A dust particle, a thermal vibration, a local concentration fluctuation. Any defect can trigger crystal nucleation. The moment is unpredictable. The location is accidental.
The final structure was already written in the geometry of atomic bonds. Tetrahedral carbon, hexagonal silicon, body-centered cubic iron. The topology of electronic orbitals imposes its law before the first bond forms. Chance chooses the moment.
Physics imposes the form.
The crystal lattice does not build through progressive accumulation. It emerges through phase transition, a threshold crossed, a broken symmetry, an order that imposes itself globally as soon as local conditions permit. The final structure pre-existed as energetic potential, a landscape of minima that the system explores until it finds one deep enough never to escape.
Ostwald had identified the rule: metastable phases appear first, then evolve toward thermodynamic equilibrium. The system explores its possibilities in order of energetic proximity. But it always arrives at the same crystal.
Except when it doesn't. Quasi-crystals, discovered by Shechtman in 1982, are ordered without being periodic. Their symmetry is real but their structure does not repeat. They occupy a space that classical classification did not foresee: neither perfect crystal nor amorphous glass, something between the two regimes. The energetic landscape contained this possibility. No one had searched for it because theory did not render it visible.
Doctrine
The designer chooses the architecture. The algorithm explores the landscape that this architecture has carved. Neither imposes the form. They negotiate with a topology that precedes them both.
We do not impose form. We negotiate with it.
Open vector
A perfect crystal is highly ordered but carries little information in Shannon's sense: its structure is entirely compressible, each cell predicts all the following ones. If learning only converges toward the crystal, AI maximizes order by minimizing information. It becomes a compression of the starting geometry, not a discovery of what it contains.
The quasi-crystal suggests another possibility: a regime where order and information coexist, where structure is not repetitive but not random either. In machine learning, this regime corresponds to generalization, the capacity to produce correct responses on unseen data, neither memorized nor extrapolated but read in the topology of the problem.
The question is not whether learning discovers or locates. The question is what type of crystal the landscape contains, and whether the architectures we build today can reach the quasi-crystal, or whether they can only produce the perfect crystal.
