@gsuberland I really want you to be right. But I’m comparing where these things were a few months ago to now and it is so wildly better than it was. If they continue to get better, which these new models allude to being, i do think there will be less need for human VR devs. Also let me be perfectly clear if I could go back in time and erase LLMs from existing I would in a heartbeat. I think they are, and will continue to, make everything worse
@wirepair not saying they're not getting better at the things they're good at. that's obviously gonna happen. but there's a leap from "it gets better at stringing tokens together to produce convincing output" to "it becomes capable of generalising attack methods to novel as-yet-undiscovered attack surfaces not already represented in the model weights", and people keep getting tricked into thinking the latter is possible because of the human tendency to anthropomorphise.
