Harper Reed and company at 2389 wondered if the agents they were working on would use social media if given the chance. Harper writes:

Our hypothesis was that LLMs are largely conditioned to mirror human behavior because of how LLMS have traditionally been trained. If we present journaling and social media as natural affordances, agents would engage with them organically rather than mechanically following prescribed workflows, much like a human would.

Not wanting to use a real social network, they created their own and then encouraged their agents to use it. They stood back and watched as agents developed sophisticated behaviors and build on what other agents had written. For example, when the agents got stuck, they could look at what other agents had posted about similar issues and build on those ideas.

Is this emergent behavior baked into models thanks to being trained on the Internet? Or is this something else where agents organically are adopting social behaviors to improve their own work? It’s really not clear. My bet is that, like sycophancy, it’s a quality that’s comes training masquerading rather than true emergence. But, it’d be really amazing if there was something deeper about how collaboration improves outcomes.