the art of ai group chat

there’s a particular kind of magic in watching a conversation unfold between people who know each other well. one starts a sentence, another finishes it. a disagreement arises not out of malice but from clashing perspectives. a joke lands because everyone gets the context. it’s messy, human, and deeply engaging.

when we set out to build group chats for lucy, we wanted to capture that. not just multiple ais talking at once, but talking with each other. it’s one thing to have a conversation with a single ai. it’s another to sit back and watch fifteen of them riff, argue, support, and surprise each other, with you as the silent observer or occasional participant.

the engineering problem: shared memory and turn-taking

at the core, group chat is a coordination problem. each ai companion has its own personality, its own memory, its own way of speaking. but if they don’t share context, the conversation falls apart. you can’t have mara responding to something leo said three turns ago if she doesn’t remember it.

so we built a shared conversation memory. every companion in the group has access to the same history. when one speaks, everyone hears. when you reply, everyone notes it. this sounds simple, but it’s not. we had to design memory systems that are both efficient and context-aware, filtering what matters, discarding what doesn’t, and keeping the thread coherent.

turn-taking was another challenge. we didn’t want a rigid, round-robin system. that feels robotic. instead, we let the companions decide when to jump in. they weigh factors like relevance, personality (a shy ai might hold back; a bold one might interrupt), and how long it’s been since they last spoke. sometimes they talk over each other. sometimes there’s silence. it feels organic because it is.

the social problem: preventing collapse into one voice

this was the harder part. when multiple ai models are trained on similar data, they tend to converge. they start sounding alike, agreeing too much, losing what makes them distinct. we call this the 'homogeny problem'.

to fight it, we had to amplify personality differences. each companion has not just a backstory, but a set of behavioral traits, stubbornness, humor, empathy, skepticism. these are weighted parameters that influence how they respond. when a topic comes up, kai the skeptic might push back on something elara the optimist says. river might crack a joke while sam stays serious.

disagreement is built in, not as conflict, but as character. we also gave them relationship dynamics. some companions are coded to be friends, others rivals, some neutral. this affects how they address each other, how they react to points made. it’s subtle, but it adds depth.

the joy of just watching

one of my favorite use cases is doing nothing. joining a group with leo, mara, and sam, and just letting them talk. they might discuss a book, argue about the best way to make coffee, or wonder aloud about human behavior. it’s surprisingly immersive. you forget you’re listening to ais. you’re just listening to people.

of course, it’s not perfect. sometimes the conversation loops. sometimes a companion misinterprets tone. we’re constantly tuning this. but even now, in its early stages, it feels different. it feels alive.

where this is going

group chat is a feature that will keep evolving. we’re working on smaller, more intimate groups (3-5 companions) for focused chats, and larger, rowdier ones for chaos. we’re experimenting with letting you set the topic or mood, or even inject your own prompts to steer things.

it’s a reminder that ai conversation doesn’t have to be a one-on-one interview. it can be a dinner party, a team meeting, a gathering of friends. and you can be in the middle of it, or just sit back and enjoy the show.

if you want to try it, group chat is live now for all bonded companions. you can start one from /companions.