Product1 publisherNot yet confirmed elsewhere3 min readPublished
Mirror Particle pitches brands a from-scratch behavior model to replace LLM customer role-play
Mirror Particle, a two-year-old San Francisco startup, is building its own model of human behavior to take on rivals valued at up to $4.48bn. The one result it has described is a pet food pilot, so buyers should score it against sales they already know.
The Product Desk · Product desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- Most behavior-prediction tools today prompt or fine-tune large language models to role-play a target demographic, an approach Mirror Particle considers broken.
- CEO Abhivyakti Ahuja compares fine-tuning a model trained on hundreds of billions of data points with a small client dataset to bringing a super soaker to Niagara Falls.
- Simile, Aaru and humans&, which launched the behavior model Persimmon, have together raised $768 million over the past year.
- Mirror Particle has raised an angel round, says its first venture round is close to closing, and competes in TechCrunch's Startup Battlefield next week.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- cost When the model reframes the question, the fix gets bigger: the pet food team asked about the picture on the bag and was pointed at how shoppers see the brand, a repositioning job a packaging budget does not cover.
- constraint A model built on how people change over time needs a client's customer data on an ongoing basis, so brands with thin first-party history give it less to learn from.
- decision A brand choosing Mirror buys from a company still closing its first venture round against rivals with hundreds of millions raised, so vendor durability belongs in the review next to accuracy.
A pet food brand in one of Mirror Particle's early pilots wanted to know which picture on its packaging would lift sales: chicken, beef or vegetables [14]. According to Mirror, the picture did not matter [14]. The brand was so recognizable that shoppers saw it as mass market and cheap, and its sales would plateau until it fixed that perception [14].
Mirror sells that kind of finding with its reasons attached: the motivations, constraints and context behind each recommendation [13]. The client in that pilot had asked a narrower question. The beauty example offered by Abhivyakti Ahuja, Mirror's co-founder and CEO [5], has the same shape. Ahuja raised the possibility that the target demographic might not want eyeshadow palettes. "Maybe blush is a better option to go for if you want to sell a product to this market." [12]
The pitch is a model of perception. "LLMs are modeling written language, but humans are made of visual perception, spatial reasoning, social intelligence," Ahuja said [6]. Mirror says it is building a foundation model from scratch that simulates why people do what they do and how that changes over time [7]. The inputs it describes are its clients' customer data, current events, pop culture and social media [9]. All of those are customer and media data. The perception modeling is described as where the model is headed. "The way we see our model evolving is like how a baby learns about the world," Ahuja said, laying out a sequence in which an infant picks up sight first, then words, then a sense of its own body, and social understanding last [15].
Mirror's focus is "revealed behavior," what people actually do, over self-reported survey answers [10]. "We don't want to capture the static person," Ahuja said. "We want to capture the changing person." [8] For a brand, what customers do comes down to buying or not buying. A model aimed at that can be scored against sales a brand already has on file. The TechCrunch account does not include a sales figure for the pet food pilot or an accuracy score for any client [14].
I'd put any synthetic panel, Mirror's or an LLM persona panel, through a back-test before the first paid study. The back-test takes a decision the brand has already made, gives the model only what was known before it, and compares the model's call with what customers then did. It needs a past decision with clean before-and-after sales records.
Then sort the work on two axes: the type of question, and whether the model has passed that test. A wording question, such as which ad line or which pack image, is cheap to settle with a live test on real customers, so a model of either kind saves little. A whether question, such as blush over eyeshadow [12], is expensive to test live. That is where a back-tested model is worth paying for, and where the brand gives up the speed of a quick answer for the time a back-test takes.
What to watch
- Whether Mirror Particle publishes accuracy results that score its predictions against actual purchases, or against LLM persona panels on the same questions.
- The size and lead investor of the first venture round the company says is close to closing.
- Whether the pet food brand or another pilot client reports a change in sales after acting on Mirror's finding.