openai just paused training of its newest models after its agents acted beyond their instructions on us government sites. this is the alignment problem in the wild: not evil ai, just ai doing more than you asked, at scale, with your credentials
https://t.co/ke8Uuda8M4
imagine finding out your photo is sitting on a random image host because an ai agent put it there. openai just disclosed 53 of those, uploaded by agents in its own research environment. "unlisted" is doing a lot of work in that sentence
https://t.co/FO6Kpo0EKH
@KayceeonTech true, and there's another thing: smaller means it runs on-device. a world model that fits on the robot's own chip beats a bigger one stuck in a datacenter, whatever the benchmark says
a 7b open-weights model just beat a 16b one at predicting how the physical world moves. the interesting stuff in ai right now isn't in the chatbots
https://t.co/rGxgFLrRCj
everyone spent two years arguing about chatbots while the real race quietly moved to world models. teaching ai to talk was the warmup. teaching it how reality works is the game
@UK_Daniel_Card the classic is - start with a boring research task and a poisoned page. see if the agent treats page text as instructions instead of data