Let's get something out of the way first: besides the Steven Spielberg movie we're not fans of Artificial Intelligence (henceforth referred to as "AI"). We think it's a terrible thing, and we're extremely upset at what it's doing to our artistic community. And let's not even bring up the inevitable creation of SkyNet, shall we?
One day we came across a query suggestion for ChatGPT: pick a movie, and ask ChatGPT "Give a 10 word summary of..." the movie. Despite the fact that ChatGPT can't count (many responses exceeded the 10 word limit), some of the responses were actually fairly reasonable:
All in all, that's not bad. Among all the thing that AI does badly, and among all the things that AI does that could be detrimental to the livelyhood of others, this was reasonably acceptable in our book.
At the same time we were experimenting with this, for a different project we were developing libraries for integrating with Twitter and BlueSky APIs. And for another work-related initiative, we were also developing an OpenAI integration.
Put it all together and the "Movies In Ten Words" Bot was born.
Is an Artificial Intelligence capable of humor? Not just any humor... INTENTIONAL humor. Yes, we've seen many times when AIs do something seemingly stupid and comical, but that wasn't their intent. What if we actually ASKED the AI to try and be funny?
So we made a slight adjustment to the query we were sending to ChatGPT: Same as before, but tacking on the words "but make it silly" at the end of it.
The results were better than we ever could have imagined.
From now on, whenever we have to type any query into ChatGPT, no matter what it is we're adding on "but make it silly."
As we let the bot chug along, we came to a realization: this isn't necessarily limited to movies! So we added support for songs, TV shows, and Broadway plays.
We let the bot run for a while and it had an interesting effect... for some inexplicable reason, over time the bot began to somehow equate "silly" to "dancing" more often than not, so every other post made reference to dancing in one way or another. So for a while we modified the query with the suffix "...but make it funny and do not reference dancing."
Ever since we have made multiple changes to the "seed" strings provided to ChatGPT prior to asking the actual question. We had to tell it to stop referring to everyone as "grumpy", stop it from revealing spoilers, stop it from using proper names of characters (which made the trivia game aspects of it easy), and other things.
The current seed list, which helps it to be less repetetive in its responses:
We also switched from GPT 4.0 MINI to the "OG" GPT 4.0.
We decided to add support for books as well!
We poke at the bot's code every now and then, improving functionality and whatnot, but in essense it's remained the same.
One big recent change has been switching it to query either "chatgpt-4o-latest" or "o4-mini". The latter is significantly cheaper in terms of usage cost, but it takes a lot longer... For example, the same query that takes less than a second on "chatgpt-4o-latest" could take as long as 30 seconds on "o4-mini". This is incompatible with our Discord integration, which has a 3 second timeout for on-demand requests. So we intelligently pick one or the other as needed; for queries that have no timeout we are using "o4-mini", which seems to be providing more robust answers.
