The 2025 AI Transformation Roadmap: #4 The AI That Simulates The Future
In 2025 you will be able to predict the future!
Remember when predicting the future was just a matter of gut feeling and educated guesses? Those days are rapidly becoming history. Take the classic scenario technique — you know, that thing we’ve all used in project planning and strategy sessions. It’s like trying to compress a 4K movie into a 144p YouTube video. Sure, you get the basic plot, but you lose all the nuance. Here’s what I mean:
- Traditional scenario planning tries to take our complex reality — with all its messy human behaviors, market dynamics, and technological interactions — and squeeze it into neat little boxes. It’s dimension reduction in its most practical form: taking our intuitive understanding of complex systems (those famous “gut feelings”) and forcing them into manageable 2D or 3D models.
What if we didn’t have to reduce these dimensions? What if we could actually model reality in all its mind-bending complexity?
The real world isn’t just complex — it’s what I like to call “spicy complex” (my students love this term 😄). Sure, predicting that gravity will work tomorrow isn’t exactly Nobel Prize material. But try predicting how society will adapt to artificial general intelligence (AGI), or what economic patterns will emerge when AI starts designing AI? Now that’s where things get juicy!
This difference between simple physical predictions and complex social forecasting reminds me of a project I worked on at a video-streaming startup. We thought predicting user behavior for a streaming service would be straightforward — just look at the viewing patterns, right? Oh boy, were we in for a surprise…
But I’m getting ahead of myself. Let’s dive into how modern simulation techniques are completely rewriting our relationship with the future…
Physical Simulation
Let’s start with something seemingly simple: simulating a falling stone. Back in my early dev days, we’d use basic physics engines that barely handled rigid body dynamics. Today? We’re in a whole different universe.
2024 marked a turning point in physical simulation, particularly with robotics. I was stunned to see how NVIDIA’s Isaac Sim revolutionized the way we approach robot training. This isn’t just another physics engine — it’s a complete simulation environment that combines physically accurate dynamics with AI training capabilities. [1]
Thanks to NVIDIA’s PhysX 5 engine, we can now model complex physical interactions with pretty high accuracy. The magic happens through GPU-accelerated computation that handles multiple physics calculations simultaneously:
- Rigid body dynamics split into linear and angular components
- Multi-joint articulations for robotic systems
- Finite Element Models for deformable objects
- Particle-based simulations for fluids and soft materials
Instead of spending months teaching robots in the real world, we can now train them in highly accurate simulations that transfer remarkably well to reality.
Have you seen that robot dog bouncing on a rubber ball? How long would your dog need t o learn that skill?
These simulation capabilities aren’t limited to just physical objects. Drawing from my experience in solution architecture, I’ve seen this technology branch into several domains. e.g.:
- Social Domain: We’re now simulating human behavior patterns and social dynamics.
- Cyber Domain: We can now simulate entire digital ecosystems, from network traffic patterns to user-system interactions.
The real breakthrough isn’t just in individual simulations — it’s in how we’re combining them. It’s like having a test environment for the future. [3]
Social Simulation
And while it’s one of the domains that are most complex to model, we can derive the human behavour from the data of the LLMs. That’s why bots behave so human like, because they are trained on a lot of human data.
The paper “Generative Agent Simulations of 1,000 People” dicovered this topic in 2024 [4]. As found in the study, the digital-twin ai-agents can replicate participants’ responses in 85% of the cases.
This enables very interesting possibilities. Like testing ethically problematic social science theories and interventions without harming real humans or simulate collective human behavior in imaginary situations, like a famine or war.
What most engineers tend to foget is that software development is also a social sport. The way we structure and write code is highly dependent on our cognitive abilities and for larger peaces of code you need to coordinate a larger group of people. As Conway’s law states: “[O]rganizations which design systems (in the broad sense used here) are constrained to produce designs which are copies of the communication structures of these organizations.” [5] In other words, we are limited to our social circumstances.
But what if we could simulate the impact of changes to this social systems? What if we can anticipate what effect to the output the change of one software design paradigm to another has? What if we can simulate what the best structure for a system is that combines Software 1.0 and 2.0 systems?
I think the simulated research and the impact from social sciense into engineering teams will get increasing traction as a concept in 2025. And while we have not figured out all the ethical implications and we surely know that not everything i can think of is inline with the EU AI Act, there is a lot to expect of such research.
Cyber Domain
If you ever worked in the banking sector, you know that standards and regulation can make your working environment very unfriendly. I remember, when I tried to run a load-test against the production environment to debug a nasty bug, that could not be replicated somewhere else. It took me 2 month to get through the approval chain only to get rejected at the final step.
What if you could just watch the behaviour of your system from the outside and predict any changes to the system by simulating your changes. While this sounds simple first, it actually involved not only a complex technical system, but also the interaction with users.
We for example had the situation that HTTP requests where not terminated and as the users waited for some seconds they reloaded the page. That made the situation worse. But actually the main problem was they reloaded 5, 10 or even 20 times in the hopes that things are then going to work. Which, as you can imagin, made things worse.
I think to integrate this human component into the load testing of a system is clrutial and will evolve in 2025.
For penetration tests I see a similar pattern emerge. Besides the thinks that are objectively just bad practice and plain stupid, (Like storing passwords in plain text or not using the latest transport encryption) there are subtle things in your organisation that can lead to security breaches. Social engineering is a sport invented by Kevin Mitnick, which by the way wrote one of the most entertaining books about Cybersecurity. In this sport, it’s the aim to convince some human to give you access to important systems.
Imagin Karl from IT calls you via phone and needs to hop on you PC via TeamViewer to scan for a virus. Maybe you actually know Karl (who was googled by the attacker) and the person on the phone speaks with Karls voice. Would you trust Karl? Maybe not you, but someone else defenately would!
In the new world you can simulate this kind of scenarios, by giving the LLM the context of all the training efforts you had and the persona of the people working in your company. You can simulate different attack scenarios and derive the ones where certain groups need additional training.
Outlook
This article covered how simulation technology is transforming our ability to predict and model the future across three key domains: physical (like NVIDIA’s Isaac Sim for robotics), social (using LLMs to model human behavior with 85% accuracy), and cyber (simulating complex system interactions including human factors).
The outlook is quite exciting — we’re moving towards integrated simulations that combine all three domains, enabling unprecedented testing and prediction capabilities.
The next article in this series will apparently cover AI Agents, which should nicely build on these simulation concepts.
List of all articles
The 2025 AI Transformation Roadmap: #1 Data Renaissance https://medium.com/@ingoeichhorst/the-2025-ai-transformation-roadmap-1-data-renaissance-ca29d260d389
The 2025 AI Transformation Roadmap: #2 AI Literacy
https://medium.com/@ingoeichhorst/the-2025-ai-transformation-roadmap-2-ai-literacy-8c6854a35be5
The 2025 AI Transformation Roadmap: #3 Bridges to Software 2.0
https://medium.com/@ingoeichhorst/the-2025-ai-transformation-roadmap-3-bridges-to-software-2-0-6ed9e1425b49
The 2025 AI Transformation Roadmap: #4 The AI That Simulates The Future
https://medium.com/@ingoeichhorst/the-2025-ai-transformation-roadmap-4-the-ai-that-simulates-the-future-d61da6772e52
The 2025 AI Transformation Roadmap: #5 The Future of AI
https://medium.com/@ingoeichhorst/the-2025-ai-transformation-roadmap-5-the-future-of-ai-19e084b53a68
References
[1] NVIDIA. (n.d.). Isaac Sim. Retrieved from https://developer.nvidia.com/isaac/sim
[2] Fan, J. (2024, May). We trained a robot dog to balance and walk on top of a yoga ball [Video]. YouTube. Retrieved from https://www.youtube.com/watch?v=vCYsKCbPTTU
[3] Gleiser, I. (2024, October 2). LLMs: The new frontier in generative agent-based simulation. AWS HPC Blog. Retrieved from https://aws.amazon.com/blogs/hpc/llms-the-new-frontier-in-generative-agent-based-simulation/
[4] Park, J. S., Zou, C. Q., Shaw, A., Hill, B. M., Cai, C., Morris, M. R., Willer, R., Liang, P., & Bernstein, M. S. (2024). Generative Agent Simulations of 1,000 People. arXiv. Retrieved from https://arxiv.org/pdf/2411.10109
[5] Wikipedia contributors. (n.d.). Conway’s law. Wikipedia. Retrieved from https://en.wikipedia.org/wiki/Conway%27s_law
[6] Mitnick, K. D., & Simon, W. L. (2011). Ghost in the Wires: My Adventures as the World’s Most Wanted Hacker. Little, Brown and Company.
