Ahh, I see one source of confusion. I’ll answer quickly first: no I don’t have any examples for narratives.
I should have made this more obvious, but my standards for what could be done this year are lower than ‘average’.
Specifically, I think by the end of this year we’ll be able to generate a tv show at all. My standards for that are more technical: enough consistency and control to generate characters and a world that is superficially a TV show. It’s on the order of difficulty of generating a movie’s worth of consistent footage. So at that point, with a human writing/directing (AI assisted or not), we could see a ‘fully’ generated TV show.
AI generating a complete ‘average’ TV show on demand I think is a few years away (1-3), assuming we don’t hit some kind of wall. I think we can probably do it with models the size we have now (~1T params) but fine-tuned to be better at writing fiction or creative direction or whatever. If we stop being able to make models bigger, the next logical thing is improving training (which is already going on with all the post-training being done with like opus 4.5 → 4.6, and gpt 5 → … → 5.4).
Maybe I think to little of ‘average’ TV. I’ve never watched much so maybe I’m not the best person to estimate it. Maybe I’m thinking of more like the 20th percentile or something.
Well you’d need to convert to a single unit and that probably doesn’t work well and would be hard to agree on (I’m reminded of MFDMM etc). But we might agree more about like buckets of TV shows (eg 5 buckets from really bad to really good) and that kind of breakpoint-y agreement is good enough.