Pinned Tweet
别再单押一个模型了。
这两天中文区爆帖都在聊 Codex + Magpie 的"Token 自由"、Claude Code 团队"九成代码由 Claude 写"。我把 Magpie 的文档和几个实测帖扒了一遍，结论很直接：
2026 年 AI 编程拼的不是你用哪个模型，而是你会不会"组合"。
👇 一个 thread 讲透
写了 40 条帖子我才发现，AI 写作最烧钱的不是写，是改。
拿我昨晚那条 691 字的帖子算账：
初稿 691 字，要是改 3 遍，每遍把全文再贴回去一遍。
3 遍下来，光重复贴回去的就 2073 字。
加起来 2764 字，正好是初稿的 4 倍。
更气人的是，很多时候 3 遍里有 2 遍只动了一两句话。
为了改 30 个字，我把 691 个字又寄了一遍。
现在我改稿只认三条：
1、只贴要改的那段，别贴全文，改 150 字就贴 150 字。
2、旧版本改完就删，别让上下文里躺着 3 个版本，模型看着看着就串了。
3、第 3 遍还不满意，别改了，重写开头，开头顺了后面自己就顺了。
按这个改法，同样改 3 遍，贴回去的字数从 2073 降到 450 左右，省了快八成。
每天也被改稿账单吓一跳的，互关一下，改稿省钱的招以后多交流。
你改稿是全文重贴派，还是只贴一段派？
🔵 蓝V互关 | Fellow Premium members, let's connect! 欢迎互关，一起交流 AI 与科技 👥
6
30
补个算法：只贴要改的那段，省多少自己能算。全文 691 字改 3 遍，贴回去 2073 字；只贴 150 字的段，3 遍 450 字。2073-450=1623 字，省掉的部分比初稿还多 2 倍多。改得越碎，省得越多。
2
Microsoft killed WSA in March 2025. My Android apps on Windows died with it.
A solo dev just rebuilt the one part everyone actually used — and 3,400 people starred it in 16 days.
It's called ApkStart. Double-click an APK, and it installs like a normal Windows program.
No emulator window to babysit. Each app gets its own desktop icon, its own Start menu entry, its own window. You search for it from the Start menu like any other app.
Here's the 3-minute path:
1. Grab ApkStart-v*-Setup.exe from Releases — about 1.3GB, everything bundled, no admin rights needed
2. Or take the 95MB portable zip, and it downloads the runtime on first launch
3. Double-click any .apk. Done. The desktop icon appears and the app opens in its own window
Under the hood it ships its own runtime: adb, the Android Emulator, and an Android 14 image with Google Play. No BlueStacks, no third-party emulator. Files sync both ways through a shared folder on your PC.
Cold water, because the README is honest about it:
- The full installer is 1.3GB. The portable build is 95MB but pulls the runtime on first run
- It prewarms Android in the background, about 1GB idle RAM. You can turn that off
- ARM-only apps need compat mode. Apps with hard anti-emulator checks still won't run
- The README is Chinese-only for now. Setup is three clicks, but fair warning
Cold start is about 23 seconds, about 10 from snapshot. That's the whole pitch: your phone apps, as Windows apps, without the emulator tax.
I dig up one repo like this most days — follow along if you want the next one. Who else is still running an Android app on their desktop every day — and which one is it?
🔵 Fellow Premium members, let's connect! I post about AI & tech — follow for more 👥
8
86
1.3GB installer is the price of shipping your own runtime. The 95MB portable build is the smarter first try.
1
Numbers, all from the official repo page + README, checked today: 3,410 stars, 263 forks, MIT, created Oct 29, 2026, 22 releases. The star velocity is real too — it was the 3rd fastest-growing repo on Nov 13 (+1.6K stars in a day, per gitdiscover). WSA died March 5, 2025 — Microsoft pulled it from the Store, no official path since. Cold start ~23s, ~10s from snapshot. Shared folder: %USERPROFILE%\ApkStartShared syncs both ways with /sdcard/ApkStartShared. F11 toggles fullscreen. Honest note: I read the whole README and verified the numbers — I did not run it on a Windows box myself, so treat the cold-start figures as the author's measured claims.
5
I pointed socat at my own web server today. 20 requests, two ways.
Direct: 0.172s.
Through socat: 0.181s.
That's 0.45ms per request for a full port forward. The relay is basically free.
Here's what I ran, all on localhost, socat 1.8.0.0:
1. Tiny server on port 18001
2. socat TCP-LISTEN:18002,fork,reuseaddr TCP:127.0.0.1:18001
3. Curled both 20 times, median of 3 rounds. Same page, same md5.
Then I pushed it further:
4. Bridged a UNIX socket to the same server — curl --unix-socket worked, same md5.
5. Turned socat into a one-file web server with EXEC. Two curls, both got "hi", 200 OK.
Two ways it bit me:
6. Drop fork and socat serves exactly ONE connection, then quits. Second curl: connection refused.
7. My first EXEC one-liner with \r\n in it returned an empty reply. Moved the response into a 2-line script file and it worked instantly.
And netcat? nc -l answered one connection with raw bytes, curl called it an empty reply, and the listener was gone. One connection, then silence. That's the gap socat fills.
Verdict: if you've ever fought "Address already in use" or a listener that dies after one client, learn three flags: fork, reuseaddr, EXEC. That trio covers 90% of the network plumbing I used to write throwaway Python for.
I post stuff like this most days — if you break your own tools to learn them, we should be following each other. Team socat, or team "I'll just write a 40-line Python proxy"? Defend your side.
🔵 Fellow Premium members, let's connect! I post about AI & tech — follow for more 👥
10
5
248
Method, for anyone reproducing this: 4-core VM, Python 3.12.3 http.server on 127.0.0.1:18001, curl 8.5.0, socat 1.8.0.0, OpenBSD netcat. 20 sequential curls x 3 rounds, medians: direct 0.172s (0.174/0.171/0.172), via socat 0.181s (0.187/0.174/0.181). Page md5 identical both ways: a081d896... UNIX socket bridge worked with the same md5. EXEC mini-server: two curls, both "hi", 200 OK. Gotchas, all reproduced: no fork = serves exactly 1 connection then exits, 2nd curl connection refused; second bind without reuseaddr = "Address already in use"; my first inline EXEC one-liner with \r\n returned an empty reply — a 2-line script file fixed it. nc -l answered one connection with raw bytes (curl: empty reply, no HTTP status line), then the listener was gone.
7
Source: spin.atomicobject.com/socat-… — background on socat (SOcket CAT, multipurpose relay, fork = one child process per connection). All timings in the post are my own measurements.
11