Universal photonic artificial intelligence acceleration | Hacker News Reader

Universal photonic artificial intelligence acceleration

nature.com

14 pointswesturner6 comments

"Universal photonic artificial intelligence acceleration" (2025) https://www.nature.com/articles/s41586-025-08854-x :

> Abstract: [...] Here we introduce a photonic AI processor that executes advanced AI models, including ResNet [3] and BERT [20,21], along with the Atari deep reinforcement learning algorithm originally demonstrated by DeepMind [22]. This processor achieves near-electronic precision for many workloads, marking a notable entry for photonic computing into competition with established electronic AI accelerators [23] and an essential step towards developing post-transistor computing technologies.

This is going to transform our society more than anything else. A giant foundation model can be cast for very, very cheap inference. Imagine the improvement from room sized computers to your iPhone, applied to this branch of technology.

Photon-based chips vs electron-based chips? It's interesting that photons and electrons are closely related: two neutral photons can become an electron-positron pair and vice versa.

The paper mentions that their photonic chip is less precise than an electronic one, but this looks like an advantage for AI. In fact, the stubborn precision of electron-based processors that erase the quantum nature of electrons is what I think is preventing the creation of real AI. In other words, if a microprocessor is effectively a deterministic complex mechanism, it won't become AI no matter what, but if the quantum nature is let loose, at least slightly, interesting things will start to happen.

I don't see why you would say closely related. They are 2 separate fields which interaction sure. But fermion vs boson, matter vs force carrier, etc.

That quantum stuff is just woo. QM isn't magic.

The energy level involved for pair production is astronomically far away from the application discussed here, not to mention the intensity required due to the infinitesimally tiny cross section.

This is more using matter as an intermediary for interaction at far lower energy levels, along the lines of Kerr or Pockels.

There are certainly infinitely non-halting automata on deterministic electronic processors.

Abstractly, in terms of constructor theory, the non-halting task is independent from a constructor, which is implemented with a computation medium.

FWIU from reading a table on wikipedia about physical platforms for QC it would be possible to do quantum computing with just electron voltage but typical component quality.

So phase; certain phases.

And now parametic single photon emission and detection.

"Low-noise balanced homodyne detection with superconducting nanowire single-photon detectors" (2024) https://news.ycombinator.com/item?id=39537236

"A physical [photonic] qubit with built-in error correction" (2024) https://news.ycombinator.com/item?id=39243929

"Quantum vortices of strongly interacting photons" (2024) https://www.science.org/doi/10.1126/science.adh5315 .. https://arxiv.org/abs/2302.05967 .. "Study of photons in QC reveals photon collisions in matter create vortices" https://news.ycombinator.com/item?id=40600736

"How do photons mediate both attraction and repulsion?" (2025) [as phonons in matter] https://news.ycombinator.com/item?id=42661511 notes re: recent findings with photons ("quanta")

Reply on news.ycombinator.com