2025 - 07 July

AI Practicals

The article Building an AI factory looked one of those “here’s a AI random tip!” articles, but actually has a lot of non-AI things that are new to me (e.g., git worktree).

Agentic Coding Recommendations

Coding with LLMs in the summer of 2025 (an update) Antirez (created of Redis) consistently has good, well-balanced takes on AI coding, especially on the practical aspects. He was one of the first very-high-level programmer that I saw write a positive takes them (while also giving some clear examples where they unexpectedly bad at the time).

The interesting part about this update to me is:

The fundamental requirement for the LLM to be used is: don’t use agents or things like editor with integrated coding agents.

I have started to try Claude Code more on one of the small, auxiliary libraries I’m working on — under 1,000 lines of source code, plenty of good examples, not the most interesting, just a library that I wish already existed — and I’ve been having mixed results. Perhaps true throwaway code is best for agent-based programming, and that I care a little too much about the quality of the whole repository, but I’m still finding that some of the updates are outstanding and 90% as good as what my manual implementation would be, and other times it’s a full step backwards and only adds bad complexity. However, I’m starting to get better results by the simple loop

  1. Pick the most relevant files by hand. This selection does not need to be super precise, but it could be fully encompassing. There’s a chance Claude finds a more relevant piece of code to use in something you don’t tell it, but don’t count on it if you care.
  2. Use a tool like files-to-prompt to concatenate the file context into one markdown file.
  3. Prompt Gemini Pro, or o3, with the task, the concatenated files. Ask it to think about solution with clear tasks to implement.
    1. (3a: remove the weird fluff and fake metrics that o3 adds to its task plan)
  4. Feed this to Claude code to work on your feature.

On detecting LLM writing

Nostalgebraist on OpenAI’s “creative writing” AI

When Sam Altman posted the story from their “new model good at creative writing”, my two reactions were

  1. I think it’s pretty good, trying pretty hard, not sure I like it
  2. (Seeing others’ underwhelmed reactions) It has clearly surpassed my current ability to write fiction, which means I’d need to defer to better writers about what exactly is mediocre about it

This anon is much better fiction writer than I am; their article is an excellent explanation of how it has latched onto a few “writing tics”, probably heavily upvoted by those running the RLHF training process, that make the output look the same as top writers to the public.

I personally don’t think there are many universal signs of AI writing, which work in all cases and contexts… even “it’s not just X… it’s Y”, despite being the most obvious one for awhile, might start to creep into real people’s writing (similar to my own uptick in em-dash usage haha). I’d rather err on the side of assuming human intent until it’s very obviously not a human.

For me, there are more clear signs in people’s code. Besides the obvious “the code has comments above every line and unicode characters that no human knows how to write”, the way I’ve started thinking about it is “Does this code make sense on its own? Or does make more sense as an LLM response?”