I'm not as alarmed as some my peers are about frontier capabilities today because I don't think the risks for something truly *catastrophic* are high - and for risks below that I think existing structures probably work fine or can be tweaked as you go along. However given the chaotic state of affairs it seems clear that some legible framework that isn't too distortive and guarantees some transparency is desirable. And for genuine future tail risks, it's reasonable to want some degree of foresight and preparation - if reasonably structured. I've spent years reviewing legislation drafts, and so my concerns are less about *whether* but more about the exact language and nature of obligations. There are many bad, onerous proposals out there, and a lot of people are more interested in winning than having something that actually works in the long run. So it's important to keep an eye on a number of things. Some (non-exhaustive) examples: - Scoping mechanisms that are well defined and not over inclusive. Capabilities based metrics still need to be developed before we can go beyond naive FLOP counts, and this has been an unaddressed problem for years (epoch.ai/gradient-updates/th…). - Recognizing the limits of what you can do at the model level (arxiv.org/abs/2606.00047) whilst acknowledging that fast diffusion is net positive for society and adaptation. This one is harder given known unknowns about the future shape of the market. - Ensuring that a regime verifying compliance with a safety framework doesn't become some sort of fishing expedition to effectively slow down R&D. You want to avoid a de facto vetocracy/actively slowing down innovation as a hidden goal. - Ensuring that threat models are clearly specified and standardized, so that everyone is measuring the same thing, and the public/experts can determine whether the right risks are tackled. See frontiermodelforum.org/techn… and frontiermodelforum.org/issue… - Regulatory structures that actually encourage rapidly changing standards/practices. Too often people assume you set the rules once and you're done; this is risky and governance should be a living thing, particularly when considering less intuitive risk conceptions like loss of control. I (unsurprisingly) quite like Google's FARO proposal: static.googleusercontent.com… - To the extent that you want audits, these should be clearly scoped, reasonable in nature, and not subject to conflicts of interest; I've worked on actual audits with Big 4 accountancies before and they're soul crushing, let's not replicate that please. - Critically, you want to ensure a vibrant, competitive frontier, and avoid anything that leads to a regulatory moat or incumbents entrenching themselves. This is why you need good external visibility over standards and expectations. I think we'll get there eventually, and there's a lot of great work happening in parallel - I just really want to avoid the blobification of AI, and going down the bleak path of nuclear, GMOs, housing, NEPA and so on - all cases where precaution overshot. If we strike the right balance we can get the Good World, but that path is narrow and politics tends to degrade quality a lot - the other alignment problem...