Il researcher Izzy Miller he cloned the own group chat that shares with friendsi, downloading over 500 thousand messages e by instructing l’AI (LLaMA, the Meta algorithm and not ChatGPT and Bing) to recreate group conversations. The result reveals a lot – but perhaps more about group chats as a social entitythan on AI.
Clone friends group chat with AI, with amazing results
Group chats between friends, on WhatsApp, Telegram, Discord or any other digital place, they are a social experiment in themselves. They often become an opportunity to organize pizza parties with friends, but they can also become a square where share memes, videos and other funny messages. Differently from what happens when you meet in person: the most shy people can become the ones who share the most updates about their life.
Izzy Miller tells The Verge that “my group chat is a lifeline and a comfort, and also a point of connection. I thought it would be fun and even sinister to replace her” using AI.
Miller then downloaded hundreds of messages from the chat he has with his friends for over seven years, since they met in college. Six people who shared a lot, condensed in the messages they sent each other. Miller explains that it took only a few weekends and just a hundred dollars to access the AI that would simulate his group chat – and his friends.
Amazing results – without too much effort
The results were not long in coming. And they were amazing. “I was really surprised by the degree to which the model has inherently learned things about who we were, not just the way we talk. He knows things about who are we dating, where did we go to schoolthe name of our house where we lived, etc.”
Miller, who has been working on large language model (LLM) technology for some time, knew his way around it. Downloading over 500 thousand messages from iMessage, trained LLaMA, Meta’s artificial intelligence – which is on the level of GPT-3, according to the expert. The code for this algorithm was leaked online last month, which made it easier for Miller to use.
Miller explains that once this type of research (simulating six different personalities) would have required months of work for an entire team of data scientists. He could do it in his spare time, for a hundred dollars.
Once the model was built, Miller created an iMessage chat with all of his friends and digital clones. This way, they were able to have a group chat with the AI that was simulating them. Miller explains that some of the speeches were so realistic that she had to search the chat to see it the AI wasn’t just copying old threads to them. But it was all original.
“There’s something delightful about capturing your friends’ voices perfectly. It’s not exactly nostalgia, because these conversations never happened. But it’s a familiar sense of joy… This has given me and my friends more hours of pure enjoyment than I ever imagined,” comments Miller.
AI in friends’ group chats: one window into the past and one into the future
Miller, however, explains that there are limits. From time to time, the difference in personality between the six members became more subtle. Furthermore, the‘AI doesn’t have a good sense of chronology: it doesn’t distinguish between the past and the present. So he would sometimes talk about old college flames like they were friends’ current girlfriends, for example.
Also, the AI doesn’t rate friends’ group chat overall, but overall message volume. Hence, as chat was more active in my college daysacts like it’s 2017. “If I ask him how old he is, he says 21 or 22,” Miller explains, “If I ask him where you are, he says he’s in the college cafeteria.”
This type of project demonstrates the potential for the AI of tomorrow. It might help fight social isolation, or push even further towards loneliness who struggles to socialize? One day, he could recreate the conversations with a loved one that no longer exists: is this a therapeutic asset or a risk not to be run? The doubts are many.
But beyond thinking ahead, Miller explains that he has also reopened a window on the past. Now he’s planning a trip to Arizona to meet up with his college friends after a long time.