Moments

GPT-5.6 Sol, Terra and Luna are three tiered models launched by OpenAI. They are not "shrinked versions" of the same model, but independent product lines for different task scenarios, the core difference lies inreasoning ability, speed and pricing:

Sol (Sun):Flagship strongest model, mainly focuses on highest-level reasoning, complex software engineering, deep research, and multi-step tool orchestration.

Terra (Earth):Balanced main model, performance close to Sol, but cost reduced by half. It is the default first choice for daily production, technical development, and content creation.

Luna (Moon):High cost-performance model, focuses on ultra-fast response and massive batch processing, the cheapest price, suitable for tasks with less strict logical requirements.

Recently, I watched quite a few math-related videos on TikTok and found many interesting problem-solving techniques and ways of thinking. I took some time at noon to make an MVP demo using Codex, quickly implementing the idea:https://math.ygbase.cn/

Just noticed that Ruby 4.0.6 has been released. 🚀

Contemplating the Self
By Yao the Striver
When thirsty, drink; when weary, sleep

Why trouble the mind with Zen and prayer?

The mundane world itself is a sacred place

Where the heart finds peace, that is nirvana

Language Is World Studio, what a great name!

Last Sunday, I planned to redesign the website homepage, hoping to remove excessive personal elements and shift the overall style more towards a media product.

So I started using Figma Make for the design. After the new homepage design came out, I was very satisfied with the overall style, but it was clearly different from the previous website style. So I continued to have Figma follow the new design direction to fully restructure the remaining pages.

Expanding from optimizing a single page to a full site overhaul, the workload increased significantly. However, with the cooperation of Figma MCP and Cursor, the page restructuring efficiency improved a lot. The original data structure was mostly reusable, and I also added some data fields based on the new design requirements to support new visuals and content displays.

After going live, I continued to make some fine adjustments and experience optimizations. During the constant revisions, I also reorganized some functionality and interaction logic. Many design ideas can only be truly validated through actual use and iteration.

This new design direction removes rounded corners and shadows, narrows content width, and aims to make the page more focused and restrained, maintaining a simple, clean overall experience.

My MacBook Pro's built-in speakers suddenly had no sound.

I checked the audio output and mute switch; all settings were correct.

Finally, I asked GPT and opened Terminal, then ran a command:

sudo killall coreaudiod

After executing it, the sound immediately played normally.

After enabling the system proxy in the VPN, local test domain names (such as *.test) suddenly can't be accessed?

Wrong move: Add a DIRECT (direct connection) rule in the proxy software.

This is useless! Because once traffic enters the proxy software, it will use public DNS to look up this local domain name. If it can't find the IP, it will immediately report an error and block it.

Correct approach: In the operating system's "Network Settings -> Proxy Bypass List (Bypass)", add *.test (or configure the Bypass bypass list in the proxy software).

Core principle:

Traffic is first received by the operating system.

Bypass (bypass): Traffic never enters the proxy software; the operating system handles it itself and uses local DNS to access directly.

DIRECT (direct connection): Traffic has already entered the proxy software; the proxy software attempts a direct connection, but usually fails at DNS resolution.

I've encountered this problem before and have always been able to solve it, but today is the first time I've fully understood it.

My uncle's company website was about to expire. To help him save on future maintenance costs and also ensure the domain and website were fully under his control, I took over.

First, I downloaded the existing website completely so that the copy, images, and other assets could be reused, making the subsequent redesign easier.

Considering that WorkBuddy's tokens are free, I started with WorkBuddy. I remembered there was an expert mode, but I didn't use it this time. It was just a regular task, no need to bother the expert.

After generating it, I checked and found that the result was still lacking compared to the live website: the homepage banner images were missing, the carousel switching effect wasn't implemented, and clicking thumbnails in the product list didn't pop up larger images. The overall fidelity was not high.

Then I tried downloading the entire site again using ChatGPT Codex. This time, the result was perfect—the pages, interactions, and details were all well reproduced.

Throughout the process, it only consumed about 2.3 RMB worth of tokens, yet saved a huge amount of time manually organizing assets and rebuilding pages.

Codex has become ChatGPT App, which is a bit hard to get used to.