Mprah-Boamey  ·  Ghana  ·  19

Physical computing,
from
first principles.

Optical neural networks in PyTorch. Wave propagation, diffractive layers, interference-based attention. The physics is verified; the hardware is the next problem.

See the work ↓ GitHub ↗ X ↗

01  Current work

001 — FLAGSHIP

PHOTEX

Optical computing

Optical neural network simulator in PyTorch. Trainable diffractive layers act as weights; wave propagation via the Angular Spectrum Method runs the forward pass; wave interference between encoded query and key fields computes self-attention physically. Simulator projections only — no physical device has been measured.

7.5M–375M tok/s projected ~1.3ns forward pass 45T params / cm³

github.com/mprahboamey/PHOTEX-clean ↗

002 — SYSTEMS

HoloWeights

Inference efficiency

Neural network weights in memory-mapped tiles inspired by holographic storage geometry. Only loads what inference actually needs — RAM pressure proportional to what's touched, not the full model. Sits next to any existing runtime.

~33.7MB RSS delta ~3.3× vs dense FP16

github.com/mprahboamey/holoweights ↗

02  Past experiments

Poincaré Experiment

Compress enough randomness, gravity, and statistical mechanics into a digital system, run it forward — eventually emergent sentience from scratch.

Failed. Learned everything.

Solid-state battery sim

Computational search for a solid-state battery candidate material. Burned $300 on compute finding out what doesn't work.

Failed. Not embarrassed.

Constructor theory AI

Stripped all wheel and rolling vocabulary from training. The system independently derived that curved contact was optimal for moving heavy objects.

It figured it out.

Non-Newtonian physics engine

Hypergraph-based discrete physics. The AI experiences physical consequences rather than predicting them. It knows what glass is.

Ongoing.

Campus accountability app

Started as me manually messaging 20 students daily. Scaled to 120 active users in under a month via printed QR-code flyers.

Died on vacation. Fair enough.

03  About

I build and simulate physical computing systems from first principles. One year of university, then stopped out to keep building. Most of the work sits at the intersection of physics, numerical methods, and machine learning. The questions I follow are the ones existing tools were not built to answer.

Right now: crossing from simulation into hardware. The physics behind PHOTEX is verified against published optics literature. The device does not exist yet. That is the next problem.

Open to collaborators, research partners, and groups working on photonic computing or unconventional inference architectures.

7 Major projects in one year
3mo To build PHOTEX and HoloWeights
120 Active users, campus app
19 Years old. Stopped out. Building.

Let's build
something.

Photonic computing, inference simulation, or research collaborations.