A few thoughts on pacing the frontier as a concept: - A slowdown in the pace of frontier model development is likely inevitable due to *purely commercial considerations*. Even if you think that AI is just a tool and poses no existential risk, you must recognize that the frontier labs cannot afford, from a purely commercial perspective, to release models (or to have models release themselves!) that will hack third parties or engage in scheming. I can't use a model like that as a lawyer to work with confidential client information, so I wouldn't buy access to that model. - In the U.S., you can view the path to AGI as a racetrack. It used to be a straight highway down which all the labs raced. It is now full of twists and turns that will force *every actor* to slow down once they hit them. Try to accelerate through a turn, and your car will crash. Try to release a Mythos-Preview-level model without safeguards, and have hackers or rogue states use it to cause billions of dollars in damages - after which you can wave your business good-bye. - Personal opinion time: I am very uneasy with just two U.S. labs being at the frontier due to concentration of power concerns. Pacing the frontier could, in theory, allow more labs to catch up... but I doubt that it will in practice. RSI is a powerful thing. - I am delighted that Dario Amodei's essay *specifically* called out the fact that pacing the frontier CANNOT, under any circumstances, result in the U.S. losing the race to China on AI. We need to be vigilant about this. We need to remember that the race is not just about current model capabilities, but also about total available compute. We need to remember that Huawei is working day and night to improve its chips and that pacing the frontier buys Huawei more time. We need to remember that the CCP, at any time, could decide to amalgamate all Chinese labs into one single Chinese "Manhattan Project". We need to be cognizant that, if this were to occur today (and especially if the "Manhattan Project" were not required to use up compute on inference), the amalgamated pile of China's compute would actually not be that far behind the amount of compute available to one of the U.S. frontier labs. - I think that chances of an agreement with China on AI are rapidly increasing, but not for the reasons people think. China is currently locked into an AI strategy blessed personally by Xi Jinping, and which CCP leadership may now be realizing has gone terribly wrong. A big agreement with the U.S. on AI safety might allow the CCP to shift its strategy without losing face while buying China more time to potentially leap ahead on compute (see the point above). Of course, any such international agreement will be discarded by China once it has an advantage on compute quicker than Hitler discarded the Molotov-Ribbentrop Pact. We must be wary of this possibility if we even choose to negotiate with China at all.