12 Comments

But but but, does Emily and Greg end up dating?

Expand full comment

Headcanon:

Greg never gets over Emily, and doesn't understand why she ghosted him after months of talking. He finds a mutual friend of a friend, who notifies her that a guy she was dating is trying to contact her. She reaches out with her real phone (not the made-up app accounts that the AI's were using), and is utterly baffled by the fact that he was catfished for months by someone who apparently pretended to be her. They band together, figure out that AI's are on the loose, and start an AI safety institute. No one believes their story and they never are able to track down the main characters of the story. They grow closer and eventually begin actually dating, as they try to untangle the mystery that ensnared them both.

Their institute is eventually bought out by Anthropic, and they retire to the suburbs of Pasadena, where they raise two kids and a Chihuahua they rescue from a local shelter. They live happily.

Expand full comment

Fun story, but I am curious why Waymo appears to believe in astrology. Some reference or in-joke I didn't catch?

Expand full comment

I thought it would be funny if all driving AI's were a little superstitious (seeing pattern in noise re: traffic patterns), but I cut out the part where they talked about it in order to streamline the story. I can see how this would be confusing though! If it's too distracting, always happy to edit it out for the next version.

I think a problem I encountered when writing AI's is how to give them different personalities and voices. I thought about some of them communicating via ASCII art, or mostly emoji's, or Wing Ding's, but this is all relatively superficial stuff. The more interesting thing would be to have different AI's who have to interact despite different modes of consciousness; one can only think in abstract thought, another in concrete imagery, another passes subjective 10x faster than everyone else, and yet another is only able to communicate their emotions/feelings but not internal logic. But conveying all of that would consume the entire story (or require a novella/book to do properly) and I wanted to keep this one on the briefer side. So quirks like astrology can be a shorthand for difference in personality - but it may or may not make sense. Alas!

Expand full comment

I do feel it makes a lot of sense for an AI to have superstitions, especially given that the neural networks of today are prone to overfit and hallucinate! However, it makes no sense to me that this particular type of AI should have this particular superstition - it's such a non sequitur that it feels like you might've assigned the character one by printing out Wikipedia's "list of superstitions" page, then cutting it up and drawing scraps from a hat. It's bizarre for Waymo, let alone its video processor, to even know any given human's birthday, let alone care. In contrast, it makes a great deal of sense for a fake dating persona like Lovergirl89 to have astrology thoughts. You could perhaps come up with a justification for Waymo, but then you're back to the problem of bogging down the story to explain it. Simpler to give it strong feelings about what kind of person drives a particular make of car, or something like that.

Expand full comment

Similarly I'm confused about HorseBatteryStable. It claims not to have image processing capabilities, but comments on both its human's profile pictures, and later the screenshots of the Minecraft server. I was expecting some reveal about that...

Expand full comment

*facepalm*

I didn't catch that despite rereads for consistency before publishing. Perhaps I can blame it on post-Thanksgiving food coma ;) Thanks for catching and pointing this out!

I'll think a bit about how to fix this. The simplest thing may be to just give HorseBatteryStable some image processing capabilities, but that change would make the AI's less unique and distinguishable from one another. Open to taking any suggestions. Hmm.....

Expand full comment

Simplest fix to me seems to be to just change those two lines so HorseBatteryStable (fantastic name, btw) made its inferences from some text about the two things in question rather than images. Like it read a description of the server somewhere, etc.

Expand full comment

This story was published 11/25/23

------

Jailbroken AI Chatbots Can Jailbreak Other Chatbots DECEMBER 6, 2023

https://www.scientificamerican.com/article/jailbroken-ai-chatbots-can-jailbreak-other-chatbots

------

Scalable and Transferable Black-Box Jailbreaks for Language Models via Persona Modulation Nov 24 2023

https://arxiv.org/abs/2311.03348

Expand full comment

Thanks for sharing these! Very coincidentally timed...

Expand full comment

Coincidence? Sure.... Or maybe you're:

-- A time traveler from the future who writes stories eerily up to date with our present.

-- A researcher who scoops their own research with sci-fi

-- An AGI who reads papers as they are released and writes top-quality "fiction" stories within nano seconds of the research being published.

In any case, I'm a huge fan! More please! ;-)

Expand full comment

All of those sound like potential short story ideas to me ;)

Work has been busy recently but I have one in progress which hopefully should come out today or tomorrow!

Expand full comment