What's this about?
A browser demo where absurdly tiny agents — 14 bytes of "brain" each — try (and mostly struggle) to navigate a 2D maze, evolving over generations. It's a playful counterpoint to trillion-parameter week: a visceral look at how little, or how much, computation a task really needs.
What exactly is a "14-byte brain"?
Here the "brain" is 14 bytes of parameters — a rounding error next to a real neural network's billions. It's essentially a tiny rule-set nudged by evolution rather than trained by gradient descent. The point isn't capability; it's watching the floor of "intelligence" — how simple a controller can be and still sometimes reach the exit.
Why does it matter?
It's a healthy reminder, on the day a 2.8T model launches, that parameter count is a means, not a virtue — and that maze-running doesn't need a frontier model. Good intuition-fuel for right-sizing models to tasks.