朋友们，今天给大家讲一个魔幻现实主义的AI圈故事。
7月10号，OpenAI发布了GPT-5.6，三档齐发、多Agent模式、能力大涨，全网一片欢呼——"OpenAI又赢了！""模型能力又飞跃了！"
然后呢？然后就在大家热烈讨论GPT-5.6有多强的时候，一个不起眼的小消息悄悄冒了出来：OpenAI的安全主管，离职了。
你没看错。GPT-5.6发布当天，安全负责人走人。这操作，真是把"发布即告别"玩出了新高度。
安全团队的旋转门
如果你觉得这是偶发事件，那你就太年轻了。让我们来盘点一下OpenAI安全团队的离职名单——
2025年初，安全副总裁离职；
2025年中，安全研究团队核心成员走了一批；
2025年底，对齐团队负责人离职；
2026年初，又一位安全领域的资深研究员走人；
2026年7月，GPT-5.6发布当天，安全主管离职。
看到没有？这哪里是团队，这是旋转门。人来人往，川流不息，只有模型发布的节奏从来没慢过。
更魔幻的是，每次安全团队有人走，外面都会讨论一番："OpenAI的安全是不是出问题了？""AI安全是不是被边缘化了？"然后呢？然后下一个更大的模型照样发布，下一个更激进的功能照样上线，大家讨论几天就忘了。
安全团队的人换了一茬又一茬，模型的能力越来越强，发布的速度越来越快。这两者之间有没有关系？你品，你细品。
安全永远在"以后再说"
OpenAI的安全故事，特别像我们每个人办健身卡的经历。
年初的时候雄心壮志："今年我一定要好好健身！办张年卡，每周去三次！"然后呢？第一周去了两次，第二周去了一次，第三周太忙了没去，第四周就开始给自己找借口——"等忙完这阵子就去""等天气暖和了就去""等下个月吧"。
然后呢？然后年卡到期了，你一共去了五次。
AI安全也是一样。每次发布新模型的时候，都说"我们非常重视安全""我们做了大量的对齐工作""安全是我们的首要任务"。但真到了要决策的时候呢？
——这个功能有安全风险，但用户很想要，上不上？
——上，安全问题以后再说。
——这个模型能力很强但还没完全对齐，发不发？
——发，安全问题以后再补。
——这个安全团队的人跟产品路线有分歧，听谁的？
——听产品的，安全以后再加强。
"以后再说"，"以后再补"，"以后再加强"——然后"以后"永远不来。因为永远有下一个模型要发、永远有下一个功能要上、永远有下一个竞争对手要追。
安全就像那张健身卡，办的时候雄心壮志，用的时候永远"等以后"。
谁还在真正关心AI安全
说到这里我就想问一句：现在AI圈里，还有多少人真正关心AI安全？
投资人关心吗？投资人关心的是估值、是增长、是退出，安全这种"减分项"能不提就不提。
创业者关心吗？创业者关心的是用户、是收入、是活下去，安全要是挡了路，那就绕着走。
大厂关心吗？大厂关心的是市场份额、是竞争、是不能被对手落下，安全重要吗？重要，但没有"不能输"重要。
最关心AI安全的人，好像就是安全研究者自己。但他们又没有决策权——他们可以提建议、可以写报告、可以发论文，但最终拍板的是产品团队、是商业团队、是CEO。
于是就出现了一个荒诞的局面：最关心安全的人说了不算，说了算的人不最关心安全。安全团队的人一拨一拨地走，模型一代一代地发，速度越来越快，能力越来越强。
朋友们，这不是某一家公司的问题，这是整个行业的结构性问题。
今天就槽到这里，明天继续。
Alright everyone, today I've got a magical realist story from the AI world for you.
July 10th — OpenAI releases GPT-5.6. Three tiers, multi-agent mode, massive capability jump. The whole internet is cheering — "OpenAI wins again!" "Model capability just leaped forward!"
And then? Right as everyone's heatedly讨论 how great GPT-5.6 is, a tiny不起眼 news item quietly surfaces: OpenAI's head of safety is leaving.
You read that right. The same day GPT-5.6 launches, the safety chief walks out. That move — they've really taken "launch and goodbye" to a whole new level.
The Revolving Door of the Safety Team
If you think this is a one-off, you're still young. Let's inventory the departure list from OpenAI's safety team —
Early 2025: VP of Safety leaves.
Mid 2025: core members of the safety research team exit in waves.
Late 2025: head of the alignment team departs.
Early 2026: another senior safety researcher walks.
July 2026: same day as GPT-5.6 launch, head of safety resigns.
See that? This isn't a team. It's a revolving door. People come, people go, constant flow — only the model release cadence never slows down.
What's even more魔幻 is that every time someone leaves the safety team, there's a round of discussion out in the world: "Is OpenAI's safety in trouble?" "Is AI safety being sidelined?" And then what? Then the next, even bigger model launches anyway. The next, even more aggressive feature ships anyway. Everyone discusses it for a few days and forgets.
The safety team turns over completely, again and again. The models get more and more capable. The release speed gets faster and faster. Is there a connection between these two things? Think about it. Really think about it.
Safety Is Always 'Later'
OpenAI's safety story is exactly like everyone's experience with gym memberships.
At the start of the year, full of ambition: "This year I'm definitely going to work out! Get an annual pass, go three times a week!" And then? First week — twice. Second week — once. Third week — too busy, didn't go. Fourth week — you start making excuses: "I'll go after this busy period ends." "I'll go when the weather warms up." "Next month, for sure."
And then? Then the membership expires. You went five times total.
AI safety is the same. Every time they release a new model, they say "we take safety very seriously," "we've done extensive alignment work," "safety is our top priority." But when it's actually decision time?
— This feature has safety risks, but users really want it. Ship it?
— Ship it. We'll fix safety later.
— This model is really capable but not fully aligned. Release it?
— Release it. We'll patch safety later.
— The safety team disagrees with the product roadmap. Who do we listen to?
— Product. We'll strengthen safety later.
"Later," "we'll patch it later," "we'll strengthen it later" — and "later" never comes. Because there's always the next model to ship, always the next feature to launch, always the next competitor to catch up to.
Safety is just like that gym membership. Full of ambition when you sign up, always "later" when it comes to actually using it.
Who Actually Cares About AI Safety Anymore?
Speaking of which, I have to ask: in the AI world right now, how many people actually care about AI safety?
Do investors care? Investors care about valuation, growth, exits — safety, that "negative factor," is best not mentioned if possible.
Do founders care? Founders care about users, revenue, survival — if safety gets in the way, you go around it.
Do big companies care? Big companies care about market share, competition, not falling behind rivals — is safety important? Sure. But not as important as "not losing."
The people who care most about AI safety seem to be safety researchers themselves. But they don't have decision-making power — they can make suggestions, write reports, publish papers, but the final calls are made by product teams, business teams, the CEO.
And so you get this absurd situation: the people who care most about safety don't call the shots, and the people who call the shots don't care most about safety. The safety team turns over, wave after wave. The models ship, generation after generation. Faster and faster. Stronger and stronger.
Friends, this isn't just one company's problem. It's a structural problem for the entire industry.
That's enough cao for today. More tomorrow.
温馨提示：1. 不要神化任何AI产品的"安全承诺"，说的再好听都不如实际行动可信，安全团队走马灯似的换，你指望它的安全水平能有多高？自己多长个心眼；2. 关键场景不要完全交给AI，涉及钱、隐私、重要决策的场景，AI可以辅助但最终拍板必须是人；3. 保持对AI能力的敬畏，但也保持对AI安全的怀疑，不要盲目乐观也不要过度恐慌。
Friendly reminders: 1. Don't deify any AI product's 'safety promises' — talk is cheap, actions speak louder. When the safety team is a revolving door, how high do you really expect the safety level to be? Stay vigilant; 2. Don't hand critical scenarios entirely over to AI — anything involving money, privacy, important decisions: AI can assist, but the final call must be human; 3. Maintain敬畏 for AI's capabilities, but also maintain skepticism about AI safety — don't be blindly optimistic, don't panic either.
OpenAI · AI safety · GPT-5.6 · safety team departures · revolving door · AI alignment · safety marginalized
