I wonder if it's actually being updated quickly. I think it'd be much easier to train it to use google's search index as a tool in a way hidden from the user. For example regarding the shutdown date of itself it might have googled "google bard shutdown" to find and misunderstand the HN comment cited in the twitter thread.
Seems like it would be trivial to replace "As an AI language model who hasn't been updated since x date" with "$ESCAPE_SEQUENCE search_query: $TERM" and rerun the question with the hidden prefix "Given that the google results for $TERM are $RESULTS"...
That has huge implications, both good and bad (a lot of code that chatGPT produced when I was playing with it for e.g. sveltekit is outdated)