Meeting new people at a large event can be a challenge. I'm part of the Digital Nomads group in Asheville and we decided to try an an experiment aimed at fostering new friends and connections as well as cultivating a sense of community and mutual support.
The basic idea was to have event attendees fill out a three question survey and then use that information to match them using AI to three other people that they should meet. Participants also provided a 'code name' that we used for privacy and to create a sense of fun and adventure.
The questions were free form text and were variations on 1) What brings you to this event? Let's call this question 'like'. 2) What are you looking for? Lets call this one 'ask' and 3) What can you contribute? Lets call this one 'offer'. The idea being that if two people claim to like similar things (hiking, technology, the outdoors, etc) they should be a strong match and if one person is asking for something the other is offering (ie. a job in marketing, lessons in sustainable agriculture, etc) then they too should be a strong match.
We ran the experiment twice with event specific versions of the questions. The first time, a Friendsgiving event, with google forms and a manual matching process and the second, a Huladay event, with an MVP web app (https://aiintrosapp.com) I created. Each response was embedded using "BAAI/bge-large-en" and subsequently "text-embedding-ada-002" and a direct cosine similarity score was calculated for all the like answers, the ask answers, and across the ask and offer answers. The nearest answer for each question became one of that person's matches. In a few cases the same person matched more than once so we took the next closest person so that everyone got exactly three different matches.
Then for each match we had GPT4/4.5 create an introduction based on the specific answers the people provided. We iterated several times on the prompt to get the right balance of brevity, compliments, excitement and tone. And finally an email was sent to each participant with the intros and code names of their three matches.
At the first event we had a deadline for participation while on the second event were able to have people answer the survey at the event and matches and emails were created automatically within a few minutes. We asked the participants to write their code name on their badges along with their real name and it was up to the participants to find their respective matches which set up a scavenger hunt type game.
We learned a few things from the exercise:
- Many if not most people found it fun and a good way start up a conversation even if with people they were not explicitly matched with.
- Participants were sometimes confused that matching was not bi-directional. That is if Chris matches with Pat. Pat may not necessarily match with Chris.
- Occasionally the matching process worked well "You matched me to my best friend." sometimes people were confused by why they matched.
- Participants did not always provide enough detail in their responses which resulted in non-specific reasons for matches.
- People wanted to fill out the survey after the deadline at the first event but at the second event some found themselves too busy and distracted to fill it out completely.
- It was often difficult to find your matches in a busy crowded event. So coordination at the event is critical.
- Sometimes someone could not make it at the last minute leaving others searching for someone that was not in attendance leading to occasional frustration.
Overall it was a lot of fun and an educational experiment and most people rated it highly in both enjoyment and on whether we should repeat it or not. It was a success in that it did help create friendships and foster conversation.
I don't know if we'll repeat it again with this group though as I believe it might work better for events with a more focused purpose or theme. People coming together for a specific reason should allow for more focused questions which would lead to more focused answers and consequently better matches. Or perhaps we need to rework and improve the questions.
Let me know what you think. Would you ever consider something like this?