A mathematical game from 1970 shows why a working system sits on an edge — and why, with AI, small errors don't add up, they compound. Precisely because it is so efficient, quality is not optional.
In 1970 the mathematician John Conway wrote down four rules simple enough for a child to follow. A grid of cells, each one alive or dead. A cell with too few living neighbours dies of loneliness; one with too many dies of overcrowding; under the right conditions, a dead cell comes to life. That's all.
Out of those four rules, a whole world unfolds. Patterns that stand still. Patterns that blink in rhythm. Patterns that travel across the field as "gliders." You can build working computers inside Conway's "Game of Life" — out of nothing but those four rules. That's the famous lesson: simple rules, complex behaviour.
But there is a second lesson, told far less often. And it matters to anyone working with AI today.
The edge between dying out and clogging up
Take the same starting point — the same randomly scattered seed — and run three copies side by side; that is exactly what you see above. Only one thing changes, and only slightly: how easily a cell survives.
Make the survival rule a touch stricter, demanding a little more for a cell to persist: the field dies out. Within a few steps everything is empty.
Make the same rule a touch more generous, so a cell no longer dies at all: the field clogs up. Everything congeals into a dense, structureless mass.
Only in between, in a narrow band, does the interesting thing happen. There — and only there — the still lifes emerge, the oscillators, the gliders. There the system carries structure, made of stable and moving parts at once.
That is the second lesson: a working system is not robust against its rules, it is acutely sensitive to them. It sits on an edge between two kinds of failure — too little and too much. Effort doesn't decide it; compute doesn't decide it. A small change to the rule decides whether order emerges or everything collapses. Researchers call this edge the "edge of chaos."
AI behaves the same way
Anyone working productively with AI is building systems like this without calling them that. A processing chain, an agent setup, a prompt with memory and tools — these are rule-bound systems, and they live in the same narrow band. One detail in the instructions, one slightly wrong dataset, a threshold set a little too high or too low, and the whole behaviour tips over. Not by a little. It tips.
This explains an experience many people have had and struggled to place: why an AI setup that ran reliably yesterday suddenly produces nonsense after a small, harmless-looking change. The system was never robust. It was sitting on the edge — and the change pushed it off.
Efficiency cuts both ways
Here is the part that turns a nice analogy into a hard imperative.
AI does work quickly and with enormous efficiency. But efficiency doesn't apply only to the good work. It applies to everything running through the system — the error included.
In a slow, manual process a small error stays small. It happens once, someone notices, it gets corrected. In a fast, automated system where step builds on step, the same error happens a thousand times before anyone looks — and it becomes the input to the next step, which in turn is the foundation of the one after that. The error doesn't add up. It compounds. And it does so with exactly the efficiency that makes the system valuable.
Efficiency, then, is not a neutral accelerator. It is an amplifier for whatever you feed it. Feed it quality, and it amplifies quality. Feed it carelessness, and it amplifies carelessness — just as fast, just as thoroughly.
From which follows something that runs against intuition. You might think that where a machine does so much so fast, precision matters less. The opposite is true. The more efficient the system, the more expensive the small error. Precision is not the luxury you allow yourself when there's time to spare. It is the price of admission to the fast lane.
Why this doesn't stay a picture
For us the edge of chaos is not just a metaphor but a working hypothesis: living, computing systems — from the brain to the memory of an AI — perform best when they operate near this critical edge. Close enough to be rich and adaptive; far enough not to tip into noise. Neuroscience describes exactly this "slightly subcritical" regime for the brain.
We study how the mechanisms of a memory architecture can be described as a small, ordered mathematical structure — and why such a system holds up precisely when it works near that edge. For us this is not a footnote but an active line of research. It is exactly why we treat quality not as polish at the end, but as a property of the rules themselves.
The imperative
The Game of Life is over fifty years old and looks like a toy. Its second lesson is more current today than ever. A system that draws value from rules is only ever as good as the rules it sits on — and it sits on a narrower edge than it looks.
Anyone using AI seriously is building such systems. The question is not whether you can get them fast and efficient — you get that anyway. The question is whether the quality is right before you switch on the amplifier. Because it amplifies both.
Collective intelligence on a cognitive architecture emerges this way: from simple rules, carefully chosen. That is exactly what we work on.
For context: the "Game of Life" is due to John Conway (1970); the "edge of chaos" comes from research on cellular automata (notably Christopher Langton). That the brain operates near this edge in a "slightly subcritical" regime is an established finding in neuroscience. Carrying this over to memory architectures is an active line of research of ours.