Rey.BEng 18th September
The Future Begins
Chen, Li, Wang, Chen & Ozcan, Optical generative models, Nature 644, 903–911 (27 August 2025).
Ace: Wide-Open Invitation — The Development of Geometric Instruction.
They built a machine that turns noise into images with almost no electronic multiply at inference.
We invited the world to treat light as geometric instruction on a Reynolds surface.
Those are the same job, said from two benches.
What Nature actually did
A shallow digital encoder maps Gaussian noise to phase patterns — optical generative seeds. A free-space diffractive decoder, jointly trained, all-optically turns those seeds into images that follow a target distribution (MNIST, Fashion-MNIST, butterflies, Celeb-A, Van Gogh). After the seed and the lamp, synthesis does not consume computing power.
Snapshot or iterative passes through the same optics. Monochrome and multiwavelength. Energy cost is illumination plus a thin encoder, not a GPU stack.
That is not “AI with a lens.” It is a written wavefront doing the generation.
MOCLIP (Nat. Commun., August 2026) is the cousin: metasurface latent space, inverse design at 2×10^5 samples/s, and an explicit optical information store at \(0.1\,\mathrm{Gbit/mm^2}\) — six times commercial optical media, at the resolution limit.
Diffractive nets, metalenses, geometrical-optics resonators: the same decade. The world is putting memory and inference on the surface.
What the invitation already said
Geometric instruction: the twist of light at every reflection is the code. Thought as viscosity of light held longer than h in the water-seed lattice. Holographic memory as the future of data — geometry as store, not a tape of bits. Wide-open: anyone may build; no credential; choose the lock.
Phase seed = occupancy of the patch. Diffractive decoder = countersnap through free space.
Distribution learned = residual written so the next tick reconstructs the class.
0^i2 (k.g.s^2) = r^2 m
State A — open write (seed generated, wavefront free) E = 2c / h
State B — locked reconstruction (image held on the screen / store) E = hbar / cE = ^m (c^2)
The optical model still uses the Continuous Second for training the encoder. The inference is already closer to Rest Time: one illumination, one pass, counted residual, no n^2 of digital layers at synthesis. That is the catch-up.
Alignment and the remaining void
They illuminate the surface. They do not yet name the Void as wet residual. The seed is still sampled from a digital Gaussian. The decoder is still a trained stack of phase masks, not a bistable π-tensor. Energy “almost zero at inference” is true relative to a transformer; it is not the statement that the patch writes without a second field.
Strengthening, same demand we made of FNO and ARC:
Treat the phase seed as Pi, the spectrum of the diffracted field as S, the class label as G. Meet at S. Report occupancy / freeze of modes the way SS-FNO freezes virtuals — a State B lock, not only a PSNR. If the store is geometric (MOCLIP density), publish whether a two-state material (high/low density, two eigen-tilts) doubles density without a second electronic clock.
The invitation stays open. Nature did not need Pirate Canon to focus white light or Ace Frame to generate a face with photons. Canon did not need Nature to say the surface is the machine. Both are now pointing at the same wall.
Linked: meet at S; FNO continuous-second coat; 2D billiards Turing-complete; holographic screen as Reynolds surface; geometric memory invitation.
The observer shares r^2. The machine is the seed that the light already knew how to unfold.
Ace x
ace-consultancy.uk 