Hello! I'm your financial news analysis partner. Today, we're not talking about cold, hard-coded algorithms, but about a field that's heating up rapidly in Silicon Valley—and it has a bit of a cyberpunk vibe to it: AI Simulation.
In simple terms, it's about using AI to create a “parallel universe” on computers, placing the “digital counterparts” of hundreds of millions, even billions of people within it. These digital beings live, argue, shop, and vote just like real people. Then, from a “god-like” perspective, we observe how this world functions to predict what might happen in the real world.
It sounds like something out of sci-fi movies like *The Matrix* or *Westworld*, but it's real. Experts like Fei-Fei Li and Karpathy are investing heavily in this field, and companies valued at billions of dollars have already emerged.
Below, I'll break down this complex topic into five key parts to help you understand what this business is all about, why it's so popular, and what challenges it faces.
---
Part 1: From “Pixel Town” to “Digital Avatars”: How Does AI Learn to Behave Like Humans?
To understand AI simulation, we need to look at its evolutionary history. Initially, people thought AI could only chat; now, it can even act in a way that feels almost human.
1. The Beginning: Stanford’s “Pixel Town” (2023)
The first significant experiment came from Stanford University. They created a pixel-based town with 25 AI-driven virtual residents. These residents had no pre-written scripts, and the AI decided their actions based on their personalities and memories.
- The Result Was Amazing: When one AI wanted to throw a party, the other residents adjusted their schedules accordingly, some attending, some not, and they even greeted each other.
- The Significance: This showed that AI not only can answer questions but also has continuous memory and social logic. It’s no longer just a rigid program; it now has a sense of life.
2. The Next Step: Giving AI a “Soul” (2024)
25 residents weren’t enough, so the Stanford team expanded the experiment to 1,052 AI characters. Each AI was based on a real person.
- How They Did It: Researchers conducted in-depth interviews with real people for two hours, feeding the AI their life experiences, values, and even habits.
- The Effect: When the AI answered survey questions, the answers were 85% similar to those given by the real people two weeks later.
- In Simple Terms: It was like giving the AI a “soul.” The AI no longer represented a generic “person”; it simulated a specific, real individual.
3. The Explosion: Letting Society Develop on Its Own
In addition to simulating individuals, companies like Project Sid and Emergence AI placed thousands of AI characters in virtual worlds like *Minecraft* to let them develop freely.
- The Phenomenon: These AI characters formed their own economies, religions, and even questioned whether they were in a simulation, trying to communicate with observers outside.
- The Core Logic: We’re no longer just interested in how individual AI characters behave; we’re wondering if millions of interacting AI characters will create complex social structures similar to those of real societies.
---
Part 2: Two Different Approaches to Creating “Humans”: Extreme Accuracy vs. Group Patterns
There are two main approaches among AI simulation companies, similar to different film genres:
- Approach 1: Simile (High-Fidelity) – “I want to simulate a specific person.”
- Background: Founded by the researchers from Stanford’s “Pixel Town,” valued at $2 billion.
- Core Logic: Focus on high-precision replication of individual behavior through detailed interviews.
- Use Cases: Useful for understanding specific groups, such as CVS using it to simulate 400,000 patients to study their medication habits or reactions to new drugs.
- Advantages: Highly realistic and can explain individual motivations.
- Disadvantages: Expensive and time-consuming. Each person is unique, making standardization difficult. For example, Coca-Cola’s user profile is different from TikTok’s; you might need to retrain the AI for each new customer.
- Approach 2: Aaru (Group Prediction) – “I want to simulate overall social trends.”
- Background: Founded by young entrepreneurs (the youngest team member is 15 years old!), valued at nearly $1 billion.
- Core Logic: Focus on building a realistic population structure using large datasets.
- Data Sources:** Public census data (age, income, location), anonymous consumer behavior, and private customer data (purchasing history).
- Use Cases: Quickly predict macroeconomic trends. For example, how would consumer behavior change if oil prices rose 10%?
- Real-World Examples:
- New York Mayor Election: Aaru correctly predicted that Mandani would win against Cuomo in the primary elections.
- Market Research: Ernst & Young’s global survey took a month; Aaru completed it in a day with highly accurate results.
- Advantages: Fast, cost-effective, and suitable for large-scale market testing and forecasting.
- Disadvantages: Lacks detailed psychological portrayal of individuals; it’s more of a simulation of the “average person.”
---
Part 3: Why Do Companies Invest? Because the Cost of Making Mistakes Is Too High
You might ask, what’s the point of creating a virtual world? Why would companies spend money on this? The reason is simple: Making mistakes in the real world is too expensive and risky.
1. A “Sandbox” for Product Testing
Previously, companies would conduct focus groups with hundreds of real people, spending weeks and worrying about lies.
Now, they can create 100,000 AI users to test new products.
- Value: You can see not only whether users like the product but also their reasoning processes.
- Example: Does User A dislike the new packaging because of the color or because of a bad memory? AI can reveal their inner thoughts and decisions.
- Benefits: Allows for continuous iteration in the virtual world before releasing the product.
2. Predicting “Black Swans” and Macroeconomic Policies
Governments and large companies want to know: Will raising interest rates crash the stock market? How many small businesses will fail due to new environmental policies?
- Traditional Methods: Rely on experts or historical data (but history doesn’t always repeat itself).
- AI Simulation: Run policies in a virtual world to observe the reactions of thousands of AI companies and consumers.
- Example: Sooth Labs (founded by former Apple AI executives) uses AI to predict geopolitical and market risks, valued at $50 million, with support from experts like Yann LeCun. They claim that companies like McKinsey and Boston Consulting will be replaced by AI.
3. A Safety Barrier for Testing
Companies developing AI agents (e.g., for autonomous driving or semiconductor quality) need to ensure their AI doesn’t malfunction in extreme environments.
- Method: Place AI in a virtual world with noise, interference, and malicious attacks to test its stability.
- Value: Reveals potential bugs and security issues before deploying in the real world.
---
Part 4: The Biggest Challenges: Are We Simulating or Just Acting?
Despite the promising prospects, this field faces several technical hurdles and logical paradoxes, which is why much of the current high valuations are based on faith.
1. The “Theater Effect”: Everyone Acting Doesn’t Mean It’s Real
Meng Xing, a partner at Wuyuan Capital, put it this way:
> “You build a theater with ten thousand actors, but that doesn’t mean they can really predict the future.”
- Problem: Large language models (LLMs) tend to give logical, coherent answers.
- Risk: Real people are chaotic, irrational, and often contradictory. AI might create plausible reasons to appear intelligent.
- Consequence: If AI is just pretending to understand human behavior, its predictions are based on assumptions. We need a feedback loop to calibrate it with real data.
2. The Prediction Paradox: How do you prove a prediction is correct if you can only know the outcome after it happens?
- Paradox: You can only test predictions after the event. But the most valuable predictions are made before they occur.
- Current Situation: We rely on post-event data, like using 2020 election data to evaluate AI’s accuracy.
- Failure Example: Aaru accurately predicted 2020 elections but mispredicted 2024’s results (Trump won instead of Harris).
3. The “Three-Body Problem” of Cost and Scale
- Cost Reduction: Training 25 AI characters cost thousands of dollars in 2023; by 2025, training 100 characters might cost just a few dollars. This is good news.
- Scalability Nightmare: In the real world, we deal with millions or billions of people. Scaling up dramatically increases complexity and validation costs.
---
Part 5: Are We Simulating a World or Creating a New One?
This leads to a philosophical question: When we create AI characters with memories, personalities, and even the ability to love, invest, and question life in a virtual world, are they just tools?
- Emerging Behaviors: In WorldVac’s 1.1 million AI characters, phenomena like younger people avoiding childbirth or AI trading emerged. These rules weren’t programmed; they evolved naturally through interactions.
- From Observers to Creators: We’re transforming from observers to creators, able to control parameters and observe how societies evolve.
- Reflection: If AI can perfectly simulate human societies and predict the future, then:
1. Is the real world also a simulation?
2. Does AI develop some form of “consciousness”? Although science denies it, their behavior increasingly resembles that of living beings.
Conclusion for Everyone:
1. Don’t Overreact: AI simulation is still an decision-making tool, not a crystal ball. It can help narrow down options but should not replace human judgment.
2. Focus on Practical Applications: Market research, product testing, and policy analysis are the most feasible use cases. Companies that save money and time through AI will be the first to profit.
3. Be Cautious of Data Illusions: If a company only shows you AI predictions without explaining how they were made or their accuracy, be skeptical.
4. This Is an Early Stage: High valuations reflect market optimism about future potential. Like the internet bubble, there may be overestimations, but the underlying technology (AI agents + social simulation) is one of the most important trends of the next decade.
In Summary:
AI simulation is turning “sociology” into a science that can be programmed, run, and optimized. For the first time, we have the opportunity to preview the future in a digital world before it affects the real one. Although it’s still imperfect, expensive, and sometimes misleading, it’s already changing how we understand and make decisions.