Working Notes: a commonplace notebook for recording & exploring ideas.
Home. Site Map. Subscribe. More at expLog. — Kunal

2026-01-18

Setting up some goals for building, learning, and writing for myself. I'm aiming to write a running commentary in these letters for at least half of the weeks this year, preferably a weekly commentary on what I've been playing with and working on.

Once a month I plan to publish an essay that's a bit more considered and thought through.

Notebook

Lots of small fixes and updates to the notebook over the past 2 weeks

• applying sentence transformer to all the posts and generating similar links
• adding a favicon made with gemini
• adding links, footnotes, mathjax support

Happily, the full generation code is still within a 1000 lines of code; once I hit 1k I'll probably try and refactor to simplify and shorten further.

Using eglot-python-preset has been a boon: I'd just been thinking about building (or having Claude build me) something like this, and I came across Mike's post on Reddit introducing eglot-python-preset which understands the uv script dependencies.

In general I've been thinking about how to make it easier to remember what I've been studying and have consistent references as I continue working for the rest of my life and evolving these notes accordingly.

Working with AI

I've been trying to figure out how to best leverage agents in a way that works for me, outside of the hype cycles. After building and deleting several attempts, I've been collecting them in a repository called djn (Djinn):

• Some of the projects that I've already started using daily
  • a day planner, published at https://knlb.dev/tools/day.html; description on Threads
  • a wrapper for generating quick CLI commands; this is something I've implemented a couple of times myself, published on GitHub with a sample demo video on Threads
    • exploring agent shell performance was fairly fascinating here; codex -- in Rust -- runs much faster than all the other alternatives
  • completely redoing my emacs configurations and building additional plugins with it, all versioned separately
    • this includes some shortcuts to launch into an agent to further customize my emacs connfiguration live and iterate on it.
    • the overarching idea is that emacs is fundamentally meant to be extremely extensible -- if I can adapt it to my workflows by just asking -- and not context switching to elisp -- just how far should can I expect to go.

I'll write up a summary of all of these projects at the end of the month.

• Other projects I kicked off and have been playing around with
  • a way to explore file history quickly to look at a specific file's history with associated commits rapidly and comfortably
    • the more I work with Claude and agents the better interfaces I would like to have; my biggest criticism of these tools is that they feel as if a UI was stuffed into a terminal forcibly
    • I think the interface I would like is probably closer to cursor but with much more emphasis on reading, annotating, maintaining notes, and agentic orchestration -- almost like scrivener
    • the way the agents work -- inference for trivial changes -- also feels remarkably inefficient and like a massive opportunity
  • bcc tools that track all forks / execs and plot them in perfetto -- I've built this with bpftrace at work and found it extremely illuminating in understanding what a program is doing.
  • tools to convert sql queries (using duckdb as the engine) into perfetto traces, I'll open source this soon as a standalone project (polyptych).

It took all of 3 hours to finish my Claude Pro subscription: I ended up jumping straight to Max to make the most of this month, I suspect I'll revert to Pro after. I have a lot of questions around the economics of this, and how long I should expect to have access to this kind of inference.

Learning Modeling & Transformers

I've currently stalled on the lectures I've been following CS336, CMU/DL Systems -- because I started over engineering the homework -- and have been having a hard time staying engaged with the material.

Instead, I'll probably go back to my original attempts while skimming the videos & lectures: implementing really tiny models to explore ideas and concepts and being able to verify I understood things; paired with published notebooks outlining everything I learned so that I can retain it.

I also experimented with feeding Claude, Gemini, and Codex all my public notes and work and asking them to help me learn faster and better; Codex was probably the most effective, though all the results were at least something I could engage with to try and learn better.

(See also No Ideas, but in Things and my plans up top to use this notebook for a lifetime of work.)

Papers

LLM Inference by David Patterson et al is clearly written, I'm learning to read papers with GPT helping on the side.

In general I still need to find a sweet spot in reading papers effectively and efficiently, as well as choosing the papers to read. Sometimes it's much more valuable to wait for a textbook, sometimes the papers that inspire chapters are much more approachable.

Books

Collecting books I started reading; at some point I'll complete these

• Non fiction

  • The Dictator's Handbook (started)

• Fiction

  • Agent of the Wild, Speaker of Tongues -- Chris Tullbane (read, would recommend)
  • All the childhood you can afford (read, nice & short)
  • The Strength of the Few, The Will of the Many (read, waiting for book 3)

• Self Help

  • Bird by Bird -- Anne Lamott (started)
  • The War of Art -- good, but maybe not enough (read)

(Clearly I have problems with starting too much without finishing, but I'm working on it.)

• notes/writing/writing_approaches
• notes/exploratory_programming
• notes/writing/writing_insights
• notes/tools_for_thought
• notes/blogroll
• notes/how/read/read_a_textbook
• notes/writing/writing
• notes/lectures/llmfromscratch/index
• notes/contents
• notes/llmprojects