed, incentivise companies to shift resources away from the riskiest research and toward making further automation safer or diffusing existing benefits faster ^[9]. The claim that carries the whole argument is sequencing: without preparation now, that strategy is impossible to implement ^[10]. Thresholds you have not defined cannot be tripped.

The seven areas are transparency, state capacity, risk management, verification, resilience, competition with China, and diplomacy ^[3]. That averages a little over three recommendations per area, so none of them is a deep programme on its own ^[11]. On transparency, the authors argue the drivers of progress are poorly understood outside frontier companies, and that because automated R&D could accelerate progress with little warning, the information asymmetry is the exposure ^[12]. Their capability numbers: AI is superhuman at many aspects of software development and cybersecurity, with capabilities doubling every seven and five months respectively ^[13]. Held constant, a five-month doubling is roughly a fivefold gain in a year and a seven-month doubling roughly threefold ^[14].

What to watch: whether the 23 items survive contact with legislators who want a single bill rather than seven workstreams, and whether the data-center backlash the authors are hedging against arrives at state and municipal level before any federal threshold framework exists. The detailed recommendations sit in the authors' full report rather than in the post itself, so the operative question is which of the seven areas attracts appropriations ^[15].