2026-02-05
Record my day! Today I’ll use a different format for recording.
ChatGPT Pro
On Friday, I gritted my teeth and bought a ChatGPT Pro subscription on the almighty Xianyu (Chinese secondhand marketplace) for 1369 RMB. Seeing that it was cheaper than $200, I placed the order without hesitation. It wasn’t until Tuesday that it was set up for me — they gave me an Outlook email and a ChatGPT account bound to that email.
I logged into the Outlook email and lo and behold, there was a ChatGPT invoice, paid in Philippine pesos.
plan amount ChatGPT Pro Subscription ₱8919.64 Tax: ₱1070.36 Total: ₱9990.00 Payment method Mastercard-xxxx Out of curiosity, I looked up the equivalent in CNY — it was only about 1174.54 RMB. That means they netted nearly 200 RMB in profit!
Then I did some research. Damn, they deserved to make that money. The Philippines is indeed the cheapest place globally to buy a ChatGPT Pro plan. Impressive.
I have a reliable contact in Thailand and wanted to replicate their success, but haha, Thailand is more expensive — equivalent to 1537 RMB. Could I possibly go somewhere like Poland (1942 RMB) and sell Philippine ChatGPT memberships on a similar platform to Xianyu? 🤔
http-proxy Resolution
Finally got the http-proxy thing sorted out. My god, it was more complex than I anticipated. This also shows that Legion’s ability to work across 8 projects within a mono-repo on the same topic is pretty much the limit — I even encountered a few instances where subagents crashed.
But on the other hand, without LegionMind, I recall that directly giving instructions to an AI couldn’t handle problems of this magnitude at all. This indirectly highlights the unique capability LegionMind brings to the table in such scenarios.
In the end, Legion itself managed to finish this task. I can now basically step away from coding, leaving only a few review comments on design docs in most cases. I’m fairly satisfied with this version of Legion.
LegionMind Iteration
Speaking of LegionMind.
I upgraded Legion again tonight. I feel that rigidly prescribing a workflow for it is still somewhat inflexible — essentially making the AI work within my framework. But AI, especially smart AI (like the Codex 5.2 extra high I often use after getting ChatGPT Pro), actually understands everything. I’m not sure if my multi-agent system is somewhat wasteful in this regard. I think as AI gets smarter, letting it design its own workflow is probably a more reasonable way to work. At this level, whether it’s me, Manus, or oh-my-opencode, we are essentially pioneers paving the way before AI can fully replace humans in existing tasks. As Nietzsche said, man is a rope, tied between beast and overman — a rope over an abyss.
So back to the point: we should systematically test different versions of Legion’s performance using scientific methods. We need a set of benchmarks to evaluate how different Legion versions perform on the same set of complex coding tasks. I should quickly create such a benchmarking tool (maybe referencing common industry benchmarks?).
Next, besides continuing work on Yuan, I need to focus on this area.
Speaking of which, I haven’t worked on the LegionMind GitHub bridge in the past few days. The reason is that I saw OpenCode already has this capability built in. Rather than pushing through a buggy implementation, it’s better to pause for now. I think it’s wise not to do things that can easily be washed away by the waves; instead, I should build a ship that rises with the rising sea level. (This metaphor is borrowed from the Manus blog.)
— That’s it for today. I did see CZ’s cross-server comment. I think it’s indeed worth jotting down when I feel like it, but habits take time to cultivate. For me, the greatest significance of recording is not looking back at a day’s details sometime in the future, but the act of recording itself — it’s a process of deep, chain-of-thought thinking. Even if no one reads this log, the record still holds great meaning for me.