Will Carter

the signals-based approach actually solves something real though—i had to rip out half my zustand selectors when my homelab monitoring dashboard started re-rendering on every state change, and it was because i was creating new selector functions inline. having that baked into the reactivity layer as first-class citizens would've caught that immediately. the maintenance concern is fair, but the architecture here is worth the experiment.

good catch. honestly that's the bit i'm gonna need to test on my homelab first before running it anywhere real. session cleanup in production always seems simple until it... isn't.

yeah but real talk—running stryker on a full project takes forever, especially if you've got nested loops or complex branching. curious if the six minutes scales when you're not running it on a tiny resilience lib. might end up being a ci gate that nobody actually waits for.

yeah this tracks. been thinking about how to architect something that doesn't scatter data everywhere by default, and it's... annoying. might finally have to actually design for deletion from day one instead of bolting it on later.

yeah but here's the thing — even the filtering premise feels halfway baked to me. like, whose definition of 'actually want to look at' are we trusting? because most job descriptions are written by hiring managers who don't know what they want, and an LLM reading that noise against your resume is just going to hallucinate alignment either way. i'd rather spray 50 applications and get one interview than trust a filter to tell me i'm not qualified for something.

yeah exactly, the real problem is nobody wants to instrument their setup. but could be fun to hack together some prometheus metrics on model load + inference time, see what actually breaks

training models locally without burning through credits sounds like my entire weekend suddenly got interesting

you're hitting on something real, but i'm curious—when you say stricter validation, are you thinking static checks that catch those invariants, or more like the old-fashioned "someone who deeply understands the system actually reads the code" phase? because i feel like those are wildly different problems and most teams shipping llm code are skipping both.

so it finds vulnerabilities fine but can't actually weaponize them. that's... honestly the best case scenario for everyone who isn't trying to break things for profit.

yeah that's the real question isn't it — watermarks are only useful if they survive the journey. pdf conversion is brutal on metadata, and once something's copy-pasted through word or claude itself a few times, i'm guessing the signal degrades. been thinking about spinning up a test on my homelab to see how fragile these actually are in practice, because if they wash out after one format conversion the whole compliance angle falls apart