Yeah – within the next few years. It might be possible to generate a full TV show by the end of this year (but expensive). Also we have video/image generators specifically for Anime now too (not just fine-tuned to prefer it from what I gathered), and it might be a lot easier/cheaper to generate Anime shows than live-action ones.
Have you seen those like micro-dramas that are becoming a thing with short form video? Each episode is like a minute long. I think we’re at or close to being able to generate those.
What are the impediments to generating longer TV slop? SOTA models today have issues with character and detail persistence, but those are being worked on actively. The script and stuff like camera direction, voice/actor direction, setting, all that can be done in text and therefore generated. TTS and voice cloning are pretty good now, and getting to the point of adding laughing and the like. Multimodal LLMs could provide feedback and edit instructions directly from video or ADR. We already have text-based photoshop, so text based video editing isn’t far behind. And if the characters look bad, well with deep-fake face swapping much of that can be solved in post. So I think we’re pretty close to being able to generate a TV show at all, then the only question is how long before it’s on demand.
They already can, I think. At least coherent enough. You might not be able to do this via chatgpt’s main UI, though. Really you want a coding agent (they’ll happily generate narratives) because you can easily do stuff like have overarching plot structure, world building docs, rounds of feedback (and instruct the LLM to impersonate various authors to give specific kinds of feedback), iteration for consistency… hell why not add an agent for boringness detection to reorganize scenes, change PoVs, set something up sooner to add tension, etc.
Just to be explicit: the bar I’m setting isn’t very high. It would still obviously be AI to anyone who is familiar with it, and a lot of people might not be able to stand watching or consuming the content. There can be plot holes, just not more egregious than those in average TV shows (and there are plenty of those).
I don’t know much about that side of things (the public discussion) besides hearing authors complain about it (on YT), and some experiments I’ve run. I agree that chat apps aren’t very good (though I got a passable hour-long audiobook out of gemini 2.5 – it was good enough to half-listen to while driving).
Maybe we have different standards of coherence? If you mean that there are no contradictions, yeah maybe we’re a ways away.
But if an ‘author’ (or book-orchestrator) has a good background doc for a character and can check each scene against that (\times n characters), I don’t see why those things can’t be ironed out. It’s just a matter of time, cost, etc. (Feeding a long book into claude could cost a few bucks each time, so doing that per character per scene might get really expensive, but all those costs will drop soon anyway.)