The ads come at different entry and exit points for every user.

They’re not referring to the YouTube ads, but the “let’s take a minute to talk about today’s sponsor nordvpn that I used on my trip to Antarctica.” This is a part of the video file itself, and it starts and ends at the same time for all users.

Pretty sure they just use timestamps from a crowdsourced database, just like sponsorblock.

Nope, it’s analyzing the sound to guess where the intro starts and ends. Turns out this is pretty simple to implement, but quite reliable. Source: worked for Plex