The first time we debuted this feature was at [Califlorence AGI](https://jasonbenn.com/community/10-the-unconference-to-community-pipeline). From my newsletter update: > The main constructive feedback we got last time was that 3 straight days of indoor large group conversations, while stimulating, leaves people feeling absolutely destroyed. I suspected that something should change when I had to wake up Christina Jenq for her own session. We ended up swapping another session out for Nap Time. > > So we mixed up the format this time. The idea was to **break people up into small groups for every meal and dispatch those groups to local cafes and restaurants** within walking distance. Advantages: > > 1. People are going on 5-20 minute walks with each other, in the sunshine, 4-6x/day > 2. It not-so-subtly gave people a taste of the Neighborhood lifestyle > 3. It massively simplified logistical complexity during the event. Last time, most organizer headaches were from coordinating food, staff, and chefs. By paying a one-time, up-front programming cost, we were able to outsource 75% of our meals > 4. As a result, we were able to reduce prices from $400 to $200 > 5. It allows us to start building relationships with local businesses like Palmyra, The Mill, Dumpling House, Arbor, Lady Falcon, etc > 6. Most importantly, attendees would get uninterrupted, high-quality conversations in quiet settings throughout the weekend, making for a much more humane pace and allowing people to build real relationships with at least 20 other attendees each. > > We could have matched people randomly and it probably still would’ve been an improvement, but I was really keen to make the groups feel like _exactly_ the people you wanted to meet. I’m inspired by SwapCard, another conference app that allows you to request 30 minute meetings with any other attendee. The problem with SwapCard is that it only works if you take an hour to read everyone else’s bio, and we just knew that most people wouldn’t care enough. > > **The solution was to ask people what would make the conference great for them, and then use GPT-4 to predict who they’d want to meet.** > > Specifically, for every possible pair of attendees, we had a script compare Person A’s goals to Person B’s expertise (which we also gathered during onboarding), and asked GPT-4 to estimate the likelihood of them having a great conversation from 0-100%. > > We also hacked together a janky website so that you could hand-request specific people, just like SwapCard. It was a big list of names, bios, and checkboxes, with the top 30% pre-selected for you. About 25% of attendees used it to tweak their matchmaking. > > The last step was to take these preferences and produce a series of small groups for everyone, taking into account people’s availability. > > I didn’t know how to do this, but apparently there’s a rich vein of literature about matching problems like this (one of which was awarded a [Nobel in 2012](https://en.wikipedia.org/wiki/Stable_marriage_problem?wprov=sfti1)). We settled on the Stable Roommates algorithm, found a nice simple implementation on GitHub, and ouallah, the groups were looking pretty great. > > We also added a few finishing touches: > > - We’d downgrade mutual preferences by 30% after each match, so that everyone would get mostly new people each meal > - We scripted the matches to roll out gradually throughout the weekend, so that people could modify their requests between meals > - We sent emails to each group with the restaurant, time, each other’s bios, and each others goals. This also let them confirm and coordinate with each other > - We sent multiple groups to each restaurant so that if one group disintegrated, any orphans could just join another group > - We only matched people that opted in (we had separate calendar invites for every meal), so that people could participate as much or as little as they wished > - And we cross-posted every group and their locations to the event’s Discord so that latecomers could crash other groups. > > The feedback was very positive, with several people congratulating us on our conference innovation and one attendee tearfully saying it would be “difficult to improve” (!). I disagree but it was charming to hear :) > > I’d love to see more event organizers copy this idea. From the organizer perspective, it’s very cool to have a script running in the background and to feel like your event is running somewhat automatically. And as an attendee, I really did feel like I got to develop actual relationships with most of the people I was excited to meet. > > Between events like these, my [Explorer’s Clubs](https://lu.ma/su9xm2rm), T-Groups, Jeffersonian Dinners, etc, I’m consistently struck by how LARGE the space of possibilities is for conversation formats, and how little of it we consistently explore. While the existing literature on collective intelligence is frankly [disappointing](https://cci.mit.edu/research/), I think AIs are actually quite well-suited for various types of coordination and matchmaking and so I feel very fortunate that I get to explore this space for my job. It’s amazing how much energy and creativity you unlock when you feel internally aligned. ## v2 improvements Implemented for [[HammingBio]]. Improvements: - Ability to update your bio - Ordering attendees alphabetically helps you find a specific person by name, but favors folks with names early in the alphabet. The solution is to order them alphabetically, starting just after your own name - Emailing everyone that neither opted in or out to a meal slot, to alert the often-confused folks that didn't understand the system, with a list of all groups and names (we did this manually at Califlorence AGI) - Improved interface ![[Pasted image 20240726111857.png]] Skipping a bit over the bios for every attendee... ![[Pasted image 20240726111916.png]]