Moving beyond current needs, as both a way to ensure that domestic policy doesn’t get stuck dealing with immediate economic, equity, and political issues, I think we should push for an ambitious intermediate goal to promote the adoption of international standards regarding high-risk future models. To that end, I would call for every country to pass laws today that will trigger a full ban on deploying or training AI systems larger than GPT-4 which have not been reviewed by an international regulatory body with authority to reject applications, starting in 2025, pending international governance regimes with mandatory review provisions for potentially dangerous applications and models. This isn’t helpful for the most obvious immediate risks and economic impacts of AI—and for exactly that reason, it’s critical as a way to ensure the tremendous future risks aren’t ignored.
I strongly agree with that.
You don’t talk much about compute caps as a lever elsewhere in the text, so I’m going to paste some passages I wrote on why I’m excited about compute-related interventions to slow down AI. (My summary on slowing AI is available on the database for AI governance researchers – if anyone is planning to work on this topic but doesn’t have access to that database, feel free to email me and I can give you access to a copy.)
Compute seems particularly suited for governance measures: it’s quantifiable, can’t be used by multiple actors at once, and we can restrict access to it. None of these three factors apply to software (so it’s unfortunate that software progress plays a more significant role for AI timelines than compute increases). Monitoring compute access is currently difficult because compute is easy to transport, and we don’t know where much of it is. Still, we could help set up a database, demand reporting from sellers, and shift compute use from physical access to cloud computing or data center access (centralizing access helps with monitoring). The ideal target state for compute governance might be some kind of “moving bright line” of maximum compute allowances for training runs. (A static cap might be too difficult to enforce because compute costs to circumvent the cap will fall over time.) The regulation could be flexible so labs with a proven safety mindset can receive authorization to go beyond the cap. More ambitiously, there’s the idea of
hardware-enabled governance mechanisms (previous terminology: “on-chip measures”). These are tamper-proof mechanisms on chips (or on the larger hardware components of compute clusters) that would allow for actions like communicating information about a chip’s location or its past activity, remote shutdown, or restricting the chip’s communication with other chips (limiting the size of a training run it could be involved in). Hardware-enabled mechanisms don’t yet exist in a tamper-proof way, but NVIDIA has chips that illustrate the concept. I’m particularly excited about hardware-enabled governance mechanisms because they’re the only idea related to slowing AI progress that could (combined with an ambitious regulatory framework) address the problem as a whole, instead of just giving us a small degree of temporary slowdown. (Hardware-enabled mechanisms would also continue to be helpful after the first aligned TAI is developed – it’s not like coordination challenges will automatically go away at the point when an aligned AI is first developed.) Widespread implementation of such mechanisms is several years away even in a best-case scenario, so it seems crucial to get started.
Onni Arne and Lennart Heim have been looking into hardware-enabled governance mechanisms. (My sense from talking to them is that when it comes to monitoring and auditing of compute, they see the most promise in measures that show a chip’s past activity, “proof of non-training.”) Yonadav Shavit also works on compute governance and seems like a great person to talk to about this.
And here’s an unfortunate caveat about how compute governance may not be sufficient to avoid an AI catastrophe:
Software progress vs. compute: I’m mostly writing my piece based on the assumption that software progress and compute growth are both important levers (with software progress being the stronger one). However, there’s a view on which algorithmic improvements are a lot jumpier than Ajeya Cotra assumes in her “2020 compute training requirements” framework. If so, and if we’re already in a compute overhang (in the sense that it’s realistic to assume that new discoveries could get us to TAI with current levels of compute), it could be tremendously important to prevent algorithmic exploration by creative ML researchers, even at lower-than-cutting-edge levels of compute. (Also, the scaling hypothesis would likely be false or at least incomplete in that particular world, and compute restrictions would matter less since building TAI would mainly require software breakthroughs.) In short, if the road to TAI is mostly through algorithmic breakthroughs, we might be in a pretty bad situation in terms of not having available promising interventions to slow down progress.
But there might still be some things to do to slow progress a little bit, such as improving information security to prevent leakage of insights from leading labs, and export controls on model weights.
I think that these are good points for technical discussions of how to implement rules, and thanks for brining them up—but I don’t really makes sense to focus on this if the question is whether or not to regulate AI or have a moratorium.
I strongly agree with that.
You don’t talk much about compute caps as a lever elsewhere in the text, so I’m going to paste some passages I wrote on why I’m excited about compute-related interventions to slow down AI. (My summary on slowing AI is available on the database for AI governance researchers – if anyone is planning to work on this topic but doesn’t have access to that database, feel free to email me and I can give you access to a copy.)
And here’s an unfortunate caveat about how compute governance may not be sufficient to avoid an AI catastrophe:
But there might still be some things to do to slow progress a little bit, such as improving information security to prevent leakage of insights from leading labs, and export controls on model weights.
I think that these are good points for technical discussions of how to implement rules, and thanks for brining them up—but I don’t really makes sense to focus on this if the question is whether or not to regulate AI or have a moratorium.