AI That Grows Shapes From a Single Cell Can Now Resize Them for Free

Imagine an AI that doesn't just draw a picture, but grows one. Starting from a single dot, it follows simple local rules — each piece only 'talks' to its neighbors — until a shape like a smiley-face emoji appears. This approach is called a neural cellular automaton, and researchers at the IT University of Copenhagen and Tufts University used it to study a very old question: how does a body change its shape as it grows? Their inspiration is a 1917 book by D'Arcy Thompson, who showed that you can turn one animal's outline into another's by stretching and squashing a grid.

The team's clever move was borrowing a trick from modern AI chatbots. When companies want a chatbot to specialize, they don't rebuild it from scratch — they apply a small adjustment called LoRA, short for low-rank adaptation, which is like turning a handful of dials instead of replacing the engine. The researchers applied the same idea to their growing system. Turning a single dial could stretch a grown shape taller; another dial made it wider. Combine two dials, and you get a size in between — or even shapes the system was never trained on.

Most striking: dials learned for one shape worked on completely different shapes. A tweak discovered while resizing a heart also resized a star. The team trained roughly 25,000 of these small adjustments and mapped the 'dials' that control size, style, and even a shape splitting into two equal halves — like a cell dividing.

The catch is that this is pure research. It works only on simple 2D pictures in a simulation, not on real organisms or commercial design software, and nobody is shipping it as a tool. But the underlying idea matters: if complex changes can be steered with a few cheap dials, then tomorrow's AI could adjust designs, simulations, or even biological experiments far more quickly — and cheaply — than retraining everything from zero.