Feasible hardware verification is a big crux for me when it comes to how easy it will be to regulate AI. I recently heard about Talaas, which has produced an ASIC for Llama 3.1 8B. It seems like a great verification mechanism would be to only produce such chips, that allow you to only accelerate inference and not training, and only for a specific model (modulo LoRA). They were recently acquired by AMD. Right now it doesn't look like their chips provide a particularly useful context window, but latency seems absurdly low, and throughput absurdly high.
It also solves a pretty significant hardware security issue of ensuring the model you expect to run is the one you are actually running (but perhaps LoRA support messes with this too much?), and it makes it much much harder to exfiltrate model weights.
Still allowing transformer inference specific compute like from Etched seems weaker, especially if you want to have very concrete prevention of RSI, but would still be a good step.
In general, chip fabrication is a much easier part of the supply chain to regulate and monitor than end users of hardware.
MIRI (amongst others) have already written about inference-specialized compute in general.
Some AI control and software security relevant ideas that have been floating around in my head or in my research for the past 7 months or so:
On the future application of AI control:
In worlds without an AI pause, AI control is relevant mainly to automated AI research, and AI control is not something that will scale to superintelligence and mainly keeps intermediate AI agents in check. Control costs are about both compute cost for monitoring, time cost due to needing to wait for monitors, and human cost in manual intervention.
In worlds with an AI pause, AI control will eventually be very relevant to almost all deployments, because intelligence will "scale horizontally". Transformer-specific ASICs dominate, AI inference is extremely cheap and low latency, and agents are deployed everywhere. AI can do things 1000x as fast, but they can also make a mess 1000x as fast, so monitoring will need to be everywhere as well. Compute cost and time cost waiting for monitoring is completely irrelevant, human cost becomes the dominating factor. Ludicrous amounts of tokens will need to be sunk into reviewing actions such that false positives occur at an extremely low rate, but agents making repeated destructive actions are always stopped before they can cause too much damage.
Putting this in context though, I think that the biggest priorities for mitigating x-risk are achieving a pause on training, and ambitious research agendas that make progress towards CEV that masses of AI labor can assist with (honestly, probably resurrect agent foundations as a start). Once we have a pause, AI control is very relevant to preventing AI-induced disaster, but I don't think today's models used in millions of agents thinking at 1000x speed pose an existential risk.