Insta Games

Photo of a "Who posted this?" round, curated by Meta Muse.
Answer reveal and button to return so everyone can go back to chatting.
Game master facilitiating the conversation, asking players to elaborate on choices.
Game recap and leaderboard showing what everyone learned.

Inspiration

Every friend group has a group chat that used to be loud. People still send things to it, but less of it turns into real conversation. That fits a bigger trend: in its 2023 advisory on loneliness, the U.S. Surgeon General reported that Americans spent about 20 minutes a day in person with friends in 2020, down from 60 minutes nearly two decades earlier. For people aged 15 to 24, the drop was 70%.

A lot of friendship has moved into DMs instead. Instagram head Adam Mosseri said in 2025 that "there are way more photos and videos shared into DMs than into stories, and way more photos and videos are shared into stories than into the feed." Mark Zuckerberg told investors that "Reels are already reshared over 4.5 billion times a day," and that "another focus is making it easier for people to connect over content." That sharing matters: in a 196-person survey, Brody & Cullen (First Monday, 2023) found that sending memes for fun was linked to higher relationship satisfaction.

The trouble is what happens after the share. A friend sends a reel, someone reacts with a laugh, and the chat moves on. The group's history piles up in the chat, but nothing ever brings it back.

So we built Insta Games, an AI game master inside Instagram DMs that turns what a friend group has already shared into short games, then steps back so the group actually talks. This fits the Bringing People Closer Together with AI track because the AI is the game: it chooses what to bring back, who to ask about it, and when to stay quiet.

What it does

Picture five college friends, Maya, Dev, Sam, Ana and Kofi, whose group chat has gone quiet.

The game master suggests a game in the quiet chat. It's only an offer, and anyone can ignore it.
Anyone taps Chaos, and in about 15 seconds the AI writes three rounds from the group's shared history and their public posts.
Everyone plays whenever they open the chat, like turn-based iMessage games. No one needs to be online at the same time.
- Who Sent This? shows a real message or reel, like "i've never lost at mario kart i just let y'all win", and everyone guesses who sent it.
- Hot Take asks an open question, like "Which F1 team is the most overrated right now?" Everyone types an answer, and the AI gives the point to the most interesting one.
At each reveal, the game master turns to the person with the story ("Dev, you sent that one. Do you really let everyone win?"), then stays quiet unless the chat dies again.

In short, the game is just the excuse. The goal is a group chat that's talking again.

How we built it

Insta Games builds a memory of the group ahead of time, then uses it to run each game:

Memory → Every shared message, reel and photo becomes a sentence embedding (all-MiniLM-L6-v2), and HDBSCAN clusters them into moments like "the 3am fire alarm." On a test group with known answers, this placed 58 of 61 messages correctly, versus 20 of 84 for keyword matching.

Vision → Meta Muse describes each chat photo in one plain sentence so photos join the memory too, and blocks photos that should never appear in a game, like IDs or private screenshots.

Interests → Muse reads each person's posts and stories and lists what they're into, citing the exact posts. Private likes and saves can hint at a topic but are never quoted.

Planner → It picks round material based on how much the players share, favoring moments that only part of the group knows about, and spreads rounds across players.

Writer + Judge → Muse writes every question and picks the best Hot Take answer. Each output is validated and retried once with the reason it failed. The right answer always comes from the data, never the model.

Timing → A scikit-learn classifier trained on the timing of real Instagram conversations (timestamps only, never message text) predicts when a chat is about to go quiet, so the game master knows when to speak.

Game server → FastAPI runs each game on Supabase, which sends live updates to every phone and keeps answers hidden until the reveal. The app is an Instagram web clone that runs in any phone's browser.

Challenges we ran into

Helpful, not creepy: An AI reading your group chat can feel like surveillance. We still caught a round quoting something a friend had only privately liked, so we moved that rule out of the prompt and into code.

Fair rounds: Players were asked to guess their own posts, and photos with faces gave the answer away. Now the sender sits out their own round, and photos of people are never used for guessing.

Clear questions: Muse's first questions crammed slang and several people's posts into one sentence. A stronger model didn't help, but rewriting our instructions raised clarity from 4.4 to 4.9 out of 5.

Silent failures: Four of us shared one live database, and several bugs quietly switched games to placeholder rounds without any error. We built an "X-ray" page that replays a game next to every AI decision to catch them.

Accomplishments that we're proud of

An AI whose real job is judgment: which memory will get the group talking, which friend has the story, and when to say nothing. It uses a timing model trained without reading a single message, privacy rules enforced in code, and a full game running on real phones through one link.