2026 年 7 月，我为什么还在用 DeepSeek
停更近一年半又回来了！这一年多时间，这个世界发生了很多变化。最大的就是AI在你生活中的出现频率越来越高，基本已经成为很多人日常不可替代的帮手。
Back after a year and a half! So much has changed — the biggest being AI showing up more and more in daily life, becoming an irreplaceable assistant for many.
今天写的这篇的开始，还是想要吐槽下A厂的Claude。的确，之前一直用Opus 4.6-4.7的Claude Code。使用效果上真的好，但也是最心累…一会儿封账号、一会儿模型不让用等等。正所谓"不作死就不会死"，A厂的"作"在AI界也算是"标杆"了！
Let me start by venting about Company A's Claude. Yes, I used Opus 4.6-4.7 in Claude Code — great results, but exhausting. Account bans, model restrictions, the works. As they say, "don't dig your own grave" — Company A's self-sabotage is legendary in the AI world!
实际，前段时间改用O厂的Codex。沉稳、谨慎的表现，让人很放心。但有时GPT5.5谨慎的让我想"骂人"。另外，每个月200刀的订阅其实开支并不小。国内模型，GLM在升级到5.2之后能力表现很强。思路宽广，有时别的模型解决不了的问题，它一下就找到新思路。所以我称呼它为"AI华强北"…哈哈！但介于之前发现过GLM家模型在代码中会自己硬编码一个用户名等事件，我对它的信任度不是很高…
Then I switched to Company O's Codex. Steady, cautious, reassuring. But sometimes GPT-5.5 is so cautious it makes me want to scream. Plus, $200/month isn't cheap. Among domestic models, GLM 5.2 is strong — broad thinking, finds novel solutions where others fail. I call it "AI's Huaqiangbei"... haha! But after the hardcoded credentials incident in GLM 4.7, my trust isn't high...
其实，上面的一些已经体现出一个问题：每个模型都有它自己的脾气、性格和擅长的领域。这也促成了我在几个月前做了一个多模型协同的系统，并在Zenodo上陆续发表了一些关于多模型协同系统的论文。主旨"AI管理AI"，以后有时间我可以聊聊这个…
This already reveals the issue: every model has its own temperament, personality, and strengths. That's what drove me to build a multi-model orchestration system months ago, and publish papers about it on Zenodo. The core idea: "AI governing AI." I'll write more about this when I have time...
说这些，其实想说下没有一个完美的模型适合每一个人。这时你就需要一个多模型协同工作流，我的工作流是：Deepseek V4和GLM5.2调研、做计划；GPT5.5调整思路、审核计划；Deepseek V4负责脏活累活；GPT5.5做最后的代码审核。而日常生活中，DS V4基本就是我的主力模型。很多日常事务管理、植物养护管理、家庭网络监控等等也都是交给它来协助。
My point: no single model is perfect for everyone. You need a multi-model workflow. Mine: DeepSeek V4 + GLM 5.2 do research and planning; GPT-5.5 adjusts and reviews; DeepSeek V4 handles the grunt work; GPT-5.5 does final code review. In daily life, DS V4 is my main model — managing daily tasks, plant care, home network monitoring, all delegated to it.
为何选择Deepseek？
Why DeepSeek?
更新稳定
现在模型更新都是三四个月就更新一次，更新前之前的模型就明显降智。这种不管用户"死活"的营销手法实在是让人觉得累。Deepseek的更新很稳。
Stable updates. Models now refresh every 3-4 months, and the previous version noticeably degrades before the new one launches. This "user be damned" marketing is exhausting. DeepSeek updates are stable.
成本够低
你学会在一个对话过程中保持对话内容一致性，那么模型的缓存命中率能保持在一个高水平的位置上。这时费用会更低。
Low cost. Once you learn to maintain conversation content consistency, cache hit rates stay high, making costs even lower.
DS V4的问题
整体给我的感受是：Deepseek太急，不像GPT5.5那样稳健。很多次，我告诉DS V4：你看清楚问题的全部之后再下手。但还是这样：一上头拿起来就干。不是DS能力不足，而是像个"愣头青"一样没考虑明白就开始上手做，遇到问题再换一个思路。对它来说事情是简单了，但我的Token在它的不断试错过程中直线上升。在发生过几次因为它拿起来就干的情况，并出现问题的情况下。我分析了下它的问题，原来这种情况叫"过早回复"。等于模型在没有了解事情全貌或者相关联机制的情况"头痛医头、脚痛医脚"，结果很多时候会出现因为修改1而影响2，然后为了修改2破坏了3的结构等等的情况。另一个伴随问题是，它在调研和收集资料时习惯凭训练数据"猜"，而不是主动去查官方文档。这让它在处理具体技术问题时容易给出过时或者不准确的方案。所以"强制深度思考协议"也强调了这一点——调研阶段优先查阅官方文档，不能只靠"我觉得"。
DS V4's issues. Overall impression: DeepSeek is too impatient, unlike GPT-5.5's steadiness. I keep telling it: "Read the full picture before diving in." But no — it jumps in headfirst. Not lack of capability, but like an eager rookie who starts before thinking it through, switching approaches only when stuck. Easy for the model, but my tokens skyrocket from its trial-and-error. I analyzed this: it's called "premature response" — the model treats symptoms without understanding the whole system. Fixing A breaks B, fixing B destroys C. Another issue: it "guesses" from training data instead of checking official docs, producing outdated or inaccurate solutions. That's why my "Mandatory Deep Thinking Protocol" emphasizes checking official docs first — no "I think so."
这种情况下，我给DS定了一个"强制深度思考协议"。让它强制提高思考层级，在任何下手改变代码、设置前做完三个阶段的工作：理解、方案、风险。这就是属于"磨刀不误砍柴工"。
So I created a "Mandatory Deep Thinking Protocol" — forcing higher thinking levels, requiring three phases before any code or config change: Understand, Plan, Risk-assess. Sharpening the axe before chopping wood.
其他DS V4的小问题
我还发现DS V4的一些小问题，如：Shell转义盲区、全半角混用、盲目重试等等。在解决前面"过早回复"问题时，顺手把这些问题也解决了。
Other minor issues: shell escaping blind spots, full/half-width character mixing, blind retries. Fixed these while addressing the premature response problem.
"强制深度思考协议"的使用，我的使用感觉DS V4已经有Opus 4.6的使用感受了。这在无论coding还是日常使用上都有着稳定的感受！
With the protocol, DS V4 now feels like Opus 4.6 — stable for both coding and daily use!
停更近一年半，回来写的第一篇居然是聊AI。也好。
First post after a year and a half, and it's about AI. Fitting.
这一年半，AI确实给我带来了很多。有了一套多模型协同系统——让不同模型各司其职，AI管理AI。利用AI的"幻觉"找灵感、做实验。在世纪猜想DBS里做了一些数学研究，通过对量子不确定性来定义经典不确定性问题，对贝叶斯拓扑分析做了深度探索……陆续发了一些DBS数学论文和多篇AI多模型协同的论文。
This year and a half, AI has truly brought me so much. A multi-model orchestration system — different models playing different roles, AI governing AI. Using AI "hallucinations" for inspiration and experiments. Mathematical research on the BSD conjecture, defining classical uncertainty through quantum uncertainty, deep Bayesian topology exploration... Published BSD math papers and multiple AI orchestration papers.
回头看，AI对我来说早就不只是一个问答工具了。它是一个时而让你抓狂、时而给你惊喜的搭档。这篇聊的"强制深度思考协议"，说起来也不过是——我摸清了它的脾气，给它立了点规矩。效果还不错。
Looking back, AI hasn't been just a Q&A tool for me for a long time. It's a partner — sometimes maddening, sometimes surprising. This "Mandatory Deep Thinking Protocol"? Just me figuring out its temperament and setting some ground rules. Works pretty well.
下篇聊聊在AI协助下搭的家庭网络系统，以及AI从"被指挥"到"自己做网管"的小实验。
Next: the home network system built with AI assistance, and the little experiment of AI going from "taking orders" to "being the network admin itself."