Anthropic CEO Dario Amodei 提出AI“放慢发展”计划，但被批评未真正涉及中国

You can’t have a real slowdown and be a total China hawk. If you are truly committed to the slowdow...

You can’t have a real slowdown and be a total China hawk. If you are truly committed to the slowdown, you have to make a deal. Robert Wright @robertwrighter I was initially encouraged by reports that Anthropic CEO Dario Amodei is proposing an AI "slowdown." But it turns out that (1) It's not much of a slowdown (no pause of training runs—and parts of the plan, as I explain below, work *against* slowing down); (2) He's still wed to his Manichaean democracy-vs-autocracy framing (see steps 2 and 3 of his three-step plan in the screenshot, and my critique of them below). Dario’s plan has its strong points. The internal monitoring of AI labs by METR-like groups would be a big advance. (This idea may also be clever swordsmanship: If OpenAI follows suit, the dual monitoring will show Anthropic to be running a tighter ship, and if OpenAI doesn't follow suit, Sam will look bad. Still, it’s good for Sam to be put on the spot like this, so I’m on team Dario here.) But any serious AI slowdown-and-governance effort should involve immediate, broad-gauged, and good-faith dialogue with China. Before I explain how Dario’s plan could impede that goal, it’s useful to recall how extreme his views on China are. Here’s what he proposed in his 2024 Machines of Loving Grace essay, as summarized in my book The God Test: ‘Amodei ponders the prospect of the most powerful AIs being in the hands of an “authoritarian bloc”—whose most formidable member, needless to say, would be China. To preempt this development, he says, the US should lead a coalition of democracies that would restrict the authoritarians’ access to key technologies and then would “use AI to achieve robust military superiority,” after which these democracies could “parlay their AI superiority into a durable advantage.” This could “lead to an ‘eternal 1991’—a world where democracies have the upper hand.” Then, Amodei says, the democratic coalition could use this stick to get these repressive autocracies to “give up competing with democracies” unless they want to “fight a superior foe.” And, for added incentive, they could be offered the carrot of access to advanced AI as a reward for finally seeing the light.’ This is a very dangerous strategy. As Dan @Hendrycks et. al. noted in the 2025 paper Superintelligence Strategy, if a superpower is getting close to reaching superintelligence, and a rival superpower believes this could confer complete dominance, the rival has a strong incentive to pre-empt this achievement via extreme measures of the kind that could escalate quickly to war, even nuclear war. (Reminder: China is within easy striking distance of Taiwanese factories that produce the world’s most advanced AI chips and aren’t allowed to export them to China.) Dario’s proposal does seem to allow for near-term dialogue with China on narrow issues and, maybe, eventually on less narrow ones. Does that mean he's starting to walk back his plan for an end-of-history US-China showdown? Apparently not. His new proposal says US AI labs *shouldn’t* slow down enough to endanger “the lead that US companies have over authoritarian regimes, chiefly the Chinese Communist Party.” Well when the lead is 3-6 months, that doesn’t permit much of a slowdown! In hopes of protecting (even, he says, expanding) this lead, Dario advocates not just keeping current restrictions on China’s access to chips and other AI tech but tightening them. So, from Beijing’s point of view, here’s the plan: There are two clubs, one for good guys and one for bad guys, and China doesn't qualify for the good guy club; and the good guys are going to amp up what China already sees as an aggressive, even belligerent, restriction of its access to tech. (Remember: China considers the land beneath the chip factories rightfully China’s and the people who make the chips rightfully Chinese—you don’t have to agree with this view to acknowledge its reality and its relevance to American diplomatic strategy.) If Trump embraced Dario’s plan, that would get his coming summit with Xi off to a bad start. The stated premise of Dario’s plan is solid: Things are moving so fast that we need to quickly engineer a slowdown of AI advance and meanwhile develop effective AI governance. But there can’t be an appreciable and lasting slowdown without China’s cooperation (which helps explain why Dario’s plan, on close reading, doesn’t entail one). And we can’t get the AI governance we need without China’s cooperation, either. This technology’s various dangers (most obviously a bioweapon-caused pandemic, but many others as well) tend to cross borders readily. In the age of AI, national security will require strong international governance. Some of Dario’s wariness is warranted. Yes, you of course need to be able to verify compliance with international agreements on mutual restraint. And that will in some cases be more challenging with AI than with nuclear arms—which is one reason we need to start working with China now, rather than do things that make working with China harder. I think Dario’s China hawkism is sincere—and not (as with some Silicon Valley figures) just a cynical talking point that supports policies he favors. But it’s still a dangerous ideology. This is a man who wants to race China to superintelligence and then use the resulting “military dominance” to bring China to its knees, threatening grave consequences unless Beijing complies with vague but extreme-sounding demands. (Does ‘give up competing with democracies’ mean regime change, or what? Could you elaborate, Dario?). When someone with an ideology like this proposes a plan that in theory could lead to eventual global cooperation, the plan should be inspected for signs that it's unlikely to actually have that effect. In this case the inspection doesn’t take very long. None of this is to say that Dario's proposal--as distinguished from the actual complete realization of it--can't bring net benefit. As I said, the plan has its strong points. Besides, misleading headlines are still headlines, and headlines about a proposed "slowdown" can move the conversation in a positive direction, perhaps even becoming a self-fulfilling prophecy. [Here's Dario's proposal: darioamodei.com/post/we-must-p… Excerpts from every chapter of my book on AI, The God Test: Artificial Intelligence and Our Coming Cosmic Reckoning, can be found here: thegodtest.net ] 🔗 View Quoted Tweet 💬 1 🔄 2 ❤️ 1 👀 1593 📊 1 ⚡