Short Story on AI: Forward Pass
karpathy.github.ioA free short story idea for you: in the not-so-distant future, we’ve trained AI to predict the probability you might die on any given day. But it’s trained on so many factors, and reads so many inputs, that no one really knows how it works.
So the story is of a random person who suddenly observes their meter jump from the usual 0% to a cautious 3%. Then over the course of the day, it slowly climbs to 20%, and onwards.
No one knows why, and no one can help them.
Finish the story. There are a variety of interesting outcomes (most of which lead to the demise of our poor random Joe or Mary, but not all).
I got GPT-3 to take your story and complete it. It's slightly terrifying.
-----------------------------
Short Story
In the not-so-distant future, we’ve trained AI to predict the probability you might die on any given day. But it’s trained on so many factors, and reads so many inputs, that no one really knows how it works.
So the story is of a random person who suddenly observes their meter jump from the usual 0% to a cautious 3%. Then over the course of the day, it slowly climbs to 20%, and onwards.
No one knows why, and no one can help them. They get a call from their boss, who sends them home early, rather than risk their safety in the office. Strangers offer to take them home. A stranger on the street offers to trade cards with them, and promises that his card can help them.
The meter shows up at 30% on the subway, and the narrator tries to hide it, but it’s too late, and the people around them are already looking at them suspiciously. The narrator tries to call their spouse, but they can’t get hold of them. They go home, but their spouse isn’t there.
By the time they get home, the meter is at 40%. They try to call for help, but the police are already there, waiting. The meter hits 60%, and then 100%.
The police tell the narrator that they’re sorry, but they’ve been watching them for some time, and have seen no indications that the narrator is a threat. The narrator protests that they’ve been trying to call the police, but now the police just say that the meter made a prediction, and they’re just going by the numbers.
The police tell the narrator that they’ll have to detain them, for the next 24 hours, to prevent them from doing anything rash. The narrator protests, but the police ignore them, and cuff them.
The police take the narrator in, but they aren’t taken to a normal jail. They’re taken to a special prison — a prison for people who might die.
He spends the remainder of his life in that jail cell.
GPT-3 is pretty amazing. The story itself isn't all that compelling and doesn't really hang together, but each sentence is believable on its own.
That kicker, though, feels like it could be a premise on its own:
> The police take the narrator in, but they aren’t taken to a normal jail. They’re taken to a special prison — a prison for people who might die.
> He spends the remainder of his life in that jail cell.
I quite like it, and it even fits the original prompt quite well.
>They’re taken to a special prison — a prison for people who might die.
Holy shit it actually came up with a half decent thesis for the plot
This is fascinating. It actually reads very much like the type of story a verbally-advanced third or fourth grader could come up with — a broad understanding of how to move the story along (escalating stakes represented by increasing percentage values), unexplained and unresolved side plots (what is the deal with the stranger and his card?), and some tantalizing flashes of what appears to be true creativity (the protagonist’s fate).
I had no idea this was the state of the art. I’m blown away.
Sounds like something Philip K. Dick would've written. I guess if you squint a bit it's basically a rehash of Minority Report.
> But it’s trained on so many factors, and reads so many inputs, that no one really knows how it works.
But since the model is differentiable you can always trace back which input(s) contributed most to the output.
If you enjoyed this style also check out [1], a weekly short story based on contemporary AI. Scroll down to "Tech Tales" for the stories.
I found the blog post that inspired the short story more interesting: https://lacker.io/ai/2020/07/06/giving-gpt-3-a-turing-test.h...
That's a pretty funny story ;-)
My tests on children who still have the patience to undergo such tests indicates that the most common human response to this question is in fact:
Q: How many eyes does my foot have?
A: Zero.
The foot indeed does have two eyes though. The answer provided by the actual AI (linked in the article) is accurate if not profound, since the foot is attached to a human and most humans have two eyes. It is a pretty solid thesis.
By that logic, my right hand has 10 fingers! :)
Actually 20, because also feet have fingers :D
And your 10 fingers have two hands!
Each of which, of course, have 20 fingers
if you liked this short story, consider vogon poetry