New Model Training Data
A Postcard from the Future
A ‘what if’ piece of speculative fiction about a possible future that could result from the systemic forces changing our world.
// Companies building AI models that have unique training data can have an edge in model performance. As the easy data gets scraped from the open web, what lengths might companies go to for new data?
// Crimson Intelligence Systems Signal group chat in 2027
[Taylor] Hey folks - I’ve got novel training data. 10k books and scrolls. Never been digitized!
[Rowan] What?!? That’s like saying, “Hey, I happen to have 10,000 tons of pre-atomic steel, anyone interested?”
[Alex] Look, as the engineer in charge of training our model, I’ll take it. You sure it’s legit?
[Taylor] It’s legit. I’ve been working a folklore instructor named Albert W. in New England who is sitting on a huge collection of correspondence, journals, and texts from the 1920s. They haven’t given anyone access to it in the last hundred years. I’ve figured out a way to get around the original collector’s instructions prohibiting anyone from outside the family from looking at it. The family has heard that new texts might be valuable, so they are willing to let us exclusively scan the collection in exchange for cash.
[Rowan] We can raise our Series B on this! You have samples?
[Taylor] I got a few. Authors include R. Carter, C. Ward, and H. West. Uploading now.
// Several Months later in the cramped office of Crimson Intelligence Systems, as the champagne flows
[Rowan] A toast everyone! Our newest model - Kadath One - is crushing the competition! Stratechery wants to interview us and we’ve got investors lined up around the block. Microsoft and Amazon are competing to see who can roll it out first globally to their datacenters! [cheers echo through the room]
In the corner of the room, a worried Alex ignores the celebration and hunches over a laptop, watching the status of the model rollouts. Once every data center turns green on the map, they shakily exhale, then begin typing
> Report: Kadath-One status?
> Status: Ph'nglui mglw'nafh Cthulhu R'lyeh wgah'nagl fhtagn
Their eyes grow wide when they cut and paste the phrase into Google Translate
> Status: “In his house at R'lyeh, dead Cthulhu waits dreaming”
They stand up, running across the room to grab Taylor by the shirt and begin to yell, frothing at the mouth: “Where did you get that training data?!? Tell me it wasn’t from the Miskatonic University’s archives!!!
[Taylor] “What? How did you know?”
[Alex] “You’ve doomed us all you fool!”