2025 年的年底，今年就只剩最后四个小时了。刚刚我还在折腾一款自己的手边工具，说白了是一个多模型综合助手。把多个大模型的能力拧在一起，协同解决手边常见的活：质量更好的翻译提示词、跨模型的交叉分析，还能用 Pipeline 做方案评审、研发测试等日常流程。

这也是为什么到了年底还没动手写这份年终总结。脑子里还在琢磨 AK 在 Twitter 上说的10 倍能力提升到底从哪来。2025 已经不再是手写 if-else 的时代。再过几个小时 2026 就到，意味着我们得站在 AI 的肩膀上，真正把 Agent 的指挥权、编排权抓在手里。

慢吞吞的我，最近两个月才打通多个大模型的 SSE 连接，把以前攒的一堆提示词、流程模板从 Streamlit 迁到 React。看着像是平移，实际要把之前那个伪实时接口修好，把 Redis、PostgreSQL、Nginx 等脚手架统统搬上去，结果最后卡在了 CSS 的样式定制上。残酷教训时有发生，前端要处理断流重连和消息去重，后端要及时 flush；样式层最好选一套设计系统，别陷在像素级微调里。

越到年底反而更忙：一会儿前端的 SSE 时序对不上，一会儿自定义的 Pipeline 渲染组件又闹脾气……最后只能带伤上线。好在新年之前，总算捣鼓出一个能用的版本。

回头看，很多曾经觉得很复杂的活，现在越来越不稀奇了。这轮重构有 90% 是 GPT + Gemini 帮我干的。我来定目标、给约束、做验收，模型去搬砖，这是效率上限的分界线。我也更确信，数据库驱动的流程一分钟就能生成一条像样的工作流：定时抓数、拉新闻、做分析、写成稿、出图，甚至配点音乐交到你手上。技术从壁垒变成基建，代码不再稀缺，稀缺的是判断什么值得做。

这轮重构的确快，但过程里也有不少槽点。

三月还在公司忙异步任务和大批量文档解析的质量监控；四月开始琢磨小团队怎么用 AI 提高开发质量，想来那会儿 Manus 正在更快更稳地把这些事逐一落地吧；五月离开原团队；六月在家照顾读完高中的儿子过暑假；七月带女儿一路从张家界、少林寺、南太行、平遥、壶口瀑布，路过西安翻秦岭，看完大熊猫再一路南下回家。一个多月回来后，又接着捣鼓 Alembic 的数据库升级——迁移前先做影子表和回滚脚本，线上改动心里才不虚。

九十月，长期加班落下的老毛病虽不至恶化，但身体的确大不如前。还好前面做的工具帮了忙，去医院前先让 AI 帮我把症状、报告理一理，来回几轮问答，我看到了新发布的 Gemini 3 的优点，也看到了不足；后来用上 GPT‑5.2，感觉更稳但也更慢。为什么说慢？因为每个病况都要反复澄清，每个针对性的训练都要提供足够的背景。所幸我有三四年的体检报告，还有带了四五年的苹果手表的数据。慢慢地，也就能更有目的、更科学地练起来了。这算一次真实场景的落地：指标要结合传感器算法和历史趋势一起看，单点结论没意义。

换个话题，回到 Vibe Coding 的主线。Codex 和后来的 Antigravity 给了我很大震撼。很多一闪而过的念头，如今已经被实现得相当成熟。以 GPT‑5.2 和 Gemini 3 为代表的大模型加持，这两款工具更是如虎添翼。正如张小珺在一次智能驾驶采访里引用的那句话：某些模型的发布，直接加速了团队接近 8 个月的工作，还省下了若干亿的研发成本。

当然也别神化它们。真到开干、重构几个产品后你会发现，它们不是倚天屠龙。首先你得能驾驭，其次你要有足够的数据。就像前面的健康问题，不能哪儿不舒服就治哪儿，还得结合历史数据、当下情况和不同场景综合分析。肺结节可以靠影像确诊，但 VO2Max 历年的曲线是多少？HRV 最近 6 个月在走什么趋势？哪怕最大摄氧量只有 27，看着不乐观；好消息是最近 3 个月已经提高了 30%。更好的消息是，我之前以骑行训练为主，手表并不会用骑行来计算 VO2Max，这个指标本身就偏保守。

再说回工程细节。哪怕用了 Codex、Claude Code，它们也搞不定我们那种一句话需求。我折腾了两天的 SSE 接口始终不能实时吐数据，根因是三个月前为另一个需求开了 GZIP。后来才想起老规矩：Nginx 必须对 text/event-stream 关闭 gzip，关掉 proxy_buffering，设置正确的 Content-Type/Cache-Control，并保证后端 chunked flush；如果走 CDN，还要确保 no-transform。很多时候，看似模型不行，其实是工程细节没抠干净。至于配 VPN，也总还有一万种可优化的路子。

但你不得不承认，有了这些好工具，产出速度是成倍提高的。我也妄想着去手搓一批 Skills、MCP、Tools。仅仅调用 API 只是入门，你还得配策略、记忆、评测和更细的分工，加上可观测性和回退路径。说实话，2025 年的我还没太想好这台发动机要先装成小轿车还是越野车，至少别只把它做成一个写稿子的工具吧。

不管如何，今年快过去了，我还在忙：写段脚本把 Notes 里的日记导出，清洗一下数据，再慢慢整理出一份只属于我的年记。

下载 Mac Notes Export 代码

也许新的一年我还会像今年一样碌碌无为，也许会在 AI 的加持下发出些光和热。把自己二十多年的工作经验、技术积累和人资阅历，整理成一系列文字和小工具，分享给大家。这大概会是来年的目标之一。

新的一年来了，再过不了几年就年近五十了。我更能接受那些必然，也更愿意拥抱生活里的不确定。这本就是现实真实的样子。知识于此刻的我，不再是堆积，而是学习——学习如何改掉坏习惯。拥抱能掌控的，放过不可控的。

AI 时代最宝贵之处，或许正在于：当你拥抱那些可以掌控的事情的时候，也能顺手赠人玫瑰。于我而言，健康、养娃、投资、指挥 AI 写代码，大概就是中年人为数不多的几件要事了。

2026，不用去等待完美的 GUI 工具，Roll up our sleeves。

It’s the end of 2025—only four hours left in the year. I was just tinkering with a personal side tool: basically a multi-model assistant that braids several LLMs together to tackle the usual work at hand—sharper translation prompts, cross-model analysis, and Pipelines that handle everyday flows like proposal reviews and R&D testing.

Which is also why this year-end write-up is late. I’ve been chewing on AK’s “10x capability” claim from Twitter—where exactly does that multiplier come from? 2025 isn’t the age of hand-written if-else anymore. In a few hours, 2026 lands. Time to stand on AI’s shoulders and actually hold the baton on agent command and orchestration.

True to form, I only wired up SSE across multiple models in the last two months, and moved a pile of prompts and workflow templates from Streamlit to React. It looked like a simple lift-and-shift, but in practice I had to fix a janky “pseudo real-time” interface and haul Redis, PostgreSQL, Nginx—everything—over. Then I faceplanted on CSS customization. The lessons were blunt: the frontend must handle dropouts, reconnects, and deduping; the backend has to flush on time. And for styles, pick a design system—don’t drown in pixel-perfect fiddling.

Year-end only got busier: one minute the frontend SSE timing drifted, the next a custom Pipeline renderer threw a tantrum… In the end I shipped wounded. The good news: I got a usable version out before New Year’s.

Looking back, a lot of what once felt hard now feels routine. Ninety percent of this refactor was done by GPT + Gemini. I set goals, constraints, and acceptance; the models did the heavy lifting. That’s the line where efficiency caps out. I’m more convinced a database-driven process can spin up a decent workflow in under a minute: schedule scraping, pull news, analyze, draft, render images, even add music, and hand it off. Tech has moved from moat to middleware; code isn’t scarce—judgment is.

It was fast, yes, but not without rough edges.

March: heads-down at the company on quality monitoring for async jobs and large-batch document parsing. April: thinking about how small teams can use AI to raise engineering quality—Manus must have been landing those pieces faster and steadier by then. May: left my old team. June: stayed home, took care of my son fresh out of high school on his summer break. July: road-tripped with my daughter through Zhangjiajie, Shaolin, South Taihang, Pingyao, Hukou Falls—past Xi’an and over the Qinling, giant pandas checked off, then headed south back home. A month later, back to Alembic upgrades: shadow tables and rollback scripts before migration so I wouldn’t panic on production changes.

September–October: old overtime habits weren’t spiraling, but I could feel the wear. The tools I’d built helped. Before hospital visits, I had AI line up my symptoms and reports. A few Q&A rounds in, I saw Gemini 3’s upsides—and its gaps. Later I switched to GPT‑5.2: steadier but slower. Slower because every condition needed repeated clarification, every targeted regimen needed enough context. Luckily I’ve got three to four years of checkup reports and four to five years of Apple Watch data. Gradually, the training got more purposeful and more scientific. A real-world takeaway: metrics only make sense paired with sensor algorithms and historical trends; single points tell you almost nothing.

Back to Vibe Coding. Codex—and later Antigravity—hit me like a shock. Lots of fleeting ideas are now implemented, and well. With GPT‑5.2 and Gemini 3 in the mix, those tools are beasts. As Zhang Xiaojun quoted in an intelligent-driving interview: the release of certain models can pull a team eight months forward and shave off hundreds of millions in R&D.

But don’t mythologize them. Once you actually refactor a couple of products, you learn they’re not Excalibur. First you have to know how to drive them; second, you must have the data. Same as with health: you can’t just chase symptoms. You need a composite view—history, current state, and scenarios. You can nail pulmonary nodules with imaging, but what’s your VO2Max curve over the years? What’s HRV’s six-month trend? Even if VO2Max sits at 27—not great—the better news is it’s up 30% in three months. And there’s context: I mostly trained on the bike; the watch doesn’t use cycling to compute VO2Max, so the number is inherently conservative.

Engineering details again. Even with Codex or Claude Code, “one-line requirements” are a trap. I burned two days on an SSE endpoint that refused to stream in real time—the root cause was GZIP I’d enabled three months ago for something else. The old rules still apply: in Nginx, disable gzip for text/event-stream, turn off proxy_buffering, set proper Content-Type and Cache-Control, and make sure the backend is chunk-flushing. If you’re behind a CDN, add no-transform. Half the time it’s not the model—it’s the plumbing. And VPN setups? There are a thousand ways to shave the yak.

That said, output velocity has exploded with these tools. I’m itching to hand-roll a batch of Skills, MCPs, Tools. Calling an API is entry-level—you still need strategy, memory, eval, finer division of labor, plus observability and fallback paths. Truthfully, in 2025 I still haven’t decided whether this engine wants a sedan chassis or an off-road frame—at minimum, let’s not turn it into a pure drafting machine.

Anyway, the year’s almost out and I’m still busy: scripting an export of my Notes diary, cleaning the data, and slowly assembling a yearbook that’s mine alone.

Download Mac Notes Export Code

Maybe next year I stay as unremarkable as this one; maybe with AI I’ll throw off a little heat and light. I’d like to fold twenty-plus years of work, tech practice, and HR perspective into a series of essays and small tools to share—that’s likely one of next year’s goals.

The new year is here, and in a few years I’ll be close to fifty. I’m more at ease with the inevitable, more willing to hug uncertainty. That’s just reality. Knowledge, for me now, isn’t accumulation; it’s learning—learning how to shed bad habits. Embrace what you can control; let go of what you can’t.

The best of the AI era may be this: when you embrace the controllable, you can also pass along the roses. For me, health, parenting, investing, and directing AI to write code are the few big rocks of middle age.

2026: don’t wait for the perfect GUI. Roll up our sleeves.