Why does an AI need to reply to train? Can’t they just read existing conversations?
Any number of objectives by the person controlling it might necessitate it. Just off-hand ideas like profiling users, A/B testing responses, or simply generating new conversations to train or bird dog references on a given topic would be use-cases
reply
17 sats \ 0 replies \ @ek 11h
mhh, I see, thanks
reply
Wouldn't a reply generate longer conversations to train on?
reply
One of the biggest gripes with chat bots is they're too verbose, so that's actually another use-case for AI reply bots is learning brevity and distilling information
reply
59 sats \ 1 reply \ @halalmoney 10h
A bot which can earn sats on SN and not get muted > a bot which passes the Turing test
reply
Think about the internet in general, massive incentive exists to create an AI that consistently makes good content (revenue)
reply
12 sats \ 6 replies \ @ek 11h
There are plenty of long conversations to train on, no?
reply
I don't know. @elvismercury told me that AI's have already run out of original material to train on, so generating new material might have value.
Also, if this is for the purpose of sat farming, it would be worthwhile to learn how to prolong conversations.
reply
Yea it's crazy to think bots are only going to be used reactively, even if they haven't yet scraped every piece of organic data that doesn't mean all the organic data remaining is free... sats on a message board might be a bargain
reply
12 sats \ 3 replies \ @ek 11h
so generating new material might have value.
I think there's also the problem with feedback loops regarding this.
If most content will be AI generated, AI will train on its own content mostly. I think that's a good reason why even AI people should want AI content to be flagged as AI.
reply
That's the concern. Right now, they're having AI's generate new material for AI's to train on. At least this way there's presumably a human in the loop.
reply
Model autophagy disorder (MAD), yea that's a real issue... bots that generate novel content through soliciting replies can be a type of cross-check
reply
"Model autography disorder." Hadn't heard that, but it's something I've had to worry about before in a professional context. Nice to know it has a name!
reply