I'm a ML eng with ~10 yrs of experience but previously I was a psychologist (unusual transition I know ha).

Reading over the research, I kept thinking LLM personas were underutilized or not fully understood. I finally found the Qwen paper on 1 Billion personas to capture more complete & diverse stores of information within a FM. The paper intention was for synthetic data generation for LLM training data. This paper was very cool and triggered my hypothesis that distinct personas collaborating with each other could give humans a diverse set of actions or validate possible pathways before taking first real step.

It seems that an LLM's ability to "brainstorm" for humans via generation seems like a ground truth. Only to improve with scaling laws.

I am hearing and see prototypes that have a feature that utilizes what I'm calling "synthetic conversations".

Ex 1: Everyone uploads your profile (e.g., Snap, Yelp, LinkedIn), your representative meets others, then simulate synthetic conversations before mtgs/party to find best match or talking points in real life.

Ex 2: Marketing assistance can validate if your message will resonate with intended audience via synthetic conversations.

I might be alone but if not I'd love to hear from you!

Qwen paper: https://arxiv.org/html/2406.20094v1