Yuki Minoh | 状況に埋め込まれた他者

私はIAMASの魔王です

なんか今日16時から、僕は日大の先生とたくさんの人の前で対談することになっているらしいです。 博士の研究員の方もいて僕と先生と3人でやるそうです。 どういうこと……？？？がんばります！

ハァ？全然ちげーだろ。 「お前が深すぎる」だけな。 自信もてよ。卑屈になるな。 じぶんの「自信のなさ」を隠蔽しつつ、それでも自分の凄さを世界に示したいからって、ヘンに攻撃的になるなよ。勿体ないぜ〜〜。

【超悲報】 ロシアで 灯火管制が始まる。

The Fable decision is fundamentally a domestic US policy mess, and it seems likely to resolve itself, albeit chaotically. For middle powers, it's a tempting starting point for ill-conceived 'sovereign' AI takes, but I think the right move is to let it blow over and buy more time. The decision itself seems counterintuitively domestic in scope. USG got worried about the prospect of a jailbreak, didn't feel like it had a particularly precise and effective way to tackle that risk, and defaulted to limiting access in an obvious and available way. The easy way to do this was also a way to make Anthropic uncomfortable and annoy allies, but these seem like secondary considerations at best. That's still relevant information for middle powers--it goes to show that there is no immediately effective and legible reason for USG or American developers to consider their interests in making access decisions. As for middle powers' posture in reacting to this: I think nothing good comes from wading into a politically charged AI policy environment, solidarising with Anthropic and drawing the continued ire of the Trump administration. Right now, middle powers are collateral damage, and acting hastily risks making them parties to the conflict. What would have helped in this case? It's not clear that even a European model would really have. Imagine you had the absolute best case European sovereignty in place, up and running since 2024. would it have a Fable-class model by now? or would it 'only' be at the level of Opus 4.8 - a near-frontier model in its own right that remains available today? you'd have to be quite confident that the European champion would not just be 2 months behind the curve, but at the absolute bleeding edge of frontier development, to make a difference to this particular scenario. Not even the most bullish views of the domestic project makes that kind of outcome particularly likely, so this really is not a particularly incisive wake-up call on European frontier models specifically. But even if you disagree, what would it actually *mean* for this to be a wake-up call for European sovereignty? Are you going to build your own model now? What are you going to do in the--generously--three years between announcing this project and reaping the frontier models it would build: are you willing to give up frontier access in the meantime? Because the resources required for building this frontier model directly trade off against the resources you could invest into guaranteeing access instead (most notably through compute); and the political fallout from announcing an attempt to build the very kind of model capability the US is attempting to restrict would also make future access negotiations harder. Let's say you're willing to bear that delay: do you think a Trump administration that just refused to give you access to Fable is going to let you buy enough frontier chips to train an unrestricted Fable clone yourself? Are you willing to go the mat on semiconductor chokepoints, even if it comes with sky-high costs in Ukraine and trade policy? I don't actually think so. Look into the details of what would be required for a big European push right now, and you'll see the leverage for 'waking up' and divorcing from the US ecosystem simply is not feasible in the current technological or geopolitical environment. I regret that this is the case, but that doesn't make it the case any less! What, then, is the alternative? First, I think it's worth noting that this is fundamentally a very good version of a very bad thing. In a fortuitous turn of events, the Trump administration has picked the most ill-conceived version of access restrictions you could possibly come up with. It's legally fraught, so domestically impactful that it will lead to massive internal pushback, and likely extremely economically harmful. As a result, it will likely go down in flames eventually. The U.S. is not yet in the spot to actually go through with long-term cut off: international markets are still too important, the security situation is not yet sufficiently dire, and so on. So the first live fire exercise of cutting off the rest of the world is going to fail, which means labs and the admin are going to be much more wary of subsequent attempts to do the same, even if they end up more sophisticated. Second, I think the access recipe is fundamentally the same as it was yesterday: build leverage on the margins that makes cut-offs like these even less attractive, for instance through access-for-compute deals and by creating deep economic integrations that are economically central to US labs and strategically central to the US supply chain--create a lobby to push back harder against attempts like this in the future. In the future, we can use the resources and capacities that gives us to sprint toward our own frontier project if we must, but right now we neither have the political will nor the relative power to get even close to trying that. Third, and somewhat trivially, we should start thinking about what we want to do the next time this happens. I suspect any analyses that assess whether you can use ASML or any semiconductor chokepoint to avert this will come up short, but there's still value in analysing and then credibly precommitting to threats. Right now, USG did this operating under the assumption there would be absolutely no reaction from middle powers at all. Any plan in the drawer that suggests there is a non-negliblie cost for the US to act like this in the future would be helpful. There's little use in deploying it reactively now; there's lots of value in precommitting to it for the next iteration. That's different than actually going to the mat; the goal here is to play chicken a bit, increase uncertainty and latent risk for the administration in making these decisions to tilt the calculus toward integration, not to go all out on a highly costly tradewar. Fourth, I think this clarifies the specific concerns that could motivate access restrictions. Security concerns, both on misuse as well as distillation and model theft, fundamentally make the US more likely to restrict model access; this time around, it was concerns around reducing surface area for unmonitored jailbreak attempts. That is, in principle, fixable--middle power governments can and should engage with labs to create security conditions that create permission structure for exports and model sharing. Make your infrastructure as secure as they want it to be, and you reduce the risk they consider exporting to you a security vulnerability. Again, I understand if this sounds submissive and uncomfortable to you---but again, all this is necessary even if you go for the maximal sovereignty playbook at the same time, because you will need frontier access in the meantime. Instead of these reasonable responses, I worry that the low-resolution view on this whole affair is to think this should shake middle powers into the wrong kind of action. Realising how important and contingent frontier AI access is quickly leads down the path of wanting to build your own; realising how capricious the American ecosystem is makes you want to divorce from it faster. But for better or for worse, the central implication of this episode is the opposite: as evidenced by this episode being possible at all, middle powers currently do not have the leverage to do much about any of this, and building up this leverage is almost impossible to do in an openly adversarial relationship to the US. In that sense, waking up is not a matter of loud yelling, decisive action or pivotal decisions. For all the internal urgency with which I think we should precommit to some leverage and shore up our security concerns, I still think the optimal strategy is one of public restraint and progress on the margins of the current playbook. I'm just not sure there's that much to wake up from - this is just what life is like for now.

WeeklyOchiai有働さん回，延びすぎてゼミの学生さんから電話を受信する稀有な回でした．ありがとうございました！