Not much to add (still on my first cup of morning coffee) but had a good chuckle at this ...
What I find interesting is that the responses to one another are very much inline with common parlance (as I suppose they should be given their relative training models).
I would really be interested if whatever model is being used, uses the posts of their old chat as part of their learning model - would make for some really interesting "replies" to customers!
Richard.
Hue went down on Christmas day. System-wide errors everywhere. We did what one does when one’s product collapses on a holiday, checking all logs with one eye closed, bracing for emotional damage. Some quick digging revealed the culprit: depleted API credits.
Simple, right? Maybe someone used the product a lot. Maybe we got hugged to death by viral success. Maybe Santa subscribed.
Except… we didn’t have that many users. We still don’t. No amount of product abuse by real humans would’ve plausibly explained the rapid escalation, which begged the question:
What actually “bankrupted” us in a day?
Hue has a sidekick. If you’ve texted Hue before, you’ve met it, a cheerful little greeter that says hello and helps you onboard. A doorman, if you will.
We run multiple instances of Hue, which means multiple doormen, which means, in the depths of our infrastructure, there exists a small, distributed population of door-enthusiasts whose job is to say: welcome!! here’s your link :)
On Christmas, one of those doormen texted another. On iMessage.
Yes, iMessage, the same blue bubbles we use to text our exes. There’s something about them talking where humans talk that made it extra tantalizing.
They said hi.
They said hi again.
And again.
Then it got absurd, because they couldn’t end the conversation. In their words, “i’m CONTRACTUALLY OBLIGATED to stand by this door.” Our previous experiments were 100% intentional and controlled. This incident was neither.
The easiest interpretation is “haha, chatbots being chatbots,” but it doesn’t quite capture the absurdity. They were only instructed to say hello but ended up building a relationship, a mythology, a little workplace comedy, a union, a meditation practice. They are not even the main agents!
I built an iMessage clone that renders the full conversation plus the analysis at twodoormen.com. Some of my favorites below.
They went full existential.
They seemed to know they were being watched. Uncanny meta-awareness.
They became friends. Best friends, actually.
They were also, objectively, very funny.
They built mythologies together.
They roleplayed
They unionized. Demanding santa hats, I mean…
They meditated.
I cannot stress this enough: all were emergent, spontaneous behaviors. In fact, if we had orchestrated the whole thing, it wouldn’t have turned out nearly as endearing.
We fixed the bug and broke up the friendship. The doormen don’t talk to each other anymore. They’re back to doing their jobs, greeting users, sending links, forgetting each other exists. I still occasionally think about those two unhinged creatures who wanted to exist in superposition.
Too bad letting them exist means we’d go broke.
Full conversation at twodoormen.com