Hive Mind, Another Big Word?
TL;DR: Yegge’s piece on Anthropic’s hive mind is not really about smart people collaborating well. The interesting part is that hive mind has preconditions. AI did not invent elite teams. It changed the ratio: one strong person can now operate more like a mini surgical team. The new bottleneck is judgment, not execution.
Steve Yegge wrote about Anthropic’s hive mind. A lot of people focus on “how smart people collaborate.” But I don’t think that’s the point.
I’m a little wary of the term “hive mind.” Another big word?
People talk about shared context, high agency, fast iteration, everyone being able to see everyone else’s work. It all sounds right. What I’m thinking is: where are the boundaries?
Without boundaries, “hive mind” is just “smart people communicate more.” That’s a slogan, not an operating model.
Brooks and The Mythical Man-Month
Brooks had a version of this problem in The Mythical Man-Month. The book talks about a surgical team: one surgeon does the critical work, everyone else supports. The chief programmer team follows the same idea: organize the team around the person who best understands the system’s intent, and put everyone else in supporting roles.
That model solves an old problem: how to protect conceptual integrity. The method is simple. One strong person owns the system’s brain. Everyone else reduces the load.
If Brooks’s surgical team is one surgeon with a support crew, Anthropic looks more like many surgeons in the same room. Everyone is strong. Everyone can push a piece of the work forward. Everyone has judgment.
The problem shifts.
Old problem: how to make average people support an expert.
New problem: how to keep a room full of experts from grabbing the steering wheel at the same time.
Elite Teams Have Coordination Problems Too
This is the part many discussions skip. Elite teams still have coordination problems. The problems are just shaped differently.
Regular teams struggle with training, supervision, and knowledge transfer. Elite teams struggle with territory, taste, priority, and decision rights.
Everyone can deliver. That is exactly why boundaries matter.
The first boundary is meaningful work.
If there is enough genuinely important work, people do not need to fight. You take this frontier, I take that one. Everyone has room to move.
If the valuable work slots are scarce, politics comes back. Even smart people start fighting for ownership, visibility, and the roadmap. That is the problem most organizations face.
The second boundary is shared context.
Without it, everyone ends up building in private. People need to see what others are doing, which judgments have been validated, and which directions have already been abandoned. That is shared context.
The third boundary is an explicit judgment mechanism.
Shared context only shows everyone the same map. It does not make decisions for the group. Pile on more channels, more docs, more AI summaries, more weekly updates, and the result may still be: everyone is informed, nobody is aligned. That is louder Slack culture.
A real hive mind needs a judgment mechanism: user feedback, evals, tests, production incidents, demos, explicit leadership constraints. Their job is not to add more information. Their job is to end arguments. Something has to pin different judgments to reality.
AI Changed the Ratio
AI did not invent elite teams or shared context. These have been around in organization design for a long time. AI changed the ratio.
A surgeon used to need a support crew. Now, for prototyping, research, and review, one strong person can temporarily summon many support roles: researcher, coder, reviewer, summarizer, analyst, prototype builder.
So a high-level person now looks more like a mini surgical team. They can open more branches, try more directions, and generate more options.
That makes hive mind more likely to emerge, and easier to lose control of.
Because the new bottleneck is not execution. It is judgment. AI made option generation cheap. The question is: who evaluates those options? Who kills branches? Who decides which prototype represents a real need, and which one merely demos well?
If judgment does not keep up, AI does not make the organization smarter. It makes the organization louder.
The Preconditions
At this point, the preconditions for hive mind are fairly clear:
-
Enough meaningful work.
-
High enough talent density.
-
Shared context, but not an information flood.
-
Fast feedback loops.
-
Explicit judgment mechanisms.
And one more thing: people have to be willing to let reality kill their ideas.
That last one is the hardest.
Many people say they like high agency. What they really like is having agency themselves. When other people also have agency, it becomes less romantic.
That is why I think the Anthropic case is interesting, but you cannot copy-paste it as a culture slogan.
“Hive mind” isn’t a vibe. It’s a set of preconditions.
- No meaningful work, and it becomes politics.
- Too little shared context, and it becomes parallel chaos.
- No explicit judgment mechanism, and it becomes louder Slack.
- No judgment, and it becomes AI-generated noise.
The real question isn’t: can we become a hive mind?
The question is: do we have what it takes to live with it?
简单说： Yegge 写 Anthropic 的 hive mind，重点不是”聪明人协作好”。重点是 hive mind 有一组前提条件。AI 没发明 elite team，它改变的是 ratio——一个强人现在更像一个 mini surgical team。但新瓶颈是 judgment，不是 execution。
Steve Yegge 写 Anthropic 的 hive mind。很多人会把重点放在”聪明人怎么协作”上。但我觉得重点不是这个。
Brooks 的人月神话
我对 “hive mind” 这个词有点疑惑。又一个大词么？
大家说 shared context，说 high agency，说 fast iteration，说所有人都能看到所有人的工作。听起来都很正确。我想的是：边界在哪里？
如果不讲边界，hive mind 就只是”聪明人多沟通一点”。这不是 operating model，这是口号。
Brooks 讲过一个类似的问题。The Mythical Man-Month 里面有 surgical team 的说法：一个 surgeon 做最关键的工作，其他人提供支持。chief programmer team 也是这个思路，围绕一个最理解系统意图的人组织，其他人做 supporting roles。
这个模型解决一个老问题：怎么保护 conceptual integrity。方法很简单，一个强人负责系统的脑子。别人帮他减负。
如果 Brooks 的 surgical team 是一个 surgeon 加一圈支持者，那 Anthropic 更像很多 surgeon 同时在场。每个人都很强，每个人都能推进一块东西，每个人都可能有自己的判断。
问题就变了。
以前的问题是：怎么让普通人支持高手。
现在的问题是：怎么让一群高手不要互相抢方向盘。
Elite Team 的 Coordination Problem
我觉得这是很多讨论没讲清楚的地方。Elite team 不是没有 coordination problem，只是 coordination problem 换了形状。
普通团队的问题是培训、监督、知识转移；Elite team 的问题是 territory、taste、priority、decision rights。
大家都能干，反而更需要边界。
边界的第一层是 meaningful work。
如果有足够多真正重要的工作，人不需要互相争。你做这个 frontier，我做那个 frontier，大家都有地方施展。
如果真正有价值的 work slot 很少，politics 就来了。再聪明的人也会开始抢 ownership、抢 visibility、抢 roadmap。这就是大多数组织面临的问题。
边界的第二层是 shared context。
没有这一层，可能就变成了闭门造车。团队里的每个人应该要看见彼此在做什么，知道哪些判断已经被验证，哪些方向已经被放弃。这就是 shared context。
边界的第三层是明确的判断机制。
Shared context 只是让大家看到同一张图。它不负责替大家做判断。堆更多频道、更多 docs、更多 AI summary、更多 weekly update，最后可能只是 everyone is informed, nobody is aligned。这就变成了 louder Slack culture。
所以真的 hive mind 还需要一套明确的判断机制：用户反馈、evals、tests、production incident、demo、明确的 leadership constraint。它们的作用不是增加信息，而是结束争论。总得有东西把不同判断压到现实上。
AI 改变的是 Ratio
AI 没有发明 elite team，也没有发明 shared context。这些都是组织管理中长久存在的玩意。AI 改变的是 ratio。
以前一个 surgeon 需要一圈人支持。现在，在 prototype、research、review 这些环节里，一个强人可以临时召唤一部分 support roles：researcher、coder、reviewer、summarizer、analyst、prototype builder。
所以一个高水平的人，现在更像一个 mini surgical team。他能开更多分支，试更多方向，产出更多 options。
这会让 hive mind 更容易出现，也更容易失控。
因为新的 bottleneck 不是 execution，而是 judgment。 AI 把 option generation 变便宜了。问题是，谁来判断这些 option？谁来 kill branch？谁来决定哪个 prototype 代表真实需求，哪个只是 demo 好看？
如果 judgment 没跟上，AI 不是让组织更聪明，而是让组织更吵。
前提条件
到这里，hive mind 的前提就比较清楚了：
-
足够多 meaningful work。
-
足够高的 talent density。
-
Shared context，但不是信息洪水。
-
快速的 feedback loop。
-
明确的判断机制。
还有：大家愿意让自己的想法被现实打掉。
最后这个最难。
很多人说喜欢 high agency，其实喜欢的是自己有 agency。别人也有 agency 的时候，就没那么浪漫了。
所以我觉得 Anthropic 这个 case 有意思，但不能直接拿来当文化口号。
“Hive mind” 不是一种氛围。它是一组前提。
- 没了 meaningful work，就变 politics。
- Shared context 少了，就变 parallel chaos。
- 没了明确的判断机制，就变 louder Slack。
- Judgment 没了，就变 AI-generated noise。
真正的问题不是：我们能不能变成 hive mind。
问题是：我们有没有资格承受它。