Notes

I don’t know what this is about yet, but that’s the point. I’m reading in [1] Rose Horowitch “The Age of Reading is Over”, a sentiment I have encountered also through Sam Harris too, but it is a good reminder to see it again. She writes that writing is the space where you realize what you actually think. She also says it is how you share this with others. This kind of also reminds me of Daniel Dennett’s thought pump but also of Socrates’s method, which elevate thoughts into reasoning. But coming back to the beginning of this idea, it’s almost like you are the first person you are trying to convince. And how is it that you don’t know what you think until you write it? Do you literally not know yet how to put it into words or are you improvising on the spot? And maybe the answer to that is in the realm of philosophy or psychology, but I think until you linearize your thought, either by speech or pen, you are only deluding yourself if you think you really know. This is very much like the Feynman line, ...

I’m sure recommendations from a coding harness will vary in quality with the parameters of the model and the context available, but if these are not quite right, there will likely always be some room for error. Several interesting issues today with my semantic notes search project. I was using opencode with GPT 5.5 Fast recently and I ended up with pod memory monitoring that was not taking advantage of kubernetes capabilities and a parallelism split that did not align with my cluster node cardinality. ...

I have been pairing with codex to add parallelization to my logseq semantic search kubernetes batch job. Codex authored this function for checking if a shard environmental variable, SHARD_INDEX, should process a particular journal date among the set of journal dates to be processed. Paraphrasing the func, to just the basic hash algo, def index_from_date(date_str, shard_count): digest = hashlib.sha256(date_str.encode("utf-8")).digest() return int.from_bytes(digest[:8], "big") % shard_count I wanted to test out of curiosity the distribution, and found it was wow pretty skewed, non uniform, ...

PLACEHOLDER Want to add a placeholder to this comment I started drafting on this video1 suggesting that the inability to hold the full context of a code base is the core limitation to agentic software engineering. I don’t think people have that either but we do just fine. I scrolled through the ffirst half of the comments and I couldn’t find the challenge that is on my mind here. But there was at least one commenter who pointed out, paraphrasing, that only leads on a team really have the context and most people are going to be just like LLMs, working with limited information. ...

Hugo suddenly serving only 404s locally while GitHub Pages worked Trying something new here. I have in this post, asked Chat GPT to turn a full troubleshooting thread into a blog post. I did not edit it–maybe except a tiny one–, because my goal was to document this particular weird problem without really spending time on writing a blogpost. Actually I think the result is pretty good, sticking very nicely to reality. It is not in my voice or my style, but that’s ok, I literally just wanted to document the thing and move on. ...

Feel like I’m not seeing what the Ralph Loop community is seeing. They say it’s simple just write a task list and run your Ralph loop until the tasks are done, easy! But most problems cannot decomposed into sub problems easily. This is South Park and Underpants Gnomes. This is Russel Peters and Baking a Cake. This is literally why waterfall project management doesn’t work as advertised. This is why pi planning con at best align on high level goals but is terrible at the decomposition. but there are clearly many many people on the internet who are fascinated by this workflow. If you already could decompose a problem in such a way, then you would just write a program after all. Am I missing something? ...

I noticed, mentioning here1 also, how images with spaces display fine on logseq will not display on github. But also that I learned about the hack to wrap paths with <> to support both without having to rename filenames like I had done in the past. Well I used codex to create a utility for this into my logseq_utils repo, and that ended up being a cool code gen complexity example. ...

if an image in logseq has the form  that displays fine on logseq but does not display on github. On github, without changing the filename you can however replace spaces with %20  but that doesn’t work on logseq. However interestingly, though according to this link I found per a google search, common mark supports one alternative that works in both logseq and github markdown ! ...

Replying to a coworker’s comments about code generation, how it is a necessary trade off between delegation and abdication. I like that he uses the phrase, for abdicating vs. delegating, that “prompts with explicit intent, drastically improve claud code’s chances at producing quality code,” emphasis on “chances”. This is to me why it’s not quite assembly since, compilation is deterministic. More on that later, but begs do question, are you in control? ...

Just hit a mini milestone today w.r.t. a small project I have been working on, to implement cosine similarity search on my logseq4 notes. I have embedded a few months worth of my notes with Qwen/Qwen3-Embedding-0.6B1 into chromadb and I have a small flask fastapi server2,3 around this, for querying. I have chunked up my notes before embedding them and therefore results show those chunks. I am serving this on ec2 using tailscale5, so I can even hit the endpoint from my phone without really exposing it publicly. ...