Benedict Patrick

Research

A chip designed around memory, a kernel for context windows, agents that have to prove their work, a world model you can play inside, and two papers under review.

04 featured, 02 papers, 02 more in the index

Phone AI chip, architecture studyArchitecture + RTL

AT1

Phones generate text slowly because they spend their time moving weights, not multiplying them. AT1 is designed around that: move fewer bytes, move them a shorter distance, and compute right beside the memory banks.

  • Verilog
  • Yosys
  • SkyWater 130 nm
  • Python simulator

Speed and energy come from a simulator calibrated to measured Snapdragon numbers, assuming 70% speculative decoding acceptance. Area comes from Yosys synthesis. No silicon yet.

faster than a phone NPU
5.5×
faster than a phone NPULlama 3.1 8B, 4k context
less energy per token
8×
less energy per tokensame run
area per weight
0.45×
area per weightternary sign select vs a W4A8 multiplier
SkyWater synthesis
130 nm
SkyWater synthesisthree Verilog blocks, all tests passing
Virtual memory for LLM contextOpen source

OpenContext

Keeps an agent's working set inside a token budget, swaps cold pages to disk behind a one line tombstone, and pages them back in when the model asks for them again.

fewer tokens, mixed agent session
74.8%
fewer tokens, mixed agent session
smaller, file swapped to disk
99.6%
smaller, file swapped to disk
median page fault
0.22 ms
median page fault
  • Python
  • MCP
  • SQLite
Source on GitHub
Verification sandbox for coding agentsIn development

OpenReason

Coding agents edit your files and hope. OpenReason lets them be wrong somewhere else: several attempts run in isolated git worktrees against your real tests, and only a diff that passed comes back.

attempts, one shared prompt prefix
K
attempts, one shared prompt prefix
writes to your tree until you apply
0
writes to your tree until you apply
agents: Claude Code, Cursor, Codex, Antigravity
4
agents: Claude Code, Cursor, Codex, Antigravity
  • Python
  • MCP
  • Ollama
A world model you can play insideIn progress

GameWM

A small diffusion model that imagines a Breakout style game I wrote, one frame at a time, from your key presses. Because I own the game, every imagined rollout is scored against the true ball, paddle and bricks, not just pixels. It beats a copy last frame baseline and a plain U-Net, tracks the ball for a median of 10 steps, and still loses it before step 50. That gap is the next thing to close.

Game, model and benchmark all written by me

median before the ball is lost
10 steps
median before the ball is lost
of rollouts still on track at step 10
53%
of rollouts still on track at step 10
peak training memory, laptop GPU
1.8 GB
peak training memory, laptop GPU
  • Diffusion
  • EDM
  • PyTorch
  • Ground truth benchmark

Papers

Both are submitted and under review, so there is no public copy yet.

Journal of Materials Processing TechnologyUnder review

When does physical structure help?

Physics informed machine learning is often proposed for metals where tests are scarce. I benchmarked it honestly: three pre registered rounds on five public alloy datasets, with every split and decision rule fixed before a single test score was seen. Structure helped only when it matched how the material varies, and a classical hot working law did just as well.

P. Benedict, Ashwini Jayachandran, A. Saidharshan

error on unseen Al 6082 lots, GPR 17.0
5.2 MPa
error on unseen Al 6082 lots, GPR 17.0
pre registered, results hashed
3 rounds
pre registered, results hashed
public datasets, aluminium and steel
5 alloys
public datasets, aluminium and steel
  • Flow curves
  • PINN
  • Conformal bands
  • Pre registration
Energy and BuildingsUnder review

Termite inspired ventilation

Underground tunnels that cool air before it enters a building, laid out like a termite mound. A reduced order model of a four room house in Chennai, searched with five optimisers across 524,288 designs, and its airflow checked against NIST CONTAM.

P. Benedict, Rajeswari Yogamalar N, Sankar Padmanabhan

design configurations searched
524,288
design configurations searched
predicted peak room temperature
32.6 °C
predicted peak room temperature
airflow agreement with CONTAM
4.2%
airflow agreement with CONTAM
  • Passive cooling
  • Earth air exchange
  • Optimisation
  • QAOA

Also in the lab

  1. 01

    PARQ

    Manuscript

    Detecting Parkinson’s disease from voice with a multitask, dual branch model.

    Manuscript with Sai Dharshan A. and Dr Ashwini, Hindustan Institute of Technology and Science.

    Reference code
  2. 02

    Cognitive Matter

    Browser demo

    A spiking neural fabric simulated live in the browser. Touch it and it learns where you touched.

    4,096 neurons on sparse wiring, 5.1 ms per frame, and a modeled power draw near zero at rest.

    Touch it

Say hello

Selective about projects. Ambitious about the ones I take: models, engines, and the tools around them.

Start a project →benedictpatrickjohn@gmail.com