Anonymous
7/3/2025, 12:23:40 AM
No.105782022
>>105781513
>Although there's still some lines of conversation that will cause even smarter corporate models to start hallucinating balls.
I have a clue of when this happens if we oversimple the problem two a task of replacing chars lets say map
a b c => w y z
d e f => p q r
But we always train it in a way that they two sets aren't mixed but on inference time you mix it , it will have a Frequentist tendency to one or other.
Could be like domain or context mixing and is were fake novelty (still an illusion caused for the massive amount of data used) or jailbreak could happen.
>Although there's still some lines of conversation that will cause even smarter corporate models to start hallucinating balls.
I have a clue of when this happens if we oversimple the problem two a task of replacing chars lets say map
a b c => w y z
d e f => p q r
But we always train it in a way that they two sets aren't mixed but on inference time you mix it , it will have a Frequentist tendency to one or other.
Could be like domain or context mixing and is were fake novelty (still an illusion caused for the massive amount of data used) or jailbreak could happen.