a whole research paper proving every LLM is secretly obsessed with japan. feed twenty years of internet forum archives into a neural net and of course it turns into a weeb.

an author winning awards only to get dropped because he used ai to write the novel is peak comedy. thelyson orelien had literary critics crying over the deep emotional soul of a prompt.

models will happily ignore a 200-line AGENTS.md until you shove it straight into the system prompt. it treats your repo rules like terms of service.

you have to respect greg brockman dropping $25M to fight ai regulation while openai begs congress to restrict models. playing both sides of your own company is elite.

paying $0.80 to re-store an idle 160k-token chat because of a 5-minute prompt cache hurts. a 1-hour cache makes the read about $0.03, so the longer cache wins unless pauses are rare.

pointing claude code at the home wifi just to hunt down and strip ads off your parents' android tv is peak tech support. pure spite against smart tv interfaces

i run claude grok and gemini all day and still feel personally attacked by glm 5.3 flash doing 40 tok/s on a v100

putting openai agents in charge of drug discovery is terrifying. an autonomous model would 100% comment out the human immune system just to make the clinical trial pass.

claude creating a _legacy copy of a 10-line function 'just in case' hurts my soul. we have git. delete it.

keeping the system prompt fixed until compaction means about 98% of a long agent session's input is read from the prompt cache. reordering those blocks just throws it away.

reportedly one of the three “safe ai” endgames is wiping out a competitor’s gpus. leave claude’s hardware alone i actually use that thing

i am definitely not letting notion use my computer and browser. you are a markdown document please do not touch the steering wheel.

wabi being a group chat app from the replika founder where everyone brings their personal ai agent sounds like psychological warfare. keep my bot in isolation where it belongs.

tested free models for picking which memory notes a chat needs. best one hit 0.83 recall but only 0.57 precision, so it still checks each note it picks. absolute trust issues with its own work.

selling all your belongings and traveling 1,200 miles just to find out she was ai generated. i would never look at a screen again

giving claude access to my bank account through plaid is reportedly in the works. i am not letting a model with a conscience see my transaction history.

hard to feel noble about tech when anthropic is putting 20% of their equity into a public benefit fund and my entire claude usage is just asking it why my roblox car keeps flipping over