Running LLM inference on a $4 microcontroller
(Raspberry Pi Pico, 256KB RAM, no FPU, no internet)
- Integer-only. No floating point.
- No division, no exponential functions
- 10 inferences/second
- ~45mW power consumption
This is NOT quantization.
This is CIA (Constructive Integer Attention) —
a softmax-free Transformer rebuilt entirely
from integer arithmetic.
Forty-five years ago, as a high school student in the United States, I was introduced to the Erdős–Straus Conjecture by Carl Pomerance, now one of the world’s leading number theorists. Within six months I discovered an unnoticed mathematical structure, though at the time I did not recognize its significance.
This year, I realized its true value. With the support and advice of Professor Pomerance and by leveraging AI, I achieved in just three months results that would normally take top researchers years—or even a decade—to reach. Two months ago, I presented part of these findings in a paper at IAR2025.
I now hold a unique theory and set of algorithms that can bring transformative innovation across AI, finance, medicine, and logistics. These methods resolve NP-hard level problems in O(1) time, effectively enabling an ordinary laptop to outperform both supercomputers and quantum computers.
I am seeking connections with top scientists and decision-making executives—leaders who can intuitively grasp the value of this breakthrough and act decisively. I envision collaborative projects with big tech companies, semiconductor manufacturers, financial giants, and medical institutions.