Journal: July 2025

July 2025 journal

Monthly retrospective

I continue studying and trying various things to make good use of LLMs. I organized some of my thoughts into slides on Knowledge work in the AI era and presented them at a small gathering.

A month of Chinese models pouring out (Kimi K2, new versions of Qwen 3).

Articles I wrote in July:

Journal

2025-07-30

My invitation finally arrived, so I installed the Comet browser. Apart from Perplexity being the default search engine, it is almost identical to Dia: “/” commands, “@,” a sidebar LLM chat panel, chat integrated into the omnibox, and so forth.

A desktop version of Observable Notebook is out.

2025-07-29

I briefly tried Firebase Studio, and the code it produced was quite good. My first impression was better than v0 or Replit. The process from entering the first prompt to deploying on Cloud Run flows naturally, and the generated code is reasonably good.

I pasted an OPML list of about 200 RSS feeds I subscribe to into Gemini 2.5 Pro and asked it to write an “About Me.” It produced the following. The prose is overblown and feels like LinkedIn, which I dislike, but it seems to summarize my interests well.

About me according to my reading list

Next I plan to use this to build an information filter, to reduce how much I read. I should also add logic that occasionally suggests something unexpected, to keep my interests from narrowing. (This is always important.)

2025-07-26

I added web support for Obsidian’s internal embeds. For example:

“Any sufficiently advanced technology is indistinguishable from magic.” —Arthur C. Clarke (original)

I have not yet implemented embedding specific sections.

I have kept a personal wiki for 24 years, starting with PhpWiki in 2001.

For a few years in the middle, I moved among various tools (Notion, Mem, MediaWiki, Semantic MediaWiki, etc.), scattering my material and never settling down. There were consequently several years when I did not keep a wiki. Looking back, my memories of those years are so much hazier and sparser than my memories of the years when I kept one consistently.

I am searching Google Cloud, various social media platforms, miscellaneous note files, internet caches, and so forth, gathering the scattered pieces one at a time.

2025-07-25

Google Opal entered public beta. It builds workflows by connecting nodes in a graph. It says US-only, but works in Korea too. Why?

2025-07-24

LLM as a reverse dictionary lookup algorithm

2025-07-23

I skimmed LeanTool and A Proposal for Safe and Hallucination-free Coding AI.

2025-07-22

I added a slide mode to the wiki and created slides on Knowledge work in the AI era. I used CSS scroll-snap, without JavaScript, but the behavior is not particularly graceful.

Sam Altman’s warning about AI risks. Daniel Dennett said this before. See Counterfeit people.

2025-07-19

I am organizing my personal knowledge management methods, but the scope seems to have become too broad because the distinction between knowledge management and an attitude toward life(?) has grown blurry.

2025-07-18

TIL: I looked up why “default value” and “sovereign default” both use “default.” It means failing to do what one should: the value when no choice is made, not repaying borrowed money, and so forth. Latin fallere (disappoint, deceive) > French defaillir (to fail) > French default > English default.

2025-07-17

I wrote about the limitations of specifications written in natural language. (Revising and expanding an example I used at the Nexon Developers Conference in 2011.)

While I was at it, I also made a CLI script to run Alloy code from the command line (Python + JPype). It works wonderfully with the Hammerspoon CLI script I made before.

2025-07-14

AWS released a development tool called Kiro. Like Cursor, it extends VSCode. Its emphasis on “spec-driven development” stands out. It resembles my experiment (Agentic coding experiment 2), with a few differences.

Scope: Kiro’s specifications start at the requirements level; I am experimenting with starting at business goals. Perhaps my scope is too broad.
Traceability: Kiro does not explicitly support traceability (a dependency graph from high-level requirements down to code).
Natural language: Kiro treats documents as natural language. Although it writes acceptance tests in EARS format, they are still natural language.

2025-07-12

I went running for the first time in 11 months. 3 km in 20 minutes 44 seconds.

2025-07-08

I studied the Iterated learning model.

I wrote an article about AI anthropomorphization.

2025-07-04

When discussing AI, we need to distinguish a model’s capabilities from those of a whole system that uses the model as a component. “We cannot completely eliminate hallucinations from LLMs of the current kind” is true, but “We cannot completely eliminate hallucinations from a system that uses an LLM as a component” is plainly false. Example: A Proposal for Safe and Hallucination-free Coding AI.

Many factors influence impressions of AI. How people use it and how difficult the tasks they assign are make a great difference. Moreover, humans still don’t know what is difficult for AI. I therefore think simple declarations like “Only people who have tried easy tasks claim AI will replace humans” are unlikely to be true.

Some are disappointed because they do not use LLMs well. Others are disappointed because they assign tasks that are easy for humans but hard for AI. Others are awed by AI because they assign tasks that are hard for humans but easy for AI. Some speak of the present; others extrapolate the current trajectory of progress to talk about next year.

2025-07-02

I had kongguksu at “Matjarang” near Seolleung Station, and it was delicious. Comparable to Jinju Hoegwan.