It had been a long time since Douyin was opened just to scroll through videos. When it finally was, there were already 32 videos forwarded by a childhood friend waiting unread. Counting it up, it had been more than half a month without really touching the app.
That was not the result of some determined attempt to quit short videos. It happened because another, even more addictive “game” had taken over: Vibe Coding.
Maybe there are no habits that are truly impossible to quit. Sometimes you just have not found a better replacement yet.
Picking Up an Old Game Project Again
A game project called My Blog, which was started last year, has recently been brought back into active development. With OpenClaw helping out, progress has been fairly steady.
There was one painful setback: when switching the system to Ubuntu, the optimized version of the game project was somehow never backed up or pushed to GitHub. After reinstalling the system and pulling the repository again, the missing work was immediately obvious.
A few days ago, after spending enough time playing around with Longxia, there were still plenty of Coding Plan uses left. Letting them expire would have been a waste, so they were put toward optimizing and pushing the game project forward.
Last year, some Godot code was still being written manually. This time, with Longxia available, manual coding was simply abandoned. It is too inefficient by comparison. Longxia, paired with a large model, still makes mistakes from time to time, but overall it is far more productive than writing everything by hand.
That does not mean coding becomes as simple as typing one sentence and accepting whatever block of code the model spits out. Since last year, even though actual hand-written code has been rare, a lot of time has gone into learning how to work with large models properly. The goal is not just “code output,” but code that is usable: aligned with the requirement, logically sound, controllable, editable, and readable by a human.
Why Computing Power Suddenly Feels Tight
Alibaba Cloud has stopped new orders for Coding Plan, with the suspension beginning on March 20. The appeal of this plan was obvious: it did not count tokens, only usage次数. That meant there was no need to worry about token consumption or the anxiety of accidentally burning through too much. Once the usage limit was reached, service was simply paused rather than billed extra, which at least protected users from unexpected charges. Even if Longxia started reciting code from GitHub, there was no need to panic about money disappearing in the background.
So where does the current pressure on computing power come from?
The short version is that demand has exploded while supply has not kept up. AI programming tools are becoming popular, and both companies and individual developers are competing for GPU resources for training and inference. At the same time, high-end chips such as H100 and A100 GPUs remain limited in supply, and manufacturers cannot expand capacity quickly enough. On top of that, the large-model arms race has pushed major companies to stockpile computing power in preparation for training stronger models.
In plain terms: demand is growing too fast, and supply cannot match it.
Alibaba Cloud’s decision to stop selling the Coding Plan is likely tied to this same pressure. The platform’s own compute resources are tight, and priority has to go to large paying customers.
The Latest OpenClaw Version Feels Noticeably Better
After upgrading to OpenClaw 2026.04.02, the command-line upgrade actually completed successfully in one go. That alone made the experience feel much smoother than expected.
The newer setup also switched to the MiniMax 2.5 model, and the efficiency improvement is obvious. People had been saying MiniMax was good to use, and this time it really did not feel like empty praise.