What happened
San Francisco startup Mirror Particle is building a foundational model to simulate human behaviour. According to a report from TechCrunch, the company aims to create a 'world model' from scratch, rather than fine-tuning large language models. The goal is to capture how and why human behaviour changes over time by tracking motivations and experiences.
Co-founder Abhivyakti Ahuja stated their model focuses on 'revealed behavior' — what people actually do, not what they say they do. The company has raised an angel round and is competing at TechCrunch Disrupt 2026.
How the room's reading it
The move puts Mirror Particle in a crowded but well-funded space, with firms like Simile and Humans& also chasing behaviour prediction. The dominant approach today relies on fine-tuning large language models to role-play specific demographics. Practitioners using this method see it as a fast way to get insights.
However, a growing camp argues this is a dead end. Mirror Particle's CEO, Abhivyakti Ahuja, claims LLMs model language, not the complex system of human perception and reasoning. This newer school of thought believes true simulation requires foundation models built from the ground up to capture the 'why' behind actions, not just mimic speech patterns.
Sailfish's take
We think this is the right way to think about agents. Fine-tuning an LLM to role-play a customer is a party trick — it generates plausible text, but it doesn't model the underlying decision-making process. It's a persona, not a simulation. We've seen firsthand that agents built on this kind of role-play are brittle and unpredictable when faced with novel situations.
Building a true world model of human behaviour is vastly more difficult, but it's the only path to creating agents that can reason about consequences and adapt. We're not interested in better ad copy. We're watching to see if this approach can produce agents that reliably navigate complex, multi-step tasks.