On one cheap AI provider, you only rent its uptime.把工作流押在一个便宜的单一模型上，你买到的是算力，租到的却是它的在线率——它崩，你就崩。
DeepSeek went down again this evening: the web app, the chat service, and both the V4.1 Flash and V4 Pro APIs were disrupted at once. The status page read "fix applied, monitoring," yet all three were still flagged as degraded. This came just 19 days after the roughly five-hour outage on October 15; by some counts it is already the 18th major disruption this year.
The mechanism is simple. The model is cheap enough and easy enough to wire in that a lot of downstream work treats it as the default backend — and many people wired in only this one. When it breaks, every workflow built on it stalls at once, no matter how good your own code is. And no SLA means no credits and no promised recovery time; all you can do is refresh the status page and wait.
Cheap is not the trap. Treating your only provider as infrastructure that never fails is. The people who ride out an outage like this already kept a second path: a fallback route to another model, a cache of the critical results, a minimal version they can still ship when it breaks.
What would change this view? DeepSeek offering an SLA with real credits and pushing its failure rate back to single digits a year. Until then, treat it like a power grid that can trip, not a light that is always on.
今天傍晚 DeepSeek 又崩了：网页、对话，连同 V4.1 Flash、V4 Pro 两个 API 一起中断，官方状态页挂出『已实施修复、正在监控』，三项服务仍显示性能下降。这距 10 月 15 日那次约 5 小时的停摆只隔了 19 天；据报道，已是今年第 18 次大面积异常。
机制不复杂：模型够便宜、接入够省事，于是大量下游把它当默认后端，很多人只接了这一个。它一中断，所有把它当唯一后端的工作流同时停摆，跟你自己的代码好不好无关。而没有 SLA，就意味着没有赔付、没有承诺的恢复时间——你能做的只有刷新状态页等。
便宜不是陷阱，把唯一的供应商当成不会断的基础设施才是。扛得住中断的人，早就留好了第二条路：另一个模型的降级通道、关键结果的缓存、它崩了也能先发出去的最小版本。
什么会改变这个判断？DeepSeek 拿出可赔付的 SLA，并把故障频率压回一年个位数。在那之前，把它当会跳闸的电网用，别当成永远亮着的灯。
Sources来源
Add WAKENEX as a preferred source on Google在 Google 把 WAKENEX 设为首选来源
READ NEXT接着读
Hedge funds cutting AI stocks is not the market revising AI.这轮 AI 股抛售，是对冲基金在给又挤又借的仓位平仓，不是市场改判了 AI。
Goldman Sachs prime-brokerage data shows hedge funds net-sold about $20 billion of tech over two weeks; five straight sessions of net selling in global AI stocks erased every net purchase made since the start of 2025. The tell is the pace: they cut existing longs about 3.5 times faster than they added new shorts. That is not a bet that demand for chips, data centers, or software is about to fall apart. It is risk reduction on a crowded, borrowed book, and Morgan Stanley logged a de-gross of roughly 7.8% of total equity exposure, the largest in its series since 2016.
高盛（Goldman Sachs）主经纪商数据显示：过去两周对冲基金净卖出约 200 亿美元科技股；连续五个交易日对全球 AI 股的净卖出，就抹平了 2025 年初以来的全部净买入。关键在节奏——他们减已有多头的速度，是新开空头的 3.5 倍。这不是在押注芯片、数据中心或软件的需求会垮，而是在给一本又挤又借的账降风险：摩根士丹利记录到约 7.8% 的总持仓净减仓，是其数据序列里 2016 年以来最大的一次。
Open note阅读短评
The agent removed the coordinating step, not the reviewing oneAgent 移走的是「协调」那一步，不是「审查」那一步
DeepSeek open-sourced a coding agent built to need as little of your attention as possible. Read its own design, and the limit it cannot remove shows up.
DeepSeek 开源了一个号称尽量少占用你注意力的编码 agent。读一读它自己的设计，那个它无法移除的上限就冒出来了。
Read essay阅读长文
Peak growth isn't peak value.见顶的是增速，不是价值。
Goldman's 'inflection' is an inflection in the growth of AI capex's contribution to US GDP — about 1.5%, peaking in Q4 2026 — not a peak in spending itself. Reading it as 'AI is about to collapse' uses a pace metric to answer a question it can't: is this investment worth it.
高盛说的“拐点”，是 AI 资本开支增速对 GDP 贡献的拐点（约 1.5%，今年四季度见顶），不是开支本身见顶——把它读成“AI 要崩了”，是在用一个增速指标，回答“这笔投入到底值不值”这个它答不了的问题。
Open note阅读短评