This video emphasizes that the potential of AI agents is far greater as 'modelers' that model complex reality to support better decision-making, rather than remaining as 'doers' that simply execute tasks. By adding simulation capabilities, AI agents can function like digital twins -- exploring diverse future scenarios, compressing learning processes, and accumulating insights to revolutionize decision-making quality for both businesses and individuals. Through examples from Renault, BMW, F1, and others, it demonstrates the real value of simulation while also addressing concerns about accuracy, cost, and cultural resistance.
1. Our Misconception About AI Agents: Execution vs. Modeling
The video points out that we are currently viewing AI agents from far too narrow a perspective. Most people are focused on AI agents performing specific tasks -- writing emails, answering customer inquiries, generating code, and so on. While agents as executors certainly provide excellent automation and efficiency, this represents only the low end of the value AI agents can deliver.
"We're pouring tokens into agents that save us a few minutes, but the trillion-dollar value lies in compressing a 10-year strategy into a 10-hour simulation."
The truly exponential growth opportunity lies in leveraging AI agents as modelers -- helping predict the future and make better decisions by simulating complex reality. Some smart companies are already achieving a quiet AI revolution through this approach, and this is emphasized as the key to creating the next trillion-dollar market.
2. Limitations of Traditional AI Agents
The concept of traditional AI agents is relatively simple: a combination of LLM (Large Language Model) + tools + instructions.
- LLM (brain): Serves as the core and brain of the agent.
- Tools (executors): Various tools that the LLM calls to perform tasks.
- Instructions (policies/constraints): Policies and constraints that tell the agent what to do and what not to do.
We typically measure how well agents combining LLMs, tools, and instructions perform actual tasks. For example, key performance metrics include tickets processed, time saved, and cost per interaction. Even approaches that bundle multiple agents into networks or teams are all based on this 'executor' concept. This excels at automation and execution but causes us to miss bigger opportunities.
3. AI Agents Become Reality Simulators!
The true potential is for AI agents to become reality simulators. This connects to the concept of 'digital twins.' In January 2025, when NVIDIA CEO Jensen Huang introduced digital twins for manufacturing and warehousing, many people focused on AI agents, but his core message was actually that digital twins are critically important for maximizing long-term productivity and AI agent utilization.
To use agents as modelers, one additional element must be added to the traditional stack: LLM + tools + instructions + simulated world.
This 'simulated world' can be visual like a 3D video game, but it can also be text or other formats modeling relevant constraints. For example, simulating a difficult stakeholder situation through a ChatGPT conversation, or practicing a breakup conversation with an ex -- both are good examples of using AI agents as reality simulators.
4. Linear Value vs. Exponential Value: Why Simulation Matters
Using AI agents as executors delivers linear time savings. Like turning a 10-minute email into 0 minutes. This is certainly great, but it cannot compare to the exponential value of revolutionizing business decision-making.
Reality simulator agents help companies simulate and explore various business timelines. In the past, this was limited to PowerPoint presentations offering one preferred option among three to the board, but AI now provides far more possibilities.
"We can now use computing power to simulate various scenarios across all the timelines we historically had to look at just two or three steps ahead, and make wiser decisions based on them."
For example, a 10-year market cycle can be compressed into a 10-hour simulation to explore 5-6 different scenarios and gain a much more useful understanding of business direction. Even a slight improvement in human decision-making ability would more than offset all the impact of execution-focused LLM agents.
5. Exponential Value Leverage of Simulation
The exponential value available when leveraging AI agents as modelers includes the following:
5.1. Alternate Timeline Exploration
You can simulate various options not just for the overall business, but also for specific scenarios.
- Simulating customer reactions to product launches
- Predicting effectiveness before spending on marketing campaigns
- Testing various code combinations before release
5.2. Time Compression
While your competitors are on their 3rd iteration, simulation time lets you be on your 300th. Enormous acceleration is possible because you're operating on simulation time rather than physical time.
"You're moving on simulation time, not wall-clock time, and you can simulate incredibly fast and discard what doesn't work."
Questions may be raised about the accuracy of such simulations, but the speaker counters:
- Leading global companies are already using this and creating enormous value.
- Even if not perfectly accurate, it's far better than not thinking at all. Even 70% accuracy can be extremely useful.
Robotics and Tesla's autonomous driving AI training are good examples. Robots learn to 'walk' thousands of times in virtual environments for rapid training, and Tesla trains its AI on simulated courses to experience countless edge cases without expensive real-world accidents.
5.3. Compounding Insights
Each simulation yields better prior knowledge, leading to nonlinear innovation.
- Discovering pricing cliffs
- Uncovering hidden customer segments
- Breakthrough product development
These insights are value that even the smartest execution agents cannot obtain. While execution agents remain at a linear value scale, AI agents as model simulators deliver nonlinear value scale.
6. Real-World Success Stories
Success stories of simulation-based AI agents already exist in abundance. Particularly interesting examples can be found mainly in the automotive industry.
- Renault: Used digital twins to reduce vehicle development time by 60%. Before actual prototype production, digital twins predict crash outcomes to set appropriate vehicle development directions.
- BMW: Built a virtual factory that simulates thousands of production line change combinations to find optimal factory operation methods.
- Formula 1: Through real-time pit strategy simulation, they find the most efficient energy distribution methods during pit crew changes, helping cars return to the race as quickly as possible.
- Ad Networks: Pre-test ad creative combinations without spending, contributing to improved return on ad spend (ROAS).
All these cases demonstrate AI agents performing the role of 'world modelers' -- given constraints, tools, and a world to operate in, they model that world and present results.
7. Counterarguments to Simulation and Solutions
Several counterarguments about the value of simulation may still exist.
7.1. "Garbage in, garbage out"
- Solution: Introduce verified calibration loops and pay attention to input data. Continuously back-test to confirm simulation results match actual performance, and if there's a significant gap, honestly evaluate and fix whether any constraints are missing from the simulation.
7.2. "It gives false confidence"
- Solution: Use simulation to constrain distributions rather than point predictions. Rather than expecting a single precise result, use multiple scenarios to understand the range of possible outcomes. Humans tend to fixate too much on point predictions, and simulation can help understand the world as a continuum of distributions.
7.3. "Compute is super pricey"
- Solution: This can be turned around with the question "How can you afford NOT to?" If simulation provides breakthrough potential, it's well worth the investment.
7.4. "Culture change is hard"
- Solution: Corporate culture must shift to reward not just shipped features but also 'decision quality' and 'disaster avoidance'. Acknowledging that this is difficult, now is the opportunity to rethink decision-making methods and how agents are used in business. Computing power can bring innovation to decision-making and future prediction, which requires deep reflection on how we think and avoid disasters.
8. How to Start: Start Small and Scale Fast
How should you begin with these simulations?
- Select One KPI: First, choose one key performance indicator (KPI) that you think you understand well and create a 'digital twin' for it. This could be customer acquisition cost or churn rate. You can start by having an LLM model it with a simple prompt, or by building a custom system.
- Understand and Manage Data: You need to precisely understand the data used for simulation and set up data refresh methods and feedback loops.
- Robust Tool Stack: Having a reliable and robust tool stack is important.
- Large enterprises: Can build enterprise stacks including data lakes, lakehouses, feature stores, simulation engines, dashboards, etc.
- Individuals: It doesn't have to be complicated. For example, if you're doing a breakup simulation with ChatGPT, the key elements are quality data about the relationship, information updated after dates (refresh), and feedback on simulation results.
Through personal examples like these, it concretely shows how simulation can be applied to real life, emphasizing that sufficient information provision and prior knowledge updates based on real-world changes are essential for useful simulations.
9. Conclusion: AI Agents That Build the Future
The speaker poses an important final question:
"If we have the ability to predict clearer futures and yet don't use it, doesn't our moral responsibility for future timelines become even greater?"
Since we now have the computing power to think of AI agents as 'worldbuilders,' we have a responsibility to think more deeply.
While everyone else sees AI agents as executors, if you think of AI agents as a way to model future reality and make better decisions, you are playing an entirely different game -- and you will be the first mover in that game.
So it's time to stop or reduce simply asking "How can AI do this task?" and instead ask "How can AI show me various futures and improve my decision-making ability?" Where is the digital twin that will prevent your next big mistake? This is precisely the question we should be asking ourselves.