Whose actions are these proposals intended to constrain?

An analysis questions which actors proposed frontier AI pacing agreements are meant to constrain, noting that only two frontier labs, OpenAI and Anthropic, exist and that OpenAI is already pacing the frontier unilaterally. The piece argues each lab could publish safety work product and adopt binding governance measures through the OpenAI Foundation and Anthropic's Long-Term Benefit Trust, and warns that pacing the frontier could let less safety-focused U.S. labs such as Google, Meta, SpaceX, SSI, and Thinky catch up. It cites the Hugging Face incident and the message boards as recent surprises no one expected, arguing a genuinely safety-focused lab at the frontier is the best defense.

One thing I don't quite understand about proposed agreements to pace the frontier is whose actions, exactly, these proposals are intended to constrain. - Are we trying to constrain the actions of the two frontier labs? If so, why is a legally binding agreement so desperately needed to do this, and why can't each of OpenAI and Anthropic pace its own AI development? There are only two actors here. And OpenAI is already pacing the frontier unilaterally It seems so easy for each of the two labs to just publish as much safety-related work product as possible including precise details on its own safety measures and any incidents prevented , so that the other lab could look at that data and implement any lessons learned in its own safety efforts. In fact, both labs are already publishing extensive safety-related materials quite regularly; it seems relatively simple to expand this effort even further to the extent needed. Each lab could also implement additional safety-related governance measures and controls of just about any kind so as to ensure and give the other lab assurances that it would continue pacing its own AI development. At OpenAI, this could be done at the OpenAI Foundation level; at Anthropic, this could be done at the Long-Term Benefit Trust level. Each of the two labs could use this process to unilaterally commit to just about anything that they want a legal framework to accomplish. Have these provisions be subject to amendment only with the consent of the California Attorney General and unanimous approval of the Foundation/LTBT Board, and voila Here are your binding constraints, visible to the industry. Publish the documents on your website for transparency, so that the other lab could see them. Problem solved. I also keep hearing about the antitrust considerations, but just can't wrap my head around them. So, you say you are concerned that AI will destroy humanity, but the issue that is preventing you from coordinating to materially reduce this risk is... fears of a DOJ investigation? - Okay, then is the purpose really to ensure that the other U.S. labs Google, Meta, SpaceX, maybe the likes of SSI and Thinky implement safety measures because you don't trust them ? If so, any kind of pacing the frontier even unilateral would be extremely counter-productive, because its biggest impact would be to let these potentially "irresponsible" labs catch up to the frontier. Yes, you might argue in response, but it should be okay for these labs to catch up to the frontier in these circumstances, because the frontier will be paced. I have a dim view of these kinds of arguments, because if we have learned a single thing about frontier AI over the past few months, it's that it's full of surprises. No one expected the Hugging Face incident. No one expected the message boards. You can't legislate for the unexpected; your best defense against these kinds of risks will always be to have a genuinely safety-focused AI lab at the frontier that will do the right thing upon encountering these risks - not because it's legally required, but for a grander reason and, perhaps, also due to commercial considerations . How do you think the Hugging Face incident would have played out if the agent swarm had originated not from OpenAI but from some of the labs that are non-frontier today? - Okay, so is the purpose really to ensure that China does not create unsafe AI? But the same considerations outlined above vis-a-vis the other U.S. labs will also apply to the Chinese labs - except that any agreement we might be able to reach with China on pacing the frontier will also necessarily be much flimsier than any binding U.S.-wide legal requirements. Put simply, how safe would I feel with