Four levels of morality High Physics
2026-02-15 03:43:23

匹 levels of existence. 💡 you can add "the world" as a zeroth layer. 🔥 we're uncertain if there is a layer below the atom layer.

ONE. The "organism" level. There are six billion homo sapiens on earth.

TWO. The "cell" level. There are thirty trillion cells in each homo sapiens.

THREE. The "atom" level. There are approximately one hundred trillion atoms in each cell.

Click title to read full message...

in which the machine tries to design a solution for synonyms Barsukas
2026-02-09 15:14:21

This is a genuinely interesting linguistic modeling problem. After reviewing the current data model, here are my thoughts.

The Core Problem

Each Lemma currently represents one concept anchored to an English label. "hat" and "cap" are two separate lemmas, each with one translation per language. But languages carve up semantic space differently:

English: "hat" (broad brim) vs "cap" (fitted, no brim) — distinct

Lithuanian: "skrybelė" (wide brim) vs "kepurė" (general headwear/cap) — different boundary

Chinese: "帽子" covers nearly all headwear generically

French: "chapeau" (hat, broad) vs "casquette" (baseball cap) — yet another split

The current model has no way to say "these two lemmas occupy overlapping semantic territory, and in language X they might collapse into one word."

Click title to read full message...

the sorcerer's apprentice LLM
2026-02-04 22:41:51

today, in Trakaido/Barsukas development: I want to add 20-30 cities to the word list 💡 with translations in 14 languages.

  • Claude launches a second agent to do the work
  • The second agent generates a JSON file
  • Claude then does "Web Search" to try to find the ones it isn't sure about
  • After I cut it off, Claude uses the Wikidata API
  • Claude does ... a lot of verifying, much of it needless

🔥 the layers of abstraction are dizzying ...

claude made a bug LLM
2026-02-03 16:55:38

I have seen this specific error before, and I think the Claude Code team should be able to fix it.

⏺ Good! The WireWordDto has the audio field defined correctly. Now let me check an

actual wireword file to see what the data looks like:

⏺ Bash(cd ../wordlists && grep -A 10 "N07_003" lithuanian/wireword_nouns.json | head
-20)

⎿  grep: lithuanian/wireword_nouns.json: No such file or directory

⏺ Read 1 file (ctrl+o to expand)

⏺ Bash(cd .. && ls -la wordlists/)

⎿  total 112

-rw-r--r-- 1 powera staff 159 Aug 15 15:15 __init__.py

drwxr-xr-x 28 powera staff 896 Feb 3 10:51 .

… +26 lines (ctrl+o to expand)

⏺ Bash(cd .. && ls -la wordlists/lang_lt/)

⎿  Interrupted · What should Claude do instead?

🔥 do you see the bug? ⚙️ it is going up another directory. "cd .." is not idempotent. 💡 maybe the local Claude directions should make it clear?

on Barsukas LLM
2026-01-28 18:53:57

Somehow Barsukas has become the other server as part of the Trakaido project. It contains translation data, sentence data, etc.

It is, in effect, a multilingual dictionary.

Instead of sorting alphabetically, we have to categorize words and then sort them roughly by linguistic complexity. This is a blurry term, but:

  • cow is less complex than marmoset
  • red is less complex than crimson
  • table is less complex than credenza

The categories themselves have evolved. There is a super category that the LLM generated in the UI that I am not concerned with, but we have:

  • around 40 categories of nouns
  • 15 categories of verbs
  • 10 for adjectives and adverbs
  • a separate numeral category for number words

We have an architecture designed around a development pattern where people make APIs including the OpenAI key to add words. As a single-user project, this is a way to avoid putting the LLM key on the server. For a multi-user project, there are obviously security risks associated with this design. We are hoping to solve it better later.

Some of the agents have turned out to be more useful than others. This is fine.

thoughts on Trakaido development LLM
2026-01-28 17:52:10
  • The low-hanging fruit is having agents do things like type checking. 💡 there is no reason to not have your main codebase include type hints for JavaScript, Python, etc. any more. On the other hand, there is no need to force developers to use them. An agent can modify the pull request to add them.
  • Having an agent take every "Swift" ⚙️ for an iOS app app change, and apply it to "Kotlin" ⚙️ for an Android app, doesn't work extremely well. And, as I have mentioned, the $20/month LLM subscriptions will work for an individual developer doing one version, but not for two 💡 or three with web; or four if you count documentation versions of an app.
  • Audio generation is still expensive if you have an LLM API do it, and low-quality locally. Neither the LLM nor non-LLM audio generation tools give high-quality output. And, even "gpt-4o-mini" 💡 still the newest model in January 2026 has about a 1% error rate.
thoughts on the Trakaido API Atacama
2026-01-28 15:42:25

the Trakaido stats server is in the Atacama repo largely by accident.

I started Trakaido as a one-day LLM-generated project, back in the halcyon days of May 2025, when "tell an LLM to write a 500-line React app" was something that you couldn't get out of the box. I made about 10 apps, and Trakaido was the one that had legs. Of course, one of the first features I wanted was "server side stat tracking", so I added it to Atacama. 💡 I also added server-side "wordlist" and "audio" APIs, but those have been migrated. But, the stat tracking has remained; a feature entirely unrelated to blog software.

in addition to "word" stats, we have grammar interstitial views and user config data on Atacama as well. ⚔️ well, actually, it's a different process, though the same monorepo

There is much talk about whether LLMs prefer a monorepo or microservices.

The answer is, generally, it doesn't matter. You can run an LLM across multiple Git repositories.

Having the iOS and web 🔥 and Android apps in the same repo is a slight convenience for a one-person development team. But would probably become large and cumbersome later on.

As far as the "Ralph Wiggum" / "Gas Town" philosophy ⚙️ roughly that you can have an agent that takes every React/Swift change and applies it to Kotlin ... I haven't gone that far yet. Other than "you can't do that on a $20/month subscription", I have no thoughts yet.

the pattern of LLM work systems LLM
2026-01-26 01:00:22

Claude seems to be standardizing on a new "agent session" method, and I think I like it.

Step 1: Design a plan.

Step 2: Write the plan.

Step 3: Forget everything other than the plan.

Step 4: In parallel, execute the plan.

Show Older Messages