Journal: March 2025
March 2025 journal
Monthly retrospective
Yoon Suk Yeol is still president.
AI and programming:
- I learned the term Agent-computer interaction.
- I wrote about Less boring technology.
- I recorded my experience conducting developer hiring interviews in which candidates were encouraged to use ChatGPT actively: Alternative ways to tech interview.
- I am gradually organizing thoughts on the bottleneck in the AI–human interaction loop.
- I thought about the fieldstone method of programming.
Reading novels:
- Finished Harry Potter and the Sorcerer’s Stone.
- Reading Harry Potter and the Chamber of Secrets.
- Reading Tomorrow, and Tomorrow, and Tomorrow.
Journal
2025-03-30
If I copy AI-generated content into my knowledge management system (wiki, etc.), is it my knowledge or not? Around 2008, I read an essay called The Extended Mind, and something in it seems relevant.
The essay argues that if someone with long-term memory loss writes everything in a notebook and always refers to it, the entries are an “extended memory.” It goes further, offering additional criteria for the notebook’s contents to constitute not merely “memory” but “extended dispositional belief.”
- The information is constantly present in the person’s life, and when a relevant situation arises, the person almost invariably consults it and acts accordingly.
- The information is readily accessible whenever the person wants to consult it.
- Once consulted, the information is accepted without question.
- The information is there because at some point in the past I trusted its contents.
Copying AI-generated content without reviewing it fails criterion 4, which in turn creates a problem with 3. Unreviewed copying and pasting is therefore a poor strategy and is likely to damage my knowledge management system.
The essay’s authors do say the last criterion is debatable, because beliefs can form unconsciously, memories can be manipulated, and so on. But in a KMS context, I think we can regard 4 as important.
2025-03-29
The head of a client company used Replit to build the site they wanted and showed it to me. A very good experience. Being able to quickly show what one wants as a working prototype reduces communication costs. Once Replit advances enough to build sites that can actually be used, I will be able to focus on more important work.
2025-03-28
I thought about something called the fieldstone method of programming.
2025-03-26
My 911-day streak of closing my Apple Watch rings was accidentally broken. Such a shame ㅜㅜ
2025-03-25
Reading Tomorrow, and Tomorrow, and Tomorrow. Set in the late 1980s through the 1990s. The male lead is a mathematician and the female lead a game developer. Mentions of real people, real gaming magazines, and games I enjoyed as a child make it a nostalgic read.
2025-03-23
I created Bottleneck of AI-Human interaction loop and am gathering related thoughts.
Insurrection ringleader Yoon Suk Yeol has still not been removed from office. Almost four months have passed.
2025-03-20
I wrote up methods I recently tried while helping several companies hire developers: Software developer job interview.
2025-03-18
I finished the final edition of Nudge: Improving Decisions About Health, Wealth, and Happiness. Now reading Tomorrow, and Tomorrow, and Tomorrow.
I updated New practices for AI-aided programming a little, adding:
- Controlling stride length.
- Creating high-quality feedback loops for AI agents.
- Distinguishing verification from validation.
2025-03-13
ChatGPT Plus allows 15 Deep Research uses a month. I planned to keep a backlog of questions so I wouldn’t forget them and use all 15. In reality, I wasn’t that curious about much, the backlog is empty, and unread reports are piling up instead. As expected, I am the bottleneck: not enough curiosity or time to read. See Generative indigestion.
2025-03-12
I finished Harry Potter and the Sorcerer’s Stone and am now reading Harry Potter and the Chamber of Secrets.
TIL:
- While looking into various things to help an acquaintance, I learned about TabNet. I also learned about Permutation feature importance along the way, though it is unnecessary in the context of using TabNet.
2025-03-11
After reading You might not need Redis, I organized some thoughts I have long had about discussions of this kind: Less boring technology.
2025-03-07
TIL:
Today’s reading:
- Superintelligence Strategy
- Some evidence on the cognitive benefits of learning to code
- The cognitive benefits of learning computer programming - A meta-analysis of transfer effects
2025-03-04
OpenAI Plus subscribers can now use Deep Research 15 times a month. Whether or not I read them, I should ask a meaningful question roughly every other day and collect the results. The bottleneck caused by my laziness and stupidity keeps getting worse. See Bottleneck of AI-Human interaction loop.
2025-03-03
When a new technology leaves gaps in existing norms, confusion arises until new norms take shape. (Cultural lag.)
For example, if you record a meeting with an AI recording app to produce minutes, should you ask the other participants for consent beforehand? What if it is an ordinary recording, without AI? What if you type notes instead of recording? Or take shorthand on paper?
(Assuming the speech-to-text module runs on a local model and no audio is sent to an outside company), there seems to be no essential difference between AI transcription and handwritten notes on paper, yet they feel quite different emotionally.
When Google first put targeted ads in the Gmail inbox, there was controversy over “privacy violations.” Since an algorithm is what “reads” the email, whether it really constitutes a privacy violation is ambiguous (Code: and other laws in cyberspace by Lawrence Lessig). In any case, it was controversial then, but now is taken completely for granted.
Apparently, radio advertising also stirred controversy over privacy in the 1920s (The attention merchants by Tim Wu). People felt that playing commercial advertisements on a radio in the home, a private space, went too far. Yet now this, too, is taken completely for granted.