Superintelligent Cellular Automata
What would have happened if China had embraced capitalism in 1949? With its vast coal reserves, which historically accounted for nearly 75% of its energy production in the mid-20th century, would market incentives have driven engineers toward a hyper-efficient, miniaturized steam engine? Could there be an entire branch of coal generators that never existed because history went a different way?
We can't answer that question today. But I believe that the tools to answer it are being built right now.
The Problem with "Top-Down" Models
For decades, we've tried to simulate the world using hand-coded equations. The United Nations' Project LINK model uses 4,500 equations and 6,000 variables to measure how policy changes in one country affect others. The Federal Reserve and the World Bank run their own versions.
None of them work particularly well. They miss the granular details that actually drive outcomes: political crises, social movements, and even human irrationality. These models don't fail because they lack equations but because no team of economists can write them down at the scale that is needed.
From Simple Grids to Emergent Life
This is the same wall that early cellular automata ran into. When they were first explored, each cell was kept simple due to compute constraints. However, the interesting thing about cellular automata is that simple grids of cells can simulate emergent behavior that arises from basic rules. Under the right condition, interesting patterns start to appear. Complex properties emerge from minor perturbations.
This, in a way, reflects the way that life originated according to the primordial soup hypothesis. In a warm pond with methane, ammonia, and water, electricity from lightning provided just enough energy to form amino acids. When these ponds flooded and mixed with one another, more and more complex structures began to form. From the most basic building blocks emerged complex properties resulting from a minor perturbation.
The Four Characteristics of an Agent
To move from a "simple" grid to a "simulated world," John Holland's work at the Santa Fe Institute is worth exploring. He identified four key characteristics of agents that enable the development of complex adaptive systems:
- Parallelism: Individual agents exist and operate in parallel.
- Hierarchy: These agents form subsystems and hierarchies, leading to further complex behaviors.
- Anticipation: Agents are individually motivated. They anticipate the future and adapt.
- Niches: Specialized agents fill gaps in the environment.
In economics, we consider governments, banks, corporations, and individuals as agents. You could consider particles as agents within the context of physics. When thinking about the primordial soup hypothesis, the ammonia, methane, and water could be considered precursors to agents that gave rise to amino acids, which started acting as agents by combining with one another to form more complex building blocks.
The initial settings and "personalities" of these agents have a massive impact as time progresses. Mutations are also another important perturbations to kick start more complex reactions.
The Upgrade: Mathematical Superintelligence
The aforementioned Project Link model, along with models used by the US Federal Reserve and the World Bank, is far from perfect despite having thousands of equations and variables. Granular details like political and social factors are not captured, similar to how early cellular automata were.
But with the emergence of advanced math AI models, alongside the ability to create more complex agents and readily available compute, I believe that it's only a matter of time before superintelligent cellular automata power a complex adaptive system that helps simulate the world in ways that were not possible before. Each automaton would be an agent whose behavior is dictated by a set of mathematical equations that a superintelligent math model captures from observations of humans. The math models would autoformalize the observations, skipping the tedious steps of hand-coding them.
By doing so, this superintelligent cellular automata would capture each agent at the microscopic, mesoscopic, and macroscopic level, representing the simulations at the individual, societal, and even universal level.
Imagine the difference between a physicist writing down Newton's laws from intuition versus an AI watching millions of objects in motion and deriving the laws of mechanics from the data itself. Now apply that to economic behavior, social dynamics, and political decision-making.
This would not only help us understand the current trajectory of the world, but also help us simulate radically different worlds to see what new breakthroughs could occur.
Discovering the Futures That Never Were
Most of our world is governed by Path Dependence, the idea that minor perturbations in the past "lock in" specific technologies.
Think of the adoption of light-water nuclear reactors or VHS. They weren't the best technologies but minor perturbations in the equilibrium tilted adoption in their favor.
A superintelligent cellular automata could rewind those moments, tweak the inputs, and model the outcomes that never happened. It would reveal breakthroughs that are invisible from inside our current trajectory. We would be able to discover the best ideas from worlds that never existed.
I can't put an exact timeline on when these simulations will start materializing, but the fact that we're making significant progress on math AI models that are getting better at autoformalization feels like a turning point worth paying attention to.