Summary of Key Points
User simulation involves using thousands of AI agents (similar to “digital humans”) to mimic the decision-making behavior of human groups, creating a sandbox that can “predict the future.” This technology can be used to forecast U.S. elections, optimize e-commerce strategies, and test AI products. Top Silicon Valley VCs have invested hundreds of millions of dollars in supporting related startups. It’s not a single technology but rather a five-layered framework ranging from data generation to group prediction. Currently, it is most rapidly being applied in the e-commerce sector, helping businesses save costs and increase revenue. While it cannot replace traditional A/B testing, it can serve as a “pre-flight check” to eliminate ineffective options. In the future, it may even enable more realistic digital twins—perhaps allowing deceased family members to “come back to life.”
What Exactly is User Simulation? An Explanation of the Five-Layer Framework
User simulation isn’t some magical way of “guessing the future”; rather, it’s a process that involves building a “digital sandbox” in five steps:
- Layer 1: Data Simulation – Fake materials used to train AI. For example, simulating users’ emails, documents, and shopping records to address the issue of insufficient real data (Google uses this to train its Gmail assistant).
- Layer 2: Interaction Simulation – Testing whether the AI will act inappropriately. Different personalities are assigned to the AI agents (e.g., irritable or polite users), and they are interacted with to see if they say anything offensive (Google uses this to test its smart speakers).
- Layer 3: Journey Simulation – Simulating the complete user behavior path. For instance, if you browse Taobao, from seeing an ad to clicking on it, hesitating, adding it to the cart, and then forgetting to pay, all these actions can be simulated by AI.
- Layer 4: In-depth Individual Simulation – Essentially “cloning” a person. Through two hours of in-depth interviews and questionnaires, AI is trained to not only mimic the person’s appearance but also their thoughts and behavior (for example, Simile from Stanford can address the issue of people saying they support the Democrats but actually voting for Trump).
- Layer 5: Group Prediction – Simulating the decisions of a group of people. By having thousands of agents interact with each other, it’s possible to predict election results and market trends (for example, Aaru can quickly replicate survey results with a correlation of 0.9).
Why is Silicon Valley Investing So Much? The Appeal of “Predicting the Future”
By 2025, user simulation has become one of the hottest fields in Silicon Valley. Companies like Simile and Aaru have raised hundreds of millions of dollars for this reason:
- Simile (founded at Stanford) – Became popular thanks to its “in-depth individual simulation” approach. They conducted experiments with virtual towns where dozens of agents interacted socially on their own; they also solved the problem of people’s inconsistency between words and actions (e.g., users saying they hate a product but still buying it).
- Aaru (macro perspective) – Collaborates with EY and can complete what would otherwise take months of research in just one day. However, it has limitations: it can only predict things that have already happened; predicting low-probability events like the mayor of Sacramento’s election is more challenging.
Silicon Valley’s investment is a bet on the ultimate ability to “predict the future.” Whoever can make AI simulate humans more accurately will gain a competitive advantage in business, politics, and other fields.
E-commerce Takes the Lead: Using User Simulation to Help Businesses Make Money
Another startup, UserApproved, chose the e-commerce sector because shopping behavior is relatively easy to simulate:
- How it works: They integrate real business data (e.g., Google Analytics, Shopify) and then use AI agents to simulate user interactions in a virtual environment (e.g., simulating users seeing a promotion, clicking on a video, and forgetting to pay).
- What the results are: They provide daily recommendations to businesses, such as “Changing three aspects today could increase revenue by $50,000” (e.g., adjusting promotional copy or optimizing video placement).
- Why it’s effective: Instead of just presenting data, they offer actionable advice at a senior management level, helping businesses avoid the cost and time wasted on unnecessary A/B testing.
Can It Replace A/B Testing? It’s a “Pre-flight Check,” Not the Endgame
Many people ask if user simulation can replace A/B testing. The answer is no, but it can make it more efficient:
- Problems with A/B testing: Large models may exaggerate their evaluations when acting as users (e.g., they might say Option B is 10 times better than Option A, but real users may not agree). Moreover, large models can be biased, and having a large sample doesn’t always matter.
- Value of user simulation: It serves as a “pre-flight check.” For example, if a business plans to test ten options, simulation can eliminate seven that are clearly ineffective, saving time, traffic, and costs before conducting actual A/B tests.
Technical Challenges and Future Possibilities: Getting AI to “Make Mistakes” on Purpose and Creating Digital Twins
There are many technical challenges in user simulation, but the potential for future developments is vast:
- Technical difficulties: Making AI “make mistakes” intentionally (e.g., making GPT less intelligent so it doesn’t always solve problems too well) and ensuring consistent behavior (e.g., preventing users from having sudden changes in behavior during simulations).
- Future prospects: This technology could be used for:
- Business decision-making rehearsals (e.g., simulating user reactions before launching a new product or preparing for investor meetings).
- Digital Twins 2.0: Current digital twins only look similar; in the future, they could also think like humans (e.g., allowing deceased family members to communicate and make decisions).
User simulation isn’t a crystal ball, but it’s becoming the “infrastructure” for decision-making—helping us find direction in uncertainty. This could be the third wave of AI advancements (following information retrieval and content generation).
(The entire text is explained in plain language to ensure non-finance professionals can easily understand it.)