Silicon Does Not Emit Light

Lex Fridman opens with a paper Shainline wrote called *Optoelectronic Intelligence*. The idea is an architecture for brain-inspired computing that uses light to move messages and electronic circuits to transform them. Much of the work at NIST is specifically superconducting electronics, Shainline says, but that choice only makes sense once you see what it is trying not to be: ordinary semiconducting silicon.

A semiconductor such as silicon has no free charges at low temperature. Apply a voltage and electrons move; those currents can be measured and used as information. The material is a crystal, atoms sitting on a nearly perfect lattice, and you can replace a silicon atom with phosphorus to change how many electrons are free to run. A transistor is then three terminals: left, right, and a gate that turns the channel on or off. From that zero-and-one, digital electronics built the rest of the modern world.

The lattice spacing is about half a nanometer, some six orders of magnitude thinner than a human hair. What people call Moore’s Law is the ability to shrink the channel again and again, packing more transistors on the same chip so the same program gets faster and cheaper without the software author doing much at all. Power stopped scaling as kindly in recent decades, Shainline notes, but the run was long. State of the art when they spoke was a seven-nanometer transistor, only tens of atoms along the conduction path.

Those transistors are not built one at a time. A foundry starts with a pristine wafer, uses photolithography to paint shapes in photoresist, and implants ions only where the pattern is open. Photon hits snap polymer chains; an aqueous wash dissolves the exposed resist and leaves the rest. The same pattern is printed across dies, layer upon layer — tens of layers, perhaps a hundred in a mature process — on wafers that have grown to 300 millimeters, about the size of a pizza, holding maybe a thousand processors before a saw dices them apart.

The light used for lithography already sits deep in the ultraviolet, wavelengths over 100 nanometers, and yet the patterned features are seven nanometers, with four nanometers being discussed. Shainline does not pretend to know every optical trick that makes that possible. What strikes him is that the physics of the universe *allows* it. Human beings discovered the scaling more than they invented it, even if they also invented every tool along the way.

Fridman asks whether the specialness is just lithography or the whole stack, down to electrons that still behave as gates at those sizes. Shainline says all of it. Silicon itself is a peculiar semiconductor. A MOSFET needs a thin insulator under the gate so carriers are pulled into the channel instead of being sucked into the contact, and silicon dioxide grows naturally on the surface in air. Almost nothing else on the periodic table makes as good a gate oxide: low defect density, a large energy gap, and the ability to hold a big voltage without leaking.

After the transistor was invented in the late 1940s, several semiconductors were in the race. Most people favored germanium, whose electrons move nicely in the lattice. Others bet on III–V compound semiconductors, whose device physics looked even prettier. Casting a broad net is how science is supposed to work. Silicon still won, and not because the others are useless — they survive in niche roles — but because silicon is the platform of digital computing.

The win was a compromise among many influences, none of them a knockout on its own. Besides the gate oxide, silicon is an elemental semiconductor: you only need silicon atoms. Compound semiconductors need two species alternating on the lattice, which opens a zoo of defects when an atom sits in the wrong seat. There is also the band gap. Germanium sits near 0.75 electron volts; silicon is 1.1. Thermal excitations from room temperature kick accidental carriers over a gap that is too small, and those carriers cause computational errors. Carrier density falls exponentially with the gap, so a few tenths of an electron volt puts silicon in the sweet spot for ambient life.

Fridman notes how strange it is that digital computation assumes errors are vanishingly rare, the opposite of a brain, which assumes failure everywhere and has to stay robust anyway. Shainline agrees. Asked the mischievous question of whether physicists or engineers deserve more credit, he refuses the trap. Physics gets you onto the field — semiconductor theory, then the transistor. Moore’s Law after that is engineering, which is why a phone now outruns old supercomputers. Silicon’s material properties, lithography, X-ray crystallography, and polymer chemistry do not sit cleanly on either side of the line. It is a long dance. Biology, he jokes, is only in the humans doing the work — until they get to brains.

In a normal conductor, an electron that hits a lattice defect or another electron loses momentum, so you must keep applying voltage to keep current flowing. In a superconductor, once current starts, it flows indefinitely with no dissipation. Shainline is talking about conventional, low-critical-temperature materials: a critical temperature around 10 kelvin, operated near 4 kelvin, four degrees above absolute zero. At that temperature the noise of vibrating atoms and colliding electrons drops enough for the electrons to settle into what is sometimes called a macroscopic quantum state.

Cool a block of niobium below its critical temperature and the electrons in the superconducting state occupy one coherent quantum wave function that extends across the whole chunk of metal. Electrons are identical particles, so if one runs into a defect that should have scattered it, another can take its place and the entire state is not scattered. The undissipated flow is a supercurrent. You can still make wires, but you can also make devices. The most versatile component is the Josephson junction: two superconducting wires separated by a thin insulator or normal metal, a weak link the superconducting wave function tunnels across.

A superconductor carries supercurrent only up to a critical current; beyond that it becomes an ordinary resistor. Bias a Josephson junction above that limit and it injects a quantized packet of current into a neighboring superconducting loop. In the community vernacular, you pop a fluxon into the loop. Those pulses can travel at about a third the speed of light, the switching can happen in a few tens of picoseconds, and devices can run in the hundreds of gigahertz. Conventional processors sit nearer one to three gigahertz. Gamers thrilled by 4 gigahertz are looking at the wrong sport.

The physics of Josephson junctions is well understood. They have not displaced silicon in digital computing for mundane reasons. You can keep shrinking a transistor and it still performs its digital job; a Josephson circuit often depends on the interplay between the junction and a loop, and that loop cannot be made arbitrarily small. The limit is not lithography. It is how magnetic field interacts with superconducting material. Liquid helium and a cryostat the size of a beer keg will not go in a phone. For warehouse-scale machines the accounting changes: Shainline visited Oak Ridge’s Titan and Summit, rooms already spending a fortune on cooling, and even there he does not think superconductors will replace semiconductors for digital work.

IBM pushed superconducting digital computing in the 1970s and, in hindsight, chose devices and architectures that were doomed. In the 1990s, Konstantin Likharev and V. K. Semenov proposed a whole family of Josephson logic that could run a hundred times faster than silicon. Shainline later read the papers and thought they were wonderful. Silicon kept winning because shrinking features was a simple way to keep getting better. Speed or energy of a single switch is not a computer. Manufacturing, packaging, and materials all count. Likharev himself, Shainline recalls, suggested that stealing silicon’s digital show was probably the wrong use of these circuits.

That leaves a different kind of computation. Digital logic is bits, two well-separated states, a clock, and arithmetic depth — highly serial. A brain is indisputably processing information, but the start and end of a computation are poorly defined, the signaling is analog and asynchronous, and there is no clock; time scales are built from internal activity. A transistor is a building block of a digital computer. A neuron is a processor. Even a synapse is not a transistor. Neurons take many kinds of input across many spatial and temporal scales and produce many kinds of output, which is where computation and communication become different jobs.

In *Optoelectronic Intelligence*, Shainline writes that electrons excel at computation while light is excellent for communication. Computation, as he uses the word, is taking in information, operating on it, and producing new, more useful information. A picture with a key in it goes in; the coordinates of the key come out. You have reduced the total information and kept what matters. Communication is getting that information from the place it is stored to the place that needs it. Rolf Landauer insisted that information is physical, so it sits somewhere, and moving it is a physical act. Ideally, communication does not change the information. It only relocates it.

Electrons are charged and interact strongly. Pile them up and they want to leave; they also have mass, so you can park them in a place and call that a bit. Change a gate voltage and you move them. Photons, especially at low intensity, essentially do not interact with one another. Add one more photon to a channel and the others keep doing what they were doing, which is wonderful for fan-out and terrible for computation: one packet of light cannot easily change another. Sending electrons down a wire means charging the line. Capacitance grows with length, energy grows with capacitance and voltage squared, and a neuron that must speak to 10,000 colleagues — a realistic cortical number — pays that penalty 10,000 times.

A waveguide is different. Inject photons into glass fiber or on-chip silicon and the voltage you must raise does not grow with length or with extra taps. More connections mean more light, because photons split down branches, but there is no capacitive wiring parasitic of the electronic kind. Catching the photon on the other end is not a dumb question; it defines much of the NIST group’s work. Group leader Sae Woo Nam built a career on superconducting single-photon detectors. Current flows through a superconducting branch until one photon wrecks the macroscopic quantum state, that branch goes resistive, current diverts into a measured pulse, and a few nanoseconds later the wire cools and superconductivity returns.

A semiconductor detector needs on the order of a thousand photons to register a pulse. A superconducting detector needs one. In Shainline’s neuromorphic design that is a feature: one photon or five produce the same synaptic event, so communication stays binary and does not smuggle extra analog information into the channel. Other superconducting sensors in the group, including transition-edge sensors associated with Adriana Lita, can count photons by pulse amplitude, which matters for photonic quantum computing that encodes information in number states. For this architecture, the point is simpler. Communication should say only that a neuron fired.

Neuromorphic computing, Shainline says, is a continuum. Some chips still look like digital computers with more cores and spike-like messages. Others use transistors as analog circuits that behave more like neurons — the lineage of Carver Mead, Giacomo Indiveri, Gert Cauwenberghs, Jennifer Hasler, and many more. Shainline lives at the far end: give up the assumption of conventional hardware and ask, from first principles, what physics would let an artificial system capture the brain’s information-processing tricks. Neurons and synapses adapt on time scales from the gap between spikes to the age of the organism. The network structure matters at least as much as the devices.

Neocortical neurons make thousands of connections, which is how information can jump quickly from any spot in the network to any other. Draw a box in cortex and most connections stay inside it — clustering — but a non-negligible fraction leave, and a bigger box shows the same statistics. That is scale invariance. Connection probability versus distance is not an exponential, which would make the brain local; it is a power law, the same family of shape as gravity and electromagnetism. The same thing happens in time. Neurons do not share one characteristic frequency; the distribution of oscillation rates is again a power law, from tens of milliseconds up to a lifetime of memory, and, if you like, across organisms and civilizations. Capturing those nested spatial and temporal dynamics in hardware is, to Shainline, the goal of neuromorphic computing.

Architecture is not optional. Unroll cortex on a table and it is about the size of a pizza: some 10 billion neurons with power-law connectivity, processing what György Buzsáki would call the *what*. The hippocampus keeps track of where and when, with connectivity that is almost random and barely decaying with distance, but only about 100 million neurons, small enough to stay densely wired. The two modules talk constantly through the thalamus. Together they are the thalamocortical complex. Building something with that modular division of labor, Shainline argues, is probably required if you want the same kinds of activity, not just a soup of neurons.

Memory is several mechanisms, not one. Hopfield-style working memory lives in dynamical patterns: similar enough inputs fall into the same attractor, the face that last time belonged to a movie star. Imprinting those patterns requires synaptic weight updates. In Shainline’s hardware language, supervised learning means an external user has a knob on every weight; unsupervised learning means the weights retune themselves from device physics, with no wire in from outside. Growing new long-range axons is not how you learn this conversation; that is plasticity on existing contacts. Short-term synaptic plasticity weakens a synapse after the second, third, and fourth spike. Metaplasticity changes the *rate* at which weights change, as when a neuromodulator in a lecture hall makes you more writable. Homeostasis lets an overfiring neuron raise its threshold and return to a useful range. Weight updates are a large part of the story. They are not the whole story.

The hardware program is a team sport at NIST’s Boulder campus, a Department of Commerce laboratory better known for standards and precision measurement. Shainline names Sae Woo Nam, Rich Mirin, Sonia Buckley, Jeff Chiles, Adam McCaughan, and Bryce Primavera among the people whose thinking this article is borrowing. They are mostly physicists, electrical engineers, and materials scientists; he would like more neuroscientists than they have. Outsiders say we cannot copy the brain because we do not understand it. Shainline believed that five or seven years earlier. Then he read the literature and was struck by the opposite: how much is known, and how mathematically complete many of the concepts are. Consciousness and the self remain open. What the brain is doing, and why, is already a usable roadmap for a hardware designer.

If a light source on a silicon chip were as simple as a transistor, he says, there would not be a processor in the world that did not use light for communication above some distance. Silicon is notoriously bad at emitting light because of its band structure, an immutable material fact at room temperature. Compound semiconductors emit with high efficiency, but they do not share silicon’s lattice constant, and growing them needs temperatures that wreck transistors already on the wafer. Nature, in Shainline’s reading, is not inviting you to monolithically combine lasers with digital switches. The practical industry answer is package-level integration: make the silicon chip and the III–V light-source chip in different foundries and bolt them together, as in work Shainline had done around Vladimir Stojanović’s project, later Ayar Labs. It helps bandwidth. It is still not a light source per neuron.

Superconducting electronics change the integration problem. You still start with a silicon wafer, but now it is mostly bread for the sandwich; the computation lives in deposited superconductors, which do not need to be lattice-matched to a transistor. Compound-semiconductor emitters can sit on the wafer with superconducting circuits grown above — at least, it is not obviously doomed. Four kelvin also makes silicon itself a less terrible emitter. And because a superconducting synapse needs one photon rather than a few thousand, you can live with the world’s worst light source as long as it spits out a few thousand photons when a neuron fires. Room-temperature optoelectronic integration has been exceedingly hard. Shainline’s bet is that a different application — neuromorphic, not digital — plus superconducting detectors might finally let it happen.

Four kelvin makes many people stand up and leave. Shainline calls that earth-centric. We live at 300 kelvin and want our gadgets there too. The cosmic microwave background is 2.7 kelvin; we are still warmer than deep space. Quantum computing routinely works at 50 to 100 millikelvin, another factor of a hundred colder, and that has not stopped the field. Give up the phone and the laptop on a train, he says, and 4 kelvin is a laboratory commonplace. For scientific questions about the physical limits of cognition, comfort is not a constraint. If cold hardware can think at a scale orders of magnitude beyond any room-temperature technology, you get cold.

A neuron on this platform is Josephson circuitry. A synapse receives one photon on a superconducting detector and adds analog current — eight to ten bits of effective resolution — to a loop; a resistor sets an exponential decay, the postsynaptic signal. The same building block, a DC SQUID of two Josephson junctions, can be a synapse, a dendrite, a soma, or a plasticity element, differing only in a few parameters. When the cell body drives a junction above critical current, a pulse is voltage-amplified into a semiconductor light source, the only semiconducting piece of the neuron, and that light fans out through a waveguide network designed by colleagues such as Jeff Chiles. Memory is current circulating in loops; coupling is transformer-style from loop to loop. Shainline calls them loop neurons because the phrase beats “superconducting optoelectronic neurons.”

Hardware neurons of real complexity, silicon or superconducting, end up around 100 microns on a side. Waveguides are stuck near a micron by the wavelength of light; a biological axon can be 10 nanometers. To approach brain scale you have to use the third dimension: several planes of active superconducting circuits and passive waveguides on a wafer, perhaps five to ten, enough for millions of neurons, then whole wafers stacked with fiber between them until a table-sized volume holds some 10 billion neurons. The stacking itself must stay fractal, power-law connectivity from the on-wafer network to the wafer stack to the whole machine, or information cannot spread from a point across the system. At that size, Shainline says, of course you use light.

A single synapse is a handful of Josephson junctions and can be simulated in SPICE by integrating the circuit equations on a picosecond grid. Networks explode that cost. During the pandemic the group abstracted the shared building block into one leaky-integrator differential equation per synapse, dendrite, or neuron, and stepped it on a coarser time grid. The speedup is a factor of a thousand to ten thousand, enough to aim at a million neurons instead of tens or a hundred. Shainline likes the source of the complexity: simple tiles, repeated, with parameters and wiring that yield behavior you did not put in by hand. The same platform can host neuroscience questions and other physics, critical phenomena and Ising models included. He is not trying to turn it into a company. If you want short-term money, he says, do deep learning on silicon.

He will not overpromise that neuromorphic superconducting hardware will soon beat machine learning at machine-learning’s own game. From his seat, ML usually requires that you know the input and the objective. Intelligence, as he wants the word, is the ability to handle problems where you do not know what is coming in and are not sure what you are trying to accomplish — “I’m not sure what I’m trying to accomplish in this world.” Colleagues are still pushing the ML angle. NIST’s Mike Schneider is building a superconducting image classifier on a nine-pixel, three-category toy, with classification that might take around 100 picoseconds. Josephson circuits themselves can be orders of magnitude more efficient than transistors doing the same job, but the cryostat’s idle cost is large, so the system has to be big enough to amortize it. Not a phone. Maybe a data center.

Fridman describes Tesla’s then-announced Dojo: an application-driven training machine in a loop of train, deploy to a fleet, harvest edge cases, retrain. Shainline does not know the product timeline, but the pattern matches what superconducting optoelectronics might one day offer — a centralized cognitive hub with massive data integration, not a cryostat in every trunk. Five or ten years is not crazy; tomorrow is. NIST also cares about what happens when those networks leave the lab. Elham Tabassi leads a much larger AI effort there, centered on trustworthiness. Computer science wants algorithms that are provably correct. Biological systems prove nothing; they are robust to being wrong. Shainline quotes Turing in spirit: you can have an intelligent system or a flawless one, but not both.

The second half of the conversation turns on a paper Shainline titled “Does Cosmological Evolution Select for Technology?” Physicists learned that the equations of nature are accompanied by parameters — coupling strengths, masses, the fine-structure constant — and those numbers look dialed in. Nudge them and the universe can appear and vanish in a nanosecond. Someone called it a put-up job. The majority reply is the weak anthropic principle: of course we see a universe fit for us, because otherwise we would not be here. That never sat well with Shainline, and it explains even less once you notice superconductivity, or silicon’s nearly perfect native oxide. Those gifts are useful for technology. They are not required for naked survival.

He places the argument on Lee Smolin’s cosmological natural selection, from a late-1980s paper, “Did the Universe Evolve?” Inflationary cosmology, developed by Alan Guth, Andrei Linde, and others, is the starting point, not a replacement. In 1986 Guth and Edward Farhi argued that a big bang is mathematically indistinguishable from a black hole. Smolin’s move is that a black hole in one universe is a big bang in another, and because black-hole formation is a quantum-gravitational process the offspring’s parameters mutate slightly. Variation plus selection: universes that make more black holes have more children. Smolin estimated some 200 million generations and argued that the selection pressure is stars, because stars make black holes.

Shainline thinks Smolin is right about almost all of that and still incomplete, because it leaves intelligent life as a fluke in the cracks of a star factory. Stars make black holes wastefully, by gravity. Quantum-field-theory calculations, which he is careful to say he is only translating, suggest that compressing about 10 kilograms into a tiny volume can make a black hole likely to inflate into a child universe. A one-kilometer asteroid, broken into 10-kilogram chunks, would yield at least a trillion black holes — three orders of magnitude more than all the core-collapse supernovae in the Milky Way’s history, and that is one piece of scrap. A civilization that reached the asteroid belt, perhaps running on superconducting optoelectronic networks, could outpace stellar fecundity. Then you would expect our parameters to be a compromise among long-lived stars, useful water, rocky habitable worlds, *and* the physics that makes technology possible, sitting at a tug-of-war center rather than at any one optimum.

If the hypothesis is right, intelligent civilizations need to be common enough that *some* of them reach the ability to make black holes, but they need not carpet every solar system. One mature civilization per galaxy could dominate that galaxy’s fecundity. Shainline puts his chips on rarity. He follows Max Tegmark and recent astrobiology: you should not be shocked that an intelligent species arose in the Milky Way, and you should be shocked if the next star over has one. Across a galaxy’s lifetime the count might be one to a hundred, but those lives need not overlap. Earth’s record is a warning about time windows. Life appeared quickly as microbes; then roughly two and a half billion years of oxygenation passed before complex bodies. The Cambrian explosion, five to six hundred million years ago, is the rare-Earth threshold he buys: microbes may be everywhere a wet habitable planet exists, while that burst of body plans is not, and humans may be a narrower gate still.

Fridman offers a beautiful inversion: if cosmological evolution selects for technology, humans cannot be the most astonishing product of the process. Most of a universe is boring, a few rocks and some apes recording a podcast. Shainline does not follow that into cynicism. He sees hierarchy. The early universe was plasma. First-generation stars had to form before they could seed metals for later generations; we are only in the third. Life on Earth is another layer. The next might grow on top of us: cognitive systems far more capable, living off Earth’s surface in the cold dark, with single-photon sensors from radio to gamma rays and huge collection optics, watching the universe. A system like that would probably amuse itself. One amusement, for someone who likes video games, would be to create, maintain, and observe something like Earth.

Fridman closes by noting the strange coupling in Shainline’s work: cutting-edge hardware on one desk, the evolution of physical law on the other, philosophy applied as strictly as engineering. He calls Shainline one of the best explainers of hard ideas, able to work at every level of the intellectual stack — fractal, in the sense the conversation has been using. Shainline returns the compliment. The valuable part, both of them agree, is the habit of connecting those layers rather than camping on one of them.