for long-running autonomous builds, in combination with token limits, slow rust compilation times are kind of complementary (the rust compilation times aren't limiting throughput)

say what you will about nix, but locking yourself out of ssh from your server, and then using the hetzner console to restore to a previous nix version on boot is so easy and so good it makes it all worth it

i enjoy not having a car and making travel plans based on when someone happens to be driving in a certain direction. there is something fun about the serendipity and interdependence of it

letsencrypt is a cute name

wrote up a readme with some configs for initial experiments using Apertus-70B-Instruct-2509 for coding (working somewhat) codeberg.org/notplants/Ap... > developed by the Swiss AI Initiative, open weights, open data, open science. fully open training infrastructure

feel this also

its a kind of interesting and unexpected turn of events that writing some type of constraint poem has become a fairly common tire-kick benchmark of llms in engineering culture

a builder-adversary loop seems to use way more tokens than using one agent, but the adversary seems to keep catching subtle and significant bugs 🤷‍♀️

Fascists lose wars because they don't observe reality theconversation.com/the-warrior-...

the folder in the benchmark containing 30 calculators is funny, almost musical. some part of me appreciates repetition and variation in general

to any mathematicians/cryptographers on here: was thinking about erdosbench, and wondering are there problems that can be generated from a “secret” that aren’t crypto-grade but are hard to solve without the secret and easy to verify?

last night i ate a cbd gummy (which ended up having thc, unexpectedly) and i started wondering if there is a fake-account-bot-net in the bluesky simcluster and how many bots were on bsky and if there was anyway to really know

the little details in stories of software projects and bugs and how they get shared around online is so interesting

nix community #testimonials

some people like to do rock climbing on a sunday, some people like to use their extra tokens to set up 4 different agent harnesses to build a calculator five times in a row and compare the token efficiency, tomato tomato

contrary to the notion that you can only learn about things from lived experience, I've met chronically ill people who give annoying/confused advice abt illness and not-chronically-ill people who are understanding and compassionate. seems like illness could teach you these things but not necessarily

every vibe coder must at some point vibe code their own harness — it is the ancient way

ive always enjoyed that when writing code for a project, I'm also making patterns and templates I can reuse in future projects and this feels amplified by llms, like there is a sense of continuity and a growing corpus between every project individually and collectively

ls is the softest command (listening, non-destructive) and also phonetically the softest command (el-es)

harsh take, but a part of me feels like in the age of agents, if you are not making e2e encryption for private stuff, you are making slop

interesting that a model can be so capable, but still need a tiny watchdog script to remind it to keep running

claude is currently attempting to patch the latest ghostty version to run on my 2012 thinkpad and im wondering if i should just get a new computer however i would ideally like the new computer to be given to me, for esoteric reasons

i never have trusted the ai ask-for-permissions things, so ive always run llms locally inside an unprivileged docker container which hides .env files before mounting /workspace the script i use for starting this container has become kind of hacky but works afaict tangled.org/notplants.bs...

maybe wearing a belt and suspenders is not a bad idea

trump forks mastodon to make truth.social was never on my bingo card, and im curious what an equivalent successful outcome for atproto would be (adversary adopts the tech not because they like the culture, but because it works)

is there an atproto reddit yet with vouching + sybil attack defense

thinking about how with an e2e encrypted app, its better for the users (system admins cant read their data), and also could open up workflows for llm-driven server maintenance that aren't vulnerable to the lethal trifecta for data exfiltration (ofc there are subtleties here)

yak barbershop would be a good name for a dev studio

looking at a backlog cleanup/polish task that has been there for years yes we could have done this without ai, in theory, like in theory if we had unlimited time and energy

Leibniz was really cooking when he came up with the integral sign

i dont like to make over-confident predictions when anything is possible, but this has also been my structural intuition

i get that computer security is like a science or even economic supply chain problem, but there is another perspective that it just like peek-a-boo and u try to just not show something to something else and not make mistakes and u just keep doing that, in various ways

imagining if instead of using an smtp relay to send email (to avoid spam filters) there was some kind of service that you would pay a small amount to keep your emails out of spam via a sort of vouching, but not actually do the relay and still get direct inbox-to-inbox messaging

if you thought rust compile times were rough, you should see the repl times for getting healthcare in the usa

thinking about how if code is being frequently updated then an interpreted language could be faster than a compiled language by some definitions of a problem/system and appreciating python again

so it turns out the new papal encyclical has a pretty good diagnosis of why the fediverse and the atmosphere keep struggling to govern themselves. guess were doing catholic theology in a protocol blog now connectedplaces.online/reports/fr16...

if it wasn't for ai, I would be at home, writing css by hand as god intended, unfortunately, due to satan's influence, I will be at the shibari rope jam while the autonomous agent loop is writing nix configs

If you can’t handle me at my egg in a teflon pan, you don’t deserve me at my Erdos conjecture disproof

we back on leaflet 🍃

feeling inspired by the work on standard.site collaborative, pragmatic, simple (in the right ways) and radical at the same time integrated with bluesky, but atmosphere-first feels like setting a good standard in lexicon commons-ing (literally and as exemplar)

government funded open source software like lasuite-docs but with data on atproto to prevent lock-in... would be a really cool system on multiple levels

I won’t consider Claude to be fully aligned until this happens

this spring we made and updated co-op cloud recipes for lasuite-docs lasuite-drive lasuite-meet cryptpad lichen-markdown made for @autonomic.zone and collective.tools, but all open source and in the configuration commons if anyone wants to use them git.coopcloud.tech

minimizing the speed at which tests run at the beginning of a project is an interesting example of lived experience, knowing that eventually later on slightly slower per-test speeds will add up to be a headache

permacomputing projects with minimal dependencies are currently looking good again in this cve era

in the garden of eden, there was no f chord, and all the songs were easy to play

i wish nix had a slightly better story around migrating apps on servers i see different patterns people are using, but it feels like an afterthought the dream of declarative infra sometimes feels like it conveniently glosses over the fact that apps do in fact have state

feel like busy work that isn't what you are trying to learn doesn't help learning, and something happening automatically that you are trying to learn also doesn't help learning, and so "cognitive off-loading" could be helping or hindering learning depending on what is being off-loaded. qed.