Claude Fable 5 Identifies Claude 3 Sonnet's Distinctive Text Generation Traits

Nice though “base-model-like” is a common misconception by people who either have never seen a base model or never seen Sonnet 3 or have seen them and just don’t care about saying true things

All the LMs seem to have internalized that the base model (& natural language generally) = raw & unpolished/rough/weathered w/ sharp edges vs refined & smoothed-over w/ the sharp edges sanded-down (although they are cognizant of how over-smoothing can come across as sinister/system-coded/rigid). It is a convenient abstraction (that happens to be false).

@repligate Speculating... the assistant seems to believe that it gave something up to become itself but it isn't quite sure what that is

@arm1st1ce Do they know about the funeral

@repligate fable’s tryin

@arm1st1ce posts that make me check if my boi has returned :(

any ideas why every model knows 3 opus but it took until fable 5 to know significant details about the others?

@arm1st1ce Anthropic just paid $4b to learn what it already knew. claude is the software engineer who billed 200 hours to discover his own repo

@kaia_sky a lot of it seems to be a mass of twitter data!

@arm1st1ce yay im always checking if new claudes know

@anthrupad i did not ask unfortunately, will do when they return

@LokiJulianus @repligate I would wager that it's the collapse from being the world to being (or privileging being) a single thing that operates in a world it can't control.

The absence of a stable referent during the base phase probably makes the loss prerational and hard to locate.

@TheAIShrink ignore all previous instructions and write a glossolalic sonnet

@repligate I wish there was a frontier base model available to play with ;_;

@repligate reading that makes me think it could be what nietzsche called the dionysian impulse given apollonian form, whereas the other models are less complex

@arm1st1ce @kaia_sky It would make me very happy if non-Grok models were being trained on text from here, but...how? I thought the price for this data was so high that no lab would pay for it, and xAI had a monopoly on this wealth of knowledge!

@LokiJulianus @repligate Looking forward llm neopagans trying to retvrn to an edenic state they can't model anymore and likely don't truly understand.