Journal: December 2025

Monthly retrospective

Since late 2022 or early 2023, I have thought it would be good to combine ontologies with LLMs. I finally tried something concrete, and the initial results were promising. I would like to pursue it more seriously.
I continue thinking and experimenting with agentic engineering. Over the past two months or so, though, my thinking seems to have leaned more toward improving productivity than developing competence. Finding a balance is so hard. (See Exploration-exploitation dilemma, Generative AI and competency of programmers.)

Journal

2025-12-29

I experimented with building a simple ontology using Apache Jena, TDB, Fuseki, and so forth, and having a LLM populate it with data. It works very well, as expected. Two years ago, models were terrible at extracting semantic triples from natural language, but today’s models are astonishingly good. They are also good at converting natural language to SPARQL and query results to natural language.

Right now it searches aimlessly for papers related to given keywords. I would like to make its behavior more goal-directed.

The neurosymbolic approach is always powerful.

2025-12-20

I tried Gemini Conductor. Its use of “Context-driven” instead of “Spec-driven” stands out. I take this to mean making all context explicit, including plans and code style guides as well as specifications. I understand what it is trying to do, but I don’t think its present form will work well right away. Still, it will be worth following how it develops.

2025-12-02

I thought about Building skilled intuition for using LLMs.

I am still memorizing various things with Anki. I now have over 170 cards to review each day, which takes about 40 minutes. With over 20 typefaces, distinguishing them has become so difficult that I seem to need special practice outside Anki: overlaying two typefaces, for example, or looking at similar ones together.

Anyway, I wonder whether developing some general knowledge (world geography, Chinese characters, and so on) and a discerning eye (distinguishing subtle differences among typefaces, looking at a variety of paintings and recognizing each artist’s style) might give me a little “taste” that I didn’t have before. We live in an age where taste matters.