rkletr

would be cool if there was a common ffi interface between js runtimes and there could just be one node: polyfill that works for everyone and then people can stop all duplicating the same fucking backcompat work

they should make a xnor xnor xnor without the 6:00-6:34

this book constantly reiterates an idea that we don't know how to do computer security, encouraging thinking that making things safe is an unsolvable problem people who have this mindset in the industry are the ones who choose to fail to apply the engineering principles that solve security problems

private follows ingestion is broken and i can't figure out why; there might not be any new posts on the feed for a little bit

"If Anyone Builds It, Everyone Dies" makes me so fucking angry as an actual software engineer everything is just so so so wrong

"computer security experts have said that computers cannot be made truly secure" which experts? what definition of secure? which security properties are they saying can and cannot be guaranteed? which ones does your argument rely on being impossible to guarantee? are these impossibilities provable?

"If Anyone Builds It, Everyone Dies" makes me so fucking angry as an actual software engineer everything is just so so so wrong

the 4th paragraph casually tells you to assume this fictional ai (that they want to convince you could exist in the real world) breaks Amdahl's law. "yeah parallel scaling laws trust me it's just like the other scaling laws" and then they start talking about a cluster... no mention of communication.

"If Anyone Builds It, Everyone Dies" makes me so fucking angry as an actual software engineer everything is just so so so wrong

it feels so good to be better at type bullshit than frontier models

gpt 6.1 sol aint got shit on me man

turns out you can just make a sandbox better than openai's in a day it's really not that hard

the nix modules that make it work on the guest are generated from the config so you can just scp it on from the host. or hell, have the repo on the guest and generate nix inside. it's not like having the code actually gives any power, the only thing breaking the nix side can do is remove capabilites

turns out you can just make a sandbox better than openai's in a day it's really not that hard

nothing here is particularly novel imo (barely took a day to setup) but basically firecracker vm everything through vsock (only socat on guest for tcp) mitmproxy-like gateway on host configured with a list of upstreams, each with policy; path whitelist for nix cache, only read git ops, etc.

turns out you can just make a sandbox better than openai's in a day it's really not that hard

i wish i could delegate all my programming to the machine god but the machine's capabilities are jagged and unfortunately lag behind in areas i consider critical

designing normal databases has probably been one of the more important things a cs degree taught me (in my case moreso gave me vocabulary and formal framework for what i already knew intuitively but still) if you don't intuitively get sql schemas i would highly recommend putting the effort in for it

maybe the reason llm benchmarks mostly suck and dont represent performance is just cause the average llm user cant stop themselves from making slop how hard can it be to just actually manually go through and audit your damn benchmark tasks before releasing it to the world

wip

after raytracer (3kuymtsx53k2y, 3lm656ah3622k) i thought it would be fun to do rasterizer i don't think i'm gonna be able to fit this in 300 after adding shading, especially since some of the tricks i used to golf this wip are evil (e.g. got rid of some opts, renamed the input from teapot.obj to i)

wip

wtf is there not a single _un_opinionated js formatter. why is the whole industry just copying one other formatter's defaults