this loop of shocked pikachu "machine undertook incorrect costly action in complex environment" followed by "it doesn't have enough context" followed by "context is nonlinearly expensive to provide"
followed by "let's give it precise inputs that let it do its job efficiently" or "let's ban it from taking costly action"
baby you're programming again. without intent this time. how disempowering
the belief that data models the environment accurately is _absurd_ and was not the intent of any dataset digitally available. the treachery of images
can we blame pokemon for this. capturing real things in a little pocket sphere. sinister
Oct 15, 2026, 12:41 AM@thumphries I mean pretraining for representation learning does work, and people have been using it successfully for years before LLMs existed, and for messy things in the real world (like images and sound and natural language) that's the only way we can model them at all
Oct 15, 2026, 12:43 AM@thumphries what's new is people thinking they can take one of those pretrained models, do some RL training, and then plug it straight into actuators without any filtering. if you did that in robotics five years ago people would think you were crazy
@bob even in those cases they are operating off things like photos, scans of photos, writing, transcription. secondary artifacts of actual things, mostly not created with the intent to wholly represent thing
@bob usually datasets are constructed to solve a narrow problem or answer specific questions, rather than to distill pure reality
Oct 15, 2026, 2:17 PM@thumphries for pretraining? not really. the standard datasets for vision like imagenet and COCO are just huge piles of photos scraped from the web, and when people do eg: object detection they start from a pretrained net trained on one of those datasets