Arthur Lin

11 Lab Scribe V2 Realtime，這個模型還蠻不錯的。我覺得你們可以去寫一個 Python wrapper，按某一個鍵，然後可以觸發它，然後再按某一個鍵可以停止錄音。基本上這是自制的 dictation，效果還蠻不錯的，因為它的模型也是支援99種語言，然後你要中英切換也都沒有問題。我現在講的話就是用 Scribe V2 轉錄的，然後你可能會覺得我墜字有點多我講話確實墜字有點多

@ArthurLinAI @wangray wispflow试用过一段时间，对比下来还是typeless丝滑一点。那你自己DIY了吗 厉害

搞 AI 的同行 +1，但有个反差：我 Dock 上没有 Claude 和 Codex 的 App 图标，就一个终端做入口。 用 Ghostty 还得专门感谢刷到 @wangray 的这个推荐，从 iTerm 迁过来体验确实有提升。 我每天就 2 个图标： 1. Ghostty 终端：Claude Code + Codex + ssh + pm2 一切的一切唯一入口 2. Typeless：语音口述 + 结构化总结，键盘打字都省了

@hiarun02 But it works on my machine 🥹

Claude Code 4.7 is insane. i know literally NOTHING about coding. ZERO. and i just built 3 fully functioning web apps in 30 minutes. http://localhost:3000/ http://localhost:8000/ http://localhost:5000/ check it out.

@dashen_wang 400k - 128k（最大輸出Token） 差不多會等於258 K的上下文窗口 所以沒有400K 確實偏小

Context 是 AI 的短期记忆，你说的每句话、它写的每行代码都在消耗额度，满了就开始出事。 Codex 实际能用的大概 258k，比宣传的 400k 少了将近三分之一，压缩功能要么触发太晚要么直接卡死，GitHub 上一堆人在骂，OpenAI 说在修。 先把任务拆小，别指望它自己处理好。

@weswinder I think ppl will realize that whether it's codex or claude code. They aren't just tools for devs. It's more like a general agent that can do a lot of things, and writing code for those agents is their way to meanfully interacting with our world.

codex is so smart but claude code absolutely destroys codex on branding you simply can’t compete with a cute crab idk why they thought the “codex” branding was a good idea for the “everything app”

artificial goblin intelligence achieved

@InvShahid @theo I think it’s indeed a great way that mainly use cc as main agent and claude use codex as sub agent They work really well

@theo Well wouldn't it be better use opus on top and make it use gpt 5.5? So it can decode the intent for gpt 5.5 and guide it what the user wants while gpt 5.5 does the coding Gpt 5.5 is better at coding its obvious or maybe the codex gives it an advantage 5.3 codex on opencode ❌

It's so hard to describe the vibe difference between Opus 4.7 and GPT 5.5 (for coding) GPT is smarter and can unblock you, but it gets stuck in stupid ways and strangles itself with context sometimes. Opus will go down the most insane paths and refuse to acknowledge obvious answers, but it understands intent better and has more taste. Whenever I use one for more than an hour, I always reach to the other to "clean up". Best part? All of this changes every few weeks 🙃

@runes_leo @wangray Cmux 其實也不錯

@manateelazycat 超好笑還真的是這樣

新型程序员 Vibe Coding 现状🤣

@FradSer 酷誒讚喔

👀👻 这个彩蛋好啊

@edgaragr The ability to stay on track for Opus 4.7 is decreased I guess. Opus 4.6 feels more coherent across the entire session?

Today Opus 4.7 had two consecutive hallucinations, even after I warned it the first time. Something is definitely wrong with this model, this has never happened to me with previous models.

@skywind3000 感覺 .7 結尾的模型都這樣 3.7 出來的時候我也不是很喜歡 現在4.7出來後用個幾天就回去 4.6 Opus 了

opus 4.7 这些说话越来越没有人味的，已经偏离通用模型了，就是一个纯 coding agent 专用模型。

@ai_xiaomu 3.7 也有遇到一樣的問題 好像 *.7都會有這樣的問題 感覺好像被詛咒了 創意寫作感覺 Opus 4.6 還是比 GPT 5.5 還要有溫度吧？

Claude Opus 4.7有个代价： 长文档检索能力从91.9%退到59.2%，创意写作温度感也降了。 Anthropic在把它打造成纯"干活机器"。 对我们意味着什么？ 用Opus写代码、做工程；用DeepSeek做长文档分析；用GPT-5.5做创意内容。 会组合，就是竞争力。这个认知差，就值钱。

@wholemars Btw do you know you can use Codex in Claude Code? So you can talk to a nicer AI Opus While when required precision you can ask Claude Code to spawn codex as sub agent to deal with tricky things or just review Claude Code's work? `codex exec "message"`

I use them both. When one is struggling I give the other a shot. Sometimes I have them both do implementations in separate work trees and then pick the best result. It’s a constant game of music chairs in terms of who has the best model any given week.

@ai_xiaomu 感覺之後會偏向漲價誒 現在便宜 AI 的日子不久了 GitHub Copilot 現在也改計價方式了 大家都在收緊預算 就 OpenAI 還真的是蠻大方的 你們覺得 AI 公司補貼開發者的日子還有多久🤔

Claude Code涨到100美金/月了，开发者社区直接炸了。 OpenAI趁机表态：Codex继续留在免费和Plus计划里。 我的看法：别纠结哪个便宜，纠结你用AI省了多少钱。 一个月100美金，帮你省掉一个初级开发的活，贵吗？

I don't think it's hype It was good, just things kinda falling apart lately.(actually everything, the harness is buggy, they ship buggy product constantly, and anthropic doesn't really listen to their user, running out of computer etc) But there were good times, was really good. Now we're having alternatives

@rezoundous Claude was mainly hype, the way Dario spoke said it all. For a brief moment performance was good. Now it's just o lot of hustle from their side.

I'm seeing huge number of people either downgrading their Claude plan, or switching to Codex. This has got to be a wake up call for Anthropic.

@edgaragr @rezoundous I think you can consider subscribing to both They work as a team really well tbh calling codex in cc with codex exec "message" Having a different perspective when trying to solve problem or debug and writing codes. Worth a shot

@rezoundous I did upgrade from $100 to $200 and the rate limit seems untouched, there is not way I do the same thing every day and the rate limit is like before upgrading. Next cycle I will switch to Codex.

I think I still use both CC and Codex I think they work well together as a team, different perspective The problem with anthropic are complicated harness failing apart, shipping buggy code half done product and move on. But tbh I think the model is genuinely good to talk to If you want to use codex, cc can call codex, cc is the main agent. The more familiar harness. When required precision, come up with a plan with CC and ask codex to verify and build based on that I think it works pretty well so far though. GPT 5.5 is indeed a great model, but I still find myself more comfortable talking to Opus Don't know what you think

Never been happier to cancel a subscription. Some people genuinely have no idea how far behind Anthropic are. At this point, if you still don’t see it… you’re just not paying attention.

At this moment if one wanna build something to solve problem I don't think we have the time as there were? I mean for example one is trying to learn python, and other with CC or whatever coding agent would be able to build a website(probably only the frontend, and flimsy) just trying to say AI is lowering the barrier, for ppl normal folk to engage and build things. Probably things would fall apart quickly but it's sometimes still really fun and each time you participate with AI to do something Cloning a repo study it? perhaps only partially one will still learn something. Repos on github are treasures now, that were distant. because settings things up really require some expertise(SWE knowledge etc) (btw opening PR to harass dev is another issue though) But it's solving ppl's problem and I find it genuinely interesting to try things Without AI I think I probably would just give up before I get rewarding. AI is making things so fun, I can explore a lot even on the surface level comparing to real dev I guess (Sry for my rough english I don't quite speak english in the place I live. things are kinda off, things don't always come out as complete and the way I intend to though)

Becoming a great AI-first software engineer takes 10 years and one day. 10 years to become a great engineer. One day to learn the coding agent. Don't skip the first part.

@nummanali Btw just curious though do you use Codex? the CLI version or GUI one? just curious though I personally find myself more enjoy the cli one because it doesn't quite collapse agent's tool use results etc

If your terminal plied, you should use yazi file browser