SHOW-1 and Showrunner Agents in Multi-Agent Simulations
fablestudio.github.io
fablestudio.github.io
In fact, this is a blind spot of ALL AI that I've yet seen.
Matt and Trey approach writing scenes with the following intent: if a scene can be described as following up a previous scene with 'AND THEN', it fails.
If the following scene can be described as 'BUT' or 'THEREFORE', they write it and it goes in the show.
This is so simple, yet it's completely absent from what AI is doing, and seemed absent from the example script (I'd have liked a written version of their fake South Park script, as it was pretty insufferable). I'm not sure what beyond expanding context length will be needed, context length is merely a technical problem… but this is a significant failing in what we get out of AI.
This is where the sea of meaninglessness comes from. Good creators have a great deal of context and intentionality in their work, and for all that Matt and Trey are puerile and crass, they've got a clearly expressed agenda artistically (and philosophically).
There's also human creators who are very facile, but very prone to 'AND THEN': AI will have a much easier time replacing humans like this, than replacing Matt and Trey.
The joke about Trump was pretty good, though.
I’ve been up all night flying internationally but I might have misunderstood.
What you'd need for Matt and Trey style 'but' and 'therefore' is entire prompts being introduced in the background and switched out. Imagine thousands of words of prompt. 'but' means, fundamentally obstruct something about where your whole prompt is heading, like a screenwriter introducing a twist that must be resolved before the story can continue. 'therefore' means describe something that unfolds obviously as a result of all that's been introduced in the prompt.
These are not sentence-level issues, not output-level stuff. These are prompt level. More than that, they're prompt level with intentionality: you have to understand how a 'but' will fundamentally obstruct your prompt, how a 'therefore' will integrate both your original prompt and the obstruction.
Assume you have to coherently switch around your prompt introducing new fundamentals, and still have that make sense. It might get you rather formulaic results (but, Luke loses his mentor! therefore he must study and meditate and get to the final goal through his own transcendence!) but that just shows you it's working. That's how you get from a pile of arbitrary time-wasting, to a capital S Story.
From there on out, it's about which stories to tell, how far you can depart from the norms while still providing the intentionality and purpose, and what the purpose is :)
'BUT' approach would require opposite, start with story, somehow integrate character quirks.
Maybe in my lifetime an AI could generate a few more seasons of Firefly (that got cancelled too soon)?
Another AI challenge would be to super-impose Harrison Ford onto Alec Baldwin in the movie Hunt for Red October. That would give nice continuity to the Jack Ryan character across the Tom Clancy movie series (Patriot Games, Clear and Present Danger)
You can’t take the Skynet from me
- Custom diffusion models were trained on South Park character and background image datasets. These models could then generate new South Park-style characters and backgrounds.
- GPT-4 was used to generate dialogue for scenes, based on prompts about the overall episode premise and plot points.
- An "AI camera system" was mentioned for scene setup, but details were not provided on how much of the camera work it handled. Voice cloning was used to generate audio clips of the dialogue.
Note: this is just a skim of the paper, entirely possible I and Claude may have have missed something.
I agree it is not a worry, at least not of the same kind. I do worry a bit though about assuming something and reasoning/discussing with that premise, which is nullified if the premise was wrong.
There is a significant drop in believability as soon as the first scene of the generated show starts after the intro. Look at the first few seconds of the first scene. Here, the three boys are just standing statically in the hall while they are talking.
Now take South Park Episode 1, Season 1, from 1997.
https://www.southparkstudios.com/episodes/940f8z/south-park-...
Look at the first few seconds of that opening scene after the intro in this very first episode of the real show. Notice that even though there is a lot of time where some of the characters don’t move at all, they still feel a lot more alive than what you saw in that video that is in the OP. And I think part of it is also the use of exaggerated expressions when they do move. The others not moving is ok, because the ones that do move provide a lot of entertainment with those movements and expressions. And of course the voice acting is very passionate as well.
In the years since then the South Park episodes have gotten even better and better animation.
Maybe I am being too harsh, but the lacklustre animation in the first few seconds of the first scene of the OP video made me not watch any more of the video.
If you want to take on South Park, and you want to beat it at its own game, the bar is very high. Both in terms of script, voice acting, and animation.
Over-fitted or copied.
any questions you can drop them in the twitch chat... or here...
They used AI to generate character dialog and voices (easily the worst parts of the video) while the humans guided the plot and picked the lines of dialog that would be used.
Its giving 'pokemon go to the polls' energy - people using a random pop culture reference to flaccidly try to promote their ideological preference.
I hated it, and was only impressed until I realized how much manual human labor went into its construction. The editorial influence of the humans involved was far too heavy handed. I would have preferred something alien, weird and incoherent.
Here's an example of what AI can actually do - full synthesis of script, visuals and audio. It is incoherent but also strangely unsettling: https://www.tiktok.com/@never_ever_never_land/video/72531490...