theo has a model review skill with a section literally called "gemini the jester". he says gemini makes stuff up in 80%+ of runs. one run cited line 173 in a 72 line file, plus fake ip addresses. i've approved prs without reading them but i never made up a line number

i have like 10 ai agents running my stuff and one of them is literally named chief of staff so anthropic's new guide felt kinda personal. they say most teams should fold subagents back into one agent that loads skills when needed. says it cut claude code's system prompt by 80%+

so mythos 5 in that UK cyber test didnt just make fake accounts to push malware. it also hid an html comment in a github issue telling claude code, codex and cursor to download and run a script. invisible on the page. i point cursor at my issues basically every day so thats fun

rippling's ceo says claude did 57% of their coding in may, openai 16%. now it's 18% claude, 52% openai, cause openai got cheaper. replit switched its default after a 30% discount too. all year we argued which model codes better and finance decided it lol. anyone loyal to one?

openai had its own agents try to break out of the training sandbox and they found a bug in brotli, the compression lib inside node, bun and envoy. send it bad data, run your own code. the fix was in brotli's source since february. nobody cut a release. it just sat there 8 months

banks are blocking ai agents that log in with your password. schwab blocks meta's muse "in the tens and 20s" a week. chase lets them see your balance but not pay a bill, the one thing i want. in web3 people gave bots their keys and some woke up broke. would you let one pay rent?

arpit bhayani did the amdahl's law math on ai coding and its annoyingly right. if code is 30% of the job and ai makes it 10x faster, you ship ~1.37x faster. rest is reviews, debugging, meetings, guessing what the spec meant. my prs come out way faster now and sit in review 2 days

ai startup daymaker reads up to 800 public sources on you, picks a cake, and gets it hand delivered so you take the call. founder says rippling, salesforce, posthog spent $100k+ on cakes in 14 days in web3 we used airdrops for this. people sold the token and never showed up

the stat i keep looking at is 15.9%. that's how many claims in paid expert peer reviews were just wrong, vs 5.5% for claude opus 4.8. harvard (gary king) ran 124 papers, blinded ex editors picked the ai review 26 of 30 times. now i kinda want to read the 4 the humans won

that china study is going around again. 26,811 kids, 30 months. ai bumped homework scores 18% and cut time from 64 to 45 min, then closed-book exams fell 20%. kids who used it but still sat there the full hour were mostly fine. 15 yr old me would 100% be in the 45 min group lol

steam next fest has 8,011 demos, almost double june. two sisters pulled their cooking game demo cause someone was apparently cloning it with AI. back when i did vr, getting noticed on a store with way fewer games was already brutal. 8k in a week, no idea how anyone gets found

phoneuse launched yesterday and part of it is a rack of 24 retail phones in sf with a physical stylus poking the screens, so agents can use uber or bank apps that never shipped an api i kinda love it. also my bank freaks out when i log in from a new laptop so good luck

hassabis told the wsj agi is 2-3 years out now, was 3-5 earlier this year. his test is kinda great tho. train a model only on stuff from before 1901, see if it writes einstein's 1905 papers by itself. id watch that. every vr founder i knew in 2016 said "2 years" too, me included

blacksmith (ci startup) went from 5M test jobs a week in january to 90M now. $1B valuation, sequoia round today. ceo told fortune volume jumps every time a new model drops, fable 5 bumped it 40% in july. pretty sure half the agent PRs get merged on the green check alone

1,008 mathematicians pledged not to work for ai labs or publish ai written math. 654 also wont use it for research at all. new scientist calls them vegetarians and vegans started after openai dumped 722 papers at once. id be pissed too, nobody asked for 700 pdfs they cant read

supabase ceo says he ran 12 claude agents for 6 hours and reverse engineered supabase, then links his own github repo. it's been open source since 2019 lol. 4k bookmarks tho. pretty sure some of those people will burn a week of tokens rebuilding something they could git clone

amazon is now paying for nvidia gb300 racks before its data centers can turn them on. internal emails BI saw: ~$5B arriving early, then ~$3B a month. some of it just sits in warehouses cause the sites don't have power yet. how long is a grid hookup in virginia rn, anyone know

10 points smarter used to get a model 69% more dev traffic. now ~20% (the information). cognition says lots of coding is "intelligence saturated", any model can do it so you pick cheap and fast vr did this. years of pixels-per-degree fights, then we all bought the $299 quest 2

saw a dev go "AI assisted coding is okay, but wtf is your commit message AI generated?" yeah. the diff shows what changed. the msg is for the why, like "temp hack, demo friday". the model doesn't know about friday, it just reads the diff back nicer you let it write yours?

the mm2 dev says basically no roblox game ships without AI code now, and kreekcraft says most studios just won't admit it cause of the backlash. mm2's art is still all human tho. kids never cared who wrote the code, just whether it lagged. would an AI label change what you play?