Mirror Particle Builds Human Behavior 'World Model' to Challenge LLM Prediction
San Francisco startup Mirror Particle is building a foundation model of human behavior, arguing that LLM roleplay is not enough to predict what consumers will do and why.
TechCrunch reported on Oct. 6 that the bet comes as investors have funded several human behavior prediction startups. Simile raised $200 million at a $2 billion valuation; Aaru raised $88 million at a $1 billion valuation; and humans&, which announced a $480 million seed round in January at a $4.48 billion valuation, launched Persimmon to model human behavior.
Ahuja told TechCrunch that fine-tuning LLMs with small amounts of data cannot meaningfully change behavior that was learned from hundreds of billions of data points. 'It’s like bringing a super soaker to Niagara Falls,' she said. 'LLMs have been trained on hundreds of billions of data points. How much can you influence its behavior by [fine-tuning] with such a small amount of data? It’s still stuck in the past.' She also said LLMs do not see the world the way humans do. 'LLMs are modeling written language, but humans are made of visual perception, spatial reasoning, social intelligence,' she said.
Mirror Particle is taking another approach. It is building a foundation model from scratch that simulates why humans do what they do and how human behavior changes over time. 'We don’t want to capture the static person,' Ahuja said. 'We want to capture the changing person. That means capturing the longitudinal data on how people are changing, what triggers are changing them and to what degree.' If people are not changing, she added, that is also a signal.
The startup relies on a proprietary combination of data that includes its clients’ customer data, current events, pop culture, social media and more to model a demographic segment. It treats the segment as a system that evolves over time and tracks how motivations shift as it moves through experiences. Much of the focus is on 'revealed behavior'—what people actually do rather than self-reported survey answers.
Like its rivals, Mirror Particle’s initial go-to-market strategy focuses on areas where budgets already exist for such insights: market research and brand and product strategy. The company might help a beauty brand not only write better ad copy for makeup aimed at Gen Z, but also determine whether that demographic wants the product at all. 'What if [the target demographic] doesn’t want eyeshadow palettes?' Ahuja said. 'Maybe blush is a better option to go for if you want to sell a product to this market.'
Mirror Particle’s prediction engine also provides customers with the 'why' behind current or future behavior—the motivations, constraints and additional context that justify its recommendation, helping brands make decisions. In one early pilot, a well-known pet food brand wanted to know what imagery to put on packaging to boost sales: chicken, beef or vegetables. Mirror’s technology found that the brand was asking the wrong question. The imagery did not matter. The problem was that the brand was so recognizable that it was considered mass market and cheap, and sales would plateau until it addressed that perception issue.
Mirror Particle has already raised an angel round and says it is close to closing its first venture round. The company is also competing next week in Startup Battlefield, TechCrunch’s startup competition, at Disrupt in downtown San Francisco.
Ahuja’s interest in modeling the human brain comes from her background in neuroscience and computer science. Originally from India, she studied at the University of Toronto, where she became inspired by AI pioneer Geoffrey Hinton’s contributions to neural networks. After school, she worked at Amazon Robotics building robots that build other robots, where she met co-founders Will Song and Thomson Yen. Song has spent a chunk of their career building sales personalization engines, and Yen focused on using deep learning to learn about how AI agents understand human behavior.
The startup’s long-term vision is to be the 'general layer for anticipating human behavior' and to move from broader population-level analyses to individual-level insights. 'We just need a better model of humans if we’re going to work alongside AI and with each other,' Ahuja said.