They trained a thing to learn mimicking the full attention distribution but only filtering the top-k (k=2048) most important attention tokens so that when the context window increases, the compute does not go up linearly but constantly for the attention->[query,key] process (it does grow up linearly in the graph because you still need to roughly scan the entire context window (which an "indexer" will do), but just very roughly here in order to speed up things, which is O(L) here).
I am a university year 2 student learning about basic mathematics and statistics related to neural networks. One thing that shocks me is that there isn't an "incremental" solution for building larger (more parameters) AI models (like GPT-4) despite having one in a smaller size e.g. GPT-3.5 (I saw the term "incremental (compiling)" nearly everywhere in the software engineering industry). I am curious how is this not possible theortically?