An instance that mirrors the accounts of another instance & associates each account with a markov bot trained on the posts of the original user, set to create new posts at the same rate as the original user.
@xtof54
Because char-RNN would be too computationally expensive & hand-crafting a separate grammar to use tracery with for each user would take too much engineer time.
@xtof54
Another reason to avoid RNNs is masto is new enough that any given user is unlikely to have produced enough text to make NN generation coherent -- i.e., markov models are strictly better for the next few years when we divide up the corpus by user.
(If we wanted to simulate an "average mastodon user" from, say, every node federated with mastodon.social, then RNNs make more sense because the scale is big enough.)
@enkiv2 would be crap anyway... and you have adversarial *user*: remember facebook's bot ? can't get its name...
@xtof54
Facebook's bot was called M and was actually just a call center.
Maybe you're thinking of Microsoft's bot, Tay?
@enkiv2 yep thats it: microsofts one :-)
@xtof54
I got onto some popular shared blocklist for spamming Tay with positive affirmations (at a rate of one affirmation every 23 seconds for eight hours). Obviously my affirmations had no effect.
@enkiv2 you did ? ;-) well, I guess alone against crowd is not enough. But we're so far to have good generation; would require lots of smart things: semantics, dialog model, common sense, world model... can't be done with just an lstm, whichever nb of layers are stacked...
@xtof54
Tay was definitely mostly a PR thing, and it got Microsoft a lot of (bad) press so it was successful. I don't think even MS's management & marketing was naive enough to believe the hype they mostly failed to manufacture around it.
@enkiv2 ok :-) would not have anyway enough data per user, would need to merge user toots or train on large twitter corpus and just adapt the bot based on the few usrr toots ;-)
@enkiv2 why Markov ?