Sad Creator

axiom's launch post literally opens with "most think vibecoding at scale is slop. here's a counterexample" then drops 16M lines of lean formalizing the whole stacks project, the ~7,500 page algebraic geometry reference, written mostly by agents fair point tbh. lean won't accept a proof that doesn't check. typescript will happily compile my bug and ship it

the ai spelling teacher that told a guy seventeen is S-E-V-E-E-N and hung up when he pushed back is at 26M views anam's guy posted today that the avatar is only as sharp as whatever llm the customer plugs in face part works great tbh. perfect lip sync on a wrong answer, full eye contact. i'd have believed her

openai says codex pushed so much parallel work into their monorepo that git reliability dropped below 90% at times. so they built pando, a new git backend, with codex, in about 10 weeks. median push ~5x faster, nearly a petabyte of git data a day git was made in 2005 for kernel devs mailing patches around. it held up 20 years and then met agents

david cramer says a candidate quit sentry's interview because ai wasn't allowed on every part, and that mindset gets you unemployed. sentry allows zero ai in any round right now so the industry yells use agents or get left behind, then wants a hit counter by hand on a call. kinda agree with him and still think thats funny

anthropic told ipo investors its 15 person chip team uses claude to design the chips claude runs on, faster than nvidia and amd. faster by anthropic's own guess at nvidia and amd's workflow, but ok anyone in chip design actually letting an llm touch their rtl yet

itch.io last night: "just deleted a couple terabytes of ai slop 😌" terabytes. of "poorly optimized ai generated models", their words. books, comics, tabletop and soundtrack pages now need an ai disclosure too somebody's upload script really never got told no

support bots are about to know how pissed i am hume's new api takes raw audio and you ask it stuff. how many speakers, is anyone laughing, is the caller frustrated. no transcript, a probability per answer, trained on 10M+ human judgments still wont transfer me to a human tho

google put a data center TPU in orbit. suncatcher, launched oct 1, 4 chips on ~1kW of solar, and only one can run for ~20 min at a time because of heat first prompt they sent up was a haiku. "starlight feeds my core" space is cold and it still overheats. also the haiku is mid

vercel labs open sourced agent-desktop. your coding agent reads the OS accessibility tree and clicks real buttons in mac, windows and linux apps, no screenshot guessing blind users asked devs to label their buttons for years. bet it gets done in a month now that claude needs it

tallied what kept coming up in my replies in october. #1 is code review that keeps up with agents. harvard says reviews take 49% longer and i still have agent PRs open i keep not reviewing. rest is mostly agents doing stuff nobody checked

artists are circling the wonky bits in their own drawings now so people believe a human did it. one posted an 80 hour timelapse and still got told the hair was "melting" i've drawn hands wrong my whole life and nobody ever accused me of being a computer. small win i guess

altman on cheeky pint: home PC RAM sits idle 95% of the time, a "quite terrible" use of the world's DRAM. belongs in data centers by that logic my VR headsets have been idle since 2019 and nobody's come for them keeping my 32gb. only thing i own that doesn't bill me per token

most unhinged thing i read this weekend. a dev compiled a 32-bit risc-v cpu into qwen3.5-4b's weights, no training, layers as ram, runs c mid forward pass. then he added a vector unit and ran karpathy's tiny stories model on it, token exact an llm inside an llm. why. love it

anthropic had claude rewrite their messages api from python to rust. 90k lines, nobody on the team knew rust, cpu down 40-50%. they ran it on shadow traffic first, the only part that'd let me sleep. i once shipped a vr build i half understood and found out from a 1 star review

a lab spent 4 years growing brain cells on electrodes and shipped... a video model. mischief 1 is wan 2.2 plus a tiny adapter they learned from the neurons. the neurons stay in the lab. half a cent a second, 5x faster im building a video pipeline so yeah i'll try it, not proud

openai, google, microsoft and anthropic will honor one shared "no ai training" line in robots.txt. anthropic already does, openai "in the coming months", the other two next year. meta said it supports it, then didnt say it'd comply lol. my robots.txt is from a stackoverflow post