Laughing🪁

14,089 Twitter followers

亚太主厨👨🏻‍🍳 @GoKiteAI | Prev: CMO @WeirdoGhostGang @outaxyz PDSA @AlibabaGroup Researcher @BlockBeatsAsia | 记录所思所感所想，观点仅代表个人

Posts

Laughing🪁

Thread

Actually, @vergex_ai acquired Minimax and even bought a 2000-square-meter office with a prime river view, where hundreds of employees are working hard. @Tintinx2021's parrot secretly told me this. twitter.com/0xLaughing/status/...

Laughing🪁

04-02

The Claude Code app has been updated, and a new thinking depth controller has been added. Maintain a medium level for daily use, and choose High or Max for challenging problems. However, before the effort parameter is implemented, you can use the prompt keyword to trigger deeper thinking: Think hard / think harder (deeper thinking) Ultrathink (deeper thinking, approximately 32K thinking tokens) The prompt keyword only applies to the current round and automatically reverts after use.

Laughing🪁

04-01

When I was in graduate school, I was working on computer vision. Back then, there were a lot of people doing posture detection and publishing papers, mostly just for show. The main commercial application was smart fitness mirrors (like Lekang and a few smaller brands). Now it feels like there are so many things we can do, even DIY at home.

Laughing🪁

03-31

Thread

Previously, I used the Codex CLI in Claude for code review and bug fixing. I just saw that OpenAI has open-sourced its plugin. Curious about the differences between using this plugin and my own setup, I did some research and concluded that I recommend customizing it, as this plugin is more geared towards beginners: 1️⃣ Standardized workflow, no need to maintain glue layers yourself. When you connect it yourself, the biggest cost is constantly maintaining scripts, commands, prompts, and parameter compatibility after connection. This plugin directly sets up common scenarios as fixed commands and subagents, reducing a lot of custom glue code. The official README explicitly provides codex:review, codex:rescue, etc. 2️⃣ Direct access to the officially packaged Codex review workflow within Claude Code. The best part isn't just being able to review, but that the review process has been established as a stable entry point. The README explicitly states: `/codex:review` is equivalent to running `/review` directly in Codex, which is more stable. 3️⃣ It adds a very useful "adversarial auditing" entry point. `/codex:adversarial-review` is not a regular review; it's specifically designed to challenge design decisions, assumptions, tradeoffs, and failure modes. It's suitable for investigating directional issues that Claude might easily miss, such as race conditions, rollbacks, authentication, and data loss. You can manually prompt for this scenario yourself. I personally use multiple agents for cross-auditing, but with the official fixed entry point, the reuse cost is much lower. 4️⃣ The backend task management is systematic. The most chaotic part when handling tasks yourself is how to keep long tasks running in the background, check progress and results, and how to continue running them. This plugin streamlines the entire process, which should improve actual development efficiency. 5️⃣ Directly reuses the existing local Codex environment, resulting in low migration costs. The official documentation states: The plugin is not an independent runtime; instead, it runs your local Codex binary and Codex app server, reusing the same configuration, authentication status, repo checkout, and local environment. 💡My own approach to using the Codex CLI: I slice the model according to different tasks and finer granularities. For complex tasks, I use GPT-5.4 xhigh fast, and for some brainstorming sessions, I use medium. This plugin doesn't seem to support this yet. twitter.com/0xLaughing/status/...

CODEX