Photonic quantum chips may leapfrog today’s AI by doing machine learning with light—ultrafast inference, 92%+ accuracy, far lower energy—while we keep betting on bigger transformers.
The story
My brain sees trends and thinks it’s linear,
Exponents sneak up and hit a nerve,
We model tomorrow with yesterday’s tools,
Then wonder why the future won’t curve
🟢 Trend: humans stink at exponential forecasting
I keep watching smart people extrapolate AI like it’s a polite straight line. Then exponential growth shows up like a surprise dentist bill. Our cognitive biases, linear mental models, availability bias toward whatever’s hyped, miscalibration, they all nudge us to bet on “more of the same.”
🟢 Signal: a post‑LLM compute paradigm
My current “fork in the path” bet is *photonic quantum chips*. Not “bigger transformers,” not “more tokens.” A different substrate and a different scaling story: optical quantum processing on-chip.
Some of these chips do machine learning computations *optically* by encoding network parameters into light. They execute matrix multiplications via programmable beamsplitters, with ultrafast inference in under half a nanosecond and over 92% accuracy, comparable to electronic hardware, with far lower energy use. Nonlinear optical function units do linear and nonlinear ops in the optical domain, with minimal electronics for readout.
🟢 Why it matters
LLMs are great at probabilistic next‑token prediction. Photonic processors are built to push huge linear algebra and nonlinearity at light speed, with different constraints and different “reasoning” affordances.
Recent experiments also show small-scale photonic quantum processors outperform classical ML algorithms on classification tasks, leveraging quantum effects for fewer errors without needing massive scale. China’s new photonic quantum chip reports over 1,000-fold acceleration via dense optical integration, aiming at high-bandwidth, low-energy AI workloads.
🟢 What I’m watching
On-chip quantum photonics integrating waveguides, single-photon sources, and high-fidelity gat
Recent Comments