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

2025-09-07

A year and a week after I wrote my last letter, I've decided to start again: I was thinking about things I enjoy doing, and I realized I love to tinker, and I enjoy writing and sharing what I've been learning and playing with. And so I findmyself typing away once more.

Once I get the hang of this I'll try and figure out how to cheaply and automaticaly cross post to Threads & Twitter, perhaps using text attachments and fancy links to share these more widely this time around.

Metaprogramming with Python

As an experiment, I've been trying to generate Python structs, enums, classes and functions dynamically for ease of use in an interactive setting -- and I constantly find myself delighted by how easy Python makes it to write programs that write programs.

The main things to use are: the type built-in function -- slightly confusingly, there's a 3-argument variant that can generate a class instead of simply identifying one. And then there's the types module that can be used to instantiate certain primitive types in Python -- I particularly like types.ModuleType which is a much easier way to generate a module than using the importlib machinery. References: type, types

At some point I need to sit down and write a small recipe book with these.

The Terminal, Emacs & AI

I'd like to use AI auto complete and chat much more frequently in my daily workflows, but haven't managed to set them up properly yet: gptel works well but it's not automatic. I've been struggling with minuet and generally imagine I can build much more intentional workflows, so I plan to start spending more time there.

The biggest feature I miss is something that can split my tmux window and take control of my primary terminal instead -- as a way to craft commands and have them show up in my actual terminal history. I suspect there's a lot more that can happen here.

While I'm really bullish on being able to use AI to augment my day to day tasks, I'm not yet bullish on AI agents all the way for anything meaningfully complex / that cannot be verified automatically all the way. That said, I can easily see why people managers would be much more comfortable with agents: as someone who's mostly been an IC I strongly rely on my ability to be able to make sense of the code.

With the volume of agents' output I have to treat at as a somewhat black box instead -- and at some point I completely expect the agents to start building incomprehensible code -- and then debugging and modifying the code will involve much more sophisticated tools and observability. But that's a letter for another time.

The other part I'd like to double down on is making all the content on this site much easier to search with a bot, particularly for cross linking. I have some ideas on how to do this, particularly with the tinier, fancier models.

The Glove80 Keyboard

I bought this keyboard as a birthday present and a way to have a portable split keyboard that I could use for work and I've fallen in love (I'm typing on this keyboard right now).

The layout editor was very convenient, and because all the keys were swappable my keyboard actually represents all the keys perfectly: including the thumb clusters and the punctuation. My favorite part has been the LEDs -- I generally find them completely extraneous on keyboards (such as on my Nuphy), but the glove uses them as battery and bluetooth connection indicators.

After dealing with the Kinesis 360's bluetooth issues, I was also fairly skeptical but the lightning quick device switching by pressing the Magic key & appropriate bluetooth connection button in Glove 80 has won me over and I mostly rely on that to switch between my personal and office laptop at the moment.

Books

I've been reading Nausea by Sartre and deeply enjoying the vivid imagery and language in the book; it's much more accessible than I'd expected though I suspect I'll probably miss the point on this first read. Reading the book has made me journal more frequently, and with much more satisfaction.

Kent also recommended Technological Revolutions and Financial Capital as a way to make sense of the current craze for AI and I can clearly see the phases she mentions playing out -- being able to surf the wave is the part I need to figure out.

Over time I expect to settle down on more singular topics with this letter, or accompany it with more focused posts as I flesh out and collapse discussions over a week's time. But I mainly intend to play.

Kunal