Hi - I answer from the OpenSmartRoute documentation: routing, the API, plans and quotas, self-hosting. Ask away, or open a support ticket if you need a person.
Grounded in the docs - follow a source before acting on it.
Mirror Particle builds a world model of human behavior instead of LLMs - OpenSmartRoute
Mirror Particle builds a world model of human behavior instead of LLMs
Mirror Particle raises capital to create an AI engine that simulates changing human motivations. The company rejects large language models in favor of a foundation model trained on longitudinal data.
Key points
Mirror Particle raised an angel round and is closing its first venture round.
The startup competes in Startup Battlefield 200 at TechCrunch Disrupt 2026.
Abhivyakti Ahuja calls relying on LLMs for human behavior prediction broken.
Mirror Particle focuses on revealed behavior rather than self-reported survey answers.
Why it matters: Engineers and managers need a better model of humans to work alongside AI and each other.
By OpenSmartRoute editorial · written through the router by writer-small
From TechCrunch AI - “Mirror Particle is building a ‘world model’ of human behavior”
Mirror Particle is building a new kind of AI engine to predict human behavior. This tool simulates changing motivations instead of using large language models. The company recently raised an angel round and is close to its first venture round. They plan to compete in Startup Battlefield 200 next week. This event happens at TechCrunch Disrupt 2026 in San Francisco.
The competition runs from October 13 to October 15, 2026. Judges are venture capitalists who decide the winner on Thursday afternoon. The winner will be announced during the main conference day. Many startups are trying to predict how humans act today. Simile raised $200 million at a $2 billion valuation recently. Aaru raised $88 million at a $1 billion valuation last year. Humans& launched Persimmon after a massive $480 million seed round. That round happened in January at a $4.48 billion valuation.
These companies all promise to predict human behavior. They are having a moment right now in the startup world. Mirror Particle believes their approach is fundamentally different. They think relying on large language models is broken. Abhivyakti Ahuja is the co-founder and CEO of Mirror Particle. She says LLMs have been trained on hundreds of billions of data points. Fine-tuning them with small amounts of data does not change their behavior much. The system remains stuck in the past according to her view.
Large language models (LLMs) are AI systems trained on vast text datasets. Engineers often prompt or fine-tune these models to role-play as specific demographics. This method tries to simulate how a group of people might react. Mirror Particle argues this approach misses how humans actually perceive the world. Humans use visual perception, spatial reasoning, and social intelligence daily. LLMs model written language instead of these human senses.
Getting insights from LLMs means seeing what humans do not notice. Abhivyakti Ahuja says this is beside the point for behavior prediction. She wants to capture a changing person rather than a static one. This means tracking longitudinal data on how people change over time. The system must identify triggers that cause these changes. It also needs to measure the degree of those changes.
If a person does not change, that signal matters too. Mirror Particle is building a foundation model from scratch. A foundation model is a large AI system trained on diverse data types. This model simulates why humans do what they do. It tracks how human behavior evolves through different experiences. The company has already raised an angel round of funding. They are close to closing their first venture round soon.
Mirror Particle competes in Startup Battlefield 200 next week. This is TechCrunch's renowned startup competition for emerging technologies. The event takes place at TechCrunch Disrupt 2026 in San Francisco. The competition focuses on companies building AI engines for consumer insights. Mirror Particle relies on a proprietary combination of data sources. These include client customer data, current events, and pop culture trends. They also use social media data to train their system.
The company thinks of its model as a system that evolves over time. It tracks how motivations shift as people move through experiences. Much of the focus is on revealed behavior. Revealed behavior is what people actually do rather than self-reported answers. Surveys often ask people what they will do in the future. People frequently say one thing but do another when faced with choices.
Mirror Particle's go-to-market strategy focuses on existing budgets for insights. Market research and brand strategy teams have money to spend here. They want to help beauty brands determine if a demographic wants a product. For example, the AI might suggest blush over eyeshadow palettes for Gen Z. It could also write better ad copy that appeals to that group.
The prediction engine provides customers with the "why" behind behavior. This includes motivations, constraints, and additional context for recommendations. Brands can make smarter decisions when they understand these factors. A well-known pet food brand wanted to know what imagery to use on packaging. They asked if chicken, beef, or vegetable images would boost sales.
Mirror's technology found the brand was asking the wrong question. The imagery itself did not matter to the decision-making process. The problem was that the brand was seen as mass market and cheap. Sales would plateau until they addressed this perception issue. This case study shows how the model identifies deeper problems than surface-level features.
Abhivyakti Ahuja compares their model evolution to how a baby learns. Babies move from vision to language to body awareness to social intelligence. Her background includes studying neuroscience and computer science at the University of Toronto. She was inspired by Geoffrey Hinton's contributions to neural networks during her studies.
After school, she worked at Amazon Robotics building robots that build other robots. There she met co-founders Will Song and Thomson Yen. Will Song spent his career building sales personalization engines. Thomson Yen focused on using deep learning to learn about AI agents understanding human behavior. Their long-term vision is to be the general layer for anticipating human behavior. They aim 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. This quote highlights the core mission of the startup. The team wants to create a system that understands human complexity better than current tools. They are challenging the status quo for human behavior prediction today.
Why it matters
Engineers need accurate data to build reliable behavior prediction systems. Managers must choose tools that offer real insights over generic text generation. Mirror Particle offers a foundation model trained on longitudinal data instead of LLMs. This shift could improve the quality of consumer insight decisions for brands.
What to do
Check if your current AI tool uses revealed behavior or self-reported surveys. Compare the cost and speed of foundation models against large language models. Look for companies that provide context on motivations behind predictions. Visit Startup Battlefield 200 in San Francisco to see Mirror Particle live demo. Talk to vendors about their data sources and model training methods.
Startup Battlefield 200 competition details for Mirror Particle
Mirror Particle is competing next week in Startup Battlefield 200. This event takes place at TechCrunch Disrupt 2026 in San Francisco. The dates are October 13, 2026, through October 15, 2026. The competition happens on the afternoon of Thursday, October 15. A group of venture capital judges decides the winner. These judges represent various investment firms. They evaluate startups based on their potential and innovation.
Mirror Particle is one of many innovative startups vetted by TechCrunch's editorial team. The company hopes to win recognition for its unique approach. Abhivyakti Ahuja, the CEO, wants to show how world models work. She will demonstrate predictions about human behavior changes. The demo will likely focus on longitudinal data tracking.
The event is meant to be shared with colleagues and peers. Attendees can get a pass at 50% off if they bring someone. This discount applies to partners or people in the startup ecosystem. TechCrunch wants attendees to make connections during the conference. Building momentum is a key goal for the organizers.
The shift from large language models to foundation models
The current status quo relies heavily on large language models (LLMs). These models are prompted or fine-tuned to role-play as a target demographic. Users ask questions about how people act in specific situations. The AI generates text based on patterns it learned from training data.
Mirror Particle thinks this approach is fundamentally broken. Abhivyakti Ahuja compares it to bringing a super soaker to Niagara Falls. She argues LLMs cannot handle the scale of human behavior prediction. These models are trained on hundreds of billions of data points. Fine-tuning them with small amounts of data changes little.
Ahuja believes LLMs model written language instead of visual perception. Humans rely on spatial reasoning and social intelligence too. Relying only on text misses critical aspects of reality. Insights based on what humans do not notice are useless for prediction.
Mirror Particle builds a foundation model from scratch. This is also called a world model in their documentation. The system simulates why humans do what they do. It tracks how behavior changes over time continuously. They want to capture the changing person, not the static one.
This shift represents a move away from text-based generation. Engineers must understand the difference between prompt engineering and simulation. Foundation models learn from longitudinal data on human evolution. They track triggers that cause behavioral shifts.
How Mirror Particle defines its world model approach
Mirror Particle focuses on capturing how people change over time. They collect longitudinal data to see what triggers changes. They measure the degree of these changes precisely. If people are not changing, that is also a signal. Stability in behavior provides important context for brands.
The system models a demographic segment as something that evolves. It tracks how motivations shift through different experiences. This view treats human behavior as dynamic rather than fixed. The model updates as new data points arrive over time.
Their goal is to reveal the reasons behind actions. They provide the why, not just the what. Brands need to understand constraints and motivations too. Context justifies every recommendation the system makes.
Ahuja notes that babies learn through vision first. Then they move to language and body awareness. Finally, they develop social intelligence. Her model follows this developmental sequence in humans. It mimics how a baby learns about the world around them.
This approach replaces static surveys with dynamic observation. Surveys ask people what they will do. The model observes what people actually do. Revealed behavior is more accurate than self-reported answers. People often lie or forget their true intentions.
Data sources including customer data and social media trends
Mirror Particle relies on a proprietary combination of data sources. It includes its clients' customer data directly from businesses. Current events feed into the model's understanding of context. Pop culture trends also influence human behavior patterns. Social media activity provides real-time behavioral signals.
The startup thinks of this as a system that evolves over time. It tracks how motivations shift as people move through experiences. This continuous evolution is key to accurate prediction. Static datasets cannot capture rapid cultural shifts.
Much of the focus is on revealed behavior rather than surveys. What people do speaks louder than what they say. Self-reported data often contradicts actual actions in practice. Brands need to know real purchase behaviors, not stated ones.
The model integrates these diverse data streams into a single view. It connects individual actions to broader societal trends. A change in social media sentiment might trigger a sales shift. The system links micro-behaviors to macro-trends automatically.
This multi-source approach replaces single-data-point analysis. Engineers must ensure data sources are relevant and up-to-date. Outdated information leads to flawed predictions about the future. Fresh data keeps the model aligned with reality.
Case study of a pet food brand packaging strategy
A well-known pet food brand wanted to know what imagery to use on packaging. They asked if chicken, beef, or vegetable images would boost sales. The team ran tests to see which ingredient photo worked best. They expected a clear winner among the three options.
Mirror's technology found the brand was asking the wrong question. The imagery itself did not matter to the decision-making process. Sales data showed no significant difference between the options.
The problem was that the brand was seen as mass market and cheap. Customers associated the product with low quality regardless of packaging. Sales would plateau until they addressed this perception issue. Changing the ingredient photo would not fix the underlying problem.
This case study shows how the model identifies deeper problems than surface-level features. The AI detected a brand positioning flaw that humans missed. It suggested shifting focus from ingredients to overall market perception. This insight could unlock new growth potential for the company.
The team realized they needed to reposition the entire product line. A simple packaging change was not enough to move the needle. They had to address the core identity of the brand first.
Why it matters for engineering teams building behavior prediction systems
Engineers need accurate data to build reliable behavior prediction systems. Managers must choose tools that offer real insights over generic text generation. Mirror Particle offers a foundation model trained on longitudinal data instead of LLMs. This shift could improve the quality of consumer insight decisions for brands.
Current tools often hallucinate plausible but incorrect reasons for behavior. A world model simulates actual human cognitive processes better. It accounts for visual and social factors that text models ignore. Engineers should test their models against real-world behavioral outcomes.
Managers care about cost, speed, and accuracy when buying AI tools. Foundation models may require more compute than fine-tuned LLMs initially. However, they deliver higher fidelity predictions over time. The long-term value outweighs the initial setup cost in many cases.
Safety is another concern when predicting human behavior. Misleading predictions can damage brand reputation or waste marketing budgets. Teams must validate model outputs against ground truth data regularly. Continuous evaluation prevents drift from actual human patterns.
How OpenSmartRoute helps
A team routing requests through OpenSmartRoute gains speed by avoiding slow LLMs for human behavior tasks. The router scores every candidate on quality, cost, and safety before sending a request. It learns from outcomes so models that answer well get more traffic automatically.
The platform keeps a models catalogue with prices and public rankings built from real traffic. A savings ledger shows what each routed request cost next to the most expensive option. This helps managers decide when to use specialized engines instead of general large language models.
Input guards spot prompt injection and personal data before a request leaves the system. Teams can set hard rules so sensitive data stays on an on-premises model always. OpenSmartRoute works with any OpenAI-compatible provider or open-weight models served locally.
What to do when evaluating new AI tools for consumer insights
Check if your current AI tool uses revealed behavior or self-reported surveys. Look for companies that prioritize actual actions over stated intentions. Compare the cost and speed of foundation models against large language models. Consider the total cost of ownership including data collection efforts.
Look for companies that provide context on motivations behind predictions. Good tools explain the why, not just the what. Ask vendors about their data sources and model training methods. Transparency helps you judge the reliability of their claims.
Visit Startup Battlefield 2026 in San Francisco to see Mirror Particle live demo. Attend TechCrunch Disrupt on October 13 through October 15. Talk to vendors about their approach to longitudinal data tracking. Ask how they handle changes in human behavior over time.
Evaluate your own predictions against actual sales or engagement metrics. Measure the accuracy of your insights over a quarter or more. If your tool fails to predict trends, it is not working well enough. Iterate on your model using fresh data and new feedback loops.
