Journal: November 2025

November 2025 journal

Monthly retrospective

This month, too, I spent much of my time thinking about agentic engineering. To put myself under pressure to think hard about it, I began spending a million won a month (OpenAI $220 + Anthropic $220 + Google 360,000 won).
In the wake of the AI-something meetup, I made an important decision.

Journal

2025-11-19

Gemini 3 is out. I previously canceled Google AI Pro, but this time subscribed to Google AI Ultra (360,000 won a month, 180,000 for the first three months). With OpenAI, Anthropic, and Google all together, that comes to about a million won a month ($220 + $220 + 360,000 won).

I briefly tried Google Antigravity.

2025-11-16

I tried ChatGPT’s group chat feature. This message appears when you join. Thoughtful.

Your personal ChatGPT memory is never used in group chats.

Related document: Piloting group chats in ChatGPT

2025-11-15

I wrote up a human–AI collaboration pattern called speculative tree search.

I resubscribed to ChatGPT, which I had canceled. This time, the Pro plan. I now spend $440 a month on Claude Max and ChatGPT Pro. One more reason to create value :)

2025-11-12

ChatGPT 5.1 is out.

I tend to act as though a change will not happen, even when I can foresee it in the near future with some confidence. Then, when it happens, it feels momentous, as though I hadn’t known, despite having anticipated it. This is so hard to break out of. There must be related cognitive biases or other theories.

Claude gave me various information, mixed with various fabrications. Among it, this seems related to status quo bias (resistance to change itself, with the present state becoming the reference point) and omission bias (preferring consequences of inaction to consequences of action).

I encountered an interesting development problem today. It seems like the kind I may encounter often in the future, so I am recording it.

Symptoms:

Fairly ordinary code: a browser user sends a request, the server receives it, calls the OpenAI API, and returns the result. Yet it worked for some users and not others.
It consistently worked for the same users and consistently failed for the others.
Clearing the cache or changing browsers did not help affected users. Switching their network from LG Uplus (home) to KT (smartphone tethering) made it work.
For affected users, the OpenAI API returned “Country, region, or territory not supported.”

Suspected cause:

The server calling OpenAI runs on Cloudflare Workers. CF Workers run on different edge networks depending on users’ geographic location or ISP.
Calls to the OpenAI API from a particular CF edge network used by LG Uplus customers in a certain area appeared to be blocked.

Solution:

I was considering OpenRouter anyway, so I resolved it by changing the network path from user → cf → openai to user → cf → openrouter → openai.

For me, this is a new type of problem that I expect to encounter from time to time as I use edge networks more actively.

2025-11-11

I made an important decision.

2025-11-04

I had a fun conversation at the AI-something meetup for the first time in two months.

2025-11-03

Several people around me use “brownfield” in a sense similar to “legacy,” so I looked up its origins. Its meaning was quite different. In the original context, “brownfield software development” is positive and “greenfield” negative. See Brownfield software development.

2025-11-01

I read Ethical implications of AI-mediated interspecies communication. A strange world that is hard to get used to.