AI
Mirror Particle Develops Foundation Model to Predict Human Behavior Without LLMs
A San Francisco startup is building a dedicated world model to forecast consumer actions and motivations, positioning itself against competitors relying on large language model adaptation.
TechCrunch — AI
Predicting human behavior has become a competitive focus for AI startups, with several recent fundraising rounds signaling strong investor appetite. Simile secured $200 million at a $2 billion valuation, Aaru raised $88 million at a $1 billion valuation, and humans& announced a $480 million seed round in January at a $4.48 billion valuation. The two-year-old firm Mirror Particle, however, is taking a distinctly different path to solve this problem.
The LLM Approach Falls Short#
Most firms in the space adapt large language models through prompting or fine-tuning to simulate specific consumer segments. Mirror Particle's CEO and co-founder Abhivyakti Ahuja argues this strategy has fundamental limits. She contends that LLMs trained on vast amounts of text data can't be meaningfully shaped by small amounts of behavioral data—a disconnect she frames as trying to redirect an enormous river with a water gun.
Beyond raw data volume, Ahuja questions whether language models capture how humans actually perceive and reason. "LLMs are modeling written language, but humans are made of visual perception, spatial reasoning, social intelligence." That gap means insights derived from LLMs often reflect patterns humans don't consciously notice, which misses the mark when forecasting actual consumer choices.
Building a Purpose-Built Foundation Model#
Mirror Particle is constructing a proprietary foundation model—what Ahuja describes as a world model—developed independently to understand why humans behave as they do and how those behaviors evolve. Rather than capturing static preferences, the system tracks longitudinal patterns: what triggers behavioral shifts, how intensely those shifts occur, and whether consistency itself signals something meaningful.
The company blends client customer data with current events, pop culture, and social media signals to model demographic segments as dynamic systems. A key focus is "revealed behavior"—documenting what people genuinely do rather than relying on self-reported survey responses.
