Jeffrey Shainline: Neuromorphic Computing and Optoelectronic Intelligence | Lex Fridman Podcast #225

Silicon offers an exceptional combination of properties—like a near-ideal native oxide (SiO₂) for gate insulation and a bandgap well-suited to room-temperature digital operation—that allowed MOSFETs to scale for decades under Moore’s Law. This was less an arbitrary engineering victory and more a case of exploiting “gifted” physics that made mass manufacturing and miniaturization possible.

Josephson junctions can operate at hundreds of gigahertz with extremely low switching energy, yet their circuits don’t scale down like transistors and must be cooled to ~4 K. When you factor in cooling overhead, manufacturing, and density limits, superconducting digital logic is not a drop-in successor to silicon for general-purpose digital computing.

Electrons interact strongly and can be localized, which is good for logic and state storage. Photons barely interact and incur no capacitive wiring penalty, which makes them excellent for long-distance, high-fan-out signaling—especially in brain-like networks where each “neuron” may need to connect to ~10,000 others.

The cortex exhibits power-law (not exponential) decay of connection probability with distance and similar scale-free statistics in temporal activity. This fractal organization underpins fast, flexible information integration across many spatial and temporal scales; Shainline argues any serious neuromorphic system must reflect these structural and dynamical principles, not just use spiking neurons superficially.

In Shainline’s architecture, an incoming photon triggers a superconducting single-photon detector that injects quantized current into a superconducting loop, encoding synaptic weight and postsynaptic signals as circulating currents with controlled decay. When accumulated input exceeds a threshold, Josephson circuitry amplifies the event and drives a light source, sending single-photon spikes through optical waveguides to thousands of downstream synapses.