This is the archived AI micro-notes from the second half of 2025. For 2026 onwards, see the updated micro-notes diary.
July 2025
Model Context Windows
I keep reminding my students to start a new chat after one or two turns. But I believe context windows will keep getting wider as models iterate rapidly.
Workflow Restructuring
Currently thinking about what to hand off to AI and what to keep for myself. The entire workflow needs constant, dynamic restructuring.
ML Training Diary Day 1-7
One ML lesson a day, no letting go, until I’ve got my own large model. Day 7 — sometimes there are still things ML can’t do, like making a proper bowl of authentic Thai boat noodles.
Spotting AI Writing
When I see “it’s not X, it’s Y” patterns, an abnormal number of parallel sentences, “in summary,” “to conclude,” plus tons of em dashes ”—” everywhere, I can tell it’s AI. Nobody in real life has the patience to write that many parallel structures, nobody’s in the mood to write proper conclusions, and punctuation is always a free-for-all.
NotebookLM
Look into NotebookLM. (Brief but genuinely recommended.)
Gemini for Serious Work
Gemini is really suited for serious work. (When reading papers.)
August 2025
OpenAI’s Shamelessness and Strength
OpenAI’s shamelessness in scraping data is exactly what makes their Deep Research and web search features insanely strong. Stuff that nobody else can find, they find.
ChatGPT’s Filler Nonsense
One reason I like ChatGPT less and less: every response ends with “Tell me and let’s get started right away!” Just do the damn work. Why so much filler.
The Truth About Tech Debt
Counterpoint: the freelance engineers I hired on Upwork left behind just as much tech debt as AI does.
Taiwan’s “Human Intelligence”
Read an article about Taiwanese companies adopting AI. Had a bitter laugh. Taiwanese companies still run on “human intelligence” more than artificial intelligence.
Using Cursor to Check My Bills
Using Cursor + Claude to audit my credit card rewards and cashback — I’m probably the only person doing this.
September–December 2025
The Twilight of Knowledge Websites
With AI going mainstream, these old-school knowledge aggregation sites — even though they’re the soil that fed large model parameters — will ultimately be marginalized by the sheer usability of LLMs.
Taiwan Won’t Be Replaced
I don’t think AI will replace humans in Taiwan. Because labor is just that cheap here.
The Pain of Recitation Errors
When designing an AI app, I keep hitting recitation errors, but users need to see the original question text to answer it! This problem has had me stuck for three days. Coming here to vent.
Claude Recognizes “Shit-Mountain Code”
This is hilarious — Claude even recognizes the term “shit-mountain code” (a Chinese dev slang for terrible legacy codebases).
Automation Gone Wrong
Witnessed an automated posting disaster firsthand. Remember to add an independent validator step in your workflow to verify before publishing. That’s how you avoid wrecks.