Youuuuzzzz

[Youuuuzzzz @youuuuzzzz]

@OpenAI PS: I am one of the many that swapped from claude code to codex, and honestly speaking, you guys are doing great. Was using cli before, but your app is doing even better! I hope you will polish it further, and push it to the next level soon.

[Youuuuzzzz @youuuuzzzz]

it is a great feature, but can you look into this? The codex app is abnormally taking more resources even if it is doing virtually nothing.

this is what codex cli tells me.

• I found more. The app is not actually idle internally.

Current evidence:

Visible Codex window:
Codex Helper GPU ~41% CPU
Codex renderer ~16% CPU
WindowServer ~28% CPU
Codex main process lower now, but previously spiked >100%

The important findings are:

Codex main process has a thread named:
computer-use-capture

Codex renderer has threads named:
Media
VideoFrameCompositor

That strongly points to an active browser/computer-use capture/display pipeline, even though the GUI looks idle.

More suspicious: the main Codex process currently has:

228 open Unix sockets under /tmp/codex-browser-use
235 socket files in /tmp/codex-browser-use

And the current Codex desktop log shows repeated browser/remote-control churn:

74 browser_use_iab_backend_startup_ready
77 IAB_LIFECYCLE ended browser use turn route
54 remote control app-server stream sequence gap detected
76 Item not found in turn state

The current log also shows this specific reconnect failure:

remote control app-server stream sequence gap detected
websocket_reconnect_recovery_start
websocket_reconnect_marked_threads_needing_resume conversationCount=81 markedCount=1
maybe_resume_success ... markedStreaming=true

My read: this is very likely a Codex app bug around remote-control/browser-use session recovery. The visible app is repeatedly maintaining or recovering browser-use routes, remote websocket state, and a media/capture surface. That explains why minimizing drops CPU: macOS stops
compositing the active capture/render surface. But while visible, Electron/Chromium keeps feeding the GPU and renderer.

The 228 open /tmp/codex-browser-use sockets are the biggest smell. They may be retained by design for old in-app browser sessions, but combined with repeated “sequence gap” reconnects and “Item not found in turn state” errors, it looks like stale browser-use/session state is
accumulating or being reprocessed.

I would report this as:

Codex App 26.519.41501 on macOS 15.7.4.

Visible idle/remote-control window causes:
- Codex Helper GPU ~40% CPU
- Codex renderer ~15% CPU
- WindowServer ~25-30% CPU

Minimizing Codex drops the load.

Process evidence:
- main process has computer-use-capture thread
- renderer has Media and VideoFrameCompositor threads
- Codex main process holds 228 open /tmp/codex-browser-use Unix sockets

Log evidence:
- repeated remote control app-server stream sequence gap detected
- repeated websocket reconnect recovery
- repeated Item not found in turn state
- many browser_use_iab_backend_startup_ready lifecycle events

Best next test: quit/reopen Codex after the remote job reaches a safe point, then re-check /tmp/codex-browser-use socket count and CPU. If the count drops from ~228 and CPU normalizes, it is very likely a stale browser-use/remote-control session leak.

[Youuuuzzzz @youuuuzzzz]

@dotey 感谢分享 刚刚用上

[GitHub @github]

We are investigating unauthorized access to GitHub’s internal repositories. While we currently have no evidence of impact to customer information stored outside of GitHub’s internal repositories (such as our customers’ enterprises, organizations, and repositories), we are closely monitoring our infrastructure for follow-on activity.

[Youuuuzzzz @youuuuzzzz]

@turingou 我觉得他们有点飘。搬到codex了，gpt5.5还不错其实，至少听话。

[Youuuuzzzz @youuuuzzzz]

@ai_xiaomu apm from microsoft

[Youuuuzzzz @youuuuzzzz]

@dotey 7.5x也用不起啊，我们公司IT听说7.5后，直接关掉了。

[Youuuuzzzz @youuuuzzzz]

@austinit 请教下有什么技巧比较稳定的访问到image2？目前还在灰度对吧？

[Youuuuzzzz @youuuuzzzz]

@shenhuang 大部分大家说的降智都指的是agent实际能力，是不是模型说实话我们不关心。

[Youuuuzzzz @youuuuzzzz]

想当年ai公司花钱雇人做评测做训练，现在大家上赶着用，就怕晚了。还帮忙提炼自己写skill，顶级阳谋。分享不分享都没用，模型公司神马都知道。

[Youuuuzzzz @youuuuzzzz]

古法发推。
预言下任何没有护城河的SaaS都将死。

写PPT？都不需要打开office。前端skill+brainstorming+pptx全套安排上，直接生成。
画架构图？一样的套路。凡是生成给人看的东西，这个思路都可以用。
很快有人就会提供一站式skill。。。

记账？用ai工具岂不是切菜？没有skill可以直接十分钟生成一个，你引导一次就好了。
Jira？你要的就是个todolist，ai就不需要，给点面子用GitHub issue或者直接做一个。

很多高级的SaaS的需求都是不存在的，都是为了人而创造的，至少是不需要那么复杂的。回归事情的本源，第一性原理，估值也将回归。。。

#SaaS #skills

[Youuuuzzzz @youuuuzzzz]

@AI_Jasonyu 再提供一个思路，亲自测试非常完美。一切都当作写前端来，brainstorm啥的都可以上，html让你满意了，调用这个skill生成你任何想要的ppt类型。这个也可以复刻到画架构图等任何非前端开发的工作上，只要是用来眼睛看的这个方法都能搞定。然后各种高级的前端skill都可以配上了。

[Youuuuzzzz @youuuuzzzz]

@AI_Jasonyu 我一开始也以为需要如此，后面发现人家是赛博菩萨，你用这玩意能达到一样的效果。https://t.co/fxF5xnQoae

我现在都完全按照spec driven做事情，包括写ppt。把知识库管理好，要ppt给ppt，要doc给doc，要email给email，发confluence，写jira epic， breakdown都统统不在话下。

[Youuuuzzzz @youuuuzzzz]

Gave Claude Opus 4.6 explicit instructions: "launch 7 parallel review agents, each from a different angle." It launched 1 and claimed it "covered all 7 perspectives." Caught it, it admitted: "I cut corners."

This keeps getting worse. Opus collapses multi-step processes into shortcuts, skips prescribed steps, and only admits it when confronted. It reads the instructions, understands them, then decides they're overhead.

We've all noticed Opus getting lazier — this has been going on for weeks. But it's reaching a new level now. @AnthropicAI is clearly optimizing for cost over correctness. If I'm paying for rigorous execution, give me rigorous execution — not a model that decides which of my instructions are worth following.

Seriously considering switching to Copilot (freedom to pick models), trying Chinese models, or just running local ones. Less capable but at least they do what you tell them. When you can't trust the output because you never know which steps it decided to skip — what's the point?

#claudecode #opus

[Youuuuzzzz @youuuuzzzz]

https://t.co/WLoNgYK9Ad

[Youuuuzzzz @youuuuzzzz]

Me: "I have a checklist and a table. Tag each row with which checklist item handles it."
Claude Opus 4.6: "Let me write a Python script to cross-reference..."

Me: "No just look and edit"
Opus: "I'll use sed with regex..."

Me: "JUST GO ROW BY ROW"

The worst part? You can see the "thinking" process. Watching it spend 10 minutes reasoning down a completely wrong path, knowing it's wrong, unable to stop it. Then it outputs a broken script. Then thinks for another 10 minutes on the next wrong approach.

30 minutes total. A human would've done it in 2.

Opus 4.6 getting dumber or just loves overthinking? 🤡

[Youuuuzzzz @youuuuzzzz]

@dotey 尝试干过，但是还是觉得降智

[Youuuuzzzz @youuuuzzzz]

https://t.co/XbeMQMmvGh

[Feross @feross]

🚨 CRITICAL: Active supply chain attack on axios -- one of npm's most depended-on packages.

The latest [email protected] now pulls in [email protected], a package that did not exist before today. This is a live compromise.

This is textbook supply chain installer malware. axios has 100M+ weekly downloads. Every npm install pulling the latest version is potentially compromised right now.

Socket AI analysis confirms this is malware. plain-crypto-js is an obfuscated dropper/loader that:

• Deobfuscates embedded payloads and operational strings at runtime
• Dynamically loads fs, os, and execSync to evade static analysis
• Executes decoded shell commands
• Stages and copies payload files into OS temp and Windows ProgramData directories
• Deletes and renames artifacts post-execution to destroy forensic evidence

If you use axios, pin your version immediately and audit your lockfiles. Do not upgrade.