You can add your own prompts. The AI must be publicly accessible and reasonably affordable. Roughly, less than $100 per movie.
This question was inspired by Scott Alexander's question: In 2028, will an AI be able to generate a full high-quality movie to a prompt?
It is possible that an AI will be able to generate full high-quality movies to some prompts but not others. My market is about which prompts it will be able to successfully follow and which it won't be able to.
P.S. to avoid any conflict of interest in resolving this market, I won't be making any predictions on it.
1,000
3.00Someone made a similar market about what prompts OpenAI's Sora will be able to accurately follow: https://manifold.markets/TheOtherKC/on-release-which-of-these-prompts-w?r=Y2FsZGVya25pZ2h0
How does this resolve if the prompt goes through stages of auto improving the prompt? Like with DALLE 3 for example, I can give it any prompt and ChatGPT makes my prompt better and more detailed before sending it to DALLE. In this case, basically any prompt no matter the length should work, if it works at all.
@calderknight wait, no? Any system that can write a whole movie script is obviously "auto improving the prompt" under the hood. So are you saying it won't count if it shows you an intermediate stage in the process?
@ErickBall this question is about which prompts the AI can successfully follow, not about which prompts the AI can successfully follow an auto-improved version of.
@calderknight So, say someone makes an ai movie generation service, and you can put in any prompt, and then it uses a language model internal to write a script from the prompt, and then it generates video and audio for everything described in the script. Would that count? Would it count if it also lets you look at the script before it does the rest?
@ErickBall Thanks for the questions. If the AI takes the prompt, writes a script, then uses that script to make the movie, that is OK.
If it presents a list of generated scripts from which I am to pick a script, I will pick one at random. If the resultant movie accurately fulfills the prompt, the answer will resolve YES.
@Mira I think I would choose whichever model I thought was most appropriate for the prompt. For example, if the prompt is an entire book I would choose the model best-suited for the task (for example, an AI built to do exactly that, or one that was supposed to be capable of it). I wouldn't necessarily test every prompt on every movie-model, but I might test every prompt on every one of several AIs. For example, if there were n SOTA movie-generating AIs, and it were affordable, I don't see any reason not to try them all (until one succeeded).
@Mira True, I think there's a cap of 100 answers, and older-James might not appreciate me putting him on the hook for $10,000. I'll change the wording in the description to "roughly 50 USD" and I'll commit to spending up to $5,000 if necessary.
But the resolution of some answers should be clear without actually running the prompt. For example, answers about prompts of length <100k words could resolve YES if the answer about a prompt of length 100k words resolved YES.