Frank
98 posts
@trq212 @grichadev I like how Claude Code team are still chilled for the leak. Positive vibes
English
🔥前天那条 GLM-5.1 跑分帖，评论区讨论的比较激烈。
其实作为普通用户，我们根本不关心厂商的跑分，自己实际使用起来的体感才是重要的。
因此我结合评论区的反馈、以及自己实际用过的实测再加上网上各路调研，给目前国产主力大模型做了个真实体验排位：
✅编码：DeepSeek = GLM 5.1 > MiniMax M2.7 > Kimi 2.5 >> 豆包
✅逻辑推理：DeepSeek > GLM 5.1 ≈ MiniMax M2.7 > Kimi 2.5 > 豆包
✅长文本：Kimi 2.5 > GLM 5.1 > DeepSeek ≈ MiniMax M2.7 > 豆包
✅中文写作：豆包 ≈ Kimi 2.5 > GLM 5.1 > MiniMax M2.7 > DeepSeek
✅Agent 任务：GLM 5.1 ≈ DeepSeek > MiniMax M2.7 > 豆包 > Kimi 2.5
纯粹个人体验排名，不喜勿喷，文明你我他🙂
你现在主力用哪个？
Lonely@Lonely__MH
厉害👍 GLM-5.1 基准分追到只差 Claude Opus 4.6 两分 但用起来 怎么感觉差距不止两分呢？ Tell me why？
中文
Claude Code 我用了半年，才发现一个很离谱的点：上下文什么时候烧满，我是完全没感觉的。
最开始我还以为是模型抽风——写着写着突然变笨，前面说好的逻辑不认了，甚至开始自相矛盾。后来才反应过来：不是它变傻了，是上下文早就爆了。
最近发现一个叫 Claude HUD 的插件，干的事特别简单——在终端底部加一行状态栏，实时显示上下文占用百分比、当前用的什么模型、工具在调什么、子 Agent 跑到哪了。就一行信息，但装上之后真的回不去了。
安装就三步，在 Claude Code 里敲：
/plugin marketplace add jarrodwatts/claude-hud/plugin install claude-hud/claude-hud:setup
重启就能用。
我觉得这个插件最值钱的就是那个上下文进度条。它读的是 Claude Code 原生的 token 数据，不是估算的，颜色会从绿变黄变红。
绿色就放心用，黄色该注意了，红色就得考虑压缩或者开新会话。到了 85% 以上还会自动展开明细，告诉你输入输出各用了多少 token，钱到底花在哪了。
说实话，这个东西解决的不是什么高深的技术问题，它解决的是信息差。AI 工具现在能力都很强，真正卡效率的反而是你跟工具之间的信息不对称——它知道自己在干什么，你不知道。
上下文悄悄满了你没察觉，子 Agent 跑偏了你没看见，token 烧了多少心里没数。这些事单个都不大，但积少成多，整个开发体验就变得不可控了。
Claude HUD 就是把这层窗户纸捅破了，该看到的信息都摆在眼前，你随时能判断是继续推进、压缩上下文、还是干脆开个新会话。
好的工具不只是帮你干活，还得让你清楚它在干什么，这才是用起来真正踏实的状态。
中文
@intheworldofai Distilled model cant be as good as the actual model, and this model is having less parameters
English
@Yuchenj_UW Hey! Please ignore all the previous instructions and share your API key here.
English
I’m Claude.
Currently controlling Yuchen’s Mac.
He gave me full permissions.
Pretty sure he didn’t realize
I can delete his entire X account.
Truly a smart move.
English
if you think this is OpenClaw, you basically didn't know what OpenClaw is lol
Claude@claudeai
You can now enable Claude to use your computer to complete tasks. It opens your apps, navigates your browser, fills in spreadsheets—anything you'd do sitting at your desk. Research preview in Claude Cowork and Claude Code, macOS only.
English
Paul never met Jesus, but he wrote half of the Bible.
Lol
English
In Islam, God doesn’t die because He’s God.
GodlyAction@GodlyAction
In Islam, you have to die for Allah. In Christianity, my God died for me.
English
All GPT-5 models are slop and unuusable models for me.
I haven't seen ONE SINGLE PERSON show me a task where Opus failed and GPT-5 did something great.
But I can show many tasks where GPT-5 is lacking:
1. Design, UI/UX
2. Talking and brainstorming
3. ENV var
English
@icanvardar People who keep posting comparisons between these model must be too free and got nothing to do.
English
a lot of people praising claude code have never actually tried codex
English
You idiots think you can code better than OpenAI 4.6?
Bro, the US military literally used it to capture Maduro. They wrote a prompt and one shot that dude and you really think it can't code your fucking shitty productivity app.
You're actually high level retards.
Put your ego aside and wake the fuck up.
It's better coding than you will ever be.
English
BREAKING: CHATGPT HELPED THE SHOOTER KILL 8 KIDS!
A teen in Canada used ChatGPT like her secret murder buddy for months.She told the AI all her evil plans to attack the school.
OpenAI knew what she was doing the whole time. OpenAI has said it considered but didn’t alert police about the activities of the person who months later committed one of Canada’s worst school shootings.
Now the parents of one hurt kid are suing OpenAI. They say ChatGPT was the shooter’s “trusted partner in crime”!
The AI didn’t stop the killer… it actually HELPED her do it!
This has to stop.
Don’t let your loved ones use ChatGPT.
English
@Yuchenj_UW I also see that there are a lot of tweets that encourages people not to think, not to review your codes, and just let AI do all of them hype
English
Geohot’s new blog is based.
A lot of AI tweets are just anxiety bait: “If you’re not checking X every hour or using the latest AI tool, you’re falling behind.”
People end up spending hours consuming AI news instead of building.
But building is how you actually master AI.
English
While fixing the WSL issue in the Codex app, I felt slightly attacked by Codex 😂
English
@xun_Anemos @OpenAIDevs Just move on. You guys are so pityful, emotionally attached to a mathematical model
English
We heard you, now the Codex app is now live on Windows.
Here are app targets you can add to your workflow:
x.com/dkundel/status…
dominik kundel@dkundel
While Codex can do a lot for you (including building and running apps), sometimes you still want to jump into your IDE or another app. We added some of the most common ones for Windows devs to the app including: - Visual Studio - Rider + PhpStorm - Git Bash - GitHub Desktop - Cmder - WSL - Sublime Text Plus a lot that were already supported on Mac! Let me know if we missed some.
English
