Journal: September 2025

Retrospective

A project I worked on over the past few months officially launched its commercial service, gained paying users, and appeared in newspapers and on TV. I hope it does well.

I found much inspiration in Dwarkesh Patel’s conversation with Richard S. Sutton (Father of RL thinks LLMs are a dead end) and Blaise Agüera y Arcas’s lecture on What Is Intelligence.

Claude Sonnet 4.5 and OpenAI Sora 2.0 were announced.

I wrote a few short articles:

Joining a small online group of good people has influenced me positively in many ways. I have been talking with LLMs more often than usual, probably because the group has made me more curious. I am grateful.

I played Assassin’s Creed: Mirage, set in ninth-century Baghdad.

Journal

2025-09-29

I read The Big World Hypothesis and its Ramifications for Artificial Intelligence. According to its argument, some problems in the world cannot be solved perfectly no matter how far an agent’s computational abilities advance (or finding and executing a perfect solution is not the ideal strategy). To me, the big world hypothesis seems to restate Herbert A. Simon’s bounded rationality in an AI context. Agents acquire and exercise ecological rationality through experience.

I organized these thoughts a little: The big world hypothesis and the bounded rationality.

Intersection of Antikythera and Sutton.

2025-09-25

Google released a Chrome DevTools MCP. So far, I have mostly used s4 for backend or CLI work. Now it might be good to try web frontends too.

2025-09-24

I am studying time series forecasting out of necessity, and it is fun. Learning driven by need is usually efficient, especially for someone like me with relatively low conscientiousness.

I visited a company where some acquaintances work and did a little pair programming.

I am still reading Supremacy: AI, ChatGPT, and the Race that will Change the World. Apparently, Nick Beckstead wrote:

…saving a life a rich country is substantially more important than saving a life in a poor country because richer countries have more innovation, and their workers are more economically productive.

I like effective altruism, but this claim does not resonate with me. I should think about why.

2025-09-21

Early this morning I watched Blaise Agüera y Arcas’s September 4, 2025 lecture, What is Intelligence?. It tied together the origins of life, evolution, intelligence, and consciousness through the single perspective of “computation.” Fascinating. The explanation was so neat (so economical) that it actually made me wary. A book of the same title is available online. I haven’t read it yet, but it will probably be my book of the year.

I studied geological time in broad outline, organizing eons (roughly hundreds of millions to billions of years) and eras (roughly tens to hundreds of millions of years), and updated human evolutionary history accordingly. Recording that the hominid lineage branched from great apes only near the end of the Cenozoic gave me a feeling of futility(?). Darwin felt grandeur.

In the process, I learned that an International Commission on Stratigraphy exists.

Next I will tackle periods (roughly millions to tens of millions of years). The familiar Jurassic is a period. My goal is to reach epochs (roughly hundreds of thousands to tens of millions of years). That will bring me to the Pleistocene, so often mentioned in evolutionary psychology.

2025-09-18

I fed in two textbooks from opposing fields (feminism and evolutionary psychology) and ran GraphRAG indexing. The bill was 100,000 won😇 LazyGraphRAG has only been published as research, not yet integrated into the code. Queries were slower than expected too (--method=DRIFT), which was a little disappointing.

Still, the approach seems good: extract a knowledge graph from text (ordinary enough so far), perform community detection, and then create summaries at the community level (this is the excellent part).

I need it for something, so I am studying Google Analytics again for the first time in years.

2025-09-17

I read How People Use ChatGPT. More than anything else, I am interested in the “privacy-preserving automated pipeline.” Having an LLM read instead of a human and classify within predefined categories (controlled vocabulary) is almost identical to a method I devised for a recent project.

2025-09-15

I met acquaintances near Euljiro 3-ga Station and chatted about LLMs.

2025-09-13

I thought about conceptual inflation in words such as “automatic,” “intelligence,” and “agent.”

I skimmed On the Theoretical Limitations of Embedding-Based Retrieval, published in August. It reminded me of contextual retrieval, which Anthropic released last year, so I read about that in more detail. It combines BM25 with embeddings. This also helped me understand BM25 a little better; previously I had only a rough idea.

2025-09-01

The project I have been working on for several months officially launched its commercial service.