Learning diary › Year 2025 › July, 2025 › 2025-07-19 [2025-07-19]

✍️source

- #agent
    - Agents built from alloys (on HN) #sec

        > Like most typical AI agents, we call the model in a loop. The idea behind an alloy is simple: instead of always calling the same model, sometimes call one and sometimes the other.

    - Ccusage: A CLI tool for analyzing Claude Code usage from local JSONL files
    - Coding with LLMs in the summer of 2025 (an update)
        - Simon Willison | Coding with LLMs in the summer of 2025 (an update)

        > When your goal is to reason with an LLM about implementing or fixing some code, you need to provide extensive information to the LLM: papers, big parts of the target code base. And a brain dump of all your understanding of what should be done.

        > You will be able to do things that are otherwise at the borders of your knowledge / expertise while learning much in the process. Yet, everything produced will follow your idea of code and product, and will be of high quality and will not random fail because of errors and shortcomings introduced by the LLM.

    - GLM-4.5: Deeper, Headier, & better than Kimi/Qwen/DeepSeek (SOTA China LLM?)
        - GLM-4.5: Reasoning, Coding, and Agentic Abililties (on HN)
            - Simon Willison | GLM-4.5: Reasoning, Coding, and Agentic Abililties
                - Simon Willison | My 2.5 year old laptop can write Space Invaders in JavaScript now (GLM-4.5 Air) (on HN)
        - Simon Willison | Qwen3-30B-A3B-Instruct-2507
    - The current hype around autonomous agents, and what actually works in production (on HN)
    - How Anthropic teams use Claude Code (on HN)
    - Cerebras launches Qwen3-235B, achieving 1.5k tokens per second (on HN)
    - Lumo: Privacy-first AI assistant (on HN)
    - Simon Willison | Announcing Toad - a universal UI for agentic coding in the terminal
    - Price Per Token – LLM API Pricing Data (on HN)
    - You can now train a 70B language model at home (on HN)
    - #disect
        - The Big LLM Architecture Comparison (on HN)
        - Transformer Explainer: LLM Transformer Model Visually Explained
        - LLM Embeddings Explained: A Visual and Intuitive Guide (on HN)
    - Getting into Flow State with Agentic Coding (on HN)
    - LLMs are bad at returning code in JSON
    - An AI tool I find useful: a review command
- #disect
    - Do you understand f-strings? (on HN) (on lobste.rs)
    - Enough unix to get by
    - LL and LR Parsing Demystified (2013)
    - Reverse proxy deep dive
    - Reverse Proxy Deep Dive Part 2: Why HTTP Parsing at the Edge Is Harder Than It Looks
    - What is X-Forwarded-For and when can you trust it?
    - The Secret Life of Fsync (2023) (on HN)
    - Programming vehicles in games (on HN) #game
    - Never write your own date parsing library (on HN)
    - How I hacked my washing machine (on HN) (on lobste.rs)
    - Making Postgres slower (on HN) #db
- #sec
    - Certificate Transparency - Part 1
    - Firefox-patch-bin, librewolf-fix-bin AUR packages contain malware (on HN)
- #sovereign
    - New comment by voxleone in "The future is not self-hosted"

        > Self-hosting isn't just about tech choices — it's about *who controls access to knowledge*.
        >
        > During the Enlightenment, owning a physical copy of a book meant intellectual freedom. You didn’t rent ideas; you had them.
        >
        > Today, most digital knowledge is hosted, locked, or streamed — *leased from platforms*, not owned. We’re in fact drifting into *digital feudalism*, where access to culture, tools, and even history depends on gatekeepers.

    - Myspace - Archiveteam
    - Anthropic tightens usage limits for Claude Code without telling users (on HN)
    - Apple bans entire dev account, no reason given (on HN)
    - Gmail's backup codes are useless to access account
    - Linux and Secure Boot certificate expiration (on HN) (on lobste.rs)
- #lang
    - Kap lang #apl
    - Kernel #lisp
- #idea
    - Covers as a way of learning music and code (on HN)

        > Playing someone else's music, making a cover, is great because it can make it objective. It gives you something to check against.

    - Why Lexing and Parsing Should Be Separate

        > A lexer recognizes the non-recursive structure of a language. Its output is a token stream.

        > A parser recognizes the recursive structure of a language. Its output is a tree.

        > Lexing with regular languages (expressions) is nice because it can be done in `O(n)` time and `O(1)` space (using automata-based techniques)

        > There is essentially one algorithm for lexing -- march forward through the input exactly once.

        > Parsing CFGs is `O(n^3)` in general, and parsing PEGs either takes `O(n)` space (packrat parsing) or exponential time (backtracking)

        > separating lexing and parsing reduces the n in your O(n^3). In any real program, there are many more characters than tokens

    - TODOs Aren't for Doing

        > (A TODO) doesn’t need to be a plan to actually do something. Instead, it’s a note about “here’s an edge case that wasn’t handled” or a suggestion for a better structure that the author didn’t make time to implement — it captures a little slice of the author’s brain and gives a little window into the rich context they had at the time they wrote the code.

    - Lookism
        - Liking What You See: A Documentary
    - Computational Tyranny
    - Why Does Raising the Retirement Age Hurt Young People? (on HN)
- #kb
    - Molab, a cloud-hosted Marimo notebook workspace #py
        - Announcing molab | marimo
        - Export notebook as markdown with output · Issue #3692 · marimo-team/marimo
            - Export to other formats - marimo
            - Publish to the web - marimo
            - marimo-team/quarto-marimo: marimo extension for Quarto ⚡ run reactive code inside Quarto
            - akshayka/test-markdown-export
    - Org tutorials (on HN)
    - hygg, a minimalist terminal-based document reader supporting PDFs, EPUBs, and text files with vim-like navigation
        - kruseio/hygg
    - Alto turns Apple Notes into a website (on HN)
- #idea
    - Wishes Upon My Demise
        - death
        - maybe I should write one for my death
    - Your Parents Did The Best Job They Knew How To Do