Our response to the US ban on Fable 5 and Mythos 5
摘要
美国政府近期对 Anthropic 的 Fable 5 和 Mythos 5 模型发布出口管制指令，禁止非美国公民（含盟国国民）访问，导致模型被迫全球下线。法律 AI 研究公司 Isaacus 对此回应称，此举凸显了依赖美国中心化 LLM 的脆弱性。Isaacus 借此重申其“AI 主权”立场，坚持模型应支持离线私有化部署（air-gapped self-hosting）且能在消费级硬件运行，以确保在面临政策管制或商业变动时，政府及企业级客户的 AI 能力不受影响。
荐读理由
针对美基大模型可能因政策突变而全球下线的极端风险，你可参考其‘模型小型化+私有化部署’的架构方案，作为规避断供风险并切入政府等高合规市场的商业壁垒。
原文
Our response to the US ban on Fable 5 and Mythos 5
Today, the US government issued an export control directive banning access to Anthropic’s latest LLMs, Fable 5 and Mythos 5, by anyone who is a foreign national, including even Anthropic’s own employees. Consequently, Anthropic has since taken Fable 5 and Mythos 5 offline globally.
From a practical perspective, this move is unlikely to affect many users, as Fable 5 was only released three days ago and Mythos 5 is only available to select partners. Nevertheless, the ripple effects will be enormous and long-lasting.
To the best of our knowledge, this is the first time the United States has issued an export control directive for LLM access, ever. This directive affects not only the United States’ closest allies, the UK, Canada, Australia, and New Zealand, but their nationals as well, regardless of whether or not they reside in the US.
In effect, any application depending on US-based LLMs is subject to being shut down at any moment as a result of an export control directive. For Fable 5 and Mythos 5, there was no forewarning; access ceased almost immediately.
For us at Isaacus, an Australian-based foundational legal AI research company operating in the US, UK, Canada, Australia, New Zealand, Singapore, and the EU, this reality only further cements our position that frontier artificial intelligence ought to be accessible to everyone, rather than concentrated in the hands of a select few frontier labs, ourselves included.
To that end, every single model we’ve released has, from day one, been available for air-gapped self-hosting. We have no plans on changing that. In fact, we’ve doubled down on AI sovereignty, actively shaping our product roadmap around delivering value over frontier general purpose AI models for legal tasks while keeping our models small enough to run on consumer hardware.
Our approach to AI sovereignty is as much commercially motivated as it is reflective of our own values. Certainly, making our models available for self-hosting is what has enabled us to, as a team of two founders until recently, land enterprise deployments serving multiple Australian government departments. But also, we simply enjoy disrupting monopolies, whether they are over data, LLMs, or anything else. Us continuing to offer self-hosting helps keep frontier intelligence accessible to the world, while, commercially speaking, also making us a thorn in the side of competitors who either refuse or, more often, due to their dependence on external LLMs, are unable to do the same.
As misused and misunderstood as the term may be, AI sovereignty is genuinely important. Anyone depending on AI for mission-critical work, no matter whether they’re a startup or a government, needs to consider what happens when that intelligence disappears. It might not necessarily be as dramatic as an export control directive; it might be as simple as an acquisition. Anthropic, for example, recently acquired Stainless, a platform previously used to automatically generate software development kits for OpenAI, Google, and (unfortunately) Isaacus, and then immediately wound it down. Luckily for all of our customers, self-hostable, air-gapped sovereign AI substantially minimizes that risk.
Given how both commercially viable and valuable sovereign AI has become, moving forward, we’re working hard to make not only our models but also our future AI-powered applications, including the Blackstone Graph and Isaacus Research platform, fully self-hostable.
这条对你有帮助吗？