Working Notes: a commonplace notebook for recording & exploring ideas.
Home. Site Map. Subscribe. More at expLog.
— Kunal
As I settle into the new job I'm spending a lot more time working with agents, and my habit of constantly maintaining markdown documents is paying off well. x (link) has been amazing -- particularly because it can look at both my history and previous command output to just give me solutions without adding much context.
The speed of execution -- particularly because I'm not forwarding what the agent is thinking -- gives me pause, and I think I want to build out more of my own agent harness even though it means relying heavily on API keys.
As I spend more time working with agents I'm spending more time maintaining running context: at the end of interesting debugging sessions, I'll often ask the agent to write up a summary for myself as a reference, including introductions to topics I'm otherwise unfamiliar with. I expect to build my own libraries.
The other part is that I'm finding myself able to tackle systems I'm fairly unfamiliar with to get results I care about, and I'm getting more comfortable with taking more bets on the agent generated code (as long as it works!) than I would have been otherwise for ad-hoc projects (still staying within my vibe limits).
One series of experiments I've been playing with is trying to have agents tackle problems alone (mostly only interacting to give them approval for commands) vs working directly with and through an agent steering agressively (or even copy pasting) -- so far I've had much more luck and speed in working with agents instead of letting them run by themselves.
Familiarity is also making me less conscious of the amount of compute actually being used behind the scenes to parse and simplify logs; I'm a bit uncomfortable about that and want to learn much more about inference.
A conversation I had with someone on the depreciating value of technical knowledge is something I find myself reflecting on regularly: their assertion was that it's almost worthless to learn at this point. I didn't quite have a counter argument then, but I do now -- even in the limit of everything being generated by AI, I still want to understand how and why things work; perhaps at some point programs will be like physics and mathematics -- fields we need to explore, experiment, and only understand partway -- and need to experiment with to understand. I already rely on empirical experiments to understand different systems, and I'm more than happy to do it in an AI first world.
At the same time, having AI generate most of the code helps me significantly reduce the time and energy I spent on understand "tax law" -- a term I've been using to describe the minutiae we've generally had to memorize over time, such as syscalls, strange CLI incantations, etc. -- and instead focus on much higher levels of complexity and principles on design. After a brief period of disorientation, I find myself much more excited at the idea of all the things I can build and learn.
Using LLMs to learn is also giving me way to learn a lot faster, with a much more personalized teacher.
As I settle into new rhythms at work I have to decide which personal projects I still have energy for: I plan to keep hacking on the variants of djn and the tools I'm building as extensions, and a lot more concentrated effort on learning new things: particularly ml systems.
I've slowly been fixing up my older websites to update what I'm working on.
On the writing front, I keep finding myself stuck between 2 essays: reflecting on my career so far, and the books that I'd recommend to someone following the same path, and then writing about How I Write Python: covering both tax law and design. I think I'd be able to feed the second book to Claude & Codex, so I expect I'll try to get this one out first.
It's always fun to see someone you've worked with on a podcast or other public show: it was nice to hear Bolin talk about his career so far. I used to rely on myles heavily through an Emacs plugin and it was the best way to navigate the monorepo.
"The Orb of Cairado" by Katherine Addison was a lovely, short read. I really enjoy how careful and soft the sentences are, and find reading the stories a pleasure -- and the fact that the main characters often show remarkable depth of character, inspite of the cost they may face:
"Time is the first thing one loses. There's always something more important than one's own pursuits. Always."
"History should be written by the same sort of people it's written about."