|
|
|
|
|
|
by Symmetry
1142 days ago
|
|
|
Right. I sort of worry that because LLMs are able to be so coherent without anything resembling a human's working memory at all, then adding either a crude working memory still working with LLM tokens or adding a more sophisticated one assembled from "chunks"[1] will let an AI based on them ascend to clearly superhuman reasoning with just an architectural improvement and no need for any more flops invested than we already put into GPT4. [1]https://en.wikipedia.org/wiki/Chunking_(psychology) |
|
|