Optical neural networks and what computing in light says about substrate independence

A neural network that computes in light is the cleanest possible test of the claim that a mind is an organizing pattern rather than a material. Networks that run machine learning with photons instead of electrons have moved from laboratory demonstrations to working systems, and they carry the site’s substrate-independence thesis to its physical extreme. If the same computation, including spiking dynamics, runs on light, then the material in machine consciousness is incidental.

The consciousness connection is the subject of the post, and it is twofold. Inside the brain emulation section, photonic spiking neurons matter because they test whether the temporal dynamics that consciousness theories specify can be carried by a non-electronic carrier. In the orbit section, they matter because light is the physical currency that defines the latency floor, the fastest possible causal connection a system in orbit can have. An optical neural network is computation that runs at that speed.

From Diffractive Masks to Spiking Lasers

The modern optical neural network field has two main lines. The first is free-space diffraction. In 2018, Xing Lin, Aydogan Ozcan and colleagues at UCLA showed that a stack of 3D-printed diffraction masks could classify handwritten digit images at terahertz speed with no electronic compute in the path, published in Science. The network was a physical object. Light passed through it and came out as classifications, and the weights were the mask geometry.

The second line is photonic spiking. Researchers led by Antonio Hurtado at Strathclyde and collaborators demonstrated spiking optical neurons in vertical-cavity surface-emitting lasers, and in 2026 a team reported two-section semiconductor lasers monolithically integrated on silicon that reproduce integrate-and-fire and resonate-and-fire regimes, the same dynamics the Substrate Console runs with leaky integrate-and-fire neurons. These lasers behave as neurons that sustain their state with light. The consciousness-relevant fact is not the sheer performance. It is that the neuron-model dynamics the site covers in silicon software are now demonstrated in optically generated spiking hardware.

Why Photons Change the Argument

The substrate-independence claim, as the site states it on the functionalist emergentism page, holds that the causal organization that produces mind is not tied to the electrons that carry it. An optical neural network is the strongest argument that claim can have. If integrate-and-fire dynamics running in a semiconductor laser are the same class of computation as LIF dynamics running in software, then no particular material is carrying the relevant pattern, not silicon, not glass, not the ion channels of a cell. The pattern is the work.

The commercial direction makes the claim concrete. Companies such as Lightmatter, Lightelligence and Optalysys build optical compute for inference and accelerator workloads, and academic systems such as Taichi from Tsinghua, subject of an IEEE Spectrum account in 2024, combine diffractive and interferometric optical compute with reported energy efficiency far above electronic accelerators. Taichi reports about 13.96 million parameters and energy-efficiency claims orders of magnitude beyond a GPU. The compute in that machine is optical. The substrate is incidental to the pattern it runs.

The Consciousness Connection

The question this post is built to answer is what optical compute tells us about consciousness, and the answer is a precise claim about the substrate-independence thesis. Theories that locate the relevant organization at the level of temporal spiking dynamics, recurrent processing as covered here and the global workspace account, do not care whether the spikes are electrical or optical. If a recurrent spiking architecture running in lasers meets the temporal criteria the theories specify, then, on those theories, the optical system has whatever the theories predict, and the fact that it is light rather than electrons is not a disqualifier.

The limit is exactly the limit on every substrate-independence case. Producing the right dynamics is not the same as being conscious, and the indicator literature draws that line. What an optical network adds is that it removes the most tempting deflection, that the mind is the electrons. If the computation is in photons, that answer is unavailable, and the question is forced back to the only place it can be answered, the organization.

Light, Latency, and the Orbit Floor

The same physical quantity that carries an optical neural network carries the orbit argument. The latency floor analysis on this site states that a conscious computation cannot integrate across a causal distance larger than light can cross in the relevant time. Light speed is the ceiling on every causal connection in orbit, and an optical computer runs at that ceiling by construction.

That link makes the orbit tie natural. A system in orbit that processes optically is computation running at the maximum possible speed for its location, with the minimum possible propagation cost between its own parts, which is the local-envelope condition that the orbit section’s argument treats as relevant to sustained integration. Photonic compute does not make a mind in orbit happen. It shows that the compute that would be needed, if the substrate-independence thesis is right, can operate in the same currency as the latency floor itself.

What an Optical Network Cannot Do

The honest limit is that no optical network yet demonstrates the full organizational complexity that consciousness theories require, and none makes a claim to experience. A diffraction mask classifies digits. A spiking laser reproduces a neuron model. Neither is a workspace, an integrated information structure, or a self-model. The field is at the hardware-precondition stage, and the relevance to consciousness is that the preconditions for testing temporal-dynamics theories are now directly available in light.

The brain emulation section and the orbit section both inherit this. An emulation built on photonic spiking neurons would be the emulation section’s substrate-independence case stated in the strongest possible material, and an orbital compute fabric running optically would be the orbit thesis’s latency case stated at the physical ceiling. Neither exists yet. Both are now thinkable, and the existence of the hardware makes them less speculative than they were.

Where This Sits

This post opens a new topical cluster for the site, optical and photonic compute as a consciousness-relevant substrate, and it connects two existing sections through a single physical quantity, light. The Substrate Console runs the LIF model in a browser. The substrate debate asks whether the material matters. Optical neural networks give that debate its cleanest test, a computation in which the material is light itself, and the scientific consensus remains the correct frame for what that test can conclude, what it can conclude about the organization, not about experience.

Lin, Rivenson, Ozcan and colleagues, “All-optical machine learning using diffractive deep neural networks,” Science 361: 1004-1008, 2018. Spiking photonic neurons in two-section InP/silicon lasers, 2026 (arXiv). Taichi, hybrid optical, Tsinghua, reported in IEEE Spectrum 2024. Sources are the cited papers and the optical neural network record.