Great analysis as usual. Any idea how different /similar the citation process is for AIOs? It feels like there’s a very different flow there, with grounding for citations being potentially independent of grounding for generative response? Any thoughts?

Google can’t find this paper:

https://www.google.com/search?q=%22Efficient+Pruning+of+Vision+Transformer+via+Interpolative+Decomposition%22

Very weird…

Hey Dan. Was reading through the first example response, and decided to check if the snippet was a summary or verbatim from the source.

https://arxiv.org/abs/2412.05721 is a paper titled “Impact of Sunglasses on One-to-Many Facial Identification Accuracy”.

A title search on Arxiv for [Efficient Pruning of Vision Transformer via Interpolative Decomposition] yields no results:

https://arxiv.org/search/?query=Efficient+Pruning+of+Vision+Transformer+via+Interpolative+Decomposition&searchtype=title&source=header

I wonder what happened there?

hmm… <br> got stripped

tags used for chunk boundaries. I doubt many people would guess that particular approach.

https://chromium.googlesource.com/chromium/src/+/refs/heads/main/third_party/blink/renderer/modules/content_extraction/document_chunker.cc#44

This is also related to persistent storage:

https://web.dev/articles/persistent-storage#how_is_permission_granted

Any indication that this is being transmitted home to the mothership?

The docs are unclear about privacy aspects – there’s a mention of “Site Engagement clients” but I’d expect that this to be analogous to cookies – a website would only be able to access the engagement metric for its own domain. That could be a very wonky assumption. Very strange that the docs don’t describe who can access this data.