聊聊GPT-5.6和Fable 5的审查闹剧：政府逐单审批不合理，付费用户可能白花钱，国际用户可能被抛弃。
评论指出，自GPT-2以来，模型安全审查不可避免，当前GPT-5.6和Fable 5等前沿模型需通过美国政府的逐客户审批，导致付费用户被延迟访问数天至数月。作者支持审查但反对永久封锁和基于公民身份的准入限制，认为政府应建立白名单机制加速预览阶段。已有用户因延迟起诉政府，国际用户可能被限制使用低规格模型。
What's more shocking is people being shocked. The …
What's more shocking is people being shocked. The model clampdown was inevitable. There's no need to blame Dario and Anthropic. This was always going to happen, arguably since GPT-2 if not earlier.
We've hit the threshold where models are strong enough at cybersecurity vulnerabilities that every release will get reviewed. That was predictable. The process has come earlier than expected, but earlier is better than too late. We need to go through this, even if the execution needs work.
I'm pretty sure GPT-5.6 and Fable 5 will eventually get approved for more general use. But the execution right now is messy. The US government can't literally approve customer by customer every time a lab releases a model. That's absurd. There should be some sort of whitelist system, get the preview phase done quickly, let companies and institutions test for vulnerabilities, then release.
As for general users, we're paying subscriptions for access to the latest models. Top-tier plans are explicitly for this. If we're now blocked for days, weeks, possibly months while waiting for approval, why would anyone outside the approval queue keep paying? Someone's already sued the government over this.
And what about international access? If you're not a US ally, are you just stuck with lower-spec models indefinitely? That's a real question with no clear answer right now.
I support the concept of a review phase to find and fix vulnerabilities before public release. But I'm not on board with permanent blocks or citizenship-based vetting to access frontier models. I'm hoping that's not where we're at yet, but I'm getting the feeling it might be eventually.