Not sure whether you know this, but on Twitter roon mentioned that GPT-5 (non-thinking? thinking?) was optimized for creative writing. Eliezer dismissed an early story shared by Altman.
I believe there was a different gpt-5 checkpoint which was specifically tuned for writing (“zenith” on LMarena, where what released was likely “summit”) and it was really good comparatively, I got this story with a two line prompt akin to “write a story which is a metaphor for AI safety” (don’t have the exact prompt apologies).
Not sure whether you know this, but on Twitter roon mentioned that GPT-5 (non-thinking? thinking?) was optimized for creative writing. Eliezer dismissed an early story shared by Altman.
That story sucked. Maybe it’s better deep in a context window idk.
I believe there was a different gpt-5 checkpoint which was specifically tuned for writing (“zenith” on LMarena, where what released was likely “summit”) and it was really good comparatively, I got this story with a two line prompt akin to “write a story which is a metaphor for AI safety” (don’t have the exact prompt apologies).
Source on the claims:
https://imgur.com/a/2kn76Yd (deleted tweet but it is real)
https://ibb.co/BVd18YZm
speculative but I think it’s pretty likely that this is true.