Working Notes: a commonplace notebook for recording & exploring ideas.
Home. Site Map. Subscribe. More at expLog.
Back in New York again, at my favourite weekend cafe.
I've been slowly iterating on TermDex, particularly to try and get all my ToDos and investigations down on paper in a way I can easily query them. Every time I make progress I think of more things I could be doing with it instead -- the trap with all my ToDo applications so far, and get somewhat derailed. Go's ergonomics don't quite gel with me yet, but I have to admit just how practical and applicable it is -- it's an easy language to set up and get running with quickly, with a very rich ecosystem.
I've also not found myself particularly productive with Go recently -- there are some reminders of Java that simply make me uneasy. After thinking about it for a bit, I started using Claude to generate Go code for me and it helped walk past a lot of minutiae I didn't particularly care about and gives me some encouragement that I can make things work.
Today I'll try and build out an old idea quickly: make a sqlite extension that can index markdown files based on frontmatter (yaml, toml or json -- inspired by Hugo) and then return the results quickly; making it very easy to generate appropriate queries and views for easy visualization.
This can be quickly combined with fzf and other mechanisms for editing. Writing this out made me realize I wanted to spend more time understanding the Unix Philosophy and seeing how well it's aged over the years -- it's a system of design that leverages the operating system it's surrounded by and seems to work particularly well for extensible design.
Another new experiment I would like to start is to use LLaMa and other models to all my emails, chat, and other daily minutiae so I can stay on top of things more easily: a large part of my day job involves remembering and contextualizing what's going on, and then intervening where I can best help out: the amount of context to maintain day to day is getting a little bit overwhelming, and interferes when I need to go deep into a specific problem.
If I can figure out ways to extract all my data into easily indexed text, and then apply a reasonably dumb model to query and aggregate it for context I suspect I'll be significantly more effective; over time if tools like this were built into work communications things like status reports and updates should become trivially cheap to accomplish without needing a lot of coordination or busy work to accomplish (if applied at larger scales).
Inspite of being so close to LLMs and helping out with building them, I'm nowhere close to internalizing all that's possible with their applications yet: realized this yet again while working on TermDex with Claude; I can afford to be significantly more ambitious as these tools open up and work through a lot of projects that seemed to big to tackle before this. A small part of me is scared at the quality of the outputs -- as the projects get too big for me to reason about personally, at some point I'll have to trust the AI to do things right; there are some uncomfortable parallels to managing other people that start developing as we extend these.
The downstream economic and technological consequences are pretty hard to predict; with just the technology that exists today there's a clear sweet spot of productivity possible that I don't see being applied particularly well at the moment which is surprising. Perhaps the unix philosophy lends itself particularly well to this because we can pull out small pieces of the problem that the AI can then build for us -- good design generalizes really well into the future.
While I've been having fun writing this letter every week, having to self-censor based on what I can talk about publicly and what I can't has been getting annoying. The practice of reflecting on everything I've learned recently is clearly extremely valuable, but I'd like to revisit the mechanics a little bit to make this more useful and thorough. I'm hoping I can use TermDex to achieve this more easily.
— Kunal