jitl
jitl on Hacker News
problem with agent? use another agent to fix problem ._. i have the architect agent review the code and do a project report & refactor plan every 3 hours.
yeah i currently only distribute sd card dual boot media where the the install process is “turn it on holding REC and pressing rear command dial” and uninstall process is “turn it off”
i use opus 5.5 to delegate to sol and luna via codex-thread-bridge (
https://github.com/saidelike/codex-thread-bridge). these days i’m experimenting with opus using luna as “hands” and it’s largely resolved the tokens per output issue with my Anthropic subscription.
¯\_(ツ)_/¯ it’s not efficient but yeah it’s the best
as i said, i did try it. smaller models are not up to the task of deep reverse engineering work.
i am 100% sure i need opus 5.5 model to build stuff like this because i try lesser models because i want my tokens to go farther and lesser models simply cannot:
https://fujimod.com
$200/mo @ 1 year = $2400
2400 gets me about 1/2 of a 128gb NVIDIA DGX Spark, so i could do a year of nonstop inference of (checks notes) quantized Qwen 3.8-27B, a model i would never choose to use. yay, such savings.
> PDF is an old format from a more innocent era
if apple parsed pdf with (js, swift, rust) none of this would be possible. the problem is memory-unsafe code.
my reverse engineering repo is private unfortunately
hopefully! i’ve got lots of good stuff for the EXR2 processor generation.
claude is also online-only so i doubt they mind
idk, over the past 2 weekends i have a repo for hacking on fujifilm camera firmware that's gone from zero to tens of gigabytes: i've got an emulator for older arm cpus like the one in the X70 (2016), a Hypervisor.framework VM for running their newest AArch64 software for X100VI (2024), a complete C re-implementation of the x70 source, many tools for injecting code and executing on live camera RAM over USB. the emulator has gone through 3 rewrites: first it was python and naive, then it was python with hot parts in rust with a more mature device interface, and now it's rust with some cold or experimental devices still python.
i really don't want to brick my camera, so the deterministic emulator stack lets me do integration tests end-to-end. if i was doing this by hand with no ai, i'd probably have to run stuff on the camera just to test it. but because ai, i get deterministic simulation testing, at the cost of needing to burn $100 of subscription on it.
ai code is more like clay you gotta continuously keep shaping over time, but you can also get very very far, very quickly if you decide the code itself doesn't matter and all you care about is results.
what do you think they train their ai on?
it takes orders of magnitude more focus to target "Market: World" vs targeting "Market: my company's employees". For example in enterprise software, you need sales people. for internal software, usually the dev build it is fine for the job.
i guess gentoo is rolling-release, vs something like debian where every package is frozen in time (+ backports), and debian does seem to largely solve the coordination problem by making everything on the system use a fixed version of some package; with rare situations allowing >1 version of the package at a time.
my take on linux distros being annoyed by their own policies is that perhaps they should change the policy. in this case, perhaps they should pretend uv-build updates once a month or something.
for example, they seem to be fine doing a quick sed on pyproject.toml (https://gitweb.gentoo.org/repo/gentoo.git/tree/dev-python/pa...) or patching stuff for compat (https://gitweb.gentoo.org/repo/gentoo.git/tree/dev-python/as...)
This is what i mean when it seems typical for a distro to monkey around with packages and/or their build system. These were just the first two packages i clicked on in dev-python. maybe others are "pure upstream", but there's definitely the capability here to just... not have the problem?
i got a debian maintainer for old IBM mainframes asking me to fix bugs and run tests for that architecture. ¯\_(ツ)_/¯ these people exist, but i'm not going to support them. feel free to fork me!
distros patch the build process when they package stuff for the distro all the time. why would they "have to" do anything? i'm not super familiar with gentoo but it seems to me like any other 90s era distro where the distro mediates the relationship between the user and the upstream developer for all packages, by applying patches and potentially doing whatever build they want to.
uv means it’s fine now, at least on the consumer side.
mise use uv
uv install
this coming from a long time devtools engineer
tanner seems not-acquireable
From the Cloudflare blog post, this quote from Kenton Varda (head of Workers):
> "Lock-in" actually hurts us — that's why we went open source
If there were truly no escape hatch from Workers, then some of our largest customers would never have signed on with us in the first place.
> [...]
> Ryan and Bert will be leading a new effort to make workerd self-hosting a first-class supported way to build and run apps using the Workers programming model. This will involve merging code and ideas from celld back into workerd. I'm incredibly excited for this work — I will personally be using it to host an instance of Cloudflare OS in my home.
they wanted to do a direct line by line port. they evaluated rust and picked go because the gc let them keep code structure.
Rust memory management is very different from GC heap memory management.
i love vim and i love vim-tmux-navigator keybinds letting me navigate quickly between panes in my terminal: ctrl-{h,j,k,l} to switch panes. i love that tui agents fit right in among my shells and vims. i love that my setup works just as well on my machine and on any remote machine over ssh or mosh. i’ve been working like this since 2012. although i did start using vscode for $DAYJOB since our devex team supports it (has good vim keybinds at least)
In my experience LLMs write mediocre React code: it works, but is the most naive implementation possible for a given task. It needs to be bullied extensively to write performant React that considers prop stability, does updates in event handlers instead of convoluted useEffect chains, etc. I haven't tried Svelte in a while but "lots of React in training data" doesn't feel like a huge boon to me.
Kafka libraries outside Java (and maybe Go) are such a disaster. I remember getting paged because all my NodeJS Kafka consumers were crashlooping from a segfault in Confluent's official Javascript client (wrapper around rdkafka C++). Why? Because Confluent changed some sort of telemetry setting on the broker. I also get segfaults in Rust using the most popular client out there (also an rdkafka wrapper) without any `unsafe` (although there were threads).
I'm using OpenCodex as a backend provider with Codex as the front-end, with the $200/mo ChatGPT subscription. Not OpenCode!
codex has an option to expose itself as an MCP. You can also use something like OpenCodex to bring Anthropic models into Codex as any other selectable model.
You can install a different keyboard if you want. I used 3rd party swiping keyboards but ended up using the system one for the last few years because it doesn't bother me.