[TylerLives]

I think LLMs are not the right tool for this. They will never have a "life of their own" because of how they function. Imo, a better approach (although much more difficult) would be to create some sort of a game (simulated world), use RL to create intelligent agents in that game, and then perhaps have LLMs translate agents' actions into words. Different agents could have drastically different personalities based on their learning algorithms and reward functions. I'm surprised something like this doesn't exist already.

[sandwichukulele]

The LLMs of today will only act as a sort of primitive neocortex, there is much else needed for life of their own which is why I said models of my own and not specifically large language models. I think of a virtual friend as a akin to dog, we need to simulate hormonal profiles and regulation feedback loop so that we can have a much richer experience like emulating the need to "sleep" with sleepy hormones that go up and down according to lighting, which then can be feed back to the system prompt to provide context as a very simple example of ways to mimic a "life of their own." One way is like you suggest, a "game" as an abstract representation of the world but the better approach is to really via biochemistry since it is our emotions that give things a lifelike feeling to it. There are more (much more) details required but that would require a public live demo for which I am not ready for