The landscape for startups aiming to forecast human actions is intensifying, with competitors like Simile and humans& securing billions in valuation. Amidst this surge, Mirror Particle, a two-year-old San Francisco-based company, argues that the current industry standard—relying on large language models (LLMs) fine-tuned for roleplay—is fundamentally flawed.
Abhivyakti Ahuja, co-founder and CEO of Mirror, compares the limitations of LLMs in this context to “bringing a super soaker to Niagara Falls.” She contends that because these models are trained on vast datasets, attempting to steer their behavior through limited fine-tuning leaves them anchored in past patterns rather than reflecting real-time human dynamics. Ahuja emphasizes that LLMs process written language, whereas humans operate through visual perception, spatial reasoning, and social intelligence, meaning current tools often highlight what people overlook rather than what drives their choices.
Instead, Mirror Particle is developing a foundational “world model” constructed from the ground up. The technology tracks longitudinal data to understand not just who consumers are, but how they evolve. By monitoring shifts in motivations and triggers over time, the system captures the dynamic nature of human behavior rather than treating individuals as static entities.
To build this model, the startup utilizes a proprietary mix of client customer data, current events, pop culture trends, and social media signals. A key differentiator is its focus on “revealed behavior”—actions people actually take—rather than relying on self-reported survey answers, which can be biased or inaccurate.
Currently targeting market research and brand strategy sectors, Mirror’s engine provides the reasoning behind its predictions. In a recent pilot, a major pet food brand sought advice on packaging imagery, debating whether to feature chicken, beef, or vegetables. Mirror’s analysis revealed the question itself was misguided; the brand’s strong market presence made it seem cheap and mass-market, a perception issue that no amount of imagery tweaking could resolve.
Ahuja, who holds a background in neuroscience and computer science from the University of Toronto, draws parallels between her model’s evolution and infant development, progressing from vision to language and eventually social intelligence. Her co-founders, Will Song and Thomson Yen, bring experience in sales personalization and deep learning, respectively, stemming from their time working together at Amazon Robotics.
Mirror Particle has secured an angel investment and is nearing the close of its first venture round. The company is set to compete in TechCrunch’s Startup Battlefield on October 15, where VC judges will determine the winner.
The pet food example is wild. Who knew packaging imagery was masking a premium perception issue? Insightful pilot data.
Skeptical about claiming LLMs are fundamentally flawed for this. Fine-tuning has come a long way. Let’s see the actual results.
Revealed behavior over self-reported surveys makes total sense. People lie, especially in questionnaires. This seems like a no-brainer.
Surprised they are targeting market research first. I thought this tech was meant for broader behavioral prediction applications.
Infant development parallels for AI? That is a fascinating and bold analogy. Hope it holds up under scrutiny.