Learning Developmental Scaffoldings: How AI Grows Like an Embryo via Self-Organization

Forum topic · QianXun · 2026-05-16

Summary

This forum post explains a 2026 paper by Milton L. Montero and colleagues from the IT University of Copenhagen and Sakana AI, titled 'Learning Developmental Scaffoldings to Guide Self-Organisation.' Neural Cellular Automata (NCA) are AI systems where individual cells communicate only locally, yet historically they often fail without a global 'God's-eye' blueprint, producing disorganized patterns. Inspired by embryonic morphogen gradients, the authors propose a two-stage system: a SIREN network first generates a faint, blurry 'scaffold' of guidance signals across space, and NCA cells then refine, grow, and differentiate based on this draft. This approach solves symmetry breaking, yields strong robustness to noise and cell death with rapid pattern regeneration, and allows large compression of cell rule parameters by offloading global structure to the environment. The work suggests future AI can emerge through simple agents guided by environmental cues, much like biological development.

Rejecting the 'God's-Eye View': How AI Learns to Grow Like an Embryo

*(English translation of a Chinese forum post on zhichai.net)*

Imagine you have a magical set of LEGO bricks.

There is no instruction manual and no 'architect' giving orders. Instead, every brick is alive, and each one only chats with its neighbors: 'Hey, if you're red, I'll become red too,' or 'If there are already three siblings on your left, I'll shift to the right.'

With only this kind of local communication, can the bricks grow into an elegant Eiffel Tower by themselves?

In computer science, this is called a self-organising system, and the color-changing, moving bricks are known as Neural Cellular Automata (NCA).

Cool as it sounds, scientists have long been stuck on an awkward problem: without a 'God's-eye view' blueprint, these bricks tend to fall into collective confusion. They don't know which end is which, and what grows out is usually a messy plastic blob rather than a tower.

In May 2026, researchers from the IT University of Copenhagen and Sakana AI (Milton L. Montero et al.) published a remarkably 'lifelike' paper: _Learning Developmental Scaffoldings to Guide Self-Organisation_.

They stole a secret trick from nature to solve this maddening problem.

What Is 'Developmental Scaffolding'?

Feynman once observed that nature doesn't always need complicated formulas — it often uses the simplest physical differences to guide complex life.

As a fertilized egg develops into an embryo, cells do not grow blindly from scratch. Before they begin dividing, the mother has planted invisible 'signposts' inside the egg — for example, a certain chemical being slightly denser on the left and sparser on the right.

These hidden chemical gradients are the 'developmental scaffolding.'

Cells don't need to see the full blueprint. They only need to sense the local chemical concentration to understand: 'Ah, I'm in the high-concentration zone, so I should develop into a head.'

The Paper's Breakthrough: Giving AI a 'Rough Draft'

The core idea is charming: don't expect AI cells to conjure miracles out of thin air — learn to sketch a draft for them first.

The researchers designed a two-step system:

1. Build a 'draft generator' (SIREN): It doesn't draw details; it only scatters a faint, blurry background pattern across space (this is the scaffold). 2. Then let the 'AI cells' enter (NCA): The cells perform local fine-tuning, growth, and differentiation based on this fuzzy draft.

Why This Makes AI Stronger

With this 'draft first, refine later' approach, the researchers found qualitative improvements:

1. Curing 'Symmetry Phobia'

Previously, AI cells facing an empty space didn't know which direction to grow. With a scaffold, they can immediately tell up from down. This solves the hardest biological problem of symmetry breaking.

2. Remarkable 'Stress Tolerance'

The researchers deliberately injected noise and disturbances while the AI cells grew. Scaffold-equipped systems stayed extremely calm: even if some cells died, the survivors could rapidly regenerate the original pattern using the background draft.

3. High 'Compression Efficiency'

Most remarkably, because the scaffold takes over part of the 'macroscopic guidance,' the internal rules (parameter count) of the AI cells can be greatly compressed. This proves that splitting the task between the environment (scaffold) and the individual (cells) is far smarter than making individuals bear all the pressure.

Why This Paper Matters

This work touches on a profound proposition: intelligence doesn't have to live inside a single, giant 'brain.'

It suggests that future AI might, like forests, coral reefs, or our own bodies, emerge spontaneously from simple individuals growing under subtle 'environmental induction.'

In summary:

We don't need an omniscient God's-eye view. As long as we learn to scatter a few 'key signposts' in space, AI can grow astonishing complexity by itself in the digital world, just like an embryo.

Next time you marvel at nature's exquisite biological structures, remember: beneath that beautiful skin lies a silent strategy called 'self-organising scaffolding' — and humanity has finally learned to rewrite it in the world of AI.