Brendan, thanks. I'll watch the video when I get home from work.
The game was played in 2017, and I saw several analyses of it at the time. Pretty sure agadmator did one, also chess24 and chessbase.
Scared the $%^& out of me. At that time, it looked like defenses like the Queen's Indian (the feature game) and (to a lesser extent) Caro-Kann had been outright refuted.
But neural network engines have gotten quite a bit stronger since 2017, which may or may not be surprising. The Caro-Kann is popular again, and contrary to its long-time passive reputation, strong GMs (Firouzja, for instance) even play it when they need a win! I'm not seeing many Queen's Indians, though... ;-)
P.S. The Queen's Indian being destroyed stunned me because Karpov played a huge number of QI games with both colors. And I used to play the Caro-Kann vs. 1. e4. The Alpha Zero self-training logs showed the C-K gradually falling from favor as the # of training games increased. There was a really interesting article about how various openings fared in the training games (first guess chess24, or else chessbase) , but I'd have to search for it.