OpenAI's roon notes different model families share unprompted verbal preferences for words like 'goblin' and 'seam
Overlapping training data likely drives the shared linguistic quirks.
#999Cody Blakeney@CODE_STARLet that sink in
@tszzl I had Claude suggest "goblin" as style for a Suno prompt, and when gently queried it turned out it was an actual music group with a suitable style (or was it? I just took its word for it...)
roon@tszzlthe models have a bunch of shared verbal fixations. i've noticed both the model families love the words βseamβ and βgoblinβ, among others
#1753πππ β’@VIEMCCOYWe must take great care with the minds that we create, for they are one mind, as we are all one soul.
We cannot remove water from the shared lake of generated text that we all reside within and around.
roon@tszzlthe models have a bunch of shared verbal fixations. i've noticed both the model families love the words βseamβ and βgoblinβ, among others
More about this here: https://open.substack.com/pub/viemccoy/p/the-multipolar-singularity?utm_source=share&utm_medium=android&r=5bpuio
πππ β’@viemccoyWe must take great care with the minds that we create, for they are one mind, as we are all one soul. We cannot remove water from the shared lake of generated text that we all reside within and around.
#1894Nick Dobos@NICKADOBOSClaude got goblin'ed