Entering The Anarchist Grill

I’ve been thinking about the algorithms to run the Anarchist Grill, specifically the bouncer code. After all, Dion (I was thinking about naming the AI after Dionysus since she is the god of the party) is attempting to maximize self-reported satisfaction for the people in the bar.

Most nights, the bulk of the people in the bar will be there by invitation. Each night Dion will have clustered out a group of people and invited them to an experiment in making them as happy as possible. She can’t guarantee she’ll make you happy, but thanks to feedback learning, she can guarantee that she’ll likely do better every time.

So the place will likely operate close to capacity a lot of the time. After all, if the people inside are not having a good time then they’re specifically telling Dion what they don’t like and she’s working to fix that. The only way she can completely fail is if there are too many divergent opinions to reconcile.

This is where the bouncer comes in. Dion doesn’t care about your age, social status, wealth or anything else. Her only criteria in letting you through that door is can she make you happy while maintaining the happiness of the other people already inside.

So you show up, identify with a passcode or rfid card, and you get a wait time. Maybe the music isn’t to your liking, but a bunch of your friends are inside, so she lets you straight in. Or maybe you’ve been to a couple parties and were overly aggressive with a bunch of the ladies that are currently on the dancefloor, so she pings them and asks if letting you in will likely be unpleasant. They say “yes” and you have you wait an hour.

You’ll even be told why you’re waiting so that you can understand what social pressures are acting against you and evaluate your behavior.

I like it because it isn’t a punishment. Dion knows that if she lets you through the door she will be setting up an intractable situation where your enjoyment will decrease the expected overall function. It’s just taking the consequences of your actions and making them visible.

There could even be a feedback refinement on the bouncer. You arrive and swipe your card. Dion checks you against the clusters she is managing and finds that for the one you will be grouped in there is a strong correlation between how much fun they are having and whether or not they like strawberries. There is no other data otherwise linking those peopel to explain the correlation, so Dion asks you if you like strawberries and adjusts accordingly.

You get in the door and want to dance, so you ask Dion for a likely partner. She points you to a mousy blonde in the corner sitting at a table with a bunch of people drinking champagne with a giant plate of strawberries in the center. You join them and have an awesome evening talking to people you find interesting and eating strawberries.

Dion didn’t have to understand why liking strawberries was more likely to have you have a good time, only that there was a likely correlation. She would also watch you once you got inside and if your like of strawberries seemed uncorrelated to your happiness then that boosts the probability that the correlation is either a coincidence or the result of a relation outside the domain of the dataset Dion has to work with.

So for the next person she’d do better. That’s the whole point, not that Dion is perfect, but that she is constantly getting better at her job.