Howie Hu

[Howie Hu @howiehu]

Codex + GPT 5.6 真的就适合看门，绝对严谨和死板的看门大爷，我一个一次性脚本的PR就半天过不去……（另：除了会看门啥也不是……）

[Howie Hu @howiehu]

I just realized that the default reasoning intensity for Kimi k3 in Kimi Code—including the default for documents—has been set to High. That explains the change in speed and reasoning time.

I wonder how much of a difference there is between High and Max?

@Kimi_Moonshot

[Howie Hu @howiehu]

Kimi K3 非常好，但是真的就是太慢了。我现在只有在 Fable 额度用完的情况下，才会选择使用 K3，它真的不适合需要快速反馈的任务。

如果需要快速反馈的话，我建议使用 Cursor 加 Grok，并启用 Composer 2.5 的 Agent 模式（探索子代理），而且都要使用 Fast 模式。

#Kimi #Fable #Cursor #Grok #Composer

[Howie Hu @howiehu]

由于现在的SOTA模型都太慢了，所以反而现在 Cursor 20 美元套餐可以使用 Grok 4.5 Fast，并且派发 Composer 2.5 Fast 的探索子代理，反而成了日常在受控环境下进行调研、做出某些修改或执行的最佳解决方案。

#Cursor #Grok #Composer

[Howie Hu @howiehu]

@LLMJunky Look, man, who ever said K3 was Fable? Even the official statement clearly admits that its user experience still lags behind Fable and GPT 5.6 Sol. I have no idea what you're talking about.

[Howie Hu @howiehu]

@orcus108 Based on your current understanding, you might not fully grasp Meituan's status in China, or just how massive this company is in terms of both scale and financial resources.

[Howie Hu @howiehu]

最近Typeless严重降智，导致语音输入经常不准，而且还乱结构化和修改我的本意。

@typelessdotcom

[Howie Hu @howiehu]

@typelessdotcom Add dark mode to the macos app please

[Howie Hu @howiehu]

今天终于顿悟了，如何让多模态 AI 帮助小孩学习。

其实这个时代的学习方式，就是让 AI 用你能够理解的方式，给你编写专项的教科书。甚至因为我的小孩爱看《鬼灭之刃》，所以我让他用《鬼灭之刃》的风格，来给我讲讲什么是 OAuth，以及它和其他认证方式的区别。

[Howie Hu @howiehu]

Dude, have you actually taken a good look at the DeepSeek paper?

This is a foundational model intended to carry the future of China's general large-scale models. It wasn't specifically designed just for writing code.

Furthermore, the official team stated very sincerely that its coding capability only surpasses Sonnet 3.5. There is still a gap between its performance and the coding abilities of models like Opus.

[Howie Hu @howiehu]

@fankaishuoai 是的废话居多还油嘴滑舌，切Codex会好很多

[Howie Hu @howiehu]

1M context windows aren’t new — GPT-5.4 and Gemini 3 Pro already have them.

But there’s a difference between “supports 1M” and “actually works at 1M.”

GPT-5.4’s MRCR accuracy: 91% at 16K → 37% at 1M. That’s Context Rot.

Gemini 3 Pro claims 10M but scores just 26% at 1M.

Claude Opus 4.6 holds 78% at 1M — 2x GPT-5.4’s score.

And Anthropic removed the long-context surcharge entirely. Flat pricing across the full 1M window. GPT-5.4 charges 2x beyond 272K.

Accuracy that doesn’t collapse + pricing that doesn’t spike = the first 1M context model that’s actually usable in daily engineering workflows.

#Opus #ClaudeCode

[Thariq @trq212]

Claude Opus 4.6 with 1M context is now included in Max, Team, and Enterprise subscriptions.

Select it with /model. It now counts against your normal plan limits, no extra usage required.

[Howie Hu @howiehu]

What can you do with $200?

1 - Highly recommend the $100 Claude Code Max subscription. This allows you to use the powerful Claude Code plus the Opus model under a fairly high workload.

2 - Sign up for the $60 Cursor Pro+ subscription. This enables you to make full use of GUI IDE tools like Cursor to accomplish things you enjoy.

3 - Then use $20 to subscribe to ChatGPT Plus. This allows you to use ChatGPT daily and the built-in Codex quota. This quota lets you use Fast mode and xhigh inference to review tasks done by Claude Code and Cursor, and it's basically impossible to use up.

4 - Finally, you'll have $20 left, which you can use to recharge OpenCode's Zen. This allows you to use OpenCode to call all the open-source and closed-source models it has filtered, and also generate APIs for other tools to use.

This way, you can make full use of the $200 and form a plan that allows you to utilize all mainstream top models and tools, with all quotas remaining unused.

用 200 美元你可以做什么？

1 - 强烈推荐 100 美元的 Claude Code Max 订阅。这样就可以让你在相当大的工作强度内，使用最强的 Claude Code 加 Opus 模型。

2 - 去注册 60 美元的 Cursor Pro+ 订阅。这能够让你充分利用 Cursor 这样的 GUI IDE 工具，完成一些你喜欢的事情。

3 - 再用 20 美元去注册 ChatGPT Plus 订阅。这可以让你日常使用 ChatGPT 和内置的 Codex 额度。这个额度可以让你使用 Fast 模式和xhigh推理，用来审查 Claude Code 和 Cursor 所做的任务，基本也用不完。

4 - 最后你还剩 20 美元，可以去充值 OpenCode 的 Zen。这样可以使用 OpenCode 来调用所有它筛选过的所有开源和闭源模型，并且还能够生成 API 给其他工具使用。

这样就充分利用了这 200 美元，并形成了一个可以用尽所有主流最强模型和工具的方案，而且所有的额度还都用不完。

#ClaudeCode #Cursor #Codex #OpenCode

[Howie Hu @howiehu]

我觉得 OpenClaw 的前景，主要是让人们看到了，如果把一整台电脑的权限都向 AI 开放，它能自由和强大到什么程度。

但是就安全性的角度来说，这就像一把双刃剑，所以它最终一定会实现在某种可靠的限定环境下，甚至有可能出现一种专门针对 AI 设计、由 AI 使用的操作系统。

反正这是我的洞见。OpenClaw 带来的可能性是让人激动的，虽然现在的这种实际宣传有些过火，但它也有好的一面，就是进一步扩大了 AI 的影响，让人们对 AI 有了进一步的科普。

The prospect of OpenClaw primarily lies in demonstrating just how free and powerful AI can become when granted full access to an entire computer.

From a security standpoint, however, this is a double-edged sword. Consequently, it will inevitably be implemented within some form of reliable, restricted environment—there may even emerge an operating system specifically designed for and used by AI.

Anyway, that’s my insight. The possibilities brought by OpenClaw are exciting. While the current hype might be a bit over the top, it has a positive side: further expanding AI’s influence and providing the public with a deeper understanding of the technology.

#OpenClaw

[Howie Hu @howiehu]

In actual use nowadays, I clearly find that Opus 4.6 can communicate and collaborate with you in a very natural way that humans can understand and engage in normal conversations.

However, GPT-5.4 still talks like a hooligan, saying things that are incomprehensible. It always mentions those symbolic texts without clarifying what exactly lies behind the words. #ClaudeCode #GPT

[Howie Hu @howiehu]

@d6g @xxx111god I feel the same way. I'm almost fed up with the greasiness and all the nonsense from 5.4.

[Howie Hu @howiehu]

@novtopro @jetbrains same issue

[Howie Hu @howiehu]

@MFA_China Never Kneel Down!