OpenAI + Anthropic 同周呼吁放慢 — Nate Herk 视角
2026 年 6 月，两家顶级 AI 实验室在同一周分别公开呼吁"放慢 AI 发展"。但 Nate 的关键洞察："我们要求暂停" ≠ "我们愿意停"。他们都在说"我们停不下来，需要别人建个裁判机构强制所有人一起停"。这是两回事。
| 日期 | 事件 | Nate 解读 |
|---|---|---|
| 几天前 | Anthropic 发布报告：呼吁"可验证的暂停/放慢方案" | 公开说"放慢 AI" |
| 6/8 | OpenAI "Built to Benefit Everyone" + 申请上市 | Sam Altman + 首席科学家联合署名 |
| 6/9-6/11 | 公开讨论 / 媒体分析 | — |
| 6/12 | US 政府要求 Anthropic 暂停 Cloud Mythos + Cloud Fable | 第一个实质政府介入 |
"The two companies racing the hardest are also the two who have turned around and said, 'Hey world, we need to slow down.' So what the heck is going on?"
| # | 目标 | 细节 | Nate 评估 |
|---|---|---|---|
| 1 | AI 自动化研究员 | AI 做大部分 AI 研究 | 预测 2028 年 3 月前实现 |
| 2 | 加速经济 + 分享收益 | 给地球每个人一个 personal AGI | ✅ 同意"不想自动化一切" / "权力不应集中" |
| 3 | 国际监管组织 | 原文：take coordinated action, including slowing frontier development when needed | ⚠️ 真正 dramatic — 要求别人强制自己慢下来 |
目标 #3 才是真信号 — AI 行业最头部的公司公开承认：单靠市场激励，他们自己停不下来。这从 Sam Altman 嘴里说出来 = 真正开始严肃了。
"They're both saying we can't stop on our own. We can't give up our lead. So somebody needs to build some sort of group or like a referee that makes all of us stop together."
"The break they're asking for isn't a break that they're willing to pull. It's a break that they want somebody else to create and hold."
"Until somebody cracks that, I think that asking for a global pause button is either a genuine cry for help or just the safest thing you can say when you already know you can't actually stop. Maybe a little bit of both."
| 类比 | 核武器 | AI 训练 | 可见度 |
|---|---|---|---|
| 原料 | 铀（mining 可查） | GPU 芯片（单供应链：英伟达主导） | ✅ 高度可见 |
| 设施 | 离心机难隐藏 | 数据中心 = 城市级电力（footprint 可见） | ✅ 高度可见 |
| 信号 | 核试验 | 大规模电力消耗 + 芯片订单 + 建筑施工 | ✅ 高度可见 |
| 供应链 | 多国可生产 | 单条供应链（Anthropic 自承：bottleneck 是 compute） | ✅ 更可控 |
Training frontier model = tens of thousands specialized chips + 一个建筑 pull 一个 small city 的电
Footprint 难藏：construction + power + massive chip orders = 都留 trail
Verification 难但可解。真正难的是激励翻转。
"A slowdown only holds if breaking it costs you more than winning ever could."
✅ 假设成立
✅ 难但可解
❌ 整个条约崩溃
"No matter how bulletproof a contract is, if it covers every single scenario, it's only as good as the person on the other side of it."
"How do you get to a world where the winning move is to respect a slowdown and you actually like think about society and like how can everyone be doing this AI thing together rather than competing?"
"The brand of AI right now is honestly terrible."
"I don't understand how the gap got so wide."
US 政府 正式要求 Anthropic 暂停 Cloud Mythos + Cloud Fable
Anthropic 声明"会恢复" · 时间：6/12 · 视频录制时事件
| 维度 | OpenAI/Anthropic 要求 | Mythos 下架 |
|---|---|---|
| 性质 | 多边条约 / 国际组织 | 单边政府行为 |
| 主体 | 公司自己要求 | 政府对公司的命令 |
| 难度 | 极高（所有国家同意） | 相对低（一个国家政府） |
| 先例 | 无 | 第一个实质政府介入 |
US 政府可以单边停一个产品 → 比国际条约容易
Nate 承诺：会做专门视频分析
"The only thing that you fully control is how good you get at this stuff and what you know about it."
"I wouldn't bet on a company. I just bet on yourself."
Pick one thing you do over and over:
不要发明新 workflow，就用 AI 做你已经在做的事
"The human role that stays valuable is judgment, taste, deciding what's actually worth doing."
"Because the moment everybody's got a personal AGI, the tool stops being the edge."
Nate 说"renting intelligence"是问题 → 本地 LLM = 自有 intelligence → 反垄断准备
LiteLLM 多 provider 路由 = 降低 vendor lock-in 风险（不仅是 cost，是 geopolitical）
自动化的是 "how"（执行）→ 不能自动化的是 "what"（决定做什么）