Post B9ZtSJBjs94OoZT9ea by ariadne@social.treehouse.systems
 (DIR) More posts by ariadne@social.treehouse.systems
 (DIR) Post #B9ZsnsVDfc2MT0IoSW by ariadne@social.treehouse.systems
       0 likes, 0 repeats
       
       @mcc it is probably for identifying tainted output because ingesting LLM output basically causes the equivalent of prion disease for an LLM, the model starts outputting totally useless junk
       
 (DIR) Post #B9ZsyqKm4Zj4vWe4aO by ariadne@social.treehouse.systems
       0 likes, 0 repeats
       
       @dysfun @mcc should i write about what happens and what it looks like and why?
       
 (DIR) Post #B9Zt5n6qSbg0x8RXDE by ariadne@social.treehouse.systems
       0 likes, 0 repeats
       
       @dysfun @mcc i mean, it is actually a kind of interesting problem
       
 (DIR) Post #B9ZtSJBjs94OoZT9ea by ariadne@social.treehouse.systems
       0 likes, 0 repeats
       
       @dysfun @mcc i mean, yes, but i feel like i could explain inference in simple enough terms that most would stop using this as anything more than a divination tool
       
 (DIR) Post #B9ZwPFnUK8U3I6MS9o by conniptions@mastodon.social
       0 likes, 0 repeats
       
       @ariadne  I for one would be delighted if you did this.
       
 (DIR) Post #B9ZyOb5ds7icwqfQci by whitequark@social.treehouse.systems
       0 likes, 0 repeats
       
       @ariadne @mcc does it, actually? i've seen conflicting reports on this topic
       
 (DIR) Post #B9a1aeP1gIBuYlAn5c by ariadne@social.treehouse.systems
       0 likes, 0 repeats
       
       @mcc @whitequark it depends on how much tainted data gets ingested.  but the theory is that the weights get biased.
       
 (DIR) Post #B9aK1MlLAmgMqMV02S by mcc@mastodon.social
       0 likes, 0 repeats
       
       @ariadne @whitequark It would have the effect of just emphasizing the existing weights further, right?What happens if you just double the weights without retraining? Like surely there's a way when interpreting the weights to just say "whatever the weights, square them"? Do results get worse? Better?
       
 (DIR) Post #B9bL0Ai6R2fLV1liim by mcc@mastodon.social
       0 likes, 0 repeats
       
       @ariadne @dysfun In my opinion, that would be a generally helpful and interesting blog post. Which I would probably not read, personally. But probably other people would like it.